fs/nfs/pagelist.c

   1 /*
   2  * linux/fs/nfs/pagelist.c
   3  *
   4  * A set of helper functions for managing NFS read and write requests.
   5  * The main purpose of these routines is to provide support for the
   6  * coalescing of several requests into a single RPC call.
   7  *
   8  * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no>
   9  *
  10  */
  11
  12 #include <linux/slab.h>
  13 #include <linux/file.h>
  14 #include <linux/sched.h>
  15 #include <linux/sunrpc/clnt.h>
  16 #include <linux/nfs.h>
  17 #include <linux/nfs3.h>
  18 #include <linux/nfs4.h>
  19 #include <linux/nfs_page.h>
  20 #include <linux/nfs_fs.h>
  21 #include <linux/nfs_mount.h>
  22 #include <linux/export.h>
  23
  24 #include "internal.h"
  25 #include "pnfs.h"
  26
  27 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  28
  29 static struct kmem_cache *nfs_page_cachep;
  30 static const struct rpc_call_ops nfs_pgio_common_ops;
  31
  32 static bool nfs_pgarray_set(struct nfs_page_array *p, unsigned int pagecount)
  33 {
  34         p->npages = pagecount;
  35         if (pagecount <= ARRAY_SIZE(p->page_array))
  36                 p->pagevec = p->page_array;
  37         else {
  38                 p->pagevec = kcalloc(pagecount, sizeof(struct page *), GFP_KERNEL);
  39                 if (!p->pagevec)
  40                         p->npages = 0;
  41         }
  42         return p->pagevec != NULL;
  43 }
  44
  45 void nfs_pgheader_init(struct nfs_pageio_descriptor *desc,
  46                        struct nfs_pgio_header *hdr,
  47                        void (*release)(struct nfs_pgio_header *hdr))
  48 {
  49         hdr->req = nfs_list_entry(desc->pg_list.next);
  50         hdr->inode = desc->pg_inode;
  51         hdr->cred = hdr->req->wb_context->cred;
  52         hdr->io_start = req_offset(hdr->req);
  53         hdr->good_bytes = desc->pg_count;
  54         hdr->dreq = desc->pg_dreq;
  55         hdr->layout_private = desc->pg_layout_private;
  56         hdr->release = release;
  57         hdr->completion_ops = desc->pg_completion_ops;
  58         if (hdr->completion_ops->init_hdr)
  59                 hdr->completion_ops->init_hdr(hdr);
  60 }
  61 EXPORT_SYMBOL_GPL(nfs_pgheader_init);
  62
  63 void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos)
  64 {
  65         spin_lock(&hdr->lock);
  66         if (pos < hdr->io_start + hdr->good_bytes) {
  67                 set_bit(NFS_IOHDR_ERROR, &hdr->flags);
  68                 clear_bit(NFS_IOHDR_EOF, &hdr->flags);
  69                 hdr->good_bytes = pos - hdr->io_start;
  70                 hdr->error = error;
  71         }
  72         spin_unlock(&hdr->lock);
  73 }
  74
  75 static inline struct nfs_page *
  76 nfs_page_alloc(void)
  77 {
  78         struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_NOIO);
  79         if (p)
  80                 INIT_LIST_HEAD(&p->wb_list);
  81         return p;
  82 }
  83
  84 static inline void
  85 nfs_page_free(struct nfs_page *p)
  86 {
  87         kmem_cache_free(nfs_page_cachep, p);
  88 }
  89
  90 static void
  91 nfs_iocounter_inc(struct nfs_io_counter *c)
  92 {
  93         atomic_inc(&c->io_count);
  94 }
  95
  96 static void
  97 nfs_iocounter_dec(struct nfs_io_counter *c)
  98 {
  99         if (atomic_dec_and_test(&c->io_count)) {
 100                 clear_bit(NFS_IO_INPROGRESS, &c->flags);
 101                 smp_mb__after_atomic();
 102                 wake_up_bit(&c->flags, NFS_IO_INPROGRESS);
 103         }
 104 }
 105
 106 static int
 107 __nfs_iocounter_wait(struct nfs_io_counter *c)
 108 {
 109         wait_queue_head_t *wq = bit_waitqueue(&c->flags, NFS_IO_INPROGRESS);
 110         DEFINE_WAIT_BIT(q, &c->flags, NFS_IO_INPROGRESS);
 111         int ret = 0;
 112
 113         do {
 114                 prepare_to_wait(wq, &q.wait, TASK_KILLABLE);
 115                 set_bit(NFS_IO_INPROGRESS, &c->flags);
 116                 if (atomic_read(&c->io_count) == 0)
 117                         break;
 118                 ret = nfs_wait_bit_killable(&q.key);
 119         } while (atomic_read(&c->io_count) != 0);
 120         finish_wait(wq, &q.wait);
 121         return ret;
 122 }
 123
 124 /**
 125  * nfs_iocounter_wait - wait for i/o to complete
 126  * @c: nfs_io_counter to use
 127  *
 128  * returns -ERESTARTSYS if interrupted by a fatal signal.
