drivers/block/mtip32xx/mtip32xx.c

   1 /*
   2  * Driver for the Micron P320 SSD
   3  *   Copyright (C) 2011 Micron Technology, Inc.
   4  *
   5  * Portions of this code were derived from works subjected to the
   6  * following copyright:
   7  *    Copyright (C) 2009 Integrated Device Technology, Inc.
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  */
  20
  21 #include <linux/pci.h>
  22 #include <linux/interrupt.h>
  23 #include <linux/ata.h>
  24 #include <linux/delay.h>
  25 #include <linux/hdreg.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/random.h>
  28 #include <linux/smp.h>
  29 #include <linux/compat.h>
  30 #include <linux/fs.h>
  31 #include <linux/module.h>
  32 #include <linux/genhd.h>
  33 #include <linux/blkdev.h>
  34 #include <linux/bio.h>
  35 #include <linux/dma-mapping.h>
  36 #include <linux/idr.h>
  37 #include <linux/kthread.h>
  38 #include <../drivers/ata/ahci.h>
  39 #include <linux/export.h>
  40 #include <linux/debugfs.h>
  41 #include "mtip32xx.h"
  42
  43 #define HW_CMD_SLOT_SZ          (MTIP_MAX_COMMAND_SLOTS * 32)
  44 #define HW_CMD_TBL_SZ           (AHCI_CMD_TBL_HDR_SZ + (MTIP_MAX_SG * 16))
  45 #define HW_CMD_TBL_AR_SZ        (HW_CMD_TBL_SZ * MTIP_MAX_COMMAND_SLOTS)
  46 #define HW_PORT_PRIV_DMA_SZ \
  47                 (HW_CMD_SLOT_SZ + HW_CMD_TBL_AR_SZ + AHCI_RX_FIS_SZ)
  48
  49 #define HOST_CAP_NZDMA          (1 << 19)
  50 #define HOST_HSORG              0xFC
  51 #define HSORG_DISABLE_SLOTGRP_INTR (1<<24)
  52 #define HSORG_DISABLE_SLOTGRP_PXIS (1<<16)
  53 #define HSORG_HWREV             0xFF00
  54 #define HSORG_STYLE             0x8
  55 #define HSORG_SLOTGROUPS        0x7
  56
  57 #define PORT_COMMAND_ISSUE      0x38
  58 #define PORT_SDBV               0x7C
  59
  60 #define PORT_OFFSET             0x100
  61 #define PORT_MEM_SIZE           0x80
  62
  63 #define PORT_IRQ_ERR \
  64         (PORT_IRQ_HBUS_ERR | PORT_IRQ_IF_ERR | PORT_IRQ_CONNECT | \
  65          PORT_IRQ_PHYRDY | PORT_IRQ_UNK_FIS | PORT_IRQ_BAD_PMP | \
  66          PORT_IRQ_TF_ERR | PORT_IRQ_HBUS_DATA_ERR | PORT_IRQ_IF_NONFATAL | \
  67          PORT_IRQ_OVERFLOW)
  68 #define PORT_IRQ_LEGACY \
  69         (PORT_IRQ_PIOS_FIS | PORT_IRQ_D2H_REG_FIS)
  70 #define PORT_IRQ_HANDLED \
  71         (PORT_IRQ_SDB_FIS | PORT_IRQ_LEGACY | \
  72          PORT_IRQ_TF_ERR | PORT_IRQ_IF_ERR | \
  73          PORT_IRQ_CONNECT | PORT_IRQ_PHYRDY)
  74 #define DEF_PORT_IRQ \
  75         (PORT_IRQ_ERR | PORT_IRQ_LEGACY | PORT_IRQ_SDB_FIS)
  76
  77 /* product numbers */
  78 #define MTIP_PRODUCT_UNKNOWN    0x00
  79 #define MTIP_PRODUCT_ASICFPGA   0x11
  80
  81 /* Device instance number, incremented each time a device is probed. */
  82 static int instance;
  83
  84 struct list_head online_list;
  85 struct list_head removing_list;
  86 spinlock_t dev_lock;
  87
  88 /*
  89  * Global variable used to hold the major block device number
  90  * allocated in mtip_init().
  91  */
  92 static int mtip_major;
  93 static struct dentry *dfs_parent;
  94 static struct dentry *dfs_device_status;
  95
  96 static u32 cpu_use[NR_CPUS];
  97
  98 static DEFINE_SPINLOCK(rssd_index_lock);
  99 static DEFINE_IDA(rssd_index_ida);
 100
 101 static int mtip_block_initialize(struct driver_data *dd);
 102
 103 #ifdef CONFIG_COMPAT
 104 struct mtip_compat_ide_task_request_s {
 105         __u8            io_ports[8];
 106         __u8            hob_ports[8];
 107         ide_reg_valid_t out_flags;
 108         ide_reg_valid_t in_flags;
 109         int             data_phase;
 110         int             req_cmd;
 111         compat_ulong_t  out_size;
 112         compat_ulong_t  in_size;
 113 };
 114 #endif
 115
 116 /*
 117  * This function check_for_surprise_removal is called
 118  * while card is removed from the system and it will
 119  * read the vendor id from the configration space
 120  *
 121  * @pdev Pointer to the pci_dev structure.
 122  *
 123  * return value
 124  *       true if device removed, else false
 125  */
 126 static bool mtip_check_surprise_removal(struct pci_dev *pdev)
 127 {
 128         u16 vendor_id = 0;
 129
 130        /* Read the vendorID from the configuration space */
 131         pci_read_config_word(pdev, 0x00, &vendor_id);
 132         if (vendor_id == 0xFFFF)
 133                 return true; /* device removed */
 134
 135         return false; /* device present */
 136 }
 137
 138 /*
 139  * This function is called for clean the pending command in the
 140  * command slot during the surprise removal of device and return
 141  * error to the upper layer.
 142  *
 143  * @dd Pointer to the DRIVER_DATA structure.
 144  *
 145  * return value
 146  *      None
 147  */
 148 static void mtip_command_cleanup(struct driver_data *dd)
 149 {
 150         int group = 0, commandslot = 0, commandindex = 0;
 151         struct mtip_cmd *command;
 152         struct mtip_port *port = dd->port;
 153         static int in_progress;
 154
 155         if (in_progress)
 156                 return;
 157
 158         in_progress = 1;
 159
 160         for (group = 0; group < 4; group++) {
 161                 for (commandslot = 0; commandslot < 32; commandslot++) {
 162                         if (!(port->allocated[group] & (1 << commandslot)))
 163                                 continue;
 164
 165                         commandindex = group << 5 | commandslot;
 166                         command = &port->commands[commandindex];
 167
 168                         if (atomic_read(&command->active)
 169                             && (command->async_callback)) {
 170                                 command->async_callback(command->async_data,
 171                                         -ENODEV);
 172                                 command->async_callback = NULL;
 173                                 command->async_data = NULL;
 174                         }
 175
 176                         dma_unmap_sg(&port->dd->pdev->dev,
 177                                 command->sg,
 178                                 command->scatter_ents,
 179                                 command->direction);
 180                 }
 181         }
 182
 183         up(&port->cmd_slot);
 184
 185         set_bit(MTIP_DDF_CLEANUP_BIT, &dd->dd_flag);
 186         in_progress = 0;
 187 }
 188
 189 /*
 190  * Obtain an empty command slot.
 191  *
 192  * This function needs to be reentrant since it could be called
 193  * at the same time on multiple CPUs. The allocation of the
 194  * command slot must be atomic.
 195  *
 196  * @port Pointer to the port data structure.
 197  *
 198  * return value
 199  *      >= 0    Index of command slot obtained.
 200  *      -1      No command slots available.
 201  */
 202 static int get_slot(struct mtip_port *port)
 203 {
 204         int slot, i;
 205         unsigned int num_command_slots = port->dd->slot_groups * 32;
 206
 207         /*
 208          * Try 10 times, because there is a small race here.
 209          *  that's ok, because it's still cheaper than a lock.
 210          *
 211          * Race: Since this section is not protected by lock, same bit
 212          * could be chosen by different process contexts running in
 213          * different processor. So instead of costly lock, we are going
 214          * with loop.
 215          */
 216         for (i = 0; i < 10; i++) {
 217                 slot = find_next_zero_bit(port->allocated,
 218                                          num_command_slots, 1);
 219                 if ((slot < num_command_slots) &&
 220                     (!test_and_set_bit(slot, port->allocated)))
 221                         return slot;
 222         }
 223         dev_warn(&port->dd->pdev->dev, "Failed to get a tag.\n");
 224
 225         if (mtip_check_surprise_removal(port->dd->pdev)) {
 226                 /* Device not present, clean outstanding commands */
 227                 mtip_command_cleanup(port->dd);
 228         }
 229         return -1;
 230 }
 231
 232 /*
 233  * Release a command slot.
 234  *
 235  * @port Pointer to the port data structure.
 236  * @tag  Tag of command to release
 237  *
 238  * return value
 239  *      None
 240  */
 241 static inline void release_slot(struct mtip_port *port, int tag)
 242 {
 243         smp_mb__before_clear_bit();
 244         clear_bit(tag, port->allocated);
 245         smp_mb__after_clear_bit();
 246 }
 247
 248 /*
 249  * Reset the HBA (without sleeping)
 250  *
 251  * @dd Pointer to the driver data structure.
 252  *
 253  * return value
 254  *      0       The reset was successful.
 255  *      -1      The HBA Reset bit did not clear.
 256  */
 257 static int mtip_hba_reset(struct driver_data *dd)
 258 {
 259         unsigned long timeout;
 260
 261         /* Set the reset bit */
 262         writel(HOST_RESET, dd->mmio + HOST_CTL);
 263
 264         /* Flush */
 265         readl(dd->mmio + HOST_CTL);
 266
 267         /* Spin for up to 2 seconds, waiting for reset acknowledgement */
 268         timeout = jiffies + msecs_to_jiffies(2000);
 269         do {
 270                 mdelay(10);
 271                 if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag))
 272                         return -1;
 273
 274         } while ((readl(dd->mmio + HOST_CTL) & HOST_RESET)
 275                  && time_before(jiffies, timeout));
 276
 277         if (readl(dd->mmio + HOST_CTL) & HOST_RESET)
 278                 return -1;
 279
 280         return 0;
 281 }
 282
 283 /*
 284  * Issue a command to the hardware.
 285  *
 286  * Set the appropriate bit in the s_active and Command Issue hardware
 287  * registers, causing hardware command processing to begin.
 288  *
 289  * @port Pointer to the port structure.
 290  * @tag  The tag of the command to be issued.
 291  *
 292  * return value
 293  *      None
 294  */
 295 static inline void mtip_issue_ncq_command(struct mtip_port *port, int tag)
 296 {
 297         int group = tag >> 5;
 298
 299         atomic_set(&port->commands[tag].active, 1);
 300
 301         /* guard SACT and CI registers */
 302         spin_lock(&port->cmd_issue_lock[group]);
 303         writel((1 << MTIP_TAG_BIT(tag)),
 304                         port->s_active[MTIP_TAG_INDEX(tag)]);
 305         writel((1 << MTIP_TAG_BIT(tag)),
 306                         port->cmd_issue[MTIP_TAG_INDEX(tag)]);
 307         spin_unlock(&port->cmd_issue_lock[group]);
 308
 309         /* Set the command's timeout value.*/
 310         port->commands[tag].comp_time = jiffies + msecs_to_jiffies(
 311                                         MTIP_NCQ_COMMAND_TIMEOUT_MS);
 312 }
 313
 314 /*
 315  * Enable/disable the reception of FIS
 316  *
 317  * @port   Pointer to the port data structure
 318  * @enable 1 to enable, 0 to disable
 319  *
 320  * return value
 321  *      Previous state: 1 enabled, 0 disabled
 322  */
 323 static int mtip_enable_fis(struct mtip_port *port, int enable)
 324 {
 325         u32 tmp;
 326
 327         /* enable FIS reception */
 328         tmp = readl(port->mmio + PORT_CMD);
 329         if (enable)
 330                 writel(tmp | PORT_CMD_FIS_RX, port->mmio + PORT_CMD);
 331         else
 332                 writel(tmp & ~PORT_CMD_FIS_RX, port->mmio + PORT_CMD);
 333
 334         /* Flush */
 335         readl(port->mmio + PORT_CMD);
 336
 337         return (((tmp & PORT_CMD_FIS_RX) == PORT_CMD_FIS_RX));
 338 }
 339
 340 /*
 341  * Enable/disable the DMA engine
 342  *
 343  * @port   Pointer to the port data structure
 344  * @enable 1 to enable, 0 to disable
 345  *
 346  * return value
 347  *      Previous state: 1 enabled, 0 disabled.
 348  */
 349 static int mtip_enable_engine(struct mtip_port *port, int enable)
 350 {
 351         u32 tmp;
 352
 353         /* enable FIS reception */
 354         tmp = readl(port->mmio + PORT_CMD);
 355         if (enable)
 356                 writel(tmp | PORT_CMD_START, port->mmio + PORT_CMD);
 357         else
 358                 writel(tmp & ~PORT_CMD_START, port->mmio + PORT_CMD);
 359
 360         readl(port->mmio + PORT_CMD);
 361         return (((tmp & PORT_CMD_START) == PORT_CMD_START));
 362 }
 363
 364 /*
 365  * Enables the port DMA engine and FIS reception.
 366  *
 367  * return value
 368  *      None
 369  */
 370 static inline void mtip_start_port(struct mtip_port *port)
 371 {
 372         /* Enable FIS reception */
 373         mtip_enable_fis(port, 1);
 374
 375         /* Enable the DMA engine */
 376         mtip_enable_engine(port, 1);
 377 }
 378
 379 /*
 380  * Deinitialize a port by disabling port interrupts, the DMA engine,
 381  * and FIS reception.
 382  *
 383  * @port Pointer to the port structure
 384  *
 385  * return value
 386  *      None
 387  */
 388 static inline void mtip_deinit_port(struct mtip_port *port)
 389 {
 390         /* Disable interrupts on this port */
 391         writel(0, port->mmio + PORT_IRQ_MASK);
 392
 393         /* Disable the DMA engine */
 394         mtip_enable_engine(port, 0);
 395
 396         /* Disable FIS reception */
 397         mtip_enable_fis(port, 0);
 398 }
 399
 400 /*
 401  * Initialize a port.
 402  *
 403  * This function deinitializes the port by calling mtip_deinit_port() and
 404  * then initializes it by setting the command header and RX FIS addresses,
 405  * clearing the SError register and any pending port interrupts before
 406  * re-enabling the default set of port interrupts.
 407  *
 408  * @port Pointer to the port structure.
 409  *
 410  * return value
 411  *      None
 412  */
 413 static void mtip_init_port(struct mtip_port *port)
 414 {
 415         int i;
 416         mtip_deinit_port(port);
 417
 418         /* Program the command list base and FIS base addresses */
 419         if (readl(port->dd->mmio + HOST_CAP) & HOST_CAP_64) {
 420                 writel((port->command_list_dma >> 16) >> 16,
 421                          port->mmio + PORT_LST_ADDR_HI);
 422                 writel((port->rxfis_dma >> 16) >> 16,
 423                          port->mmio + PORT_FIS_ADDR_HI);
 424         }
 425
 426         writel(port->command_list_dma & 0xFFFFFFFF,
 427                         port->mmio + PORT_LST_ADDR);
 428         writel(port->rxfis_dma & 0xFFFFFFFF, port->mmio + PORT_FIS_ADDR);
 429
 430         /* Clear SError */
 431         writel(readl(port->mmio + PORT_SCR_ERR), port->mmio + PORT_SCR_ERR);
 432
 433         /* reset the completed registers.*/
 434         for (i = 0; i < port->dd->slot_groups; i++)
 435                 writel(0xFFFFFFFF, port->completed[i]);
 436
 437         /* Clear any pending interrupts for this port */
 438         writel(readl(port->mmio + PORT_IRQ_STAT), port->mmio + PORT_IRQ_STAT);
 439
 440         /* Clear any pending interrupts on the HBA. */
 441         writel(readl(port->dd->mmio + HOST_IRQ_STAT),
 442                                         port->dd->mmio + HOST_IRQ_STAT);
 443
 444         /* Enable port interrupts */
 445         writel(DEF_PORT_IRQ, port->mmio + PORT_IRQ_MASK);
 446 }
 447
 448 /*
 449  * Restart a port
 450  *
 451  * @port Pointer to the port data structure.
 452  *
 453  * return value
 454  *      None
 455  */
 456 static void mtip_restart_port(struct mtip_port *port)
 457 {
 458         unsigned long timeout;
 459
 460         /* Disable the DMA engine */
 461         mtip_enable_engine(port, 0);
 462
 463         /* Chip quirk: wait up to 500ms for PxCMD.CR == 0 */
 464         timeout = jiffies + msecs_to_jiffies(500);
 465         while ((readl(port->mmio + PORT_CMD) & PORT_CMD_LIST_ON)
 466                  && time_before(jiffies, timeout))
 467                 ;
 468
 469         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 470                 return;
 471
 472         /*
 473          * Chip quirk: escalate to hba reset if
 474          * PxCMD.CR not clear after 500 ms
 475          */
 476         if (readl(port->mmio + PORT_CMD) & PORT_CMD_LIST_ON) {
 477                 dev_warn(&port->dd->pdev->dev,
 478                         "PxCMD.CR not clear, escalating reset\n");
 479
 480                 if (mtip_hba_reset(port->dd))
 481                         dev_err(&port->dd->pdev->dev,
 482                                 "HBA reset escalation failed.\n");
 483
 484                 /* 30 ms delay before com reset to quiesce chip */
 485                 mdelay(30);
 486         }
 487
 488         dev_warn(&port->dd->pdev->dev, "Issuing COM reset\n");
 489
 490         /* Set PxSCTL.DET */
 491         writel(readl(port->mmio + PORT_SCR_CTL) |
 492                          1, port->mmio + PORT_SCR_CTL);
 493         readl(port->mmio + PORT_SCR_CTL);
 494
 495         /* Wait 1 ms to quiesce chip function */
 496         timeout = jiffies + msecs_to_jiffies(1);
 497         while (time_before(jiffies, timeout))
 498                 ;
 499
 500         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 501                 return;
 502
 503         /* Clear PxSCTL.DET */
 504         writel(readl(port->mmio + PORT_SCR_CTL) & ~1,
 505                          port->mmio + PORT_SCR_CTL);
 506         readl(port->mmio + PORT_SCR_CTL);
 507
 508         /* Wait 500 ms for bit 0 of PORT_SCR_STS to be set */
 509         timeout = jiffies + msecs_to_jiffies(500);
 510         while (((readl(port->mmio + PORT_SCR_STAT) & 0x01) == 0)
 511                          && time_before(jiffies, timeout))
 512                 ;
 513
 514         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
 515                 return;
 516
 517         if ((readl(port->mmio + PORT_SCR_STAT) & 0x01) == 0)
 518                 dev_warn(&port->dd->pdev->dev,
 519                         "COM reset failed\n");
 520
 521         mtip_init_port(port);
 522         mtip_start_port(port);
 523
 524 }
 525
 526 static int mtip_device_reset(struct driver_data *dd)
 527 {
 528         int rv = 0;
 529
 530         if (mtip_check_surprise_removal(dd->pdev))
 531                 return 0;
 532
 533         if (mtip_hba_reset(dd) < 0)
 534                 rv = -EFAULT;
 535
 536         mdelay(1);
 537         mtip_init_port(dd->port);
 538         mtip_start_port(dd->port);
 539
 540         /* Enable interrupts on the HBA. */
 541         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
 542                                         dd->mmio + HOST_CTL);
 543         return rv;
 544 }
 545
 546 /*
 547  * Helper function for tag logging
 548  */
 549 static void print_tags(struct driver_data *dd,
 550                         char *msg,
 551                         unsigned long *tagbits,
 552                         int cnt)
 553 {
 554         unsigned char tagmap[128];
 555         int group, tagmap_len = 0;
 556
 557         memset(tagmap, 0, sizeof(tagmap));
 558         for (group = SLOTBITS_IN_LONGS; group > 0; group--)
 559                 tagmap_len = sprintf(tagmap + tagmap_len, "%016lX ",
 560                                                 tagbits[group-1]);
 561         dev_warn(&dd->pdev->dev,
 562                         "%d command(s) %s: tagmap [%s]", cnt, msg, tagmap);
 563 }
 564
 565 /*
 566  * Called periodically to see if any read/write commands are
 567  * taking too long to complete.
 568  *
 569  * @data Pointer to the PORT data structure.
 570  *
 571  * return value
 572  *      None
 573  */
 574 static void mtip_timeout_function(unsigned long int data)
 575 {
 576         struct mtip_port *port = (struct mtip_port *) data;
 577         struct host_to_dev_fis *fis;
 578         struct mtip_cmd *command;
 579         int tag, cmdto_cnt = 0;
 580         unsigned int bit, group;
 581         unsigned int num_command_slots;
 582         unsigned long to, tagaccum[SLOTBITS_IN_LONGS];
 583
 584         if (unlikely(!port))
 585                 return;
 586
 587         if (test_bit(MTIP_DDF_RESUME_BIT, &port->dd->dd_flag)) {
 588                 mod_timer(&port->cmd_timer,
 589                         jiffies + msecs_to_jiffies(30000));
 590                 return;
 591         }
 592         /* clear the tag accumulator */
 593         memset(tagaccum, 0, SLOTBITS_IN_LONGS * sizeof(long));
 594         num_command_slots = port->dd->slot_groups * 32;
 595
 596         for (tag = 0; tag < num_command_slots; tag++) {
 597                 /*
 598                  * Skip internal command slot as it has
 599                  * its own timeout mechanism
 600                  */
 601                 if (tag == MTIP_TAG_INTERNAL)
 602                         continue;
 603
 604                 if (atomic_read(&port->commands[tag].active) &&
 605                    (time_after(jiffies, port->commands[tag].comp_time))) {
 606                         group = tag >> 5;
 607                         bit = tag & 0x1F;
 608
 609                         command = &port->commands[tag];
 610                         fis = (struct host_to_dev_fis *) command->command;
 611
 612                         set_bit(tag, tagaccum);
 613                         cmdto_cnt++;
 614                         if (cmdto_cnt == 1)
 615                                 set_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags);
 616
 617                         /*
 618                          * Clear the completed bit. This should prevent
 619                          *  any interrupt handlers from trying to retire
 620                          *  the command.
