Merge remote-tracking branch 'lsk/v3.10/topic/gator' into linux-linaro-lsk
[firefly-linux-kernel-4.4.55.git] / drivers / edac / edac_mc.c
1 /*
2  * edac_mc kernel module
3  * (C) 2005, 2006 Linux Networx (http://lnxi.com)
4  * This file may be distributed under the terms of the
5  * GNU General Public License.
6  *
7  * Written by Thayne Harbaugh
8  * Based on work by Dan Hollis <goemon at anime dot net> and others.
9  *      http://www.anime.net/~goemon/linux-ecc/
10  *
11  * Modified by Dave Peterson and Doug Thompson
12  *
13  */
14
15 #include <linux/module.h>
16 #include <linux/proc_fs.h>
17 #include <linux/kernel.h>
18 #include <linux/types.h>
19 #include <linux/smp.h>
20 #include <linux/init.h>
21 #include <linux/sysctl.h>
22 #include <linux/highmem.h>
23 #include <linux/timer.h>
24 #include <linux/slab.h>
25 #include <linux/jiffies.h>
26 #include <linux/spinlock.h>
27 #include <linux/list.h>
28 #include <linux/ctype.h>
29 #include <linux/edac.h>
30 #include <linux/bitops.h>
31 #include <asm/uaccess.h>
32 #include <asm/page.h>
33 #include <asm/edac.h>
34 #include "edac_core.h"
35 #include "edac_module.h"
36
37 #define CREATE_TRACE_POINTS
38 #define TRACE_INCLUDE_PATH ../../include/ras
39 #include <ras/ras_event.h>
40
41 /* lock to memory controller's control array */
42 static DEFINE_MUTEX(mem_ctls_mutex);
43 static LIST_HEAD(mc_devices);
44
45 /*
46  * Used to lock EDAC MC to just one module, avoiding two drivers e. g.
47  *      apei/ghes and i7core_edac to be used at the same time.
48  */
49 static void const *edac_mc_owner;
50
51 static struct bus_type mc_bus[EDAC_MAX_MCS];
52
53 unsigned edac_dimm_info_location(struct dimm_info *dimm, char *buf,
54                                  unsigned len)
55 {
56         struct mem_ctl_info *mci = dimm->mci;
57         int i, n, count = 0;
58         char *p = buf;
59
60         for (i = 0; i < mci->n_layers; i++) {
61                 n = snprintf(p, len, "%s %d ",
62                               edac_layer_name[mci->layers[i].type],
63                               dimm->location[i]);
64                 p += n;
65                 len -= n;
66                 count += n;
67                 if (!len)
68                         break;
69         }
70
71         return count;
72 }
73
74 #ifdef CONFIG_EDAC_DEBUG
75
76 static void edac_mc_dump_channel(struct rank_info *chan)
77 {
78         edac_dbg(4, "  channel->chan_idx = %d\n", chan->chan_idx);
79         edac_dbg(4, "    channel = %p\n", chan);
80         edac_dbg(4, "    channel->csrow = %p\n", chan->csrow);
81         edac_dbg(4, "    channel->dimm = %p\n", chan->dimm);
82 }
83
84 static void edac_mc_dump_dimm(struct dimm_info *dimm, int number)
85 {
86         char location[80];
87
88         edac_dimm_info_location(dimm, location, sizeof(location));
89
90         edac_dbg(4, "%s%i: %smapped as virtual row %d, chan %d\n",
91                  dimm->mci->csbased ? "rank" : "dimm",
92                  number, location, dimm->csrow, dimm->cschannel);
93         edac_dbg(4, "  dimm = %p\n", dimm);
94         edac_dbg(4, "  dimm->label = '%s'\n", dimm->label);
95         edac_dbg(4, "  dimm->nr_pages = 0x%x\n", dimm->nr_pages);
96         edac_dbg(4, "  dimm->grain = %d\n", dimm->grain);
97         edac_dbg(4, "  dimm->nr_pages = 0x%x\n", dimm->nr_pages);
98 }
99
100 static void edac_mc_dump_csrow(struct csrow_info *csrow)
101 {
102         edac_dbg(4, "csrow->csrow_idx = %d\n", csrow->csrow_idx);
103         edac_dbg(4, "  csrow = %p\n", csrow);
104         edac_dbg(4, "  csrow->first_page = 0x%lx\n", csrow->first_page);