 129  * Otherwise returns 0 once the io_count hits 0.
 130  */
 131 int
 132 nfs_iocounter_wait(struct nfs_io_counter *c)
 133 {
 134         if (atomic_read(&c->io_count) == 0)
 135                 return 0;
 136         return __nfs_iocounter_wait(c);
 137 }
 138
 139 /*
 140  * nfs_page_group_lock - lock the head of the page group
 141  * @req - request in group that is to be locked
 142  * @nonblock - if true don't block waiting for lock
 143  *
 144  * this lock must be held if modifying the page group list
 145  *
 146  * returns result from wait_on_bit_lock: 0 on success, < 0 on error
 147  */
 148 int
 149 nfs_page_group_lock(struct nfs_page *req, bool nonblock)
 150 {
 151         struct nfs_page *head = req->wb_head;
 152         int ret;
 153
 154         WARN_ON_ONCE(head != head->wb_head);
 155
 156         do {
 157                 ret = wait_on_bit_lock(&head->wb_flags, PG_HEADLOCK,
 158                         TASK_UNINTERRUPTIBLE);
 159         } while (!nonblock && ret != 0);
 160
 161         WARN_ON_ONCE(ret > 0);
 162         return ret;
 163 }
 164
 165 /*
 166  * nfs_page_group_unlock - unlock the head of the page group
 167  * @req - request in group that is to be unlocked
 168  */
 169 void
 170 nfs_page_group_unlock(struct nfs_page *req)
 171 {
 172         struct nfs_page *head = req->wb_head;
 173
 174         WARN_ON_ONCE(head != head->wb_head);
 175
 176         smp_mb__before_atomic();
 177         clear_bit(PG_HEADLOCK, &head->wb_flags);
 178         smp_mb__after_atomic();
 179         wake_up_bit(&head->wb_flags, PG_HEADLOCK);
 180 }
 181
 182 /*
 183  * nfs_page_group_sync_on_bit_locked
 184  *
 185  * must be called with page group lock held
 186  */
 187 static bool
 188 nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
 189 {
 190         struct nfs_page *head = req->wb_head;
 191         struct nfs_page *tmp;
 192
 193         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_flags));
 194         WARN_ON_ONCE(test_and_set_bit(bit, &req->wb_flags));
 195
 196         tmp = req->wb_this_page;
 197         while (tmp != req) {
 198                 if (!test_bit(bit, &tmp->wb_flags))
 199                         return false;
 200                 tmp = tmp->wb_this_page;
 201         }
 202
 203         /* true! reset all bits */
 204         tmp = req;
 205         do {
 206                 clear_bit(bit, &tmp->wb_flags);
 207                 tmp = tmp->wb_this_page;
 208         } while (tmp != req);
 209
 210         return true;
 211 }
 212
 213 /*
 214  * nfs_page_group_sync_on_bit - set bit on current request, but only
 215  *   return true if the bit is set for all requests in page group
 216  * @req - request in page group
 217  * @bit - PG_* bit that is used to sync page group
 218  */
 219 bool nfs_page_group_sync_on_bit(struct nfs_page *req, unsigned int bit)
 220 {
 221         bool ret;
 222
 223         nfs_page_group_lock(req, false);
 224         ret = nfs_page_group_sync_on_bit_locked(req, bit);
 225         nfs_page_group_unlock(req);
 226
 227         return ret;
 228 }
 229
 230 /*
 231  * nfs_page_group_init - Initialize the page group linkage for @req
 232  * @req - a new nfs request
 233  * @prev - the previous request in page group, or NULL if @req is the first
 234  *         or only request in the group (the head).
 235  */
 236 static inline void
 237 nfs_page_group_init(struct nfs_page *req, struct nfs_page *prev)
 238 {
 239         WARN_ON_ONCE(prev == req);
 240
 241         if (!prev) {
 242                 /* a head request */
 243                 req->wb_head = req;
 244                 req->wb_this_page = req;
 245         } else {
 246                 /* a subrequest */
 247                 WARN_ON_ONCE(prev->wb_this_page != prev->wb_head);
 248                 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &prev->wb_head->wb_flags));
 249                 req->wb_head = prev->wb_head;
 250                 req->wb_this_page = prev->wb_this_page;
 251                 prev->wb_this_page = req;
 252
 253                 /* All subrequests take a ref on the head request until
 254                  * nfs_page_group_destroy is called */
 255                 kref_get(&req->wb_head->wb_kref);
 256
 257                 /* grab extra ref if head request has extra ref from
 258                  * the write/commit path to handle handoff between write
 259                  * and commit lists */
 260                 if (test_bit(PG_INODE_REF, &prev->wb_head->wb_flags)) {
 261                         set_bit(PG_INODE_REF, &req->wb_flags);
 262                         kref_get(&req->wb_kref);
 263                 }
 264         }
 265 }
 266
 267 /*
 268  * nfs_page_group_destroy - sync the destruction of page groups
 269  * @req - request that no longer needs the page group
 270  *
 271  * releases the page group reference from each member once all
 272  * members have called this function.