 621                          */
 622                         writel(1 << bit, port->completed[group]);
 623
 624                         /* Call the async completion callback. */
 625                         if (likely(command->async_callback))
 626                                 command->async_callback(command->async_data,
 627                                                          -EIO);
 628                         command->async_callback = NULL;
 629                         command->comp_func = NULL;
 630
 631                         /* Unmap the DMA scatter list entries */
 632                         dma_unmap_sg(&port->dd->pdev->dev,
 633                                         command->sg,
 634                                         command->scatter_ents,
 635                                         command->direction);
 636
 637                         /*
 638                          * Clear the allocated bit and active tag for the
 639                          * command.
 640                          */
 641                         atomic_set(&port->commands[tag].active, 0);
 642                         release_slot(port, tag);
 643
 644                         up(&port->cmd_slot);
 645                 }
 646         }
 647
 648         if (cmdto_cnt) {
 649                 print_tags(port->dd, "timed out", tagaccum, cmdto_cnt);
 650                 if (!test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags)) {
 651                         mtip_device_reset(port->dd);
 652                         wake_up_interruptible(&port->svc_wait);
 653                 }
 654                 clear_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags);
 655         }
 656
 657         if (port->ic_pause_timer) {
 658                 to  = port->ic_pause_timer + msecs_to_jiffies(1000);
 659                 if (time_after(jiffies, to)) {
 660                         if (!test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags)) {
 661                                 port->ic_pause_timer = 0;
 662                                 clear_bit(MTIP_PF_SE_ACTIVE_BIT, &port->flags);
 663                                 clear_bit(MTIP_PF_DM_ACTIVE_BIT, &port->flags);
 664                                 clear_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
 665                                 wake_up_interruptible(&port->svc_wait);
 666                         }
 667
 668
 669                 }
 670         }
 671
 672         /* Restart the timer */
 673         mod_timer(&port->cmd_timer,
 674                 jiffies + msecs_to_jiffies(MTIP_TIMEOUT_CHECK_PERIOD));
 675 }
 676
 677 /*
 678  * IO completion function.
 679  *
 680  * This completion function is called by the driver ISR when a
 681  * command that was issued by the kernel completes. It first calls the
 682  * asynchronous completion function which normally calls back into the block
 683  * layer passing the asynchronous callback data, then unmaps the
 684  * scatter list associated with the completed command, and finally
 685  * clears the allocated bit associated with the completed command.
 686  *
 687  * @port   Pointer to the port data structure.
 688  * @tag    Tag of the command.
 689  * @data   Pointer to driver_data.
 690  * @status Completion status.
 691  *
 692  * return value
 693  *      None
 694  */
 695 static void mtip_async_complete(struct mtip_port *port,
 696                                 int tag,
 697                                 void *data,
 698                                 int status)
 699 {
 700         struct mtip_cmd *command;
 701         struct driver_data *dd = data;
 702         int cb_status = status ? -EIO : 0;
 703
 704         if (unlikely(!dd) || unlikely(!port))
 705                 return;
 706
 707         command = &port->commands[tag];
 708
 709         if (unlikely(status == PORT_IRQ_TF_ERR)) {
 710                 dev_warn(&port->dd->pdev->dev,
 711                         "Command tag %d failed due to TFE\n", tag);
 712         }
 713
 714         /* Upper layer callback */
 715         if (likely(command->async_callback))
 716                 command->async_callback(command->async_data, cb_status);
 717
 718         command->async_callback = NULL;
 719         command->comp_func = NULL;
 720
 721         /* Unmap the DMA scatter list entries */
 722         dma_unmap_sg(&dd->pdev->dev,
 723                 command->sg,
 724                 command->scatter_ents,
 725                 command->direction);
 726
 727         /* Clear the allocated and active bits for the command */
 728         atomic_set(&port->commands[tag].active, 0);
 729         release_slot(port, tag);
 730
 731         if (unlikely(command->unaligned))
 732                 up(&port->cmd_slot_unal);
 733         else
 734                 up(&port->cmd_slot);
 735 }
 736
 737 /*
 738  * Internal command completion callback function.
 739  *
 740  * This function is normally called by the driver ISR when an internal
 741  * command completed. This function signals the command completion by
 742  * calling complete().
 743  *
 744  * @port   Pointer to the port data structure.
 745  * @tag    Tag of the command that has completed.
 746  * @data   Pointer to a completion structure.
 747  * @status Completion status.
 748  *
 749  * return value
 750  *      None
 751  */
 752 static void mtip_completion(struct mtip_port *port,
 753                             int tag,
 754                             void *data,
 755                             int status)
 756 {
 757         struct mtip_cmd *command = &port->commands[tag];
 758         struct completion *waiting = data;
 759         if (unlikely(status == PORT_IRQ_TF_ERR))
 760                 dev_warn(&port->dd->pdev->dev,
 761                         "Internal command %d completed with TFE\n", tag);
 762
 763         command->async_callback = NULL;
 764         command->comp_func = NULL;
 765
 766         complete(waiting);
 767 }
 768
 769 static void mtip_null_completion(struct mtip_port *port,
 770                             int tag,
 771                             void *data,
 772                             int status)
 773 {
 774         return;
 775 }
 776
 777 static int mtip_read_log_page(struct mtip_port *port, u8 page, u16 *buffer,
 778                                 dma_addr_t buffer_dma, unsigned int sectors);
 779 static int mtip_get_smart_attr(struct mtip_port *port, unsigned int id,
 780                                                 struct smart_attr *attrib);
 781 /*
 782  * Handle an error.
 783  *
 784  * @dd Pointer to the DRIVER_DATA structure.
 785  *
 786  * return value
 787  *      None
 788  */
 789 static void mtip_handle_tfe(struct driver_data *dd)
 790 {
 791         int group, tag, bit, reissue, rv;
 792         struct mtip_port *port;
 793         struct mtip_cmd  *cmd;
 794         u32 completed;
 795         struct host_to_dev_fis *fis;
 796         unsigned long tagaccum[SLOTBITS_IN_LONGS];
 797         unsigned int cmd_cnt = 0;
 798         unsigned char *buf;
 799         char *fail_reason = NULL;
 800         int fail_all_ncq_write = 0, fail_all_ncq_cmds = 0;
 801
 802         dev_warn(&dd->pdev->dev, "Taskfile error\n");
 803
 804         port = dd->port;
 805
 806         /* Stop the timer to prevent command timeouts. */
 807         del_timer(&port->cmd_timer);
 808         set_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags);
 809
 810         if (test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags) &&
 811                         test_bit(MTIP_TAG_INTERNAL, port->allocated)) {
 812                 cmd = &port->commands[MTIP_TAG_INTERNAL];
 813                 dbg_printk(MTIP_DRV_NAME " TFE for the internal command\n");
 814
 815                 atomic_inc(&cmd->active); /* active > 1 indicates error */
 816                 if (cmd->comp_data && cmd->comp_func) {
 817                         cmd->comp_func(port, MTIP_TAG_INTERNAL,
 818                                         cmd->comp_data, PORT_IRQ_TF_ERR);
 819                 }
 820                 goto handle_tfe_exit;
 821         }
 822
 823         /* clear the tag accumulator */
 824         memset(tagaccum, 0, SLOTBITS_IN_LONGS * sizeof(long));
 825
 826         /* Loop through all the groups */
 827         for (group = 0; group < dd->slot_groups; group++) {
 828                 completed = readl(port->completed[group]);
 829
 830                 /* clear completed status register in the hardware.*/
 831                 writel(completed, port->completed[group]);
 832
 833                 /* Process successfully completed commands */
 834                 for (bit = 0; bit < 32 && completed; bit++) {
 835                         if (!(completed & (1<<bit)))
 836                                 continue;
 837                         tag = (group << 5) + bit;
 838
 839                         /* Skip the internal command slot */
 840                         if (tag == MTIP_TAG_INTERNAL)
 841                                 continue;
 842
 843                         cmd = &port->commands[tag];
 844                         if (likely(cmd->comp_func)) {
 845                                 set_bit(tag, tagaccum);
 846                                 cmd_cnt++;
 847                                 atomic_set(&cmd->active, 0);
 848                                 cmd->comp_func(port,
 849                                          tag,
 850                                          cmd->comp_data,
 851                                          0);
 852                         } else {
 853                                 dev_err(&port->dd->pdev->dev,
 854                                         "Missing completion func for tag %d",
 855                                         tag);
 856                                 if (mtip_check_surprise_removal(dd->pdev)) {
 857                                         mtip_command_cleanup(dd);
 858                                         /* don't proceed further */
 859                                         return;
 860                                 }
 861                         }
 862                 }
 863         }
 864
 865         print_tags(dd, "completed (TFE)", tagaccum, cmd_cnt);
 866
 867         /* Restart the port */
 868         mdelay(20);
 869         mtip_restart_port(port);
 870
 871         /* Trying to determine the cause of the error */
 872         rv = mtip_read_log_page(dd->port, ATA_LOG_SATA_NCQ,
 873                                 dd->port->log_buf,
 874                                 dd->port->log_buf_dma, 1);
 875         if (rv) {
 876                 dev_warn(&dd->pdev->dev,
 877                         "Error in READ LOG EXT (10h) command\n");
 878                 /* non-critical error, don't fail the load */
 879         } else {
 880                 buf = (unsigned char *)dd->port->log_buf;
 881                 if (buf[259] & 0x1) {
 882                         dev_info(&dd->pdev->dev,
 883                                 "Write protect bit is set.\n");
 884                         set_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag);
 885                         fail_all_ncq_write = 1;
 886                         fail_reason = "write protect";
 887                 }
 888                 if (buf[288] == 0xF7) {
 889                         dev_info(&dd->pdev->dev,
 890                                 "Exceeded Tmax, drive in thermal shutdown.\n");
 891                         set_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag);
 892                         fail_all_ncq_cmds = 1;
 893                         fail_reason = "thermal shutdown";
 894                 }
 895                 if (buf[288] == 0xBF) {
 896                         dev_info(&dd->pdev->dev,
 897                                 "Drive indicates rebuild has failed.\n");
 898                         fail_all_ncq_cmds = 1;
 899                         fail_reason = "rebuild failed";
 900                 }
 901         }
 902
 903         /* clear the tag accumulator */
 904         memset(tagaccum, 0, SLOTBITS_IN_LONGS * sizeof(long));
 905
 906         /* Loop through all the groups */
 907         for (group = 0; group < dd->slot_groups; group++) {
 908                 for (bit = 0; bit < 32; bit++) {
 909                         reissue = 1;
 910                         tag = (group << 5) + bit;
 911                         cmd = &port->commands[tag];
 912
 913                         /* If the active bit is set re-issue the command */
 914                         if (atomic_read(&cmd->active) == 0)
 915                                 continue;
 916
 917                         fis = (struct host_to_dev_fis *)cmd->command;
 918
 919                         /* Should re-issue? */
 920                         if (tag == MTIP_TAG_INTERNAL ||
 921                             fis->command == ATA_CMD_SET_FEATURES)
 922                                 reissue = 0;
 923                         else {
 924                                 if (fail_all_ncq_cmds ||
 925                                         (fail_all_ncq_write &&
 926                                         fis->command == ATA_CMD_FPDMA_WRITE)) {
 927                                         dev_warn(&dd->pdev->dev,
 928                                         "  Fail: %s w/tag %d [%s].\n",
 929                                         fis->command == ATA_CMD_FPDMA_WRITE ?
 930                                                 "write" : "read",
 931                                         tag,
 932                                         fail_reason != NULL ?
 933                                                 fail_reason : "unknown");
 934                                         atomic_set(&cmd->active, 0);
 935                                         if (cmd->comp_func) {
 936                                                 cmd->comp_func(port, tag,
 937                                                         cmd->comp_data,
 938                                                         -ENODATA);
 939                                         }
 940                                         continue;
 941                                 }
 942                         }
 943
 944                         /*
 945                          * First check if this command has
 946                          *  exceeded its retries.
 947                          */
 948                         if (reissue && (cmd->retries-- > 0)) {
 949
 950                                 set_bit(tag, tagaccum);
 951
 952                                 /* Re-issue the command. */
 953                                 mtip_issue_ncq_command(port, tag);
 954
 955                                 continue;
 956                         }
 957
 958                         /* Retire a command that will not be reissued */
 959                         dev_warn(&port->dd->pdev->dev,
 960                                 "retiring tag %d\n", tag);
 961                         atomic_set(&cmd->active, 0);
 962
 963                         if (cmd->comp_func)
 964                                 cmd->comp_func(
 965                                         port,
 966                                         tag,
 967                                         cmd->comp_data,
 968                                         PORT_IRQ_TF_ERR);
 969                         else
 970                                 dev_warn(&port->dd->pdev->dev,
 971                                         "Bad completion for tag %d\n",
 972                                         tag);
 973                 }
 974         }
 975         print_tags(dd, "reissued (TFE)", tagaccum, cmd_cnt);
 976
 977 handle_tfe_exit:
 978         /* clear eh_active */
 979         clear_bit(MTIP_PF_EH_ACTIVE_BIT, &port->flags);
 980         wake_up_interruptible(&port->svc_wait);
 981
 982         mod_timer(&port->cmd_timer,
 983                  jiffies + msecs_to_jiffies(MTIP_TIMEOUT_CHECK_PERIOD));
 984 }
 985
 986 /*
 987  * Handle a set device bits interrupt
 988  */
 989 static inline void mtip_workq_sdbfx(struct mtip_port *port, int group,
 990                                                         u32 completed)
 991 {
 992         struct driver_data *dd = port->dd;
 993         int tag, bit;
 994         struct mtip_cmd *command;
 995
 996         if (!completed) {
 997                 WARN_ON_ONCE(!completed);
 998                 return;
 999         }
1000         /* clear completed status register in the hardware.*/
1001         writel(completed, port->completed[group]);
1002
1003         /* Process completed commands. */
1004         for (bit = 0; (bit < 32) && completed; bit++) {
1005                 if (completed & 0x01) {
1006                         tag = (group << 5) | bit;
1007
1008                         /* skip internal command slot. */
1009                         if (unlikely(tag == MTIP_TAG_INTERNAL))
1010                                 continue;
1011
1012                         command = &port->commands[tag];
1013                         /* make internal callback */
1014                         if (likely(command->comp_func)) {
1015                                 command->comp_func(
1016                                         port,
1017                                         tag,
1018                                         command->comp_data,
1019                                         0);
1020                         } else {
1021                                 dev_warn(&dd->pdev->dev,
1022                                         "Null completion "
1023                                         "for tag %d",
1024                                         tag);
1025
1026                                 if (mtip_check_surprise_removal(
1027                                         dd->pdev)) {
1028                                         mtip_command_cleanup(dd);
1029                                         return;
1030                                 }
1031                         }
1032                 }
1033                 completed >>= 1;
1034         }
1035
1036         /* If last, re-enable interrupts */
1037         if (atomic_dec_return(&dd->irq_workers_active) == 0)
1038                 writel(0xffffffff, dd->mmio + HOST_IRQ_STAT);
1039 }
1040
1041 /*
1042  * Process legacy pio and d2h interrupts
1043  */
1044 static inline void mtip_process_legacy(struct driver_data *dd, u32 port_stat)
1045 {
1046         struct mtip_port *port = dd->port;
1047         struct mtip_cmd *cmd = &port->commands[MTIP_TAG_INTERNAL];
1048
1049         if (test_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags) &&
1050             (cmd != NULL) && !(readl(port->cmd_issue[MTIP_TAG_INTERNAL])
1051                 & (1 << MTIP_TAG_INTERNAL))) {
1052                 if (cmd->comp_func) {
1053                         cmd->comp_func(port,
1054                                 MTIP_TAG_INTERNAL,
1055                                 cmd->comp_data,
1056                                 0);
1057                         return;
1058                 }
1059         }
1060
1061         return;
1062 }
1063
1064 /*
1065  * Demux and handle errors
1066  */
1067 static inline void mtip_process_errors(struct driver_data *dd, u32 port_stat)
1068 {
1069         if (likely(port_stat & (PORT_IRQ_TF_ERR | PORT_IRQ_IF_ERR)))
1070                 mtip_handle_tfe(dd);
1071
1072         if (unlikely(port_stat & PORT_IRQ_CONNECT)) {
1073                 dev_warn(&dd->pdev->dev,
1074                         "Clearing PxSERR.DIAG.x\n");
1075                 writel((1 << 26), dd->port->mmio + PORT_SCR_ERR);
1076         }
1077
1078         if (unlikely(port_stat & PORT_IRQ_PHYRDY)) {
1079                 dev_warn(&dd->pdev->dev,
1080                         "Clearing PxSERR.DIAG.n\n");
1081                 writel((1 << 16), dd->port->mmio + PORT_SCR_ERR);
1082         }
1083
1084         if (unlikely(port_stat & ~PORT_IRQ_HANDLED)) {
1085                 dev_warn(&dd->pdev->dev,
1086                         "Port stat errors %x unhandled\n",
1087                         (port_stat & ~PORT_IRQ_HANDLED));
1088         }
1089 }
1090
1091 static inline irqreturn_t mtip_handle_irq(struct driver_data *data)
1092 {
1093         struct driver_data *dd = (struct driver_data *) data;
1094         struct mtip_port *port = dd->port;
1095         u32 hba_stat, port_stat;
1096         int rv = IRQ_NONE;
1097         int do_irq_enable = 1, i, workers;
1098         struct mtip_work *twork;
1099
1100         hba_stat = readl(dd->mmio + HOST_IRQ_STAT);
1101         if (hba_stat) {
1102                 rv = IRQ_HANDLED;
1103
1104                 /* Acknowledge the interrupt status on the port.*/
1105                 port_stat = readl(port->mmio + PORT_IRQ_STAT);
1106                 writel(port_stat, port->mmio + PORT_IRQ_STAT);
1107
1108                 /* Demux port status */
1109                 if (likely(port_stat & PORT_IRQ_SDB_FIS)) {
1110                         do_irq_enable = 0;
1111                         WARN_ON_ONCE(atomic_read(&dd->irq_workers_active) != 0);
1112
1113                         /* Start at 1: group zero is always local? */
1114                         for (i = 0, workers = 0; i < MTIP_MAX_SLOT_GROUPS;
1115                                                                         i++) {
1116                                 twork = &dd->work[i];
1117                                 twork->completed = readl(port->completed[i]);
1118                                 if (twork->completed)
1119                                         workers++;
1120                         }
1121
1122                         atomic_set(&dd->irq_workers_active, workers);
1123                         if (workers) {
1124                                 for (i = 1; i < MTIP_MAX_SLOT_GROUPS; i++) {
1125                                         twork = &dd->work[i];
1126                                         if (twork->completed)
1127                                                 queue_work_on(
1128                                                         twork->cpu_binding,
1129                                                         dd->isr_workq,
1130                                                         &twork->work);
1131                                 }
1132
1133                                 if (likely(dd->work[0].completed))
1134                                         mtip_workq_sdbfx(port, 0,
1135                                                         dd->work[0].completed);
1136
1137                         } else {
1138                                 /*
1139                                  * Chip quirk: SDB interrupt but nothing
1140                                  * to complete
1141                                  */
1142                                 do_irq_enable = 1;
1143                         }
1144                 }
1145
1146                 if (unlikely(port_stat & PORT_IRQ_ERR)) {
1147                         if (unlikely(mtip_check_surprise_removal(dd->pdev))) {
1148                                 mtip_command_cleanup(dd);
1149                                 /* don't proceed further */
1150                                 return IRQ_HANDLED;
1151                         }
1152                         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
1153                                                         &dd->dd_flag))
1154                                 return rv;
1155
1156                         mtip_process_errors(dd, port_stat & PORT_IRQ_ERR);
1157                 }
1158
1159                 if (unlikely(port_stat & PORT_IRQ_LEGACY))
1160                         mtip_process_legacy(dd, port_stat & PORT_IRQ_LEGACY);
1161         }
1162
1163         /* acknowledge interrupt */
1164         if (unlikely(do_irq_enable))
1165                 writel(hba_stat, dd->mmio + HOST_IRQ_STAT);
1166
1167         return rv;
1168 }
1169
1170 /*
1171  * HBA interrupt subroutine.
1172  *
1173  * @irq         IRQ number.
1174  * @instance    Pointer to the driver data structure.
1175  *
1176  * return value
1177  *      IRQ_HANDLED     A HBA interrupt was pending and handled.
1178  *      IRQ_NONE        This interrupt was not for the HBA.
1179  */
1180 static irqreturn_t mtip_irq_handler(int irq, void *instance)
1181 {
1182         struct driver_data *dd = instance;
1183
1184         return mtip_handle_irq(dd);
1185 }
1186
1187 static void mtip_issue_non_ncq_command(struct mtip_port *port, int tag)
1188 {
1189         atomic_set(&port->commands[tag].active, 1);
1190         writel(1 << MTIP_TAG_BIT(tag),
1191                 port->cmd_issue[MTIP_TAG_INDEX(tag)]);
1192 }
1193
1194 static bool mtip_pause_ncq(struct mtip_port *port,
1195                                 struct host_to_dev_fis *fis)
1196 {
1197         struct host_to_dev_fis *reply;
1198         unsigned long task_file_data;
1199
1200         reply = port->rxfis + RX_FIS_D2H_REG;
1201         task_file_data = readl(port->mmio+PORT_TFDATA);
1202
1203         if (fis->command == ATA_CMD_SEC_ERASE_UNIT)
1204                 clear_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1205
1206         if ((task_file_data & 1))
1207                 return false;
1208
1209         if (fis->command == ATA_CMD_SEC_ERASE_PREP) {
1210                 set_bit(MTIP_PF_SE_ACTIVE_BIT, &port->flags);
1211                 set_bit(MTIP_DDF_SEC_LOCK_BIT, &port->dd->dd_flag);
1212                 port->ic_pause_timer = jiffies;
1213                 return true;
1214         } else if ((fis->command == ATA_CMD_DOWNLOAD_MICRO) &&
1215                                         (fis->features == 0x03)) {
1216                 set_bit(MTIP_PF_DM_ACTIVE_BIT, &port->flags);
1217                 port->ic_pause_timer = jiffies;
1218                 return true;
1219         } else if ((fis->command == ATA_CMD_SEC_ERASE_UNIT) ||
1220                 ((fis->command == 0xFC) &&
1221                         (fis->features == 0x27 || fis->features == 0x72 ||
1222                          fis->features == 0x62 || fis->features == 0x26))) {
1223                 /* Com reset after secure erase or lowlevel format */
1224                 mtip_restart_port(port);
1225                 return false;
1226         }
1227
1228         return false;
1229 }
1230
1231 /*
1232  * Wait for port to quiesce
1233  *
1234  * @port    Pointer to port data structure
1235  * @timeout Max duration to wait (ms)
1236  *
1237  * return value
1238  *      0       Success
1239  *      -EBUSY  Commands still active
1240  */
1241 static int mtip_quiesce_io(struct mtip_port *port, unsigned long timeout)
1242 {
1243         unsigned long to;
1244         unsigned int n;
1245         unsigned int active = 1;
1246
1247         to = jiffies + msecs_to_jiffies(timeout);
1248         do {
1249                 if (test_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags) &&
1250                         test_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags)) {
1251                         msleep(20);
1252                         continue; /* svc thd is actively issuing commands */
1253                 }
1254                 if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
1255                         return -EFAULT;
1256                 /*
1257                  * Ignore s_active bit 0 of array element 0.