105         edac_dbg(4, "  csrow->last_page = 0x%lx\n", csrow->last_page);
106         edac_dbg(4, "  csrow->page_mask = 0x%lx\n", csrow->page_mask);
107         edac_dbg(4, "  csrow->nr_channels = %d\n", csrow->nr_channels);
108         edac_dbg(4, "  csrow->channels = %p\n", csrow->channels);
109         edac_dbg(4, "  csrow->mci = %p\n", csrow->mci);
110 }
111
112 static void edac_mc_dump_mci(struct mem_ctl_info *mci)
113 {
114         edac_dbg(3, "\tmci = %p\n", mci);
115         edac_dbg(3, "\tmci->mtype_cap = %lx\n", mci->mtype_cap);
116         edac_dbg(3, "\tmci->edac_ctl_cap = %lx\n", mci->edac_ctl_cap);
117         edac_dbg(3, "\tmci->edac_cap = %lx\n", mci->edac_cap);
118         edac_dbg(4, "\tmci->edac_check = %p\n", mci->edac_check);
119         edac_dbg(3, "\tmci->nr_csrows = %d, csrows = %p\n",
120                  mci->nr_csrows, mci->csrows);
121         edac_dbg(3, "\tmci->nr_dimms = %d, dimms = %p\n",
122                  mci->tot_dimms, mci->dimms);
123         edac_dbg(3, "\tdev = %p\n", mci->pdev);
124         edac_dbg(3, "\tmod_name:ctl_name = %s:%s\n",
125                  mci->mod_name, mci->ctl_name);
126         edac_dbg(3, "\tpvt_info = %p\n\n", mci->pvt_info);
127 }
128
129 #endif                          /* CONFIG_EDAC_DEBUG */
130
131 /*
132  * keep those in sync with the enum mem_type
133  */
134 const char *edac_mem_types[] = {
135         "Empty csrow",
136         "Reserved csrow type",
137         "Unknown csrow type",
138         "Fast page mode RAM",
139         "Extended data out RAM",
140         "Burst Extended data out RAM",
141         "Single data rate SDRAM",
142         "Registered single data rate SDRAM",
143         "Double data rate SDRAM",
144         "Registered Double data rate SDRAM",
145         "Rambus DRAM",
146         "Unbuffered DDR2 RAM",
147         "Fully buffered DDR2",
148         "Registered DDR2 RAM",
149         "Rambus XDR",
150         "Unbuffered DDR3 RAM",
151         "Registered DDR3 RAM",
152 };
153 EXPORT_SYMBOL_GPL(edac_mem_types);
154
155 /**
156  * edac_align_ptr - Prepares the pointer offsets for a single-shot allocation
157  * @p:          pointer to a pointer with the memory offset to be used. At
158  *              return, this will be incremented to point to the next offset
159  * @size:       Size of the data structure to be reserved
160  * @n_elems:    Number of elements that should be reserved
161  *
162  * If 'size' is a constant, the compiler will optimize this whole function
163  * down to either a no-op or the addition of a constant to the value of '*p'.
164  *
165  * The 'p' pointer is absolutely needed to keep the proper advancing
166  * further in memory to the proper offsets when allocating the struct along
167  * with its embedded structs, as edac_device_alloc_ctl_info() does it
168  * above, for example.
169  *
170  * At return, the pointer 'p' will be incremented to be used on a next call
171  * to this function.
172  */
173 void *edac_align_ptr(void **p, unsigned size, int n_elems)
174 {
175         unsigned align, r;
176         void *ptr = *p;
177
178         *p += size * n_elems;
179
180         /*
181          * 'p' can possibly be an unaligned item X such that sizeof(X) is
182          * 'size'.  Adjust 'p' so that its alignment is at least as
183          * stringent as what the compiler would provide for X and return
184          * the aligned result.
185          * Here we assume that the alignment of a "long long" is the most
186          * stringent alignment that the compiler will ever provide by default.
187          * As far as I know, this is a reasonable assumption.