 273  */
 274 static void
 275 nfs_page_group_destroy(struct kref *kref)
 276 {
 277         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 278         struct nfs_page *tmp, *next;
 279
 280         /* subrequests must release the ref on the head request */
 281         if (req->wb_head != req)
 282                 nfs_release_request(req->wb_head);
 283
 284         if (!nfs_page_group_sync_on_bit(req, PG_TEARDOWN))
 285                 return;
 286
 287         tmp = req;
 288         do {
 289                 next = tmp->wb_this_page;
 290                 /* unlink and free */
 291                 tmp->wb_this_page = tmp;
 292                 tmp->wb_head = tmp;
 293                 nfs_free_request(tmp);
 294                 tmp = next;
 295         } while (tmp != req);
 296 }
 297
 298 /**
 299  * nfs_create_request - Create an NFS read/write request.
 300  * @ctx: open context to use
 301  * @page: page to write
 302  * @last: last nfs request created for this page group or NULL if head
 303  * @offset: starting offset within the page for the write
 304  * @count: number of bytes to read/write
 305  *
 306  * The page must be locked by the caller. This makes sure we never
 307  * create two different requests for the same page.
 308  * User should ensure it is safe to sleep in this function.
 309  */
 310 struct nfs_page *
 311 nfs_create_request(struct nfs_open_context *ctx, struct page *page,
 312                    struct nfs_page *last, unsigned int offset,
 313                    unsigned int count)
 314 {
 315         struct nfs_page         *req;
 316         struct nfs_lock_context *l_ctx;
 317
 318         if (test_bit(NFS_CONTEXT_BAD, &ctx->flags))
 319                 return ERR_PTR(-EBADF);
 320         /* try to allocate the request struct */
 321         req = nfs_page_alloc();
 322         if (req == NULL)
 323                 return ERR_PTR(-ENOMEM);
 324
 325         /* get lock context early so we can deal with alloc failures */
 326         l_ctx = nfs_get_lock_context(ctx);
 327         if (IS_ERR(l_ctx)) {
 328                 nfs_page_free(req);
 329                 return ERR_CAST(l_ctx);
 330         }
 331         req->wb_lock_context = l_ctx;
 332         nfs_iocounter_inc(&l_ctx->io_count);
 333
 334         /* Initialize the request struct. Initially, we assume a
 335          * long write-back delay. This will be adjusted in
 336          * update_nfs_request below if the region is not locked. */
 337         req->wb_page    = page;
 338         req->wb_index   = page_file_index(page);
 339         page_cache_get(page);
 340         req->wb_offset  = offset;
 341         req->wb_pgbase  = offset;
 342         req->wb_bytes   = count;
 343         req->wb_context = get_nfs_open_context(ctx);
 344         kref_init(&req->wb_kref);
 345         nfs_page_group_init(req, last);
 346         return req;
 347 }
 348
 349 /**
 350  * nfs_unlock_request - Unlock request and wake up sleepers.
 351  * @req:
 352  */
 353 void nfs_unlock_request(struct nfs_page *req)
 354 {
 355         if (!NFS_WBACK_BUSY(req)) {
 356                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 357                 BUG();
 358         }
 359         smp_mb__before_atomic();
 360         clear_bit(PG_BUSY, &req->wb_flags);
 361         smp_mb__after_atomic();
 362         wake_up_bit(&req->wb_flags, PG_BUSY);
 363 }
 364
 365 /**
 366  * nfs_unlock_and_release_request - Unlock request and release the nfs_page
 367  * @req:
 368  */
 369 void nfs_unlock_and_release_request(struct nfs_page *req)
 370 {
 371         nfs_unlock_request(req);
 372         nfs_release_request(req);
 373 }
 374
 375 /*
 376  * nfs_clear_request - Free up all resources allocated to the request
 377  * @req:
 378  *
 379  * Release page and open context resources associated with a read/write
 380  * request after it has completed.
 381  */
 382 static void nfs_clear_request(struct nfs_page *req)
 383 {
 384         struct page *page = req->wb_page;
 385         struct nfs_open_context *ctx = req->wb_context;
 386         struct nfs_lock_context *l_ctx = req->wb_lock_context;
 387
 388         if (page != NULL) {
 389                 page_cache_release(page);
 390                 req->wb_page = NULL;
 391         }
 392         if (l_ctx != NULL) {
 393                 nfs_iocounter_dec(&l_ctx->io_count);
 394                 nfs_put_lock_context(l_ctx);
 395                 req->wb_lock_context = NULL;
 396         }
 397         if (ctx != NULL) {
 398                 put_nfs_open_context(ctx);
 399                 req->wb_context = NULL;
 400         }
 401 }
 402
 403 /**
 404  * nfs_release_request - Release the count on an NFS read/write request
 405  * @req: request to release
 406  *
 407  * Note: Should never be called with the spinlock held!