1258                  * This bit will always be set
1259                  */
1260                 active = readl(port->s_active[0]) & 0xFFFFFFFE;
1261                 for (n = 1; n < port->dd->slot_groups; n++)
1262                         active |= readl(port->s_active[n]);
1263
1264                 if (!active)
1265                         break;
1266
1267                 msleep(20);
1268         } while (time_before(jiffies, to));
1269
1270         return active ? -EBUSY : 0;
1271 }
1272
1273 /*
1274  * Execute an internal command and wait for the completion.
1275  *
1276  * @port    Pointer to the port data structure.
1277  * @fis     Pointer to the FIS that describes the command.
1278  * @fis_len  Length in WORDS of the FIS.
1279  * @buffer  DMA accessible for command data.
1280  * @buf_len  Length, in bytes, of the data buffer.
1281  * @opts    Command header options, excluding the FIS length
1282  *             and the number of PRD entries.
1283  * @timeout Time in ms to wait for the command to complete.
1284  *
1285  * return value
1286  *      0        Command completed successfully.
1287  *      -EFAULT  The buffer address is not correctly aligned.
1288  *      -EBUSY   Internal command or other IO in progress.
1289  *      -EAGAIN  Time out waiting for command to complete.
1290  */
1291 static int mtip_exec_internal_command(struct mtip_port *port,
1292                                         struct host_to_dev_fis *fis,
1293                                         int fis_len,
1294                                         dma_addr_t buffer,
1295                                         int buf_len,
1296                                         u32 opts,
1297                                         gfp_t atomic,
1298                                         unsigned long timeout)
1299 {
1300         struct mtip_cmd_sg *command_sg;
1301         DECLARE_COMPLETION_ONSTACK(wait);
1302         int rv = 0, ready2go = 1;
1303         struct mtip_cmd *int_cmd = &port->commands[MTIP_TAG_INTERNAL];
1304         unsigned long to;
1305         struct driver_data *dd = port->dd;
1306
1307         /* Make sure the buffer is 8 byte aligned. This is asic specific. */
1308         if (buffer & 0x00000007) {
1309                 dev_err(&dd->pdev->dev, "SG buffer is not 8 byte aligned\n");
1310                 return -EFAULT;
1311         }
1312
1313         to = jiffies + msecs_to_jiffies(timeout);
1314         do {
1315                 ready2go = !test_and_set_bit(MTIP_TAG_INTERNAL,
1316                                                 port->allocated);
1317                 if (ready2go)
1318                         break;
1319                 mdelay(100);
1320         } while (time_before(jiffies, to));
1321         if (!ready2go) {
1322                 dev_warn(&dd->pdev->dev,
1323                         "Internal cmd active. new cmd [%02X]\n", fis->command);
1324                 return -EBUSY;
1325         }
1326         set_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1327         port->ic_pause_timer = 0;
1328
1329         clear_bit(MTIP_PF_SE_ACTIVE_BIT, &port->flags);
1330         clear_bit(MTIP_PF_DM_ACTIVE_BIT, &port->flags);
1331
1332         if (atomic == GFP_KERNEL) {
1333                 if (fis->command != ATA_CMD_STANDBYNOW1) {
1334                         /* wait for io to complete if non atomic */
1335                         if (mtip_quiesce_io(port, 5000) < 0) {
1336                                 dev_warn(&dd->pdev->dev,
1337                                         "Failed to quiesce IO\n");
1338                                 release_slot(port, MTIP_TAG_INTERNAL);
1339                                 clear_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1340                                 wake_up_interruptible(&port->svc_wait);
1341                                 return -EBUSY;
1342                         }
1343                 }
1344
1345                 /* Set the completion function and data for the command. */
1346                 int_cmd->comp_data = &wait;
1347                 int_cmd->comp_func = mtip_completion;
1348
1349         } else {
1350                 /* Clear completion - we're going to poll */
1351                 int_cmd->comp_data = NULL;
1352                 int_cmd->comp_func = mtip_null_completion;
1353         }
1354
1355         /* Copy the command to the command table */
1356         memcpy(int_cmd->command, fis, fis_len*4);
1357
1358         /* Populate the SG list */
1359         int_cmd->command_header->opts =
1360                  __force_bit2int cpu_to_le32(opts | fis_len);
1361         if (buf_len) {
1362                 command_sg = int_cmd->command + AHCI_CMD_TBL_HDR_SZ;
1363
1364                 command_sg->info =
1365                         __force_bit2int cpu_to_le32((buf_len-1) & 0x3FFFFF);
1366                 command_sg->dba =
1367                         __force_bit2int cpu_to_le32(buffer & 0xFFFFFFFF);
1368                 command_sg->dba_upper =
1369                         __force_bit2int cpu_to_le32((buffer >> 16) >> 16);
1370
1371                 int_cmd->command_header->opts |=
1372                         __force_bit2int cpu_to_le32((1 << 16));
1373         }
1374
1375         /* Populate the command header */
1376         int_cmd->command_header->byte_count = 0;
1377
1378         /* Issue the command to the hardware */
1379         mtip_issue_non_ncq_command(port, MTIP_TAG_INTERNAL);
1380
1381         if (atomic == GFP_KERNEL) {
1382                 /* Wait for the command to complete or timeout. */
1383                 if (wait_for_completion_interruptible_timeout(
1384                                 &wait,
1385                                 msecs_to_jiffies(timeout)) <= 0) {
1386                         if (rv == -ERESTARTSYS) { /* interrupted */
1387                                 dev_err(&dd->pdev->dev,
1388                                         "Internal command [%02X] was interrupted after %lu ms\n",
1389                                         fis->command, timeout);
1390                                 rv = -EINTR;
1391                                 goto exec_ic_exit;
1392                         } else if (rv == 0) /* timeout */
1393                                 dev_err(&dd->pdev->dev,
1394                                         "Internal command did not complete [%02X] within timeout of  %lu ms\n",
1395                                         fis->command, timeout);
1396                         else
1397                                 dev_err(&dd->pdev->dev,
1398                                         "Internal command [%02X] wait returned code [%d] after %lu ms - unhandled\n",
1399                                         fis->command, rv, timeout);
1400
1401                         if (mtip_check_surprise_removal(dd->pdev) ||
1402                                 test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
1403                                                 &dd->dd_flag)) {
1404                                 dev_err(&dd->pdev->dev,
1405                                         "Internal command [%02X] wait returned due to SR\n",
1406                                         fis->command);
1407                                 rv = -ENXIO;
1408                                 goto exec_ic_exit;
1409                         }
1410                         mtip_device_reset(dd); /* recover from timeout issue */
1411                         rv = -EAGAIN;
1412                         goto exec_ic_exit;
1413                 }
1414         } else {
1415                 u32 hba_stat, port_stat;
1416
1417                 /* Spin for <timeout> checking if command still outstanding */
1418                 timeout = jiffies + msecs_to_jiffies(timeout);
1419                 while ((readl(port->cmd_issue[MTIP_TAG_INTERNAL])
1420                                 & (1 << MTIP_TAG_INTERNAL))
1421                                 && time_before(jiffies, timeout)) {
1422                         if (mtip_check_surprise_removal(dd->pdev)) {
1423                                 rv = -ENXIO;
1424                                 goto exec_ic_exit;
1425                         }
1426                         if ((fis->command != ATA_CMD_STANDBYNOW1) &&
1427                                 test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
1428                                                 &dd->dd_flag)) {
1429                                 rv = -ENXIO;
1430                                 goto exec_ic_exit;
1431                         }
1432                         port_stat = readl(port->mmio + PORT_IRQ_STAT);
1433                         if (!port_stat)
1434                                 continue;
1435
1436                         if (port_stat & PORT_IRQ_ERR) {
1437                                 dev_err(&dd->pdev->dev,
1438                                         "Internal command [%02X] failed\n",
1439                                         fis->command);
1440                                 mtip_device_reset(dd);
1441                                 rv = -EIO;
1442                                 goto exec_ic_exit;
1443                         } else {
1444                                 writel(port_stat, port->mmio + PORT_IRQ_STAT);
1445                                 hba_stat = readl(dd->mmio + HOST_IRQ_STAT);
1446                                 if (hba_stat)
1447                                         writel(hba_stat,
1448                                                 dd->mmio + HOST_IRQ_STAT);
1449                         }
1450                         break;
1451                 }
1452         }
1453
1454         if (readl(port->cmd_issue[MTIP_TAG_INTERNAL])
1455                         & (1 << MTIP_TAG_INTERNAL)) {
1456                 rv = -ENXIO;
1457                 if (!test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)) {
1458                         mtip_device_reset(dd);
1459                         rv = -EAGAIN;
1460                 }
1461         }
1462 exec_ic_exit:
1463         /* Clear the allocated and active bits for the internal command. */
1464         atomic_set(&int_cmd->active, 0);
1465         release_slot(port, MTIP_TAG_INTERNAL);
1466         if (rv >= 0 && mtip_pause_ncq(port, fis)) {
1467                 /* NCQ paused */
1468                 return rv;
1469         }
1470         clear_bit(MTIP_PF_IC_ACTIVE_BIT, &port->flags);
1471         wake_up_interruptible(&port->svc_wait);
1472
1473         return rv;
1474 }
1475
1476 /*
1477  * Byte-swap ATA ID strings.
1478  *
1479  * ATA identify data contains strings in byte-swapped 16-bit words.
1480  * They must be swapped (on all architectures) to be usable as C strings.
1481  * This function swaps bytes in-place.
1482  *
1483  * @buf The buffer location of the string
1484  * @len The number of bytes to swap
1485  *
1486  * return value
1487  *      None
1488  */
1489 static inline void ata_swap_string(u16 *buf, unsigned int len)
1490 {
1491         int i;
1492         for (i = 0; i < (len/2); i++)
1493                 be16_to_cpus(&buf[i]);
1494 }
1495
1496 static void mtip_set_timeout(struct driver_data *dd,
1497                                         struct host_to_dev_fis *fis,
1498                                         unsigned int *timeout, u8 erasemode)
1499 {
1500         switch (fis->command) {
1501         case ATA_CMD_DOWNLOAD_MICRO:
1502                 *timeout = 120000; /* 2 minutes */
1503                 break;
1504         case ATA_CMD_SEC_ERASE_UNIT:
1505         case 0xFC:
1506                 if (erasemode)
1507                         *timeout = ((*(dd->port->identify + 90) * 2) * 60000);
1508                 else
1509                         *timeout = ((*(dd->port->identify + 89) * 2) * 60000);
1510                 break;
1511         case ATA_CMD_STANDBYNOW1:
1512                 *timeout = 120000;  /* 2 minutes */
1513                 break;
1514         case 0xF7:
1515         case 0xFA:
1516                 *timeout = 60000;  /* 60 seconds */
1517                 break;
1518         case ATA_CMD_SMART:
1519                 *timeout = 15000;  /* 15 seconds */
1520                 break;
1521         default:
1522                 *timeout = MTIP_IOCTL_COMMAND_TIMEOUT_MS;
1523                 break;
1524         }
1525 }
1526
1527 /*
1528  * Request the device identity information.
1529  *
1530  * If a user space buffer is not specified, i.e. is NULL, the
1531  * identify information is still read from the drive and placed
1532  * into the identify data buffer (@e port->identify) in the
1533  * port data structure.
1534  * When the identify buffer contains valid identify information @e
1535  * port->identify_valid is non-zero.
1536  *
1537  * @port         Pointer to the port structure.
1538  * @user_buffer  A user space buffer where the identify data should be
1539  *                    copied.
1540  *
1541  * return value
1542  *      0       Command completed successfully.
1543  *      -EFAULT An error occurred while coping data to the user buffer.
1544  *      -1      Command failed.
1545  */
1546 static int mtip_get_identify(struct mtip_port *port, void __user *user_buffer)
1547 {
1548         int rv = 0;
1549         struct host_to_dev_fis fis;
1550
1551         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &port->dd->dd_flag))
1552                 return -EFAULT;
1553
1554         /* Build the FIS. */
1555         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1556         fis.type        = 0x27;
1557         fis.opts        = 1 << 7;
1558         fis.command     = ATA_CMD_ID_ATA;
1559
1560         /* Set the identify information as invalid. */
1561         port->identify_valid = 0;
1562
1563         /* Clear the identify information. */
1564         memset(port->identify, 0, sizeof(u16) * ATA_ID_WORDS);
1565
1566         /* Execute the command. */
1567         if (mtip_exec_internal_command(port,
1568                                 &fis,
1569                                 5,
1570                                 port->identify_dma,
1571                                 sizeof(u16) * ATA_ID_WORDS,
1572                                 0,
1573                                 GFP_KERNEL,
1574                                 MTIP_INTERNAL_COMMAND_TIMEOUT_MS)
1575                                 < 0) {
1576                 rv = -1;
1577                 goto out;
1578         }
1579
1580         /*
1581          * Perform any necessary byte-swapping.  Yes, the kernel does in fact
1582          * perform field-sensitive swapping on the string fields.
1583          * See the kernel use of ata_id_string() for proof of this.
1584          */
1585 #ifdef __LITTLE_ENDIAN
1586         ata_swap_string(port->identify + 27, 40);  /* model string*/
1587         ata_swap_string(port->identify + 23, 8);   /* firmware string*/
1588         ata_swap_string(port->identify + 10, 20);  /* serial# string*/
1589 #else
1590         {
1591                 int i;
1592                 for (i = 0; i < ATA_ID_WORDS; i++)
1593                         port->identify[i] = le16_to_cpu(port->identify[i]);
1594         }
1595 #endif
1596
1597 #ifdef MTIP_TRIM /* Disabling TRIM support temporarily */
1598         /* Demux ID.DRAT & ID.RZAT to determine trim support */
1599         if (port->identify[69] & (1 << 14) && port->identify[69] & (1 << 5))
1600                 port->dd->trim_supp = true;
1601         else
1602 #endif
1603                 port->dd->trim_supp = false;
1604
1605         /* Set the identify buffer as valid. */
1606         port->identify_valid = 1;
1607
1608         if (user_buffer) {
1609                 if (copy_to_user(
1610                         user_buffer,
1611                         port->identify,
1612                         ATA_ID_WORDS * sizeof(u16))) {
1613                         rv = -EFAULT;
1614                         goto out;
1615                 }
1616         }
1617
1618 out:
1619         return rv;
1620 }
1621
1622 /*
1623  * Issue a standby immediate command to the device.
1624  *
1625  * @port Pointer to the port structure.
1626  *
1627  * return value
1628  *      0       Command was executed successfully.
1629  *      -1      An error occurred while executing the command.
1630  */
1631 static int mtip_standby_immediate(struct mtip_port *port)
1632 {
1633         int rv;
1634         struct host_to_dev_fis  fis;
1635         unsigned long start;
1636         unsigned int timeout;
1637
1638         /* Build the FIS. */
1639         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1640         fis.type        = 0x27;
1641         fis.opts        = 1 << 7;
1642         fis.command     = ATA_CMD_STANDBYNOW1;
1643
1644         mtip_set_timeout(port->dd, &fis, &timeout, 0);
1645
1646         start = jiffies;
1647         rv = mtip_exec_internal_command(port,
1648                                         &fis,
1649                                         5,
1650                                         0,
1651                                         0,
1652                                         0,
1653                                         GFP_ATOMIC,
1654                                         timeout);
1655         dbg_printk(MTIP_DRV_NAME "Time taken to complete standby cmd: %d ms\n",
1656                         jiffies_to_msecs(jiffies - start));
1657         if (rv)
1658                 dev_warn(&port->dd->pdev->dev,
1659                         "STANDBY IMMEDIATE command failed.\n");
1660
1661         return rv;
1662 }
1663
1664 /*
1665  * Issue a READ LOG EXT command to the device.
1666  *
1667  * @port        pointer to the port structure.
1668  * @page        page number to fetch
1669  * @buffer      pointer to buffer
1670  * @buffer_dma  dma address corresponding to @buffer
1671  * @sectors     page length to fetch, in sectors
1672  *
1673  * return value
1674  *      @rv     return value from mtip_exec_internal_command()
1675  */
1676 static int mtip_read_log_page(struct mtip_port *port, u8 page, u16 *buffer,
1677                                 dma_addr_t buffer_dma, unsigned int sectors)
1678 {
1679         struct host_to_dev_fis fis;
1680
1681         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1682         fis.type        = 0x27;
1683         fis.opts        = 1 << 7;
1684         fis.command     = ATA_CMD_READ_LOG_EXT;
1685         fis.sect_count  = sectors & 0xFF;
1686         fis.sect_cnt_ex = (sectors >> 8) & 0xFF;
1687         fis.lba_low     = page;
1688         fis.lba_mid     = 0;
1689         fis.device      = ATA_DEVICE_OBS;
1690
1691         memset(buffer, 0, sectors * ATA_SECT_SIZE);
1692
1693         return mtip_exec_internal_command(port,
1694                                         &fis,
1695                                         5,
1696                                         buffer_dma,
1697                                         sectors * ATA_SECT_SIZE,
1698                                         0,
1699                                         GFP_ATOMIC,
1700                                         MTIP_INTERNAL_COMMAND_TIMEOUT_MS);
1701 }
1702
1703 /*
1704  * Issue a SMART READ DATA command to the device.
1705  *
1706  * @port        pointer to the port structure.
1707  * @buffer      pointer to buffer
1708  * @buffer_dma  dma address corresponding to @buffer
1709  *
1710  * return value
1711  *      @rv     return value from mtip_exec_internal_command()
1712  */
1713 static int mtip_get_smart_data(struct mtip_port *port, u8 *buffer,
1714                                         dma_addr_t buffer_dma)
1715 {
1716         struct host_to_dev_fis fis;
1717
1718         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1719         fis.type        = 0x27;
1720         fis.opts        = 1 << 7;
1721         fis.command     = ATA_CMD_SMART;
1722         fis.features    = 0xD0;
1723         fis.sect_count  = 1;
1724         fis.lba_mid     = 0x4F;
1725         fis.lba_hi      = 0xC2;
1726         fis.device      = ATA_DEVICE_OBS;
1727
1728         return mtip_exec_internal_command(port,
1729                                         &fis,
1730                                         5,
1731                                         buffer_dma,
1732                                         ATA_SECT_SIZE,
1733                                         0,
1734                                         GFP_ATOMIC,
1735                                         15000);
1736 }
1737
1738 /*
1739  * Get the value of a smart attribute
1740  *
1741  * @port        pointer to the port structure
1742  * @id          attribute number
1743  * @attrib      pointer to return attrib information corresponding to @id
1744  *
1745  * return value
1746  *      -EINVAL NULL buffer passed or unsupported attribute @id.
1747  *      -EPERM  Identify data not valid, SMART not supported or not enabled
1748  */
1749 static int mtip_get_smart_attr(struct mtip_port *port, unsigned int id,
1750                                                 struct smart_attr *attrib)
1751 {
1752         int rv, i;
1753         struct smart_attr *pattr;
1754
1755         if (!attrib)
1756                 return -EINVAL;
1757
1758         if (!port->identify_valid) {
1759                 dev_warn(&port->dd->pdev->dev, "IDENTIFY DATA not valid\n");
1760                 return -EPERM;
1761         }
1762         if (!(port->identify[82] & 0x1)) {
1763                 dev_warn(&port->dd->pdev->dev, "SMART not supported\n");
1764                 return -EPERM;
1765         }
1766         if (!(port->identify[85] & 0x1)) {
1767                 dev_warn(&port->dd->pdev->dev, "SMART not enabled\n");
1768                 return -EPERM;
1769         }
1770
1771         memset(port->smart_buf, 0, ATA_SECT_SIZE);
1772         rv = mtip_get_smart_data(port, port->smart_buf, port->smart_buf_dma);
1773         if (rv) {
1774                 dev_warn(&port->dd->pdev->dev, "Failed to ge SMART data\n");
1775                 return rv;
1776         }
1777
1778         pattr = (struct smart_attr *)(port->smart_buf + 2);
1779         for (i = 0; i < 29; i++, pattr++)
1780                 if (pattr->attr_id == id) {
1781                         memcpy(attrib, pattr, sizeof(struct smart_attr));
1782                         break;
1783                 }
1784
1785         if (i == 29) {
1786                 dev_warn(&port->dd->pdev->dev,
1787                         "Query for invalid SMART attribute ID\n");
1788                 rv = -EINVAL;
1789         }
1790
1791         return rv;
1792 }
1793
1794 /*
1795  * Trim unused sectors
1796  *
1797  * @dd          pointer to driver_data structure
1798  * @lba         starting lba
1799  * @len         # of 512b sectors to trim
1800  *
1801  * return value
1802  *      -ENOMEM         Out of dma memory
1803  *      -EINVAL         Invalid parameters passed in, trim not supported
1804  *      -EIO            Error submitting trim request to hw
1805  */
1806 static int mtip_send_trim(struct driver_data *dd, unsigned int lba,
1807                                 unsigned int len)
1808 {
1809         int i, rv = 0;
1810         u64 tlba, tlen, sect_left;
1811         struct mtip_trim_entry *buf;
1812         dma_addr_t dma_addr;
1813         struct host_to_dev_fis fis;
1814
1815         if (!len || dd->trim_supp == false)
1816                 return -EINVAL;
1817
1818         /* Trim request too big */
1819         WARN_ON(len > (MTIP_MAX_TRIM_ENTRY_LEN * MTIP_MAX_TRIM_ENTRIES));
1820
1821         /* Trim request not aligned on 4k boundary */
1822         WARN_ON(len % 8 != 0);
1823
1824         /* Warn if vu_trim structure is too big */
1825         WARN_ON(sizeof(struct mtip_trim) > ATA_SECT_SIZE);
1826
1827         /* Allocate a DMA buffer for the trim structure */
1828         buf = dmam_alloc_coherent(&dd->pdev->dev, ATA_SECT_SIZE, &dma_addr,
1829                                                                 GFP_KERNEL);
1830         if (!buf)
1831                 return -ENOMEM;
1832         memset(buf, 0, ATA_SECT_SIZE);
1833
1834         for (i = 0, sect_left = len, tlba = lba;
1835                         i < MTIP_MAX_TRIM_ENTRIES && sect_left;
1836                         i++) {
1837                 tlen = (sect_left >= MTIP_MAX_TRIM_ENTRY_LEN ?