188          */
189         if (size > sizeof(long))
190                 align = sizeof(long long);
191         else if (size > sizeof(int))
192                 align = sizeof(long);
193         else if (size > sizeof(short))
194                 align = sizeof(int);
195         else if (size > sizeof(char))
196                 align = sizeof(short);
197         else
198                 return (char *)ptr;
199
200         r = (unsigned long)p % align;
201
202         if (r == 0)
203                 return (char *)ptr;
204
205         *p += align - r;
206
207         return (void *)(((unsigned long)ptr) + align - r);
208 }
209
210 static void _edac_mc_free(struct mem_ctl_info *mci)
211 {
212         int i, chn, row;
213         struct csrow_info *csr;
214         const unsigned int tot_dimms = mci->tot_dimms;
215         const unsigned int tot_channels = mci->num_cschannel;
216         const unsigned int tot_csrows = mci->nr_csrows;
217
218         if (mci->dimms) {
219                 for (i = 0; i < tot_dimms; i++)
220                         kfree(mci->dimms[i]);
221                 kfree(mci->dimms);
222         }
223         if (mci->csrows) {
224                 for (row = 0; row < tot_csrows; row++) {
225                         csr = mci->csrows[row];
226                         if (csr) {
227                                 if (csr->channels) {
228                                         for (chn = 0; chn < tot_channels; chn++)
229                                                 kfree(csr->channels[chn]);
230                                         kfree(csr->channels);
231                                 }
232                                 kfree(csr);
233                         }
234                 }
235                 kfree(mci->csrows);
236         }
237         kfree(mci);
238 }
239
240 /**
241  * edac_mc_alloc: Allocate and partially fill a struct mem_ctl_info structure
242  * @mc_num:             Memory controller number
243  * @n_layers:           Number of MC hierarchy layers
244  * layers:              Describes each layer as seen by the Memory Controller
245  * @size_pvt:           size of private storage needed
246  *
247  *
248  * Everything is kmalloc'ed as one big chunk - more efficient.
249  * Only can be used if all structures have the same lifetime - otherwise
250  * you have to allocate and initialize your own structures.
251  *
252  * Use edac_mc_free() to free mc structures allocated by this function.
253  *
254  * NOTE: drivers handle multi-rank memories in different ways: in some
255  * drivers, one multi-rank memory stick is mapped as one entry, while, in
256  * others, a single multi-rank memory stick would be mapped into several
257  * entries. Currently, this function will allocate multiple struct dimm_info
258  * on such scenarios, as grouping the multiple ranks require drivers change.
259  *
260  * Returns:
261  *      On failure: NULL
262  *      On success: struct mem_ctl_info pointer
263  */
264 struct mem_ctl_info *edac_mc_alloc(unsigned mc_num,
265                                    unsigned n_layers,
266                                    struct edac_mc_layer *layers,
267                                    unsigned sz_pvt)
268 {
269         struct mem_ctl_info *mci;
270         struct edac_mc_layer *layer;
271         struct csrow_info *csr;
272         struct rank_info *chan;
273         struct dimm_info *dimm;
274         u32 *ce_per_layer[EDAC_MAX_LAYERS], *ue_per_layer[EDAC_MAX_LAYERS];
275         unsigned pos[EDAC_MAX_LAYERS];
276         unsigned size, tot_dimms = 1, count = 1;
277         unsigned tot_csrows = 1, tot_channels = 1, tot_errcount = 0;
278         void *pvt, *p, *ptr = NULL;
279         int i, j, row, chn, n, len, off;
280         bool per_rank = false;
281
282         BUG_ON(n_layers > EDAC_MAX_LAYERS || n_layers == 0);
283         /*
284          * Calculate the total amount of dimms and csrows/cschannels while
285          * in the old API emulation mode
286          */
287         for (i = 0; i < n_layers; i++) {
288                 tot_dimms *= layers[i].size;
289                 if (layers[i].is_virt_csrow)
290                         tot_csrows *= layers[i].size;
291                 else
292                         tot_channels *= layers[i].size;
293
294                 if (layers[i].type == EDAC_MC_LAYER_CHIP_SELECT)
295                         per_rank = true;
296         }
297
298         /* Figure out the offsets of the various items from the start of an mc
299          * structure.  We want the alignment of each item to be at least as
300          * stringent as what the compiler would provide if we could simply
301          * hardcode everything into a single struct.
302          */
303         mci = edac_align_ptr(&ptr, sizeof(*mci), 1);
304         layer = edac_align_ptr(&ptr, sizeof(*layer), n_layers);
305         for (i = 0; i < n_layers; i++) {
306                 count *= layers[i].size;
307                 edac_dbg(4, "errcount layer %d size %d\n", i, count);
308                 ce_per_layer[i] = edac_align_ptr(&ptr, sizeof(u32), count);
309                 ue_per_layer[i] = edac_align_ptr(&ptr, sizeof(u32), count);
310                 tot_errcount += 2 * count;
311         }
312
313         edac_dbg(4, "allocating %d error counters\n", tot_errcount);
314         pvt = edac_align_ptr(&ptr, sz_pvt, 1);
315         size = ((unsigned long)pvt) + sz_pvt;
316
317         edac_dbg(1, "allocating %u bytes for mci data (%d %s, %d csrows/channels)\n",
318                  size,
319                  tot_dimms,
320                  per_rank ? "ranks" : "dimms",
321                  tot_csrows * tot_channels);
322
323         mci = kzalloc(size, GFP_KERNEL);
324         if (mci == NULL)
325                 return NULL;
326
327         /* Adjust pointers so they point within the memory we just allocated
328          * rather than an imaginary chunk of memory located at address 0.