 408  */
 409 void nfs_free_request(struct nfs_page *req)
 410 {
 411         WARN_ON_ONCE(req->wb_this_page != req);
 412
 413         /* extra debug: make sure no sync bits are still set */
 414         WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
 415         WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags));
 416         WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags));
 417         WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags));
 418         WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags));
 419
 420         /* Release struct file and open context */
 421         nfs_clear_request(req);
 422         nfs_page_free(req);
 423 }
 424
 425 void nfs_release_request(struct nfs_page *req)
 426 {
 427         kref_put(&req->wb_kref, nfs_page_group_destroy);
 428 }
 429
 430 /**
 431  * nfs_wait_on_request - Wait for a request to complete.
 432  * @req: request to wait upon.
 433  *
 434  * Interruptible by fatal signals only.
 435  * The user is responsible for holding a count on the request.
 436  */
 437 int
 438 nfs_wait_on_request(struct nfs_page *req)
 439 {
 440         return wait_on_bit_io(&req->wb_flags, PG_BUSY,
 441                               TASK_UNINTERRUPTIBLE);
 442 }
 443
 444 /*
 445  * nfs_generic_pg_test - determine if requests can be coalesced
 446  * @desc: pointer to descriptor
 447  * @prev: previous request in desc, or NULL
 448  * @req: this request
 449  *
 450  * Returns zero if @req can be coalesced into @desc, otherwise it returns
 451  * the size of the request.
 452  */
 453 size_t nfs_generic_pg_test(struct nfs_pageio_descriptor *desc,
 454                            struct nfs_page *prev, struct nfs_page *req)
 455 {
 456         if (desc->pg_count > desc->pg_bsize) {
 457                 /* should never happen */
 458                 WARN_ON_ONCE(1);
 459                 return 0;
 460         }
 461
 462         return min(desc->pg_bsize - desc->pg_count, (size_t)req->wb_bytes);
 463 }
 464 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
 465
 466 struct nfs_pgio_header *nfs_pgio_header_alloc(const struct nfs_rw_ops *ops)
 467 {
 468         struct nfs_pgio_header *hdr = ops->rw_alloc_header();
 469
 470         if (hdr) {
 471                 INIT_LIST_HEAD(&hdr->pages);
 472                 spin_lock_init(&hdr->lock);
 473                 hdr->rw_ops = ops;
 474         }
 475         return hdr;
 476 }
 477 EXPORT_SYMBOL_GPL(nfs_pgio_header_alloc);
 478
 479 /*
 480  * nfs_pgio_header_free - Free a read or write header
 481  * @hdr: The header to free
 482  */
 483 void nfs_pgio_header_free(struct nfs_pgio_header *hdr)
 484 {
 485         hdr->rw_ops->rw_free_header(hdr);
 486 }
 487 EXPORT_SYMBOL_GPL(nfs_pgio_header_free);
 488
 489 /**
 490  * nfs_pgio_data_destroy - make @hdr suitable for reuse
 491  *
 492  * Frees memory and releases refs from nfs_generic_pgio, so that it may
 493  * be called again.
 494  *
 495  * @hdr: A header that has had nfs_generic_pgio called
 496  */
 497 void nfs_pgio_data_destroy(struct nfs_pgio_header *hdr)
 498 {
 499         put_nfs_open_context(hdr->args.context);
 500         if (hdr->page_array.pagevec != hdr->page_array.page_array)
 501                 kfree(hdr->page_array.pagevec);
 502 }
 503 EXPORT_SYMBOL_GPL(nfs_pgio_data_destroy);
 504
 505 /**
 506  * nfs_pgio_rpcsetup - Set up arguments for a pageio call
 507  * @hdr: The pageio hdr
 508  * @count: Number of bytes to read
 509  * @offset: Initial offset
 510  * @how: How to commit data (writes only)
 511  * @cinfo: Commit information for the call (writes only)
 512  */
 513 static void nfs_pgio_rpcsetup(struct nfs_pgio_header *hdr,
 514                               unsigned int count, unsigned int offset,
 515                               int how, struct nfs_commit_info *cinfo)
 516 {
 517         struct nfs_page *req = hdr->req;
 518
 519         /* Set up the RPC argument and reply structs
 520          * NB: take care not to mess about with hdr->commit et al. */
 521
 522         hdr->args.fh     = NFS_FH(hdr->inode);
 523         hdr->args.offset = req_offset(req) + offset;
 524         /* pnfs_set_layoutcommit needs this */
 525         hdr->mds_offset = hdr->args.offset;
 526         hdr->args.pgbase = req->wb_pgbase + offset;
 527         hdr->args.pages  = hdr->page_array.pagevec;
 528         hdr->args.count  = count;
 529         hdr->args.context = get_nfs_open_context(req->wb_context);
 530         hdr->args.lock_context = req->wb_lock_context;
 531         hdr->args.stable  = NFS_UNSTABLE;
 532         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 533         case 0:
 534                 break;
 535         case FLUSH_COND_STABLE:
 536                 if (nfs_reqs_to_commit(cinfo))
 537                         break;
 538         default:
 539                 hdr->args.stable = NFS_FILE_SYNC;
 540         }
 541
 542         hdr->res.fattr   = &hdr->fattr;
 543         hdr->res.count   = count;
 544         hdr->res.eof     = 0;
 545         hdr->res.verf    = &hdr->verf;
 546         nfs_fattr_init(&hdr->fattr);
 547 }
 548
 549 /**
 550  * nfs_pgio_prepare - Prepare pageio hdr to go over the wire