1838                                         MTIP_MAX_TRIM_ENTRY_LEN :
1839                                         sect_left);
1840                 buf[i].lba = __force_bit2int cpu_to_le32(tlba);
1841                 buf[i].range = __force_bit2int cpu_to_le16(tlen);
1842                 tlba += tlen;
1843                 sect_left -= tlen;
1844         }
1845         WARN_ON(sect_left != 0);
1846
1847         /* Build the fis */
1848         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1849         fis.type       = 0x27;
1850         fis.opts       = 1 << 7;
1851         fis.command    = 0xfb;
1852         fis.features   = 0x60;
1853         fis.sect_count = 1;
1854         fis.device     = ATA_DEVICE_OBS;
1855
1856         if (mtip_exec_internal_command(dd->port,
1857                                         &fis,
1858                                         5,
1859                                         dma_addr,
1860                                         ATA_SECT_SIZE,
1861                                         0,
1862                                         GFP_KERNEL,
1863                                         MTIP_TRIM_TIMEOUT_MS) < 0)
1864                 rv = -EIO;
1865
1866         dmam_free_coherent(&dd->pdev->dev, ATA_SECT_SIZE, buf, dma_addr);
1867         return rv;
1868 }
1869
1870 /*
1871  * Get the drive capacity.
1872  *
1873  * @dd      Pointer to the device data structure.
1874  * @sectors Pointer to the variable that will receive the sector count.
1875  *
1876  * return value
1877  *      1 Capacity was returned successfully.
1878  *      0 The identify information is invalid.
1879  */
1880 static bool mtip_hw_get_capacity(struct driver_data *dd, sector_t *sectors)
1881 {
1882         struct mtip_port *port = dd->port;
1883         u64 total, raw0, raw1, raw2, raw3;
1884         raw0 = port->identify[100];
1885         raw1 = port->identify[101];
1886         raw2 = port->identify[102];
1887         raw3 = port->identify[103];
1888         total = raw0 | raw1<<16 | raw2<<32 | raw3<<48;
1889         *sectors = total;
1890         return (bool) !!port->identify_valid;
1891 }
1892
1893 /*
1894  * Display the identify command data.
1895  *
1896  * @port Pointer to the port data structure.
1897  *
1898  * return value
1899  *      None
1900  */
1901 static void mtip_dump_identify(struct mtip_port *port)
1902 {
1903         sector_t sectors;
1904         unsigned short revid;
1905         char cbuf[42];
1906
1907         if (!port->identify_valid)
1908                 return;
1909
1910         strlcpy(cbuf, (char *)(port->identify+10), 21);
1911         dev_info(&port->dd->pdev->dev,
1912                 "Serial No.: %s\n", cbuf);
1913
1914         strlcpy(cbuf, (char *)(port->identify+23), 9);
1915         dev_info(&port->dd->pdev->dev,
1916                 "Firmware Ver.: %s\n", cbuf);
1917
1918         strlcpy(cbuf, (char *)(port->identify+27), 41);
1919         dev_info(&port->dd->pdev->dev, "Model: %s\n", cbuf);
1920
1921         if (mtip_hw_get_capacity(port->dd, &sectors))
1922                 dev_info(&port->dd->pdev->dev,
1923                         "Capacity: %llu sectors (%llu MB)\n",
1924                          (u64)sectors,
1925                          ((u64)sectors) * ATA_SECT_SIZE >> 20);
1926
1927         pci_read_config_word(port->dd->pdev, PCI_REVISION_ID, &revid);
1928         switch (revid & 0xFF) {
1929         case 0x1:
1930                 strlcpy(cbuf, "A0", 3);
1931                 break;
1932         case 0x3:
1933                 strlcpy(cbuf, "A2", 3);
1934                 break;
1935         default:
1936                 strlcpy(cbuf, "?", 2);
1937                 break;
1938         }
1939         dev_info(&port->dd->pdev->dev,
1940                 "Card Type: %s\n", cbuf);
1941 }
1942
1943 /*
1944  * Map the commands scatter list into the command table.
1945  *
1946  * @command Pointer to the command.
1947  * @nents Number of scatter list entries.
1948  *
1949  * return value
1950  *      None
1951  */
1952 static inline void fill_command_sg(struct driver_data *dd,
1953                                 struct mtip_cmd *command,
1954                                 int nents)
1955 {
1956         int n;
1957         unsigned int dma_len;
1958         struct mtip_cmd_sg *command_sg;
1959         struct scatterlist *sg = command->sg;
1960
1961         command_sg = command->command + AHCI_CMD_TBL_HDR_SZ;
1962
1963         for (n = 0; n < nents; n++) {
1964                 dma_len = sg_dma_len(sg);
1965                 if (dma_len > 0x400000)
1966                         dev_err(&dd->pdev->dev,
1967                                 "DMA segment length truncated\n");
1968                 command_sg->info = __force_bit2int
1969                         cpu_to_le32((dma_len-1) & 0x3FFFFF);
1970                 command_sg->dba = __force_bit2int
1971                         cpu_to_le32(sg_dma_address(sg));
1972                 command_sg->dba_upper = __force_bit2int
1973                         cpu_to_le32((sg_dma_address(sg) >> 16) >> 16);
1974                 command_sg++;
1975                 sg++;
1976         }
1977 }
1978
1979 /*
1980  * @brief Execute a drive command.
1981  *
1982  * return value 0 The command completed successfully.
1983  * return value -1 An error occurred while executing the command.
1984  */
1985 static int exec_drive_task(struct mtip_port *port, u8 *command)
1986 {
1987         struct host_to_dev_fis  fis;
1988         struct host_to_dev_fis *reply = (port->rxfis + RX_FIS_D2H_REG);
1989
1990         /* Build the FIS. */
1991         memset(&fis, 0, sizeof(struct host_to_dev_fis));
1992         fis.type        = 0x27;
1993         fis.opts        = 1 << 7;
1994         fis.command     = command[0];
1995         fis.features    = command[1];
1996         fis.sect_count  = command[2];
1997         fis.sector      = command[3];
1998         fis.cyl_low     = command[4];
1999         fis.cyl_hi      = command[5];
2000         fis.device      = command[6] & ~0x10; /* Clear the dev bit*/
2001
2002         dbg_printk(MTIP_DRV_NAME " %s: User Command: cmd %x, feat %x, nsect %x, sect %x, lcyl %x, hcyl %x, sel %x\n",
2003                 __func__,
2004                 command[0],
2005                 command[1],
2006                 command[2],
2007                 command[3],
2008                 command[4],
2009                 command[5],
2010                 command[6]);
2011
2012         /* Execute the command. */
2013         if (mtip_exec_internal_command(port,
2014                                  &fis,
2015                                  5,
2016                                  0,
2017                                  0,
2018                                  0,
2019                                  GFP_KERNEL,
2020                                  MTIP_IOCTL_COMMAND_TIMEOUT_MS) < 0) {
2021                 return -1;
2022         }
2023
2024         command[0] = reply->command; /* Status*/
2025         command[1] = reply->features; /* Error*/
2026         command[4] = reply->cyl_low;
2027         command[5] = reply->cyl_hi;
2028
2029         dbg_printk(MTIP_DRV_NAME " %s: Completion Status: stat %x, err %x , cyl_lo %x cyl_hi %x\n",
2030                 __func__,
2031                 command[0],
2032                 command[1],
2033                 command[4],
2034                 command[5]);
2035
2036         return 0;
2037 }
2038
2039 /*
2040  * @brief Execute a drive command.
2041  *
2042  * @param port Pointer to the port data structure.
2043  * @param command Pointer to the user specified command parameters.
2044  * @param user_buffer Pointer to the user space buffer where read sector
2045  *                   data should be copied.
2046  *
2047  * return value 0 The command completed successfully.
2048  * return value -EFAULT An error occurred while copying the completion
2049  *                 data to the user space buffer.
2050  * return value -1 An error occurred while executing the command.
2051  */
2052 static int exec_drive_command(struct mtip_port *port, u8 *command,
2053                                 void __user *user_buffer)
2054 {
2055         struct host_to_dev_fis  fis;
2056         struct host_to_dev_fis *reply;
2057         u8 *buf = NULL;
2058         dma_addr_t dma_addr = 0;
2059         int rv = 0, xfer_sz = command[3];
2060
2061         if (xfer_sz) {
2062                 if (!user_buffer)
2063                         return -EFAULT;
2064
2065                 buf = dmam_alloc_coherent(&port->dd->pdev->dev,
2066                                 ATA_SECT_SIZE * xfer_sz,
2067                                 &dma_addr,
2068                                 GFP_KERNEL);
2069                 if (!buf) {
2070                         dev_err(&port->dd->pdev->dev,
2071                                 "Memory allocation failed (%d bytes)\n",
2072                                 ATA_SECT_SIZE * xfer_sz);
2073                         return -ENOMEM;
2074                 }
2075                 memset(buf, 0, ATA_SECT_SIZE * xfer_sz);
2076         }
2077
2078         /* Build the FIS. */
2079         memset(&fis, 0, sizeof(struct host_to_dev_fis));
2080         fis.type        = 0x27;
2081         fis.opts        = 1 << 7;
2082         fis.command     = command[0];
2083         fis.features    = command[2];
2084         fis.sect_count  = command[3];
2085         if (fis.command == ATA_CMD_SMART) {
2086                 fis.sector      = command[1];
2087                 fis.cyl_low     = 0x4F;
2088                 fis.cyl_hi      = 0xC2;
2089         }
2090
2091         if (xfer_sz)
2092                 reply = (port->rxfis + RX_FIS_PIO_SETUP);
2093         else
2094                 reply = (port->rxfis + RX_FIS_D2H_REG);
2095
2096         dbg_printk(MTIP_DRV_NAME
2097                 " %s: User Command: cmd %x, sect %x, "
2098                 "feat %x, sectcnt %x\n",
2099                 __func__,
2100                 command[0],
2101                 command[1],
2102                 command[2],
2103                 command[3]);
2104
2105         /* Execute the command. */
2106         if (mtip_exec_internal_command(port,
2107                                 &fis,
2108                                  5,
2109                                  (xfer_sz ? dma_addr : 0),
2110                                  (xfer_sz ? ATA_SECT_SIZE * xfer_sz : 0),
2111                                  0,
2112                                  GFP_KERNEL,
2113                                  MTIP_IOCTL_COMMAND_TIMEOUT_MS)
2114                                  < 0) {
2115                 rv = -EFAULT;
2116                 goto exit_drive_command;
2117         }
2118
2119         /* Collect the completion status. */
2120         command[0] = reply->command; /* Status*/
2121         command[1] = reply->features; /* Error*/
2122         command[2] = reply->sect_count;
2123
2124         dbg_printk(MTIP_DRV_NAME
2125                 " %s: Completion Status: stat %x, "
2126                 "err %x, nsect %x\n",
2127                 __func__,
2128                 command[0],
2129                 command[1],
2130                 command[2]);
2131
2132         if (xfer_sz) {
2133                 if (copy_to_user(user_buffer,
2134                                  buf,
2135                                  ATA_SECT_SIZE * command[3])) {
2136                         rv = -EFAULT;
2137                         goto exit_drive_command;
2138                 }
2139         }
2140 exit_drive_command:
2141         if (buf)
2142                 dmam_free_coherent(&port->dd->pdev->dev,
2143                                 ATA_SECT_SIZE * xfer_sz, buf, dma_addr);
2144         return rv;
2145 }
2146
2147 /*
2148  *  Indicates whether a command has a single sector payload.
2149  *
2150  *  @command passed to the device to perform the certain event.
2151  *  @features passed to the device to perform the certain event.
2152  *
2153  *  return value
2154  *      1       command is one that always has a single sector payload,
2155  *              regardless of the value in the Sector Count field.
2156  *      0       otherwise
2157  *
2158  */
2159 static unsigned int implicit_sector(unsigned char command,
2160                                     unsigned char features)
2161 {
2162         unsigned int rv = 0;
2163
2164         /* list of commands that have an implicit sector count of 1 */
2165         switch (command) {
2166         case ATA_CMD_SEC_SET_PASS:
2167         case ATA_CMD_SEC_UNLOCK:
2168         case ATA_CMD_SEC_ERASE_PREP:
2169         case ATA_CMD_SEC_ERASE_UNIT:
2170         case ATA_CMD_SEC_FREEZE_LOCK:
2171         case ATA_CMD_SEC_DISABLE_PASS:
2172         case ATA_CMD_PMP_READ:
2173         case ATA_CMD_PMP_WRITE:
2174                 rv = 1;
2175                 break;
2176         case ATA_CMD_SET_MAX:
2177                 if (features == ATA_SET_MAX_UNLOCK)
2178                         rv = 1;
2179                 break;
2180         case ATA_CMD_SMART:
2181                 if ((features == ATA_SMART_READ_VALUES) ||
2182                                 (features == ATA_SMART_READ_THRESHOLDS))
2183                         rv = 1;
2184                 break;
2185         case ATA_CMD_CONF_OVERLAY:
2186                 if ((features == ATA_DCO_IDENTIFY) ||
2187                                 (features == ATA_DCO_SET))
2188                         rv = 1;
2189                 break;
2190         }
2191         return rv;
2192 }
2193
2194 /*
2195  * Executes a taskfile
2196  * See ide_taskfile_ioctl() for derivation
2197  */
2198 static int exec_drive_taskfile(struct driver_data *dd,
2199                                void __user *buf,
2200                                ide_task_request_t *req_task,
2201                                int outtotal)
2202 {
2203         struct host_to_dev_fis  fis;
2204         struct host_to_dev_fis *reply;
2205         u8 *outbuf = NULL;
2206         u8 *inbuf = NULL;
2207         dma_addr_t outbuf_dma = 0;
2208         dma_addr_t inbuf_dma = 0;
2209         dma_addr_t dma_buffer = 0;
2210         int err = 0;
2211         unsigned int taskin = 0;
2212         unsigned int taskout = 0;
2213         u8 nsect = 0;
2214         unsigned int timeout;
2215         unsigned int force_single_sector;
2216         unsigned int transfer_size;
2217         unsigned long task_file_data;
2218         int intotal = outtotal + req_task->out_size;
2219         int erasemode = 0;
2220
2221         taskout = req_task->out_size;
2222         taskin = req_task->in_size;
2223         /* 130560 = 512 * 0xFF*/
2224         if (taskin > 130560 || taskout > 130560) {
2225                 err = -EINVAL;
2226                 goto abort;
2227         }
2228
2229         if (taskout) {
2230                 outbuf = kzalloc(taskout, GFP_KERNEL);
2231                 if (outbuf == NULL) {
2232                         err = -ENOMEM;
2233                         goto abort;
2234                 }
2235                 if (copy_from_user(outbuf, buf + outtotal, taskout)) {
2236                         err = -EFAULT;
2237                         goto abort;
2238                 }
2239                 outbuf_dma = pci_map_single(dd->pdev,
2240                                          outbuf,
2241                                          taskout,
2242                                          DMA_TO_DEVICE);
2243                 if (outbuf_dma == 0) {
2244                         err = -ENOMEM;
2245                         goto abort;
2246                 }
2247                 dma_buffer = outbuf_dma;
2248         }
2249
2250         if (taskin) {
2251                 inbuf = kzalloc(taskin, GFP_KERNEL);
2252                 if (inbuf == NULL) {
2253                         err = -ENOMEM;
2254                         goto abort;
2255                 }
2256
2257                 if (copy_from_user(inbuf, buf + intotal, taskin)) {
2258                         err = -EFAULT;
2259                         goto abort;
2260                 }
2261                 inbuf_dma = pci_map_single(dd->pdev,
2262                                          inbuf,
2263                                          taskin, DMA_FROM_DEVICE);
2264                 if (inbuf_dma == 0) {
2265                         err = -ENOMEM;
2266                         goto abort;
2267                 }
2268                 dma_buffer = inbuf_dma;
2269         }
2270
2271         /* only supports PIO and non-data commands from this ioctl. */
2272         switch (req_task->data_phase) {
2273         case TASKFILE_OUT:
2274                 nsect = taskout / ATA_SECT_SIZE;
2275                 reply = (dd->port->rxfis + RX_FIS_PIO_SETUP);
2276                 break;
2277         case TASKFILE_IN:
2278                 reply = (dd->port->rxfis + RX_FIS_PIO_SETUP);
2279                 break;
2280         case TASKFILE_NO_DATA:
2281                 reply = (dd->port->rxfis + RX_FIS_D2H_REG);
2282                 break;
2283         default:
2284                 err = -EINVAL;
2285                 goto abort;
2286         }
2287
2288         /* Build the FIS. */
2289         memset(&fis, 0, sizeof(struct host_to_dev_fis));
2290
2291         fis.type        = 0x27;
2292         fis.opts        = 1 << 7;
2293         fis.command     = req_task->io_ports[7];
2294         fis.features    = req_task->io_ports[1];
2295         fis.sect_count  = req_task->io_ports[2];
2296         fis.lba_low     = req_task->io_ports[3];
2297         fis.lba_mid     = req_task->io_ports[4];
2298         fis.lba_hi      = req_task->io_ports[5];
2299          /* Clear the dev bit*/
2300         fis.device      = req_task->io_ports[6] & ~0x10;
2301
2302         if ((req_task->in_flags.all == 0) && (req_task->out_flags.all & 1)) {
2303                 req_task->in_flags.all  =
2304                         IDE_TASKFILE_STD_IN_FLAGS |
2305                         (IDE_HOB_STD_IN_FLAGS << 8);
2306                 fis.lba_low_ex          = req_task->hob_ports[3];
2307                 fis.lba_mid_ex          = req_task->hob_ports[4];
2308                 fis.lba_hi_ex           = req_task->hob_ports[5];
2309                 fis.features_ex         = req_task->hob_ports[1];
2310                 fis.sect_cnt_ex         = req_task->hob_ports[2];
2311
2312         } else {
2313                 req_task->in_flags.all = IDE_TASKFILE_STD_IN_FLAGS;
2314         }
2315
2316         force_single_sector = implicit_sector(fis.command, fis.features);
2317
2318         if ((taskin || taskout) && (!fis.sect_count)) {
2319                 if (nsect)
2320                         fis.sect_count = nsect;
2321                 else {
2322                         if (!force_single_sector) {
2323                                 dev_warn(&dd->pdev->dev,
2324                                         "data movement but "
2325                                         "sect_count is 0\n");
2326                                         err = -EINVAL;
2327                                         goto abort;
2328                         }
2329                 }
2330         }
2331
2332         dbg_printk(MTIP_DRV_NAME
2333                 " %s: cmd %x, feat %x, nsect %x,"
2334                 " sect/lbal %x, lcyl/lbam %x, hcyl/lbah %x,"
2335                 " head/dev %x\n",
2336                 __func__,
2337                 fis.command,
2338                 fis.features,
2339                 fis.sect_count,
2340                 fis.lba_low,
2341                 fis.lba_mid,
2342                 fis.lba_hi,
2343                 fis.device);
2344
2345         /* check for erase mode support during secure erase.*/
2346         if ((fis.command == ATA_CMD_SEC_ERASE_UNIT) && outbuf &&
2347                                         (outbuf[0] & MTIP_SEC_ERASE_MODE)) {
2348                 erasemode = 1;
2349         }
2350
2351         mtip_set_timeout(dd, &fis, &timeout, erasemode);
2352
2353         /* Determine the correct transfer size.*/
2354         if (force_single_sector)
2355                 transfer_size = ATA_SECT_SIZE;
2356         else
2357                 transfer_size = ATA_SECT_SIZE * fis.sect_count;
2358
2359         /* Execute the command.*/
2360         if (mtip_exec_internal_command(dd->port,
2361                                  &fis,
2362                                  5,
2363                                  dma_buffer,
2364                                  transfer_size,
2365                                  0,
2366                                  GFP_KERNEL,
2367                                  timeout) < 0) {
2368                 err = -EIO;
2369                 goto abort;
2370         }
2371
2372         task_file_data = readl(dd->port->mmio+PORT_TFDATA);
2373
2374         if ((req_task->data_phase == TASKFILE_IN) && !(task_file_data & 1)) {
2375                 reply = dd->port->rxfis + RX_FIS_PIO_SETUP;
2376                 req_task->io_ports[7] = reply->control;
2377         } else {
2378                 reply = dd->port->rxfis + RX_FIS_D2H_REG;
2379                 req_task->io_ports[7] = reply->command;
2380         }
2381
2382         /* reclaim the DMA buffers.*/
2383         if (inbuf_dma)
2384                 pci_unmap_single(dd->pdev, inbuf_dma,
2385                         taskin, DMA_FROM_DEVICE);
2386         if (outbuf_dma)
2387                 pci_unmap_single(dd->pdev, outbuf_dma,
2388                         taskout, DMA_TO_DEVICE);
2389         inbuf_dma  = 0;
2390         outbuf_dma = 0;
2391
2392         /* return the ATA registers to the caller.*/
2393         req_task->io_ports[1] = reply->features;
2394         req_task->io_ports[2] = reply->sect_count;
2395         req_task->io_ports[3] = reply->lba_low;
2396         req_task->io_ports[4] = reply->lba_mid;
2397         req_task->io_ports[5] = reply->lba_hi;
2398         req_task->io_ports[6] = reply->device;
2399
2400         if (req_task->out_flags.all & 1)  {
2401
2402                 req_task->hob_ports[3] = reply->lba_low_ex;
2403                 req_task->hob_ports[4] = reply->lba_mid_ex;
2404                 req_task->hob_ports[5] = reply->lba_hi_ex;
2405                 req_task->hob_ports[1] = reply->features_ex;
2406                 req_task->hob_ports[2] = reply->sect_cnt_ex;
2407         }
2408         dbg_printk(MTIP_DRV_NAME
2409                 " %s: Completion: stat %x,"
2410                 "err %x, sect_cnt %x, lbalo %x,"
2411                 "lbamid %x, lbahi %x, dev %x\n",
2412                 __func__,
2413                 req_task->io_ports[7],
2414                 req_task->io_ports[1],
2415                 req_task->io_ports[2],
2416                 req_task->io_ports[3],
2417                 req_task->io_ports[4],
2418                 req_task->io_ports[5],
2419                 req_task->io_ports[6]);
2420
2421         if (taskout) {
2422                 if (copy_to_user(buf + outtotal, outbuf, taskout)) {
2423                         err = -EFAULT;
2424                         goto abort;
2425                 }
2426         }
2427         if (taskin) {
2428                 if (copy_to_user(buf + intotal, inbuf, taskin)) {
2429                         err = -EFAULT;
2430                         goto abort;
2431                 }
2432         }
2433 abort:
2434         if (inbuf_dma)
2435                 pci_unmap_single(dd->pdev, inbuf_dma,
2436                                         taskin, DMA_FROM_DEVICE);
2437         if (outbuf_dma)
2438                 pci_unmap_single(dd->pdev, outbuf_dma,
2439                                         taskout, DMA_TO_DEVICE);
2440         kfree(outbuf);
2441         kfree(inbuf);
2442
2443         return err;
2444 }
2445
2446 /*
2447  * Handle IOCTL calls from the Block Layer.