 551  * @task: The current task
 552  * @calldata: pageio header to prepare
 553  */
 554 static void nfs_pgio_prepare(struct rpc_task *task, void *calldata)
 555 {
 556         struct nfs_pgio_header *hdr = calldata;
 557         int err;
 558         err = NFS_PROTO(hdr->inode)->pgio_rpc_prepare(task, hdr);
 559         if (err)
 560                 rpc_exit(task, err);
 561 }
 562
 563 int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_header *hdr,
 564                       const struct rpc_call_ops *call_ops, int how, int flags)
 565 {
 566         struct rpc_task *task;
 567         struct rpc_message msg = {
 568                 .rpc_argp = &hdr->args,
 569                 .rpc_resp = &hdr->res,
 570                 .rpc_cred = hdr->cred,
 571         };
 572         struct rpc_task_setup task_setup_data = {
 573                 .rpc_client = clnt,
 574                 .task = &hdr->task,
 575                 .rpc_message = &msg,
 576                 .callback_ops = call_ops,
 577                 .callback_data = hdr,
 578                 .workqueue = nfsiod_workqueue,
 579                 .flags = RPC_TASK_ASYNC | flags,
 580         };
 581         int ret = 0;
 582
 583         hdr->rw_ops->rw_initiate(hdr, &msg, &task_setup_data, how);
 584
 585         dprintk("NFS: %5u initiated pgio call "
 586                 "(req %s/%llu, %u bytes @ offset %llu)\n",
 587                 hdr->task.tk_pid,
 588                 hdr->inode->i_sb->s_id,
 589                 (unsigned long long)NFS_FILEID(hdr->inode),
 590                 hdr->args.count,
 591                 (unsigned long long)hdr->args.offset);
 592
 593         task = rpc_run_task(&task_setup_data);
 594         if (IS_ERR(task)) {
 595                 ret = PTR_ERR(task);
 596                 goto out;
 597         }
 598         if (how & FLUSH_SYNC) {
 599                 ret = rpc_wait_for_completion_task(task);
 600                 if (ret == 0)
 601                         ret = task->tk_status;
 602         }
 603         rpc_put_task(task);
 604 out:
 605         return ret;
 606 }
 607 EXPORT_SYMBOL_GPL(nfs_initiate_pgio);
 608
 609 /**
 610  * nfs_pgio_error - Clean up from a pageio error
 611  * @desc: IO descriptor
 612  * @hdr: pageio header
 613  */
 614 static int nfs_pgio_error(struct nfs_pageio_descriptor *desc,
 615                           struct nfs_pgio_header *hdr)
 616 {
 617         set_bit(NFS_IOHDR_REDO, &hdr->flags);
 618         nfs_pgio_data_destroy(hdr);
 619         hdr->completion_ops->completion(hdr);
 620         desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 621         return -ENOMEM;
 622 }
 623
 624 /**
 625  * nfs_pgio_release - Release pageio data
 626  * @calldata: The pageio header to release
 627  */
 628 static void nfs_pgio_release(void *calldata)
 629 {
 630         struct nfs_pgio_header *hdr = calldata;
 631         if (hdr->rw_ops->rw_release)
 632                 hdr->rw_ops->rw_release(hdr);
 633         nfs_pgio_data_destroy(hdr);
 634         hdr->completion_ops->completion(hdr);
 635 }
 636
 637 /**
 638  * nfs_pageio_init - initialise a page io descriptor
 639  * @desc: pointer to descriptor
 640  * @inode: pointer to inode
 641  * @doio: pointer to io function
 642  * @bsize: io block size
 643  * @io_flags: extra parameters for the io function
 644  */
 645 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 646                      struct inode *inode,
 647                      const struct nfs_pageio_ops *pg_ops,
 648                      const struct nfs_pgio_completion_ops *compl_ops,
 649                      const struct nfs_rw_ops *rw_ops,
 650                      size_t bsize,
 651                      int io_flags)
 652 {
 653         INIT_LIST_HEAD(&desc->pg_list);
 654         desc->pg_bytes_written = 0;
 655         desc->pg_count = 0;
 656         desc->pg_bsize = bsize;
 657         desc->pg_base = 0;
 658         desc->pg_moreio = 0;
 659         desc->pg_recoalesce = 0;
 660         desc->pg_inode = inode;
 661         desc->pg_ops = pg_ops;
 662         desc->pg_completion_ops = compl_ops;
 663         desc->pg_rw_ops = rw_ops;
 664         desc->pg_ioflags = io_flags;
 665         desc->pg_error = 0;
 666         desc->pg_lseg = NULL;
 667         desc->pg_dreq = NULL;
 668         desc->pg_layout_private = NULL;
 669 }
 670 EXPORT_SYMBOL_GPL(nfs_pageio_init);
 671
 672 /**
 673  * nfs_pgio_result - Basic pageio error handling
 674  * @task: The task that ran
 675  * @calldata: Pageio header to check
 676  */
 677 static void nfs_pgio_result(struct rpc_task *task, void *calldata)
 678 {
 679         struct nfs_pgio_header *hdr = calldata;
 680         struct inode *inode = hdr->inode;
 681
 682         dprintk("NFS: %s: %5u, (status %d)\n", __func__,
 683                 task->tk_pid, task->tk_status);
 684
 685         if (hdr->rw_ops->rw_done(task, hdr, inode) != 0)
 686                 return;
 687         if (task->tk_status < 0)
 688                 nfs_set_pgio_error(hdr, task->tk_status, hdr->args.offset);
 689         else
 690                 hdr->rw_ops->rw_result(task, hdr);
 691 }
 692
 693 /*
 694  * Create an RPC task for the given read or write request and kick it.