2448  *
2449  * This function is called by the Block Layer when it receives an IOCTL
2450  * command that it does not understand. If the IOCTL command is not supported
2451  * this function returns -ENOTTY.
2452  *
2453  * @dd  Pointer to the driver data structure.
2454  * @cmd IOCTL command passed from the Block Layer.
2455  * @arg IOCTL argument passed from the Block Layer.
2456  *
2457  * return value
2458  *      0       The IOCTL completed successfully.
2459  *      -ENOTTY The specified command is not supported.
2460  *      -EFAULT An error occurred copying data to a user space buffer.
2461  *      -EIO    An error occurred while executing the command.
2462  */
2463 static int mtip_hw_ioctl(struct driver_data *dd, unsigned int cmd,
2464                          unsigned long arg)
2465 {
2466         switch (cmd) {
2467         case HDIO_GET_IDENTITY:
2468         {
2469                 if (copy_to_user((void __user *)arg, dd->port->identify,
2470                                                 sizeof(u16) * ATA_ID_WORDS))
2471                         return -EFAULT;
2472                 break;
2473         }
2474         case HDIO_DRIVE_CMD:
2475         {
2476                 u8 drive_command[4];
2477
2478                 /* Copy the user command info to our buffer. */
2479                 if (copy_from_user(drive_command,
2480                                          (void __user *) arg,
2481                                          sizeof(drive_command)))
2482                         return -EFAULT;
2483
2484                 /* Execute the drive command. */
2485                 if (exec_drive_command(dd->port,
2486                                          drive_command,
2487                                          (void __user *) (arg+4)))
2488                         return -EIO;
2489
2490                 /* Copy the status back to the users buffer. */
2491                 if (copy_to_user((void __user *) arg,
2492                                          drive_command,
2493                                          sizeof(drive_command)))
2494                         return -EFAULT;
2495
2496                 break;
2497         }
2498         case HDIO_DRIVE_TASK:
2499         {
2500                 u8 drive_command[7];
2501
2502                 /* Copy the user command info to our buffer. */
2503                 if (copy_from_user(drive_command,
2504                                          (void __user *) arg,
2505                                          sizeof(drive_command)))
2506                         return -EFAULT;
2507
2508                 /* Execute the drive command. */
2509                 if (exec_drive_task(dd->port, drive_command))
2510                         return -EIO;
2511
2512                 /* Copy the status back to the users buffer. */
2513                 if (copy_to_user((void __user *) arg,
2514                                          drive_command,
2515                                          sizeof(drive_command)))
2516                         return -EFAULT;
2517
2518                 break;
2519         }
2520         case HDIO_DRIVE_TASKFILE: {
2521                 ide_task_request_t req_task;
2522                 int ret, outtotal;
2523
2524                 if (copy_from_user(&req_task, (void __user *) arg,
2525                                         sizeof(req_task)))
2526                         return -EFAULT;
2527
2528                 outtotal = sizeof(req_task);
2529
2530                 ret = exec_drive_taskfile(dd, (void __user *) arg,
2531                                                 &req_task, outtotal);
2532
2533                 if (copy_to_user((void __user *) arg, &req_task,
2534                                                         sizeof(req_task)))
2535                         return -EFAULT;
2536
2537                 return ret;
2538         }
2539
2540         default:
2541                 return -EINVAL;
2542         }
2543         return 0;
2544 }
2545
2546 /*
2547  * Submit an IO to the hw
2548  *
2549  * This function is called by the block layer to issue an io
2550  * to the device. Upon completion, the callback function will
2551  * be called with the data parameter passed as the callback data.
2552  *
2553  * @dd       Pointer to the driver data structure.
2554  * @start    First sector to read.
2555  * @nsect    Number of sectors to read.
2556  * @nents    Number of entries in scatter list for the read command.
2557  * @tag      The tag of this read command.
2558  * @callback Pointer to the function that should be called
2559  *           when the read completes.
2560  * @data     Callback data passed to the callback function
2561  *           when the read completes.
2562  * @dir      Direction (read or write)
2563  *
2564  * return value
2565  *      None
2566  */
2567 static void mtip_hw_submit_io(struct driver_data *dd, sector_t sector,
2568                               int nsect, int nents, int tag, void *callback,
2569                               void *data, int dir, int unaligned)
2570 {
2571         struct host_to_dev_fis  *fis;
2572         struct mtip_port *port = dd->port;
2573         struct mtip_cmd *command = &port->commands[tag];
2574         int dma_dir = (dir == READ) ? DMA_FROM_DEVICE : DMA_TO_DEVICE;
2575         u64 start = sector;
2576
2577         /* Map the scatter list for DMA access */
2578         nents = dma_map_sg(&dd->pdev->dev, command->sg, nents, dma_dir);
2579
2580         command->scatter_ents = nents;
2581
2582         command->unaligned = unaligned;
2583         /*
2584          * The number of retries for this command before it is
2585          * reported as a failure to the upper layers.
2586          */
2587         command->retries = MTIP_MAX_RETRIES;
2588
2589         /* Fill out fis */
2590         fis = command->command;
2591         fis->type        = 0x27;
2592         fis->opts        = 1 << 7;
2593         fis->command     =
2594                 (dir == READ ? ATA_CMD_FPDMA_READ : ATA_CMD_FPDMA_WRITE);
2595         fis->lba_low     = start & 0xFF;
2596         fis->lba_mid     = (start >> 8) & 0xFF;
2597         fis->lba_hi      = (start >> 16) & 0xFF;
2598         fis->lba_low_ex  = (start >> 24) & 0xFF;
2599         fis->lba_mid_ex  = (start >> 32) & 0xFF;
2600         fis->lba_hi_ex   = (start >> 40) & 0xFF;
2601         fis->device      = 1 << 6;
2602         fis->features    = nsect & 0xFF;
2603         fis->features_ex = (nsect >> 8) & 0xFF;
2604         fis->sect_count  = ((tag << 3) | (tag >> 5));
2605         fis->sect_cnt_ex = 0;
2606         fis->control     = 0;
2607         fis->res2        = 0;
2608         fis->res3        = 0;
2609         fill_command_sg(dd, command, nents);
2610
2611         if (unaligned)
2612                 fis->device |= 1 << 7;
2613
2614         /* Populate the command header */
2615         command->command_header->opts =
2616                         __force_bit2int cpu_to_le32(
2617                                 (nents << 16) | 5 | AHCI_CMD_PREFETCH);
2618         command->command_header->byte_count = 0;
2619
2620         /*
2621          * Set the completion function and data for the command
2622          * within this layer.
2623          */
2624         command->comp_data = dd;
2625         command->comp_func = mtip_async_complete;
2626         command->direction = dma_dir;
2627
2628         /*
2629          * Set the completion function and data for the command passed
2630          * from the upper layer.
2631          */
2632         command->async_data = data;
2633         command->async_callback = callback;
2634
2635         /*
2636          * To prevent this command from being issued
2637          * if an internal command is in progress or error handling is active.
2638          */
2639         if (port->flags & MTIP_PF_PAUSE_IO) {
2640                 set_bit(tag, port->cmds_to_issue);
2641                 set_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags);
2642                 return;
2643         }
2644
2645         /* Issue the command to the hardware */
2646         mtip_issue_ncq_command(port, tag);
2647
2648         return;
2649 }
2650
2651 /*
2652  * Release a command slot.
2653  *
2654  * @dd  Pointer to the driver data structure.
2655  * @tag Slot tag
2656  *
2657  * return value
2658  *      None
2659  */
2660 static void mtip_hw_release_scatterlist(struct driver_data *dd, int tag,
2661                                                                 int unaligned)
2662 {
2663         struct semaphore *sem = unaligned ? &dd->port->cmd_slot_unal :
2664                                                         &dd->port->cmd_slot;
2665         release_slot(dd->port, tag);
2666         up(sem);
2667 }
2668
2669 /*
2670  * Obtain a command slot and return its associated scatter list.
2671  *
2672  * @dd  Pointer to the driver data structure.
2673  * @tag Pointer to an int that will receive the allocated command
2674  *            slot tag.
2675  *
2676  * return value
2677  *      Pointer to the scatter list for the allocated command slot
2678  *      or NULL if no command slots are available.
2679  */
2680 static struct scatterlist *mtip_hw_get_scatterlist(struct driver_data *dd,
2681                                                    int *tag, int unaligned)
2682 {
2683         struct semaphore *sem = unaligned ? &dd->port->cmd_slot_unal :
2684                                                         &dd->port->cmd_slot;
2685
2686         /*
2687          * It is possible that, even with this semaphore, a thread
2688          * may think that no command slots are available. Therefore, we
2689          * need to make an attempt to get_slot().
2690          */
2691         down(sem);
2692         *tag = get_slot(dd->port);
2693
2694         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag))) {
2695                 up(sem);
2696                 return NULL;
2697         }
2698         if (unlikely(*tag < 0)) {
2699                 up(sem);
2700                 return NULL;
2701         }
2702
2703         return dd->port->commands[*tag].sg;
2704 }
2705
2706 /*
2707  * Sysfs status dump.
2708  *
2709  * @dev  Pointer to the device structure, passed by the kernrel.
2710  * @attr Pointer to the device_attribute structure passed by the kernel.
2711  * @buf  Pointer to the char buffer that will receive the stats info.
2712  *
2713  * return value
2714  *      The size, in bytes, of the data copied into buf.
2715  */
2716 static ssize_t mtip_hw_show_status(struct device *dev,
2717                                 struct device_attribute *attr,
2718                                 char *buf)
2719 {
2720         struct driver_data *dd = dev_to_disk(dev)->private_data;
2721         int size = 0;
2722
2723         if (test_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag))
2724                 size += sprintf(buf, "%s", "thermal_shutdown\n");
2725         else if (test_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag))
2726                 size += sprintf(buf, "%s", "write_protect\n");
2727         else
2728                 size += sprintf(buf, "%s", "online\n");
2729
2730         return size;
2731 }
2732
2733 static DEVICE_ATTR(status, S_IRUGO, mtip_hw_show_status, NULL);
2734
2735 /* debugsfs entries */
2736
2737 static ssize_t show_device_status(struct device_driver *drv, char *buf)
2738 {
2739         int size = 0;
2740         struct driver_data *dd, *tmp;
2741         unsigned long flags;
2742         char id_buf[42];
2743         u16 status = 0;
2744
2745         spin_lock_irqsave(&dev_lock, flags);
2746         size += sprintf(&buf[size], "Devices Present:\n");
2747         list_for_each_entry_safe(dd, tmp, &online_list, online_list) {
2748                 if (dd->pdev) {
2749                         if (dd->port &&
2750                             dd->port->identify &&
2751                             dd->port->identify_valid) {
2752                                 strlcpy(id_buf,
2753                                         (char *) (dd->port->identify + 10), 21);
2754                                 status = *(dd->port->identify + 141);
2755                         } else {
2756                                 memset(id_buf, 0, 42);
2757                                 status = 0;
2758                         }
2759
2760                         if (dd->port &&
2761                             test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags)) {
2762                                 size += sprintf(&buf[size],
2763                                         " device %s %s (ftl rebuild %d %%)\n",
2764                                         dev_name(&dd->pdev->dev),
2765                                         id_buf,
2766                                         status);
2767                         } else {
2768                                 size += sprintf(&buf[size],
2769                                         " device %s %s\n",
2770                                         dev_name(&dd->pdev->dev),
2771                                         id_buf);
2772                         }
2773                 }
2774         }
2775
2776         size += sprintf(&buf[size], "Devices Being Removed:\n");
2777         list_for_each_entry_safe(dd, tmp, &removing_list, remove_list) {
2778                 if (dd->pdev) {
2779                         if (dd->port &&
2780                             dd->port->identify &&
2781                             dd->port->identify_valid) {
2782                                 strlcpy(id_buf,
2783                                         (char *) (dd->port->identify+10), 21);
2784                                 status = *(dd->port->identify + 141);
2785                         } else {
2786                                 memset(id_buf, 0, 42);
2787                                 status = 0;
2788                         }
2789
2790                         if (dd->port &&
2791                             test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags)) {
2792                                 size += sprintf(&buf[size],
2793                                         " device %s %s (ftl rebuild %d %%)\n",
2794                                         dev_name(&dd->pdev->dev),
2795                                         id_buf,
2796                                         status);
2797                         } else {
2798                                 size += sprintf(&buf[size],
2799                                         " device %s %s\n",
2800                                         dev_name(&dd->pdev->dev),
2801                                         id_buf);
2802                         }
2803                 }
2804         }
2805         spin_unlock_irqrestore(&dev_lock, flags);
2806
2807         return size;
2808 }
2809
2810 static ssize_t mtip_hw_read_device_status(struct file *f, char __user *ubuf,
2811                                                 size_t len, loff_t *offset)
2812 {
2813         int size = *offset;
2814         char buf[MTIP_DFS_MAX_BUF_SIZE];
2815
2816         if (!len || *offset)
2817                 return 0;
2818
2819         size += show_device_status(NULL, buf);
2820
2821         *offset = size <= len ? size : len;
2822         size = copy_to_user(ubuf, buf, *offset);
2823         if (size)
2824                 return -EFAULT;
2825
2826         return *offset;
2827 }
2828
2829 static ssize_t mtip_hw_read_registers(struct file *f, char __user *ubuf,
2830                                   size_t len, loff_t *offset)
2831 {
2832         struct driver_data *dd =  (struct driver_data *)f->private_data;
2833         char buf[MTIP_DFS_MAX_BUF_SIZE];
2834         u32 group_allocated;
2835         int size = *offset;
2836         int n;
2837
2838         if (!len || size)
2839                 return 0;
2840
2841         size += sprintf(&buf[size], "H/ S ACTive      : [ 0x");
2842
2843         for (n = dd->slot_groups-1; n >= 0; n--)
2844                 size += sprintf(&buf[size], "%08X ",
2845                                          readl(dd->port->s_active[n]));
2846
2847         size += sprintf(&buf[size], "]\n");
2848         size += sprintf(&buf[size], "H/ Command Issue : [ 0x");
2849
2850         for (n = dd->slot_groups-1; n >= 0; n--)
2851                 size += sprintf(&buf[size], "%08X ",
2852                                         readl(dd->port->cmd_issue[n]));
2853
2854         size += sprintf(&buf[size], "]\n");
2855         size += sprintf(&buf[size], "H/ Completed     : [ 0x");
2856
2857         for (n = dd->slot_groups-1; n >= 0; n--)
2858                 size += sprintf(&buf[size], "%08X ",
2859                                 readl(dd->port->completed[n]));
2860
2861         size += sprintf(&buf[size], "]\n");
2862         size += sprintf(&buf[size], "H/ PORT IRQ STAT : [ 0x%08X ]\n",
2863                                 readl(dd->port->mmio + PORT_IRQ_STAT));
2864         size += sprintf(&buf[size], "H/ HOST IRQ STAT : [ 0x%08X ]\n",
2865                                 readl(dd->mmio + HOST_IRQ_STAT));
2866         size += sprintf(&buf[size], "\n");
2867
2868         size += sprintf(&buf[size], "L/ Allocated     : [ 0x");
2869
2870         for (n = dd->slot_groups-1; n >= 0; n--) {
2871                 if (sizeof(long) > sizeof(u32))
2872                         group_allocated =
2873                                 dd->port->allocated[n/2] >> (32*(n&1));
2874                 else
2875                         group_allocated = dd->port->allocated[n];
2876                 size += sprintf(&buf[size], "%08X ", group_allocated);
2877         }
2878         size += sprintf(&buf[size], "]\n");
2879
2880         size += sprintf(&buf[size], "L/ Commands in Q : [ 0x");
2881
2882         for (n = dd->slot_groups-1; n >= 0; n--) {
2883                 if (sizeof(long) > sizeof(u32))
2884                         group_allocated =
2885                                 dd->port->cmds_to_issue[n/2] >> (32*(n&1));
2886                 else
2887                         group_allocated = dd->port->cmds_to_issue[n];
2888                 size += sprintf(&buf[size], "%08X ", group_allocated);
2889         }
2890         size += sprintf(&buf[size], "]\n");
2891
2892         *offset = size <= len ? size : len;
2893         size = copy_to_user(ubuf, buf, *offset);
2894         if (size)
2895                 return -EFAULT;
2896
2897         return *offset;
2898 }
2899
2900 static ssize_t mtip_hw_read_flags(struct file *f, char __user *ubuf,
2901                                   size_t len, loff_t *offset)
2902 {
2903         struct driver_data *dd =  (struct driver_data *)f->private_data;
2904         char buf[MTIP_DFS_MAX_BUF_SIZE];
2905         int size = *offset;
2906
2907         if (!len || size)
2908                 return 0;
2909
2910         size += sprintf(&buf[size], "Flag-port : [ %08lX ]\n",
2911                                                         dd->port->flags);
2912         size += sprintf(&buf[size], "Flag-dd   : [ %08lX ]\n",
2913                                                         dd->dd_flag);
2914
2915         *offset = size <= len ? size : len;
2916         size = copy_to_user(ubuf, buf, *offset);
2917         if (size)
2918                 return -EFAULT;
2919
2920         return *offset;
2921 }
2922
2923 static const struct file_operations mtip_device_status_fops = {
2924         .owner  = THIS_MODULE,
2925         .open   = simple_open,
2926         .read   = mtip_hw_read_device_status,
2927         .llseek = no_llseek,
2928 };
2929
2930 static const struct file_operations mtip_regs_fops = {
2931         .owner  = THIS_MODULE,
2932         .open   = simple_open,
2933         .read   = mtip_hw_read_registers,
2934         .llseek = no_llseek,
2935 };
2936
2937 static const struct file_operations mtip_flags_fops = {
2938         .owner  = THIS_MODULE,
2939         .open   = simple_open,
2940         .read   = mtip_hw_read_flags,
2941         .llseek = no_llseek,
2942 };
2943
2944 /*
2945  * Create the sysfs related attributes.
2946  *
2947  * @dd   Pointer to the driver data structure.
2948  * @kobj Pointer to the kobj for the block device.
2949  *
2950  * return value
2951  *      0       Operation completed successfully.
2952  *      -EINVAL Invalid parameter.
2953  */
2954 static int mtip_hw_sysfs_init(struct driver_data *dd, struct kobject *kobj)
2955 {
2956         if (!kobj || !dd)
2957                 return -EINVAL;
2958
2959         if (sysfs_create_file(kobj, &dev_attr_status.attr))
2960                 dev_warn(&dd->pdev->dev,
2961                         "Error creating 'status' sysfs entry\n");
2962         return 0;
2963 }
2964
2965 /*
2966  * Remove the sysfs related attributes.
2967  *
2968  * @dd   Pointer to the driver data structure.
2969  * @kobj Pointer to the kobj for the block device.
2970  *
2971  * return value
2972  *      0       Operation completed successfully.
2973  *      -EINVAL Invalid parameter.
2974  */
2975 static int mtip_hw_sysfs_exit(struct driver_data *dd, struct kobject *kobj)
2976 {
2977         if (!kobj || !dd)
2978                 return -EINVAL;
2979
2980         sysfs_remove_file(kobj, &dev_attr_status.attr);
2981
2982         return 0;
2983 }
2984
2985 static int mtip_hw_debugfs_init(struct driver_data *dd)
2986 {
2987         if (!dfs_parent)
2988                 return -1;
2989
2990         dd->dfs_node = debugfs_create_dir(dd->disk->disk_name, dfs_parent);
2991         if (IS_ERR_OR_NULL(dd->dfs_node)) {
2992                 dev_warn(&dd->pdev->dev,
2993                         "Error creating node %s under debugfs\n",
2994                                                 dd->disk->disk_name);
2995                 dd->dfs_node = NULL;
2996                 return -1;
2997         }
2998
2999         debugfs_create_file("flags", S_IRUGO, dd->dfs_node, dd,
3000                                                         &mtip_flags_fops);
3001         debugfs_create_file("registers", S_IRUGO, dd->dfs_node, dd,
3002                                                         &mtip_regs_fops);
3003
3004         return 0;
3005 }
3006
3007 static void mtip_hw_debugfs_exit(struct driver_data *dd)
3008 {
3009         if (dd->dfs_node)
3010                 debugfs_remove_recursive(dd->dfs_node);
3011 }
3012
3013
3014 /*
3015  * Perform any init/resume time hardware setup
3016  *
3017  * @dd Pointer to the driver data structure.