 695  * The page must have been locked by the caller.
 696  *
 697  * It may happen that the page we're passed is not marked dirty.
 698  * This is the case if nfs_updatepage detects a conflicting request
 699  * that has been written but not committed.
 700  */
 701 int nfs_generic_pgio(struct nfs_pageio_descriptor *desc,
 702                      struct nfs_pgio_header *hdr)
 703 {
 704         struct nfs_page         *req;
 705         struct page             **pages;
 706         struct list_head *head = &desc->pg_list;
 707         struct nfs_commit_info cinfo;
 708         unsigned int pagecount;
 709
 710         pagecount = nfs_page_array_len(desc->pg_base, desc->pg_count);
 711         if (!nfs_pgarray_set(&hdr->page_array, pagecount))
 712                 return nfs_pgio_error(desc, hdr);
 713
 714         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
 715         pages = hdr->page_array.pagevec;
 716         while (!list_empty(head)) {
 717                 req = nfs_list_entry(head->next);
 718                 nfs_list_remove_request(req);
 719                 nfs_list_add_request(req, &hdr->pages);
 720                 *pages++ = req->wb_page;
 721         }
 722
 723         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
 724             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
 725                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
 726
 727         /* Set up the argument struct */
 728         nfs_pgio_rpcsetup(hdr, desc->pg_count, 0, desc->pg_ioflags, &cinfo);
 729         desc->pg_rpc_callops = &nfs_pgio_common_ops;
 730         return 0;
 731 }
 732 EXPORT_SYMBOL_GPL(nfs_generic_pgio);
 733
 734 static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc)
 735 {
 736         struct nfs_pgio_header *hdr;
 737         int ret;
 738
 739         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
 740         if (!hdr) {
 741                 desc->pg_completion_ops->error_cleanup(&desc->pg_list);
 742                 return -ENOMEM;
 743         }
 744         nfs_pgheader_init(desc, hdr, nfs_pgio_header_free);
 745         ret = nfs_generic_pgio(desc, hdr);
 746         if (ret == 0)
 747                 ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode),
 748                                         hdr, desc->pg_rpc_callops,
 749                                         desc->pg_ioflags, 0);
 750         return ret;
 751 }
 752
 753 static bool nfs_match_open_context(const struct nfs_open_context *ctx1,
 754                 const struct nfs_open_context *ctx2)
 755 {
 756         return ctx1->cred == ctx2->cred && ctx1->state == ctx2->state;
 757 }
 758
 759 static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
 760                 const struct nfs_lock_context *l2)
 761 {
 762         return l1->lockowner.l_owner == l2->lockowner.l_owner
 763                 && l1->lockowner.l_pid == l2->lockowner.l_pid;
 764 }
 765
 766 /**
 767  * nfs_can_coalesce_requests - test two requests for compatibility
 768  * @prev: pointer to nfs_page
 769  * @req: pointer to nfs_page
 770  *
 771  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
 772  * page data area they describe is contiguous, and that their RPC
 773  * credentials, NFSv4 open state, and lockowners are the same.
 774  *
 775  * Return 'true' if this is the case, else return 'false'.