3018  *
3019  * return value
3020  *      None
3021  */
3022 static inline void hba_setup(struct driver_data *dd)
3023 {
3024         u32 hwdata;
3025         hwdata = readl(dd->mmio + HOST_HSORG);
3026
3027         /* interrupt bug workaround: use only 1 IS bit.*/
3028         writel(hwdata |
3029                 HSORG_DISABLE_SLOTGRP_INTR |
3030                 HSORG_DISABLE_SLOTGRP_PXIS,
3031                 dd->mmio + HOST_HSORG);
3032 }
3033
3034 static int mtip_device_unaligned_constrained(struct driver_data *dd)
3035 {
3036         return (dd->pdev->device == P420M_DEVICE_ID ? 1 : 0);
3037 }
3038
3039 /*
3040  * Detect the details of the product, and store anything needed
3041  * into the driver data structure.  This includes product type and
3042  * version and number of slot groups.
3043  *
3044  * @dd Pointer to the driver data structure.
3045  *
3046  * return value
3047  *      None
3048  */
3049 static void mtip_detect_product(struct driver_data *dd)
3050 {
3051         u32 hwdata;
3052         unsigned int rev, slotgroups;
3053
3054         /*
3055          * HBA base + 0xFC [15:0] - vendor-specific hardware interface
3056          * info register:
3057          * [15:8] hardware/software interface rev#
3058          * [   3] asic-style interface
3059          * [ 2:0] number of slot groups, minus 1 (only valid for asic-style).
3060          */
3061         hwdata = readl(dd->mmio + HOST_HSORG);
3062
3063         dd->product_type = MTIP_PRODUCT_UNKNOWN;
3064         dd->slot_groups = 1;
3065
3066         if (hwdata & 0x8) {
3067                 dd->product_type = MTIP_PRODUCT_ASICFPGA;
3068                 rev = (hwdata & HSORG_HWREV) >> 8;
3069                 slotgroups = (hwdata & HSORG_SLOTGROUPS) + 1;
3070                 dev_info(&dd->pdev->dev,
3071                         "ASIC-FPGA design, HS rev 0x%x, "
3072                         "%i slot groups [%i slots]\n",
3073                          rev,
3074                          slotgroups,
3075                          slotgroups * 32);
3076
3077                 if (slotgroups > MTIP_MAX_SLOT_GROUPS) {
3078                         dev_warn(&dd->pdev->dev,
3079                                 "Warning: driver only supports "
3080                                 "%i slot groups.\n", MTIP_MAX_SLOT_GROUPS);
3081                         slotgroups = MTIP_MAX_SLOT_GROUPS;
3082                 }
3083                 dd->slot_groups = slotgroups;
3084                 return;
3085         }
3086
3087         dev_warn(&dd->pdev->dev, "Unrecognized product id\n");
3088 }
3089
3090 /*
3091  * Blocking wait for FTL rebuild to complete
3092  *
3093  * @dd Pointer to the DRIVER_DATA structure.
3094  *
3095  * return value
3096  *      0       FTL rebuild completed successfully
3097  *      -EFAULT FTL rebuild error/timeout/interruption
3098  */
3099 static int mtip_ftl_rebuild_poll(struct driver_data *dd)
3100 {
3101         unsigned long timeout, cnt = 0, start;
3102
3103         dev_warn(&dd->pdev->dev,
3104                 "FTL rebuild in progress. Polling for completion.\n");
3105
3106         start = jiffies;
3107         timeout = jiffies + msecs_to_jiffies(MTIP_FTL_REBUILD_TIMEOUT_MS);
3108
3109         do {
3110                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
3111                                 &dd->dd_flag)))
3112                         return -EFAULT;
3113                 if (mtip_check_surprise_removal(dd->pdev))
3114                         return -EFAULT;
3115
3116                 if (mtip_get_identify(dd->port, NULL) < 0)
3117                         return -EFAULT;
3118
3119                 if (*(dd->port->identify + MTIP_FTL_REBUILD_OFFSET) ==
3120                         MTIP_FTL_REBUILD_MAGIC) {
3121                         ssleep(1);
3122                         /* Print message every 3 minutes */
3123                         if (cnt++ >= 180) {
3124                                 dev_warn(&dd->pdev->dev,
3125                                 "FTL rebuild in progress (%d secs).\n",
3126                                 jiffies_to_msecs(jiffies - start) / 1000);
3127                                 cnt = 0;
3128                         }
3129                 } else {
3130                         dev_warn(&dd->pdev->dev,
3131                                 "FTL rebuild complete (%d secs).\n",
3132                         jiffies_to_msecs(jiffies - start) / 1000);
3133                         mtip_block_initialize(dd);
3134                         return 0;
3135                 }
3136                 ssleep(10);
3137         } while (time_before(jiffies, timeout));
3138
3139         /* Check for timeout */
3140         dev_err(&dd->pdev->dev,
3141                 "Timed out waiting for FTL rebuild to complete (%d secs).\n",
3142                 jiffies_to_msecs(jiffies - start) / 1000);
3143         return -EFAULT;
3144 }
3145
3146 /*
3147  * service thread to issue queued commands
3148  *
3149  * @data Pointer to the driver data structure.
3150  *
3151  * return value
3152  *      0
3153  */
3154
3155 static int mtip_service_thread(void *data)
3156 {
3157         struct driver_data *dd = (struct driver_data *)data;
3158         unsigned long slot, slot_start, slot_wrap;
3159         unsigned int num_cmd_slots = dd->slot_groups * 32;
3160         struct mtip_port *port = dd->port;
3161
3162         while (1) {
3163                 /*
3164                  * the condition is to check neither an internal command is
3165                  * is in progress nor error handling is active
3166                  */
3167                 wait_event_interruptible(port->svc_wait, (port->flags) &&
3168                         !(port->flags & MTIP_PF_PAUSE_IO));
3169
3170                 if (kthread_should_stop())
3171                         break;
3172
3173                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
3174                                 &dd->dd_flag)))
3175                         break;
3176
3177                 set_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags);
3178                 if (test_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags)) {
3179                         slot = 1;
3180                         /* used to restrict the loop to one iteration */
3181                         slot_start = num_cmd_slots;
3182                         slot_wrap = 0;
3183                         while (1) {
3184                                 slot = find_next_bit(port->cmds_to_issue,
3185                                                 num_cmd_slots, slot);
3186                                 if (slot_wrap == 1) {
3187                                         if ((slot_start >= slot) ||
3188                                                 (slot >= num_cmd_slots))
3189                                                 break;
3190                                 }
3191                                 if (unlikely(slot_start == num_cmd_slots))
3192                                         slot_start = slot;
3193
3194                                 if (unlikely(slot == num_cmd_slots)) {
3195                                         slot = 1;
3196                                         slot_wrap = 1;
3197                                         continue;
3198                                 }
3199
3200                                 /* Issue the command to the hardware */
3201                                 mtip_issue_ncq_command(port, slot);
3202
3203                                 clear_bit(slot, port->cmds_to_issue);
3204                         }
3205
3206                         clear_bit(MTIP_PF_ISSUE_CMDS_BIT, &port->flags);
3207                 } else if (test_bit(MTIP_PF_REBUILD_BIT, &port->flags)) {
3208                         if (!mtip_ftl_rebuild_poll(dd))
3209                                 set_bit(MTIP_DDF_REBUILD_FAILED_BIT,
3210                                                         &dd->dd_flag);
3211                         clear_bit(MTIP_PF_REBUILD_BIT, &port->flags);
3212                 }
3213                 clear_bit(MTIP_PF_SVC_THD_ACTIVE_BIT, &port->flags);
3214
3215                 if (test_bit(MTIP_PF_SVC_THD_STOP_BIT, &port->flags))
3216                         break;
3217         }
3218         return 0;
3219 }
3220
3221 /*
3222  * Called once for each card.
3223  *
3224  * @dd Pointer to the driver data structure.
3225  *
3226  * return value
3227  *      0 on success, else an error code.
3228  */
3229 static int mtip_hw_init(struct driver_data *dd)
3230 {
3231         int i;
3232         int rv;
3233         unsigned int num_command_slots;
3234         unsigned long timeout, timetaken;
3235         unsigned char *buf;
3236         struct smart_attr attr242;
3237
3238         dd->mmio = pcim_iomap_table(dd->pdev)[MTIP_ABAR];
3239
3240         mtip_detect_product(dd);
3241         if (dd->product_type == MTIP_PRODUCT_UNKNOWN) {
3242                 rv = -EIO;
3243                 goto out1;
3244         }
3245         num_command_slots = dd->slot_groups * 32;
3246
3247         hba_setup(dd);
3248
3249         dd->port = kzalloc_node(sizeof(struct mtip_port), GFP_KERNEL,
3250                                 dd->numa_node);
3251         if (!dd->port) {
3252                 dev_err(&dd->pdev->dev,
3253                         "Memory allocation: port structure\n");
3254                 return -ENOMEM;
3255         }
3256
3257         /* Continue workqueue setup */
3258         for (i = 0; i < MTIP_MAX_SLOT_GROUPS; i++)
3259                 dd->work[i].port = dd->port;
3260
3261         /* Enable unaligned IO constraints for some devices */
3262         if (mtip_device_unaligned_constrained(dd))
3263                 dd->unal_qdepth = MTIP_MAX_UNALIGNED_SLOTS;
3264         else
3265                 dd->unal_qdepth = 0;
3266
3267         /* Counting semaphore to track command slot usage */
3268         sema_init(&dd->port->cmd_slot, num_command_slots - 1 - dd->unal_qdepth);
3269         sema_init(&dd->port->cmd_slot_unal, dd->unal_qdepth);
3270
3271         /* Spinlock to prevent concurrent issue */
3272         for (i = 0; i < MTIP_MAX_SLOT_GROUPS; i++)
3273                 spin_lock_init(&dd->port->cmd_issue_lock[i]);
3274
3275         /* Set the port mmio base address. */
3276         dd->port->mmio  = dd->mmio + PORT_OFFSET;
3277         dd->port->dd    = dd;
3278
3279         /* Allocate memory for the command list. */
3280         dd->port->command_list =
3281                 dmam_alloc_coherent(&dd->pdev->dev,
3282                         HW_PORT_PRIV_DMA_SZ + (ATA_SECT_SIZE * 4),
3283                         &dd->port->command_list_dma,
3284                         GFP_KERNEL);
3285         if (!dd->port->command_list) {
3286                 dev_err(&dd->pdev->dev,
3287                         "Memory allocation: command list\n");
3288                 rv = -ENOMEM;
3289                 goto out1;
3290         }
3291
3292         /* Clear the memory we have allocated. */
3293         memset(dd->port->command_list,
3294                 0,
3295                 HW_PORT_PRIV_DMA_SZ + (ATA_SECT_SIZE * 4));
3296
3297         /* Setup the addresse of the RX FIS. */
3298         dd->port->rxfis     = dd->port->command_list + HW_CMD_SLOT_SZ;
3299         dd->port->rxfis_dma = dd->port->command_list_dma + HW_CMD_SLOT_SZ;
3300
3301         /* Setup the address of the command tables. */
3302         dd->port->command_table   = dd->port->rxfis + AHCI_RX_FIS_SZ;
3303         dd->port->command_tbl_dma = dd->port->rxfis_dma + AHCI_RX_FIS_SZ;
3304
3305         /* Setup the address of the identify data. */
3306         dd->port->identify     = dd->port->command_table +
3307                                         HW_CMD_TBL_AR_SZ;
3308         dd->port->identify_dma = dd->port->command_tbl_dma +
3309                                         HW_CMD_TBL_AR_SZ;
3310
3311         /* Setup the address of the sector buffer - for some non-ncq cmds */
3312         dd->port->sector_buffer = (void *) dd->port->identify + ATA_SECT_SIZE;
3313         dd->port->sector_buffer_dma = dd->port->identify_dma + ATA_SECT_SIZE;
3314
3315         /* Setup the address of the log buf - for read log command */
3316         dd->port->log_buf = (void *)dd->port->sector_buffer  + ATA_SECT_SIZE;
3317         dd->port->log_buf_dma = dd->port->sector_buffer_dma + ATA_SECT_SIZE;
3318
3319         /* Setup the address of the smart buf - for smart read data command */
3320         dd->port->smart_buf = (void *)dd->port->log_buf  + ATA_SECT_SIZE;
3321         dd->port->smart_buf_dma = dd->port->log_buf_dma + ATA_SECT_SIZE;
3322
3323
3324         /* Point the command headers at the command tables. */
3325         for (i = 0; i < num_command_slots; i++) {
3326                 dd->port->commands[i].command_header =
3327                                         dd->port->command_list +
3328                                         (sizeof(struct mtip_cmd_hdr) * i);
3329                 dd->port->commands[i].command_header_dma =
3330                                         dd->port->command_list_dma +
3331                                         (sizeof(struct mtip_cmd_hdr) * i);
3332
3333                 dd->port->commands[i].command =
3334                         dd->port->command_table + (HW_CMD_TBL_SZ * i);
3335                 dd->port->commands[i].command_dma =
3336                         dd->port->command_tbl_dma + (HW_CMD_TBL_SZ * i);
3337
3338                 if (readl(dd->mmio + HOST_CAP) & HOST_CAP_64)
3339                         dd->port->commands[i].command_header->ctbau =
3340                         __force_bit2int cpu_to_le32(
3341                         (dd->port->commands[i].command_dma >> 16) >> 16);
3342                 dd->port->commands[i].command_header->ctba =
3343                         __force_bit2int cpu_to_le32(
3344                         dd->port->commands[i].command_dma & 0xFFFFFFFF);
3345
3346                 /*
3347                  * If this is not done, a bug is reported by the stock
3348                  * FC11 i386. Due to the fact that it has lots of kernel
3349                  * debugging enabled.
3350                  */
3351                 sg_init_table(dd->port->commands[i].sg, MTIP_MAX_SG);
3352
3353                 /* Mark all commands as currently inactive.*/
3354                 atomic_set(&dd->port->commands[i].active, 0);
3355         }
3356
3357         /* Setup the pointers to the extended s_active and CI registers. */
3358         for (i = 0; i < dd->slot_groups; i++) {
3359                 dd->port->s_active[i] =
3360                         dd->port->mmio + i*0x80 + PORT_SCR_ACT;
3361                 dd->port->cmd_issue[i] =
3362                         dd->port->mmio + i*0x80 + PORT_COMMAND_ISSUE;
3363                 dd->port->completed[i] =
3364                         dd->port->mmio + i*0x80 + PORT_SDBV;
3365         }
3366
3367         timetaken = jiffies;
3368         timeout = jiffies + msecs_to_jiffies(30000);
3369         while (((readl(dd->port->mmio + PORT_SCR_STAT) & 0x0F) != 0x03) &&
3370                  time_before(jiffies, timeout)) {
3371                 mdelay(100);
3372         }
3373         if (unlikely(mtip_check_surprise_removal(dd->pdev))) {
3374                 timetaken = jiffies - timetaken;
3375                 dev_warn(&dd->pdev->dev,
3376                         "Surprise removal detected at %u ms\n",
3377                         jiffies_to_msecs(timetaken));
3378                 rv = -ENODEV;
3379                 goto out2 ;
3380         }
3381         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag))) {
3382                 timetaken = jiffies - timetaken;
3383                 dev_warn(&dd->pdev->dev,
3384                         "Removal detected at %u ms\n",
3385                         jiffies_to_msecs(timetaken));
3386                 rv = -EFAULT;
3387                 goto out2;
3388         }
3389
3390         /* Conditionally reset the HBA. */
3391         if (!(readl(dd->mmio + HOST_CAP) & HOST_CAP_NZDMA)) {
3392                 if (mtip_hba_reset(dd) < 0) {
3393                         dev_err(&dd->pdev->dev,
3394                                 "Card did not reset within timeout\n");
3395                         rv = -EIO;
3396                         goto out2;
3397                 }
3398         } else {
3399                 /* Clear any pending interrupts on the HBA */
3400                 writel(readl(dd->mmio + HOST_IRQ_STAT),
3401                         dd->mmio + HOST_IRQ_STAT);
3402         }
3403
3404         mtip_init_port(dd->port);
3405         mtip_start_port(dd->port);
3406
3407         /* Setup the ISR and enable interrupts. */
3408         rv = devm_request_irq(&dd->pdev->dev,
3409                                 dd->pdev->irq,
3410                                 mtip_irq_handler,
3411                                 IRQF_SHARED,
3412                                 dev_driver_string(&dd->pdev->dev),
3413                                 dd);
3414
3415         if (rv) {
3416                 dev_err(&dd->pdev->dev,
3417                         "Unable to allocate IRQ %d\n", dd->pdev->irq);
3418                 goto out2;
3419         }
3420         irq_set_affinity_hint(dd->pdev->irq, get_cpu_mask(dd->isr_binding));
3421
3422         /* Enable interrupts on the HBA. */
3423         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
3424                                         dd->mmio + HOST_CTL);
3425
3426         init_timer(&dd->port->cmd_timer);
3427         init_waitqueue_head(&dd->port->svc_wait);
3428
3429         dd->port->cmd_timer.data = (unsigned long int) dd->port;
3430         dd->port->cmd_timer.function = mtip_timeout_function;
3431         mod_timer(&dd->port->cmd_timer,
3432                 jiffies + msecs_to_jiffies(MTIP_TIMEOUT_CHECK_PERIOD));
3433
3434
3435         if (test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)) {
3436                 rv = -EFAULT;
3437                 goto out3;
3438         }
3439
3440         if (mtip_get_identify(dd->port, NULL) < 0) {
3441                 rv = -EFAULT;
3442                 goto out3;
3443         }
3444
3445         if (*(dd->port->identify + MTIP_FTL_REBUILD_OFFSET) ==
3446                 MTIP_FTL_REBUILD_MAGIC) {
3447                 set_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags);
3448                 return MTIP_FTL_REBUILD_MAGIC;
3449         }
3450         mtip_dump_identify(dd->port);
3451
3452         /* check write protect, over temp and rebuild statuses */
3453         rv = mtip_read_log_page(dd->port, ATA_LOG_SATA_NCQ,
3454                                 dd->port->log_buf,
3455                                 dd->port->log_buf_dma, 1);
3456         if (rv) {
3457                 dev_warn(&dd->pdev->dev,
3458                         "Error in READ LOG EXT (10h) command\n");
3459                 /* non-critical error, don't fail the load */
3460         } else {
3461                 buf = (unsigned char *)dd->port->log_buf;
3462                 if (buf[259] & 0x1) {
3463                         dev_info(&dd->pdev->dev,
3464                                 "Write protect bit is set.\n");
3465                         set_bit(MTIP_DDF_WRITE_PROTECT_BIT, &dd->dd_flag);
3466                 }
3467                 if (buf[288] == 0xF7) {
3468                         dev_info(&dd->pdev->dev,
3469                                 "Exceeded Tmax, drive in thermal shutdown.\n");
3470                         set_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag);
3471                 }
3472                 if (buf[288] == 0xBF) {
3473                         dev_info(&dd->pdev->dev,
3474                                 "Drive indicates rebuild has failed.\n");
3475                         /* TODO */
3476                 }
3477         }
3478
3479         /* get write protect progess */
3480         memset(&attr242, 0, sizeof(struct smart_attr));
3481         if (mtip_get_smart_attr(dd->port, 242, &attr242))
3482                 dev_warn(&dd->pdev->dev,
3483                                 "Unable to check write protect progress\n");
3484         else
3485                 dev_info(&dd->pdev->dev,
3486                                 "Write protect progress: %u%% (%u blocks)\n",
3487                                 attr242.cur, le32_to_cpu(attr242.data));
3488         return rv;
3489
3490 out3:
3491         del_timer_sync(&dd->port->cmd_timer);
3492
3493         /* Disable interrupts on the HBA. */
3494         writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3495                         dd->mmio + HOST_CTL);
3496
3497         /* Release the IRQ. */
3498         irq_set_affinity_hint(dd->pdev->irq, NULL);
3499         devm_free_irq(&dd->pdev->dev, dd->pdev->irq, dd);
3500
3501 out2:
3502         mtip_deinit_port(dd->port);
3503
3504         /* Free the command/command header memory. */
3505         dmam_free_coherent(&dd->pdev->dev,
3506                                 HW_PORT_PRIV_DMA_SZ + (ATA_SECT_SIZE * 4),
3507                                 dd->port->command_list,
3508                                 dd->port->command_list_dma);
3509 out1:
3510         /* Free the memory allocated for the for structure. */
3511         kfree(dd->port);
3512
3513         return rv;
3514 }
3515
3516 /*
3517  * Called to deinitialize an interface.
3518  *
3519  * @dd Pointer to the driver data structure.
3520  *
3521  * return value
3522  *      0
3523  */
3524 static int mtip_hw_exit(struct driver_data *dd)
3525 {
3526         /*
3527          * Send standby immediate (E0h) to the drive so that it
3528          * saves its state.
3529          */
3530         if (!test_bit(MTIP_DDF_CLEANUP_BIT, &dd->dd_flag)) {
3531
3532                 if (!test_bit(MTIP_PF_REBUILD_BIT, &dd->port->flags))
3533                         if (mtip_standby_immediate(dd->port))
3534                                 dev_warn(&dd->pdev->dev,
3535                                         "STANDBY IMMEDIATE failed\n");
3536
3537                 /* de-initialize the port. */
3538                 mtip_deinit_port(dd->port);
3539
3540                 /* Disable interrupts on the HBA. */
3541                 writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3542                                 dd->mmio + HOST_CTL);
3543         }
3544
3545         del_timer_sync(&dd->port->cmd_timer);
3546
3547         /* Release the IRQ. */
3548         irq_set_affinity_hint(dd->pdev->irq, NULL);
3549         devm_free_irq(&dd->pdev->dev, dd->pdev->irq, dd);
3550
3551         /* Free the command/command header memory. */
3552         dmam_free_coherent(&dd->pdev->dev,
3553                         HW_PORT_PRIV_DMA_SZ + (ATA_SECT_SIZE * 4),
3554                         dd->port->command_list,
3555                         dd->port->command_list_dma);
3556         /* Free the memory allocated for the for structure. */
3557         kfree(dd->port);
3558
3559         return 0;
3560 }
3561
3562 /*
3563  * Issue a Standby Immediate command to the device.
3564  *
3565  * This function is called by the Block Layer just before the
3566  * system powers off during a shutdown.
3567  *
3568  * @dd Pointer to the driver data structure.
3569  *
3570  * return value
3571  *      0
3572  */
3573 static int mtip_hw_shutdown(struct driver_data *dd)
3574 {
3575         /*
3576          * Send standby immediate (E0h) to the drive so that it
3577          * saves its state.
3578          */
3579         mtip_standby_immediate(dd->port);
3580
3581         return 0;
3582 }
3583
3584 /*
3585  * Suspend function
3586  *
3587  * This function is called by the Block Layer just before the
3588  * system hibernates.