 776  */
 777 static bool nfs_can_coalesce_requests(struct nfs_page *prev,
 778                                       struct nfs_page *req,
 779                                       struct nfs_pageio_descriptor *pgio)
 780 {
 781         size_t size;
 782
 783         if (prev) {
 784                 if (!nfs_match_open_context(req->wb_context, prev->wb_context))
 785                         return false;
 786                 if (req->wb_context->dentry->d_inode->i_flock != NULL &&
 787                     !nfs_match_lock_context(req->wb_lock_context,
 788                                             prev->wb_lock_context))
 789                         return false;
 790                 if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
 791                         return false;
 792         }
 793         size = pgio->pg_ops->pg_test(pgio, prev, req);
 794         WARN_ON_ONCE(size > req->wb_bytes);
 795         if (size && size < req->wb_bytes)
 796                 req->wb_bytes = size;
 797         return size > 0;
 798 }
 799
 800 /**
 801  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
 802  * @desc: destination io descriptor
 803  * @req: request
 804  *
 805  * Returns true if the request 'req' was successfully coalesced into the
 806  * existing list of pages 'desc'.
 807  */
 808 static int nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
 809                                      struct nfs_page *req)
 810 {
 811         struct nfs_page *prev = NULL;
 812         if (desc->pg_count != 0) {
 813                 prev = nfs_list_entry(desc->pg_list.prev);
 814         } else {
 815                 if (desc->pg_ops->pg_init)
 816                         desc->pg_ops->pg_init(desc, req);
 817                 desc->pg_base = req->wb_pgbase;
 818         }
 819         if (!nfs_can_coalesce_requests(prev, req, desc))
 820                 return 0;
 821         nfs_list_remove_request(req);
 822         nfs_list_add_request(req, &desc->pg_list);
 823         desc->pg_count += req->wb_bytes;
 824         return 1;
 825 }
 826
 827 /*
 828  * Helper for nfs_pageio_add_request and nfs_pageio_complete
 829  */
 830 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
 831 {
 832         if (!list_empty(&desc->pg_list)) {
 833                 int error = desc->pg_ops->pg_doio(desc);
 834                 if (error < 0)
 835                         desc->pg_error = error;
 836                 else
 837                         desc->pg_bytes_written += desc->pg_count;
 838         }
 839         if (list_empty(&desc->pg_list)) {
 840                 desc->pg_count = 0;
 841                 desc->pg_base = 0;
 842         }
 843 }
 844
 845 /**
 846  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
 847  * @desc: destination io descriptor
 848  * @req: request
 849  *
 850  * This may split a request into subrequests which are all part of the
 851  * same page group.
 852  *
 853  * Returns true if the request 'req' was successfully coalesced into the
 854  * existing list of pages 'desc'.
 855  */
 856 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 857                            struct nfs_page *req)
 858 {
 859         struct nfs_page *subreq;
 860         unsigned int bytes_left = 0;
 861         unsigned int offset, pgbase;
 862         int ret;
 863
 864         ret = nfs_page_group_lock(req, true);
 865         if (ret < 0) {
 866                 desc->pg_error = ret;
 867                 return 0;
 868         }
 869
 870         subreq = req;
 871         bytes_left = subreq->wb_bytes;
 872         offset = subreq->wb_offset;
 873         pgbase = subreq->wb_pgbase;
 874
 875         do {
 876                 if (!nfs_pageio_do_add_request(desc, subreq)) {
 877                         /* make sure pg_test call(s) did nothing */
 878                         WARN_ON_ONCE(subreq->wb_bytes != bytes_left);
 879                         WARN_ON_ONCE(subreq->wb_offset != offset);
 880                         WARN_ON_ONCE(subreq->wb_pgbase != pgbase);
 881
 882                         nfs_page_group_unlock(req);
 883                         desc->pg_moreio = 1;
 884                         nfs_pageio_doio(desc);
 885                         if (desc->pg_error < 0)
 886                                 return 0;
 887                         if (desc->pg_recoalesce)
 888                                 return 0;
 889                         /* retry add_request for this subreq */
 890                         ret = nfs_page_group_lock(req, true);
 891                         if (ret < 0) {
 892                                 desc->pg_error = ret;
 893                                 return 0;
 894                         }
 895                         continue;
 896                 }
 897
 898                 /* check for buggy pg_test call(s) */
 899                 WARN_ON_ONCE(subreq->wb_bytes + subreq->wb_pgbase > PAGE_SIZE);
 900                 WARN_ON_ONCE(subreq->wb_bytes > bytes_left);
 901                 WARN_ON_ONCE(subreq->wb_bytes == 0);
 902
 903                 bytes_left -= subreq->wb_bytes;
 904                 offset += subreq->wb_bytes;
 905                 pgbase += subreq->wb_bytes;
 906
 907                 if (bytes_left) {
 908                         subreq = nfs_create_request(req->wb_context,
 909                                         req->wb_page,
 910                                         subreq, pgbase, bytes_left);