3589  *
3590  * @dd Pointer to the driver data structure.
3591  *
3592  * return value
3593  *      0       Suspend was successful
3594  *      -EFAULT Suspend was not successful
3595  */
3596 static int mtip_hw_suspend(struct driver_data *dd)
3597 {
3598         /*
3599          * Send standby immediate (E0h) to the drive
3600          * so that it saves its state.
3601          */
3602         if (mtip_standby_immediate(dd->port) != 0) {
3603                 dev_err(&dd->pdev->dev,
3604                         "Failed standby-immediate command\n");
3605                 return -EFAULT;
3606         }
3607
3608         /* Disable interrupts on the HBA.*/
3609         writel(readl(dd->mmio + HOST_CTL) & ~HOST_IRQ_EN,
3610                         dd->mmio + HOST_CTL);
3611         mtip_deinit_port(dd->port);
3612
3613         return 0;
3614 }
3615
3616 /*
3617  * Resume function
3618  *
3619  * This function is called by the Block Layer as the
3620  * system resumes.
3621  *
3622  * @dd Pointer to the driver data structure.
3623  *
3624  * return value
3625  *      0       Resume was successful
3626  *      -EFAULT Resume was not successful
3627  */
3628 static int mtip_hw_resume(struct driver_data *dd)
3629 {
3630         /* Perform any needed hardware setup steps */
3631         hba_setup(dd);
3632
3633         /* Reset the HBA */
3634         if (mtip_hba_reset(dd) != 0) {
3635                 dev_err(&dd->pdev->dev,
3636                         "Unable to reset the HBA\n");
3637                 return -EFAULT;
3638         }
3639
3640         /*
3641          * Enable the port, DMA engine, and FIS reception specific
3642          * h/w in controller.
3643          */
3644         mtip_init_port(dd->port);
3645         mtip_start_port(dd->port);
3646
3647         /* Enable interrupts on the HBA.*/
3648         writel(readl(dd->mmio + HOST_CTL) | HOST_IRQ_EN,
3649                         dd->mmio + HOST_CTL);
3650
3651         return 0;
3652 }
3653
3654 /*
3655  * Helper function for reusing disk name
3656  * upon hot insertion.
3657  */
3658 static int rssd_disk_name_format(char *prefix,
3659                                  int index,
3660                                  char *buf,
3661                                  int buflen)
3662 {
3663         const int base = 'z' - 'a' + 1;
3664         char *begin = buf + strlen(prefix);
3665         char *end = buf + buflen;
3666         char *p;
3667         int unit;
3668
3669         p = end - 1;
3670         *p = '\0';
3671         unit = base;
3672         do {
3673                 if (p == begin)
3674                         return -EINVAL;
3675                 *--p = 'a' + (index % unit);
3676                 index = (index / unit) - 1;
3677         } while (index >= 0);
3678
3679         memmove(begin, p, end - p);
3680         memcpy(buf, prefix, strlen(prefix));
3681
3682         return 0;
3683 }
3684
3685 /*
3686  * Block layer IOCTL handler.
3687  *
3688  * @dev Pointer to the block_device structure.
3689  * @mode ignored
3690  * @cmd IOCTL command passed from the user application.
3691  * @arg Argument passed from the user application.
3692  *
3693  * return value
3694  *      0        IOCTL completed successfully.
3695  *      -ENOTTY  IOCTL not supported or invalid driver data
3696  *                 structure pointer.
3697  */
3698 static int mtip_block_ioctl(struct block_device *dev,
3699                             fmode_t mode,
3700                             unsigned cmd,
3701                             unsigned long arg)
3702 {
3703         struct driver_data *dd = dev->bd_disk->private_data;
3704
3705         if (!capable(CAP_SYS_ADMIN))
3706                 return -EACCES;
3707
3708         if (!dd)
3709                 return -ENOTTY;
3710
3711         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)))
3712                 return -ENOTTY;
3713
3714         switch (cmd) {
3715         case BLKFLSBUF:
3716                 return -ENOTTY;
3717         default:
3718                 return mtip_hw_ioctl(dd, cmd, arg);
3719         }
3720 }
3721
3722 #ifdef CONFIG_COMPAT
3723 /*
3724  * Block layer compat IOCTL handler.
3725  *
3726  * @dev Pointer to the block_device structure.
3727  * @mode ignored
3728  * @cmd IOCTL command passed from the user application.
3729  * @arg Argument passed from the user application.
3730  *
3731  * return value
3732  *      0        IOCTL completed successfully.
3733  *      -ENOTTY  IOCTL not supported or invalid driver data
3734  *                 structure pointer.
3735  */
3736 static int mtip_block_compat_ioctl(struct block_device *dev,
3737                             fmode_t mode,
3738                             unsigned cmd,
3739                             unsigned long arg)
3740 {
3741         struct driver_data *dd = dev->bd_disk->private_data;
3742
3743         if (!capable(CAP_SYS_ADMIN))
3744                 return -EACCES;
3745
3746         if (!dd)
3747                 return -ENOTTY;
3748
3749         if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag)))
3750                 return -ENOTTY;
3751
3752         switch (cmd) {
3753         case BLKFLSBUF:
3754                 return -ENOTTY;
3755         case HDIO_DRIVE_TASKFILE: {
3756                 struct mtip_compat_ide_task_request_s __user *compat_req_task;
3757                 ide_task_request_t req_task;
3758                 int compat_tasksize, outtotal, ret;
3759
3760                 compat_tasksize =
3761                         sizeof(struct mtip_compat_ide_task_request_s);
3762
3763                 compat_req_task =
3764                         (struct mtip_compat_ide_task_request_s __user *) arg;
3765
3766                 if (copy_from_user(&req_task, (void __user *) arg,
3767                         compat_tasksize - (2 * sizeof(compat_long_t))))
3768                         return -EFAULT;
3769
3770                 if (get_user(req_task.out_size, &compat_req_task->out_size))
3771                         return -EFAULT;
3772
3773                 if (get_user(req_task.in_size, &compat_req_task->in_size))
3774                         return -EFAULT;
3775
3776                 outtotal = sizeof(struct mtip_compat_ide_task_request_s);
3777
3778                 ret = exec_drive_taskfile(dd, (void __user *) arg,
3779                                                 &req_task, outtotal);
3780
3781                 if (copy_to_user((void __user *) arg, &req_task,
3782                                 compat_tasksize -
3783                                 (2 * sizeof(compat_long_t))))
3784                         return -EFAULT;
3785
3786                 if (put_user(req_task.out_size, &compat_req_task->out_size))
3787                         return -EFAULT;
3788
3789                 if (put_user(req_task.in_size, &compat_req_task->in_size))
3790                         return -EFAULT;
3791
3792                 return ret;
3793         }
3794         default:
3795                 return mtip_hw_ioctl(dd, cmd, arg);
3796         }
3797 }
3798 #endif
3799
3800 /*
3801  * Obtain the geometry of the device.
3802  *
3803  * You may think that this function is obsolete, but some applications,
3804  * fdisk for example still used CHS values. This function describes the
3805  * device as having 224 heads and 56 sectors per cylinder. These values are
3806  * chosen so that each cylinder is aligned on a 4KB boundary. Since a
3807  * partition is described in terms of a start and end cylinder this means
3808  * that each partition is also 4KB aligned. Non-aligned partitions adversely
3809  * affects performance.
3810  *
3811  * @dev Pointer to the block_device strucutre.
3812  * @geo Pointer to a hd_geometry structure.
3813  *
3814  * return value
3815  *      0       Operation completed successfully.
3816  *      -ENOTTY An error occurred while reading the drive capacity.
3817  */
3818 static int mtip_block_getgeo(struct block_device *dev,
3819                                 struct hd_geometry *geo)
3820 {
3821         struct driver_data *dd = dev->bd_disk->private_data;
3822         sector_t capacity;
3823
3824         if (!dd)
3825                 return -ENOTTY;
3826
3827         if (!(mtip_hw_get_capacity(dd, &capacity))) {
3828                 dev_warn(&dd->pdev->dev,
3829                         "Could not get drive capacity.\n");
3830                 return -ENOTTY;
3831         }
3832
3833         geo->heads = 224;
3834         geo->sectors = 56;
3835         sector_div(capacity, (geo->heads * geo->sectors));
3836         geo->cylinders = capacity;
3837         return 0;
3838 }
3839
3840 /*
3841  * Block device operation function.
3842  *
3843  * This structure contains pointers to the functions required by the block
3844  * layer.
3845  */
3846 static const struct block_device_operations mtip_block_ops = {
3847         .ioctl          = mtip_block_ioctl,
3848 #ifdef CONFIG_COMPAT
3849         .compat_ioctl   = mtip_block_compat_ioctl,
3850 #endif
3851         .getgeo         = mtip_block_getgeo,
3852         .owner          = THIS_MODULE
3853 };
3854
3855 /*
3856  * Block layer make request function.
3857  *
3858  * This function is called by the kernel to process a BIO for
3859  * the P320 device.
3860  *
3861  * @queue Pointer to the request queue. Unused other than to obtain
3862  *              the driver data structure.
3863  * @bio   Pointer to the BIO.
3864  *
3865  */
3866 static void mtip_make_request(struct request_queue *queue, struct bio *bio)
3867 {
3868         struct driver_data *dd = queue->queuedata;
3869         struct scatterlist *sg;
3870         struct bio_vec *bvec;
3871         int i, nents = 0;
3872         int tag = 0, unaligned = 0;
3873
3874         if (unlikely(dd->dd_flag & MTIP_DDF_STOP_IO)) {
3875                 if (unlikely(test_bit(MTIP_DDF_REMOVE_PENDING_BIT,
3876                                                         &dd->dd_flag))) {
3877                         bio_endio(bio, -ENXIO);
3878                         return;
3879                 }
3880                 if (unlikely(test_bit(MTIP_DDF_OVER_TEMP_BIT, &dd->dd_flag))) {
3881                         bio_endio(bio, -ENODATA);
3882                         return;
3883                 }
3884                 if (unlikely(test_bit(MTIP_DDF_WRITE_PROTECT_BIT,
3885                                                         &dd->dd_flag) &&
3886                                 bio_data_dir(bio))) {
3887                         bio_endio(bio, -ENODATA);
3888                         return;
3889                 }
3890                 if (unlikely(test_bit(MTIP_DDF_SEC_LOCK_BIT, &dd->dd_flag))) {
3891                         bio_endio(bio, -ENODATA);
3892                         return;
3893                 }
3894         }
3895
3896         if (unlikely(bio->bi_rw & REQ_DISCARD)) {
3897                 bio_endio(bio, mtip_send_trim(dd, bio->bi_sector,
3898                                                 bio_sectors(bio)));
3899                 return;
3900         }
3901
3902         if (unlikely(!bio_has_data(bio))) {
3903                 blk_queue_flush(queue, 0);
3904                 bio_endio(bio, 0);
3905                 return;
3906         }
3907
3908         if (bio_data_dir(bio) == WRITE && bio_sectors(bio) <= 64 &&
3909                                                         dd->unal_qdepth) {
3910                 if (bio->bi_sector % 8 != 0) /* Unaligned on 4k boundaries */
3911                         unaligned = 1;
3912                 else if (bio_sectors(bio) % 8 != 0) /* Aligned but not 4k/8k */
3913                         unaligned = 1;
3914         }
3915
3916         sg = mtip_hw_get_scatterlist(dd, &tag, unaligned);
3917         if (likely(sg != NULL)) {
3918                 blk_queue_bounce(queue, &bio);
3919
3920                 if (unlikely((bio)->bi_vcnt > MTIP_MAX_SG)) {
3921                         dev_warn(&dd->pdev->dev,
3922                                 "Maximum number of SGL entries exceeded\n");
3923                         bio_io_error(bio);
3924                         mtip_hw_release_scatterlist(dd, tag, unaligned);
3925                         return;
3926                 }
3927
3928                 /* Create the scatter list for this bio. */
3929                 bio_for_each_segment(bvec, bio, i) {
3930                         sg_set_page(&sg[nents],
3931                                         bvec->bv_page,
3932                                         bvec->bv_len,
3933                                         bvec->bv_offset);
3934                         nents++;
3935                 }
3936
3937                 /* Issue the read/write. */
3938                 mtip_hw_submit_io(dd,
3939                                 bio->bi_sector,
3940                                 bio_sectors(bio),
3941                                 nents,
3942                                 tag,
3943                                 bio_endio,
3944                                 bio,
3945                                 bio_data_dir(bio),
3946                                 unaligned);
3947         } else
3948                 bio_io_error(bio);
3949 }
3950
3951 /*
3952  * Block layer initialization function.
3953  *
3954  * This function is called once by the PCI layer for each P320
3955  * device that is connected to the system.
3956  *
3957  * @dd Pointer to the driver data structure.
3958  *
3959  * return value
3960  *      0 on success else an error code.
3961  */
3962 static int mtip_block_initialize(struct driver_data *dd)
3963 {
3964         int rv = 0, wait_for_rebuild = 0;
3965         sector_t capacity;
3966         unsigned int index = 0;
3967         struct kobject *kobj;
3968         unsigned char thd_name[16];
3969
3970         if (dd->disk)
3971                 goto skip_create_disk; /* hw init done, before rebuild */
3972
3973         /* Initialize the protocol layer. */
3974         wait_for_rebuild = mtip_hw_init(dd);
3975         if (wait_for_rebuild < 0) {
3976                 dev_err(&dd->pdev->dev,
3977                         "Protocol layer initialization failed\n");
3978                 rv = -EINVAL;
3979                 goto protocol_init_error;
3980         }
3981
3982         dd->disk = alloc_disk_node(MTIP_MAX_MINORS, dd->numa_node);
3983         if (dd->disk  == NULL) {
3984                 dev_err(&dd->pdev->dev,
3985                         "Unable to allocate gendisk structure\n");
3986                 rv = -EINVAL;
3987                 goto alloc_disk_error;
3988         }
3989
3990         /* Generate the disk name, implemented same as in sd.c */
3991         do {
3992                 if (!ida_pre_get(&rssd_index_ida, GFP_KERNEL))
3993                         goto ida_get_error;
3994
3995                 spin_lock(&rssd_index_lock);
3996                 rv = ida_get_new(&rssd_index_ida, &index);
3997                 spin_unlock(&rssd_index_lock);
3998         } while (rv == -EAGAIN);
3999
4000         if (rv)
4001                 goto ida_get_error;
4002
4003         rv = rssd_disk_name_format("rssd",
4004                                 index,
4005                                 dd->disk->disk_name,
4006                                 DISK_NAME_LEN);
4007         if (rv)
4008                 goto disk_index_error;
4009
4010         dd->disk->driverfs_dev  = &dd->pdev->dev;
4011         dd->disk->major         = dd->major;
4012         dd->disk->first_minor   = dd->instance * MTIP_MAX_MINORS;
4013         dd->disk->fops          = &mtip_block_ops;
4014         dd->disk->private_data  = dd;
4015         dd->index               = index;
4016
4017         /*
4018          * if rebuild pending, start the service thread, and delay the block
4019          * queue creation and add_disk()
4020          */
4021         if (wait_for_rebuild == MTIP_FTL_REBUILD_MAGIC)
4022                 goto start_service_thread;
4023
4024 skip_create_disk:
4025         /* Allocate the request queue. */
4026         dd->queue = blk_alloc_queue_node(GFP_KERNEL, dd->numa_node);
4027         if (dd->queue == NULL) {
4028                 dev_err(&dd->pdev->dev,
4029                         "Unable to allocate request queue\n");
4030                 rv = -ENOMEM;
4031                 goto block_queue_alloc_init_error;
4032         }
4033
4034         /* Attach our request function to the request queue. */
4035         blk_queue_make_request(dd->queue, mtip_make_request);
4036
4037         dd->disk->queue         = dd->queue;
4038         dd->queue->queuedata    = dd;
4039
4040         /* Set device limits. */
4041         set_bit(QUEUE_FLAG_NONROT, &dd->queue->queue_flags);
4042         blk_queue_max_segments(dd->queue, MTIP_MAX_SG);
4043         blk_queue_physical_block_size(dd->queue, 4096);
4044         blk_queue_max_hw_sectors(dd->queue, 0xffff);
4045         blk_queue_max_segment_size(dd->queue, 0x400000);
4046         blk_queue_io_min(dd->queue, 4096);
4047         blk_queue_bounce_limit(dd->queue, dd->pdev->dma_mask);
4048
4049         /*
4050          * write back cache is not supported in the device. FUA depends on
4051          * write back cache support, hence setting flush support to zero.
4052          */
4053         blk_queue_flush(dd->queue, 0);
4054
4055         /* Signal trim support */
4056         if (dd->trim_supp == true) {
4057                 set_bit(QUEUE_FLAG_DISCARD, &dd->queue->queue_flags);
4058                 dd->queue->limits.discard_granularity = 4096;
4059                 blk_queue_max_discard_sectors(dd->queue,
4060                         MTIP_MAX_TRIM_ENTRY_LEN * MTIP_MAX_TRIM_ENTRIES);
4061                 dd->queue->limits.discard_zeroes_data = 0;
4062         }
4063
4064         /* Set the capacity of the device in 512 byte sectors. */
4065         if (!(mtip_hw_get_capacity(dd, &capacity))) {
4066                 dev_warn(&dd->pdev->dev,
4067                         "Could not read drive capacity\n");
4068                 rv = -EIO;
4069                 goto read_capacity_error;
4070         }
4071         set_capacity(dd->disk, capacity);
4072
4073         /* Enable the block device and add it to /dev */
4074         add_disk(dd->disk);
4075
4076         /*
4077          * Now that the disk is active, initialize any sysfs attributes
4078          * managed by the protocol layer.
4079          */
4080         kobj = kobject_get(&disk_to_dev(dd->disk)->kobj);
4081         if (kobj) {
4082                 mtip_hw_sysfs_init(dd, kobj);
4083                 kobject_put(kobj);
4084         }
4085         mtip_hw_debugfs_init(dd);
4086
4087         if (dd->mtip_svc_handler) {
4088                 set_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag);
4089                 return rv; /* service thread created for handling rebuild */
4090         }
4091
4092 start_service_thread:
4093         sprintf(thd_name, "mtip_svc_thd_%02d", index);
4094         dd->mtip_svc_handler = kthread_create_on_node(mtip_service_thread,
4095                                                 dd, dd->numa_node, thd_name);
4096
4097         if (IS_ERR(dd->mtip_svc_handler)) {
4098                 dev_err(&dd->pdev->dev, "service thread failed to start\n");
4099                 dd->mtip_svc_handler = NULL;
4100                 rv = -EFAULT;
4101                 goto kthread_run_error;
4102         }
4103         wake_up_process(dd->mtip_svc_handler);
4104         if (wait_for_rebuild == MTIP_FTL_REBUILD_MAGIC)
4105                 rv = wait_for_rebuild;
4106
4107         return rv;
4108
4109 kthread_run_error:
4110         mtip_hw_debugfs_exit(dd);
4111
4112         /* Delete our gendisk. This also removes the device from /dev */
4113         del_gendisk(dd->disk);
4114
4115 read_capacity_error:
4116         blk_cleanup_queue(dd->queue);
4117
4118 block_queue_alloc_init_error:
4119 disk_index_error:
4120         spin_lock(&rssd_index_lock);
4121         ida_remove(&rssd_index_ida, index);
4122         spin_unlock(&rssd_index_lock);
4123
4124 ida_get_error:
4125         put_disk(dd->disk);
4126
4127 alloc_disk_error:
4128         mtip_hw_exit(dd); /* De-initialize the protocol layer. */
4129
4130 protocol_init_error:
4131         return rv;
4132 }
4133
4134 /*
4135  * Block layer deinitialization function.
4136  *
4137  * Called by the PCI layer as each P320 device is removed.
4138  *
4139  * @dd Pointer to the driver data structure.
4140  *
4141  * return value
4142  *      0
4143  */
4144 static int mtip_block_remove(struct driver_data *dd)
4145 {
4146         struct kobject *kobj;
4147
4148         if (dd->mtip_svc_handler) {
4149                 set_bit(MTIP_PF_SVC_THD_STOP_BIT, &dd->port->flags);
4150                 wake_up_interruptible(&dd->port->svc_wait);
4151                 kthread_stop(dd->mtip_svc_handler);
4152         }
4153
4154         /* Clean up the sysfs attributes, if created */
4155         if (test_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag)) {
4156                 kobj = kobject_get(&disk_to_dev(dd->disk)->kobj);
4157                 if (kobj) {
4158                         mtip_hw_sysfs_exit(dd, kobj);
4159                         kobject_put(kobj);
4160                 }
4161         }
4162         mtip_hw_debugfs_exit(dd);
4163
4164         /*
4165          * Delete our gendisk structure. This also removes the device
4166          * from /dev
4167          */
4168         if (dd->disk) {
4169                 if (dd->disk->queue)
4170                         del_gendisk(dd->disk);
4171                 else
4172                         put_disk(dd->disk);
4173         }
4174
4175         spin_lock(&rssd_index_lock);
4176         ida_remove(&rssd_index_ida, dd->index);
4177         spin_unlock(&rssd_index_lock);
4178
4179         blk_cleanup_queue(dd->queue);
4180         dd->disk  = NULL;
4181         dd->queue = NULL;
4182
4183         /* De-initialize the protocol layer. */
4184         mtip_hw_exit(dd);
4185
4186         return 0;
4187 }
4188
4189 /*
4190  * Function called by the PCI layer when just before the
4191  * machine shuts down.
4192  *
4193  * If a protocol layer shutdown function is present it will be called
4194  * by this function.
4195  *
4196  * @dd Pointer to the driver data structure.
4197  *
4198  * return value
4199  *      0
4200  */
4201 static int mtip_block_shutdown(struct driver_data *dd)
4202 {
4203         /* Delete our gendisk structure, and cleanup the blk queue. */
4204         if (dd->disk) {
4205                 dev_info(&dd->pdev->dev,
4206                         "Shutting down %s ...\n", dd->disk->disk_name);
4207
4208                 if (dd->disk->queue) {
4209                         del_gendisk(dd->disk);
4210                         blk_cleanup_queue(dd->queue);
4211                 } else
4212                         put_disk(dd->disk);
4213                 dd->disk  = NULL;
4214                 dd->queue = NULL;
4215         }
4216
4217         spin_lock(&rssd_index_lock);
4218         ida_remove(&rssd_index_ida, dd->index);
4219         spin_unlock(&rssd_index_lock);
4220
4221         mtip_hw_shutdown(dd);
4222         return 0;
4223 }
4224
4225 static int mtip_block_suspend(struct driver_data *dd)
4226 {
4227         dev_info(&dd->pdev->dev,
4228                 "Suspending %s ...\n", dd->disk->disk_name);
4229         mtip_hw_suspend(dd);
4230         return 0;
4231 }
4232
4233 static int mtip_block_resume(struct driver_data *dd)
4234 {
4235         dev_info(&dd->pdev->dev, "Resuming %s ...\n",
4236                 dd->disk->disk_name);
4237         mtip_hw_resume(dd);
4238         return 0;
4239 }
4240
4241 static void drop_cpu(int cpu)
4242 {
4243         cpu_use[cpu]--;
4244 }
4245
4246 static int get_least_used_cpu_on_node(int node)
4247 {
4248         int cpu, least_used_cpu, least_cnt;
4249         const struct cpumask *node_mask;
4250
4251         node_mask = cpumask_of_node(node);
4252         least_used_cpu = cpumask_first(node_mask);
4253         least_cnt = cpu_use[least_used_cpu];
4254         cpu = least_used_cpu;
4255
4256         for_each_cpu(cpu, node_mask) {
4257                 if (cpu_use[cpu] < least_cnt) {
4258                         least_used_cpu = cpu;
4259                         least_cnt = cpu_use[cpu];
4260                 }
4261         }
4262         cpu_use[least_used_cpu]++;
4263         return least_used_cpu;
4264 }
4265
4266 /* Helper for selecting a node in round robin mode */
4267 static inline int mtip_get_next_rr_node(void)
4268 {
4269         static int next_node = -1;
4270
4271         if (next_node == -1) {
4272                 next_node = first_online_node;
4273                 return next_node;
4274         }
4275
4276         next_node = next_online_node(next_node);
4277         if (next_node == MAX_NUMNODES)
4278                 next_node = first_online_node;
4279         return next_node;
4280 }
4281
4282 static DEFINE_HANDLER(0);
4283 static DEFINE_HANDLER(1);
4284 static DEFINE_HANDLER(2);
4285 static DEFINE_HANDLER(3);
4286 static DEFINE_HANDLER(4);
4287 static DEFINE_HANDLER(5);
4288 static DEFINE_HANDLER(6);
4289 static DEFINE_HANDLER(7);
4290
4291 static void mtip_disable_link_opts(struct driver_data *dd, struct pci_dev *pdev)
4292 {
4293         int pos;
4294         unsigned short pcie_dev_ctrl;
4295
4296         pos = pci_find_capability(pdev, PCI_CAP_ID_EXP);
4297         if (pos) {
4298                 pci_read_config_word(pdev,
4299                         pos + PCI_EXP_DEVCTL,
4300                         &pcie_dev_ctrl);
4301                 if (pcie_dev_ctrl & (1 << 11) ||
4302                     pcie_dev_ctrl & (1 << 4)) {
4303                         dev_info(&dd->pdev->dev,
4304                                 "Disabling ERO/No-Snoop on bridge device %04x:%04x\n",
4305                                         pdev->vendor, pdev->device);
4306                         pcie_dev_ctrl &= ~(PCI_EXP_DEVCTL_NOSNOOP_EN |
4307                                                 PCI_EXP_DEVCTL_RELAX_EN);
4308                         pci_write_config_word(pdev,
4309                                 pos + PCI_EXP_DEVCTL,
4310                                 pcie_dev_ctrl);
4311                 }
4312         }
4313 }
4314
4315 static void mtip_fix_ero_nosnoop(struct driver_data *dd, struct pci_dev *pdev)
4316 {
4317         /*
4318          * This workaround is specific to AMD/ATI chipset with a PCI upstream
4319          * device with device id 0x5aXX
4320          */
4321         if (pdev->bus && pdev->bus->self) {
4322                 if (pdev->bus->self->vendor == PCI_VENDOR_ID_ATI &&
4323                     ((pdev->bus->self->device & 0xff00) == 0x5a00)) {
4324                         mtip_disable_link_opts(dd, pdev->bus->self);
4325                 } else {
4326                         /* Check further up the topology */
4327                         struct pci_dev *parent_dev = pdev->bus->self;
4328                         if (parent_dev->bus &&
4329                                 parent_dev->bus->parent &&
4330                                 parent_dev->bus->parent->self &&
4331                                 parent_dev->bus->parent->self->vendor ==
4332                                          PCI_VENDOR_ID_ATI &&
4333                                 (parent_dev->bus->parent->self->device &
4334                                         0xff00) == 0x5a00) {
4335                                 mtip_disable_link_opts(dd,
4336                                         parent_dev->bus->parent->self);
4337                         }
4338                 }
4339         }
4340 }
4341
4342 /*
4343  * Called for each supported PCI device detected.
4344  *
4345  * This function allocates the private data structure, enables the
4346  * PCI device and then calls the block layer initialization function.
4347  *
4348  * return value
4349  *      0 on success else an error code.
4350  */
4351 static int mtip_pci_probe(struct pci_dev *pdev,
4352                         const struct pci_device_id *ent)
4353 {
4354         int rv = 0;
4355         struct driver_data *dd = NULL;
4356         char cpu_list[256];
4357         const struct cpumask *node_mask;
4358         int cpu, i = 0, j = 0;
4359         int my_node = NUMA_NO_NODE;
4360         unsigned long flags;
4361
4362         /* Allocate memory for this devices private data. */
4363         my_node = pcibus_to_node(pdev->bus);
4364         if (my_node != NUMA_NO_NODE) {
4365                 if (!node_online(my_node))
4366                         my_node = mtip_get_next_rr_node();
4367         } else {
4368                 dev_info(&pdev->dev, "Kernel not reporting proximity, choosing a node\n");
4369                 my_node = mtip_get_next_rr_node();
4370         }
4371         dev_info(&pdev->dev, "NUMA node %d (closest: %d,%d, probe on %d:%d)\n",
4372                 my_node, pcibus_to_node(pdev->bus), dev_to_node(&pdev->dev),
4373                 cpu_to_node(smp_processor_id()), smp_processor_id());
4374
4375         dd = kzalloc_node(sizeof(struct driver_data), GFP_KERNEL, my_node);
4376         if (dd == NULL) {
4377                 dev_err(&pdev->dev,
4378                         "Unable to allocate memory for driver data\n");
4379                 return -ENOMEM;
4380         }
4381
4382         /* Attach the private data to this PCI device.  */
4383         pci_set_drvdata(pdev, dd);
4384
4385         rv = pcim_enable_device(pdev);
4386         if (rv < 0) {
4387                 dev_err(&pdev->dev, "Unable to enable device\n");
4388                 goto iomap_err;
4389         }
4390
4391         /* Map BAR5 to memory. */
4392         rv = pcim_iomap_regions(pdev, 1 << MTIP_ABAR, MTIP_DRV_NAME);
4393         if (rv < 0) {
4394                 dev_err(&pdev->dev, "Unable to map regions\n");
4395                 goto iomap_err;
4396         }
4397
4398         if (!pci_set_dma_mask(pdev, DMA_BIT_MASK(64))) {
4399                 rv = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
4400
4401                 if (rv) {
4402                         rv = pci_set_consistent_dma_mask(pdev,
4403                                                 DMA_BIT_MASK(32));
4404                         if (rv) {
4405                                 dev_warn(&pdev->dev,
4406                                         "64-bit DMA enable failed\n");
4407                                 goto setmask_err;
4408                         }
4409                 }
4410         }
4411
4412         /* Copy the info we may need later into the private data structure. */
4413         dd->major       = mtip_major;
4414         dd->instance    = instance;
4415         dd->pdev        = pdev;
4416         dd->numa_node   = my_node;
4417
4418         INIT_LIST_HEAD(&dd->online_list);
4419         INIT_LIST_HEAD(&dd->remove_list);
4420
4421         memset(dd->workq_name, 0, 32);
4422         snprintf(dd->workq_name, 31, "mtipq%d", dd->instance);
4423
4424         dd->isr_workq = create_workqueue(dd->workq_name);
4425         if (!dd->isr_workq) {
4426                 dev_warn(&pdev->dev, "Can't create wq %d\n", dd->instance);
4427                 rv = -ENOMEM;
4428                 goto block_initialize_err;
4429         }
4430
4431         memset(cpu_list, 0, sizeof(cpu_list));
4432
4433         node_mask = cpumask_of_node(dd->numa_node);
4434         if (!cpumask_empty(node_mask)) {
4435                 for_each_cpu(cpu, node_mask)
4436                 {
4437                         snprintf(&cpu_list[j], 256 - j, "%d ", cpu);
4438                         j = strlen(cpu_list);
4439                 }
4440
4441                 dev_info(&pdev->dev, "Node %d on package %d has %d cpu(s): %s\n",
4442                         dd->numa_node,
4443                         topology_physical_package_id(cpumask_first(node_mask)),
4444                         nr_cpus_node(dd->numa_node),
4445                         cpu_list);
4446         } else
4447                 dev_dbg(&pdev->dev, "mtip32xx: node_mask empty\n");
4448
4449         dd->isr_binding = get_least_used_cpu_on_node(dd->numa_node);
4450         dev_info(&pdev->dev, "Initial IRQ binding node:cpu %d:%d\n",
4451                 cpu_to_node(dd->isr_binding), dd->isr_binding);
4452
4453         /* first worker context always runs in ISR */
4454         dd->work[0].cpu_binding = dd->isr_binding;
4455         dd->work[1].cpu_binding = get_least_used_cpu_on_node(dd->numa_node);
4456         dd->work[2].cpu_binding = get_least_used_cpu_on_node(dd->numa_node);
4457         dd->work[3].cpu_binding = dd->work[0].cpu_binding;
4458         dd->work[4].cpu_binding = dd->work[1].cpu_binding;
4459         dd->work[5].cpu_binding = dd->work[2].cpu_binding;
4460         dd->work[6].cpu_binding = dd->work[2].cpu_binding;
4461         dd->work[7].cpu_binding = dd->work[1].cpu_binding;
4462
4463         /* Log the bindings */
4464         for_each_present_cpu(cpu) {
4465                 memset(cpu_list, 0, sizeof(cpu_list));
4466                 for (i = 0, j = 0; i < MTIP_MAX_SLOT_GROUPS; i++) {
4467                         if (dd->work[i].cpu_binding == cpu) {
4468                                 snprintf(&cpu_list[j], 256 - j, "%d ", i);
4469                                 j = strlen(cpu_list);
4470                         }
4471                 }
4472                 if (j)
4473                         dev_info(&pdev->dev, "CPU %d: WQs %s\n", cpu, cpu_list);
4474         }
4475
4476         INIT_WORK(&dd->work[0].work, mtip_workq_sdbf0);
4477         INIT_WORK(&dd->work[1].work, mtip_workq_sdbf1);
4478         INIT_WORK(&dd->work[2].work, mtip_workq_sdbf2);
4479         INIT_WORK(&dd->work[3].work, mtip_workq_sdbf3);
4480         INIT_WORK(&dd->work[4].work, mtip_workq_sdbf4);
4481         INIT_WORK(&dd->work[5].work, mtip_workq_sdbf5);
4482         INIT_WORK(&dd->work[6].work, mtip_workq_sdbf6);
4483         INIT_WORK(&dd->work[7].work, mtip_workq_sdbf7);
4484
4485         pci_set_master(pdev);
4486         rv = pci_enable_msi(pdev);
4487         if (rv) {
4488                 dev_warn(&pdev->dev,
4489                         "Unable to enable MSI interrupt.\n");
4490                 goto block_initialize_err;
4491         }
4492
4493         mtip_fix_ero_nosnoop(dd, pdev);
4494
4495         /* Initialize the block layer. */
4496         rv = mtip_block_initialize(dd);
4497         if (rv < 0) {
4498                 dev_err(&pdev->dev,
4499                         "Unable to initialize block layer\n");
4500                 goto block_initialize_err;
4501         }
4502
4503         /*
4504          * Increment the instance count so that each device has a unique
4505          * instance number.
4506          */
4507         instance++;
4508         if (rv != MTIP_FTL_REBUILD_MAGIC)
4509                 set_bit(MTIP_DDF_INIT_DONE_BIT, &dd->dd_flag);
4510         else
4511                 rv = 0; /* device in rebuild state, return 0 from probe */
4512
4513         /* Add to online list even if in ftl rebuild */
4514         spin_lock_irqsave(&dev_lock, flags);
4515         list_add(&dd->online_list, &online_list);
4516         spin_unlock_irqrestore(&dev_lock, flags);
4517
4518         goto done;
4519
4520 block_initialize_err:
4521         pci_disable_msi(pdev);
4522         if (dd->isr_workq) {
4523                 flush_workqueue(dd->isr_workq);
4524                 destroy_workqueue(dd->isr_workq);
4525                 drop_cpu(dd->work[0].cpu_binding);
4526                 drop_cpu(dd->work[1].cpu_binding);
4527                 drop_cpu(dd->work[2].cpu_binding);
4528         }
4529 setmask_err:
4530         pcim_iounmap_regions(pdev, 1 << MTIP_ABAR);
4531
4532 iomap_err:
4533         kfree(dd);
4534         pci_set_drvdata(pdev, NULL);
4535         return rv;
4536 done:
4537         return rv;
4538 }
4539
4540 /*
4541  * Called for each probed device when the device is removed or the
4542  * driver is unloaded.
4543  *
4544  * return value
4545  *      None
4546  */
4547 static void mtip_pci_remove(struct pci_dev *pdev)
4548 {
4549         struct driver_data *dd = pci_get_drvdata(pdev);
4550         int counter = 0;
4551         unsigned long flags;
4552
4553         set_bit(MTIP_DDF_REMOVE_PENDING_BIT, &dd->dd_flag);
4554
4555         spin_lock_irqsave(&dev_lock, flags);
4556         list_del_init(&dd->online_list);
4557         list_add(&dd->remove_list, &removing_list);
4558         spin_unlock_irqrestore(&dev_lock, flags);
4559
4560         if (mtip_check_surprise_removal(pdev)) {
4561                 while (!test_bit(MTIP_DDF_CLEANUP_BIT, &dd->dd_flag)) {
4562                         counter++;
4563                         msleep(20);
4564                         if (counter == 10) {
4565                                 /* Cleanup the outstanding commands */
4566                                 mtip_command_cleanup(dd);
4567                                 break;
4568                         }
4569                 }
4570         }
4571
4572         /* Clean up the block layer. */
4573         mtip_block_remove(dd);
4574
4575         if (dd->isr_workq) {
4576                 flush_workqueue(dd->isr_workq);
4577                 destroy_workqueue(dd->isr_workq);
4578                 drop_cpu(dd->work[0].cpu_binding);
4579                 drop_cpu(dd->work[1].cpu_binding);
4580                 drop_cpu(dd->work[2].cpu_binding);
4581         }
4582
4583         pci_disable_msi(pdev);
4584
4585         spin_lock_irqsave(&dev_lock, flags);
4586         list_del_init(&dd->remove_list);
4587         spin_unlock_irqrestore(&dev_lock, flags);
4588
4589         kfree(dd);
4590         pcim_iounmap_regions(pdev, 1 << MTIP_ABAR);
4591 }
4592
4593 /*
4594  * Called for each probed device when the device is suspended.
4595  *
4596  * return value
4597  *      0  Success
4598  *      <0 Error
4599  */
4600 static int mtip_pci_suspend(struct pci_dev *pdev, pm_message_t mesg)
4601 {
4602         int rv = 0;
4603         struct driver_data *dd = pci_get_drvdata(pdev);
4604
4605         if (!dd) {
4606                 dev_err(&pdev->dev,
4607                         "Driver private datastructure is NULL\n");
4608                 return -EFAULT;
4609         }
4610
4611         set_bit(MTIP_DDF_RESUME_BIT, &dd->dd_flag);
4612
4613         /* Disable ports & interrupts then send standby immediate */
4614         rv = mtip_block_suspend(dd);
4615         if (rv < 0) {
4616                 dev_err(&pdev->dev,
4617                         "Failed to suspend controller\n");
4618                 return rv;
4619         }
4620
4621         /*
4622          * Save the pci config space to pdev structure &
4623          * disable the device
4624          */
4625         pci_save_state(pdev);
4626         pci_disable_device(pdev);
4627
4628         /* Move to Low power state*/
4629         pci_set_power_state(pdev, PCI_D3hot);
4630
4631         return rv;
4632 }
4633
4634 /*
4635  * Called for each probed device when the device is resumed.
4636  *
4637  * return value
4638  *      0  Success
4639  *      <0 Error
4640  */
4641 static int mtip_pci_resume(struct pci_dev *pdev)
4642 {
4643         int rv = 0;
4644         struct driver_data *dd;
4645
4646         dd = pci_get_drvdata(pdev);
4647         if (!dd) {
4648                 dev_err(&pdev->dev,
4649                         "Driver private datastructure is NULL\n");
4650                 return -EFAULT;
4651         }
4652
4653         /* Move the device to active State */
4654         pci_set_power_state(pdev, PCI_D0);
4655
4656         /* Restore PCI configuration space */
4657         pci_restore_state(pdev);
4658
4659         /* Enable the PCI device*/
4660         rv = pcim_enable_device(pdev);
4661         if (rv < 0) {
4662                 dev_err(&pdev->dev,
4663                         "Failed to enable card during resume\n");
4664                 goto err;
4665         }
4666         pci_set_master(pdev);
4667
4668         /*
4669          * Calls hbaReset, initPort, & startPort function
4670          * then enables interrupts
4671          */
4672         rv = mtip_block_resume(dd);
4673         if (rv < 0)
4674                 dev_err(&pdev->dev, "Unable to resume\n");
4675
4676 err:
4677         clear_bit(MTIP_DDF_RESUME_BIT, &dd->dd_flag);
4678
4679         return rv;
4680 }
4681
4682 /*
4683  * Shutdown routine
4684  *
4685  * return value
4686  *      None
4687  */
4688 static void mtip_pci_shutdown(struct pci_dev *pdev)
4689 {
4690         struct driver_data *dd = pci_get_drvdata(pdev);
4691         if (dd)
4692                 mtip_block_shutdown(dd);
4693 }
4694
4695 /* Table of device ids supported by this driver. */
4696 static DEFINE_PCI_DEVICE_TABLE(mtip_pci_tbl) = {
4697         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320H_DEVICE_ID) },
4698         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320M_DEVICE_ID) },
4699         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P320S_DEVICE_ID) },
4700         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P325M_DEVICE_ID) },
4701         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P420H_DEVICE_ID) },
4702         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P420M_DEVICE_ID) },
4703         { PCI_DEVICE(PCI_VENDOR_ID_MICRON, P425M_DEVICE_ID) },
4704         { 0 }
4705 };
4706
4707 /* Structure that describes the PCI driver functions. */
4708 static struct pci_driver mtip_pci_driver = {
4709         .name                   = MTIP_DRV_NAME,
4710         .id_table               = mtip_pci_tbl,
4711         .probe                  = mtip_pci_probe,
4712         .remove                 = mtip_pci_remove,
4713         .suspend                = mtip_pci_suspend,
4714         .resume                 = mtip_pci_resume,
4715         .shutdown               = mtip_pci_shutdown,
4716 };
4717
4718 MODULE_DEVICE_TABLE(pci, mtip_pci_tbl);
4719
4720 /*
4721  * Module initialization function.
4722  *
4723  * Called once when the module is loaded. This function allocates a major
4724  * block device number to the Cyclone devices and registers the PCI layer
4725  * of the driver.
4726  *
4727  * Return value
4728  *      0 on success else error code.
4729  */
4730 static int __init mtip_init(void)
4731 {
4732         int error;
4733
4734         pr_info(MTIP_DRV_NAME " Version " MTIP_DRV_VERSION "\n");
4735
4736         spin_lock_init(&dev_lock);
4737
4738         INIT_LIST_HEAD(&online_list);
4739         INIT_LIST_HEAD(&removing_list);
4740
4741         /* Allocate a major block device number to use with this driver. */
4742         error = register_blkdev(0, MTIP_DRV_NAME);
4743         if (error <= 0) {
4744                 pr_err("Unable to register block device (%d)\n",
4745                 error);
4746                 return -EBUSY;
4747         }
4748         mtip_major = error;
4749
4750         dfs_parent = debugfs_create_dir("rssd", NULL);
4751         if (IS_ERR_OR_NULL(dfs_parent)) {
4752                 pr_warn("Error creating debugfs parent\n");
4753                 dfs_parent = NULL;
4754         }
4755         if (dfs_parent) {
4756                 dfs_device_status = debugfs_create_file("device_status",
4757                                         S_IRUGO, dfs_parent, NULL,
4758                                         &mtip_device_status_fops);
4759                 if (IS_ERR_OR_NULL(dfs_device_status)) {
4760                         pr_err("Error creating device_status node\n");
4761                         dfs_device_status = NULL;
4762                 }
4763         }
4764
4765         /* Register our PCI operations. */
4766         error = pci_register_driver(&mtip_pci_driver);
4767         if (error) {
4768                 debugfs_remove(dfs_parent);
4769                 unregister_blkdev(mtip_major, MTIP_DRV_NAME);
4770         }
4771
4772         return error;
4773 }
4774
4775 /*
4776  * Module de-initialization function.
4777  *
4778  * Called once when the module is unloaded. This function deallocates
4779  * the major block device number allocated by mtip_init() and
4780  * unregisters the PCI layer of the driver.
4781  *
4782  * Return value
4783  *      none
4784  */
4785 static void __exit mtip_exit(void)
4786 {
4787         /* Release the allocated major block device number. */
4788         unregister_blkdev(mtip_major, MTIP_DRV_NAME);
4789
4790         /* Unregister the PCI driver. */
4791         pci_unregister_driver(&mtip_pci_driver);
4792
4793         debugfs_remove_recursive(dfs_parent);
4794 }
4795
4796 MODULE_AUTHOR("Micron Technology, Inc");
4797 MODULE_DESCRIPTION("Micron RealSSD PCIe Block Driver");
4798 MODULE_LICENSE("GPL");
4799 MODULE_VERSION(MTIP_DRV_VERSION);
4800
4801 module_init(mtip_init);
4802 module_exit(mtip_exit);