 911                         if (IS_ERR(subreq))
 912                                 goto err_ptr;
 913                         nfs_lock_request(subreq);
 914                         subreq->wb_offset  = offset;
 915                         subreq->wb_index = req->wb_index;
 916                 }
 917         } while (bytes_left > 0);
 918
 919         nfs_page_group_unlock(req);
 920         return 1;
 921 err_ptr:
 922         desc->pg_error = PTR_ERR(subreq);
 923         nfs_page_group_unlock(req);
 924         return 0;
 925 }
 926
 927 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
 928 {
 929         LIST_HEAD(head);
 930
 931         do {
 932                 list_splice_init(&desc->pg_list, &head);
 933                 desc->pg_bytes_written -= desc->pg_count;
 934                 desc->pg_count = 0;
 935                 desc->pg_base = 0;
 936                 desc->pg_recoalesce = 0;
 937                 desc->pg_moreio = 0;
 938
 939                 while (!list_empty(&head)) {
 940                         struct nfs_page *req;
 941
 942                         req = list_first_entry(&head, struct nfs_page, wb_list);
 943                         nfs_list_remove_request(req);
 944                         if (__nfs_pageio_add_request(desc, req))
 945                                 continue;
 946                         if (desc->pg_error < 0)
 947                                 return 0;
 948                         break;
 949                 }
 950         } while (desc->pg_recoalesce);
 951         return 1;
 952 }
 953
 954 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
 955                 struct nfs_page *req)
 956 {
 957         int ret;
 958
 959         do {
 960                 ret = __nfs_pageio_add_request(desc, req);
 961                 if (ret)
 962                         break;
 963                 if (desc->pg_error < 0)
 964                         break;
 965                 ret = nfs_do_recoalesce(desc);
 966         } while (ret);
 967         return ret;
 968 }
 969
 970 /*
 971  * nfs_pageio_resend - Transfer requests to new descriptor and resend
 972  * @hdr - the pgio header to move request from
 973  * @desc - the pageio descriptor to add requests to
 974  *
 975  * Try to move each request (nfs_page) from @hdr to @desc then attempt
 976  * to send them.
 977  *
 978  * Returns 0 on success and < 0 on error.
 979  */
 980 int nfs_pageio_resend(struct nfs_pageio_descriptor *desc,
 981                       struct nfs_pgio_header *hdr)
 982 {
 983         LIST_HEAD(failed);
 984
 985         desc->pg_dreq = hdr->dreq;
 986         while (!list_empty(&hdr->pages)) {
 987                 struct nfs_page *req = nfs_list_entry(hdr->pages.next);
 988
 989                 nfs_list_remove_request(req);
 990                 if (!nfs_pageio_add_request(desc, req))
 991                         nfs_list_add_request(req, &failed);
 992         }
 993         nfs_pageio_complete(desc);
 994         if (!list_empty(&failed)) {
 995                 list_move(&failed, &hdr->pages);
 996                 return -EIO;
 997         }
 998         return 0;
 999 }
1000 EXPORT_SYMBOL_GPL(nfs_pageio_resend);
1001
1002 /**
1003  * nfs_pageio_complete - Complete I/O on an nfs_pageio_descriptor
1004  * @desc: pointer to io descriptor
1005  */
1006 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
1007 {
1008         for (;;) {
1009                 nfs_pageio_doio(desc);
1010                 if (!desc->pg_recoalesce)
1011                         break;
1012                 if (!nfs_do_recoalesce(desc))
1013                         break;
1014         }
1015 }
1016
1017 /**
1018  * nfs_pageio_cond_complete - Conditional I/O completion
1019  * @desc: pointer to io descriptor
1020  * @index: page index
1021  *
1022  * It is important to ensure that processes don't try to take locks
1023  * on non-contiguous ranges of pages as that might deadlock. This
1024  * function should be called before attempting to wait on a locked
1025  * nfs_page. It will complete the I/O if the page index 'index'
1026  * is not contiguous with the existing list of pages in 'desc'.
1027  */
1028 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
1029 {
1030         if (!list_empty(&desc->pg_list)) {
1031                 struct nfs_page *prev = nfs_list_entry(desc->pg_list.prev);
1032                 if (index != prev->wb_index + 1)
1033                         nfs_pageio_complete(desc);
1034         }
1035 }
1036
1037 int __init nfs_init_nfspagecache(void)
1038 {
1039         nfs_page_cachep = kmem_cache_create("nfs_page",
1040                                             sizeof(struct nfs_page),
1041                                             0, SLAB_HWCACHE_ALIGN,
1042                                             NULL);
1043         if (nfs_page_cachep == NULL)
1044                 return -ENOMEM;
1045
1046         return 0;
1047 }
1048
1049 void nfs_destroy_nfspagecache(void)
1050 {
1051         kmem_cache_destroy(nfs_page_cachep);
1052 }
1053
1054 static const struct rpc_call_ops nfs_pgio_common_ops = {
1055         .rpc_call_prepare = nfs_pgio_prepare,
1056         .rpc_call_done = nfs_pgio_result,
1057         .rpc_release = nfs_pgio_release,
1058 };
1059
1060 const struct nfs_pageio_ops nfs_pgio_rw_ops = {
1061         .pg_test = nfs_generic_pg_test,
1062         .pg_doio = nfs_generic_pg_pgios,
1063 };