329          */
330         layer = (struct edac_mc_layer *)(((char *)mci) + ((unsigned long)layer));
331         for (i = 0; i < n_layers; i++) {
332                 mci->ce_per_layer[i] = (u32 *)((char *)mci + ((unsigned long)ce_per_layer[i]));
333                 mci->ue_per_layer[i] = (u32 *)((char *)mci + ((unsigned long)ue_per_layer[i]));
334         }
335         pvt = sz_pvt ? (((char *)mci) + ((unsigned long)pvt)) : NULL;
336
337         /* setup index and various internal pointers */
338         mci->mc_idx = mc_num;
339         mci->tot_dimms = tot_dimms;
340         mci->pvt_info = pvt;
341         mci->n_layers = n_layers;
342         mci->layers = layer;
343         memcpy(mci->layers, layers, sizeof(*layer) * n_layers);
344         mci->nr_csrows = tot_csrows;
345         mci->num_cschannel = tot_channels;
346         mci->csbased = per_rank;
347
348         /*
349          * Alocate and fill the csrow/channels structs
350          */
351         mci->csrows = kcalloc(tot_csrows, sizeof(*mci->csrows), GFP_KERNEL);
352         if (!mci->csrows)
353                 goto error;
354         for (row = 0; row < tot_csrows; row++) {
355                 csr = kzalloc(sizeof(**mci->csrows), GFP_KERNEL);
356                 if (!csr)
357                         goto error;
358                 mci->csrows[row] = csr;
359                 csr->csrow_idx = row;
360                 csr->mci = mci;
361                 csr->nr_channels = tot_channels;
362                 csr->channels = kcalloc(tot_channels, sizeof(*csr->channels),
363                                         GFP_KERNEL);
364                 if (!csr->channels)
365                         goto error;
366
367                 for (chn = 0; chn < tot_channels; chn++) {
368                         chan = kzalloc(sizeof(**csr->channels), GFP_KERNEL);
369                         if (!chan)
370                                 goto error;
371                         csr->channels[chn] = chan;
372                         chan->chan_idx = chn;
373                         chan->csrow = csr;
374                 }
375         }
376
377         /*
378          * Allocate and fill the dimm structs
379          */
380         mci->dimms  = kcalloc(tot_dimms, sizeof(*mci->dimms), GFP_KERNEL);
381         if (!mci->dimms)
382                 goto error;
383
384         memset(&pos, 0, sizeof(pos));
385         row = 0;
386         chn = 0;
387         for (i = 0; i < tot_dimms; i++) {
388                 chan = mci->csrows[row]->channels[chn];
389                 off = EDAC_DIMM_OFF(layer, n_layers, pos[0], pos[1], pos[2]);
390                 if (off < 0 || off >= tot_dimms) {
391                         edac_mc_printk(mci, KERN_ERR, "EDAC core bug: EDAC_DIMM_OFF is trying to do an illegal data access\n");
392                         goto error;
393                 }
394
395                 dimm = kzalloc(sizeof(**mci->dimms), GFP_KERNEL);
396                 if (!dimm)
397                         goto error;
398                 mci->dimms[off] = dimm;
399                 dimm->mci = mci;
400
401                 /*
402                  * Copy DIMM location and initialize it.
403                  */
404                 len = sizeof(dimm->label);
405                 p = dimm->label;
406                 n = snprintf(p, len, "mc#%u", mc_num);
407                 p += n;
408                 len -= n;
409                 for (j = 0; j < n_layers; j++) {
410                         n = snprintf(p, len, "%s#%u",
411                                      edac_layer_name[layers[j].type],
412                                      pos[j]);
413                         p += n;
414                         len -= n;
415                         dimm->location[j] = pos[j];
416
417                         if (len <= 0)
418                                 break;
419                 }
420
421                 /* Link it to the csrows old API data */
422                 chan->dimm = dimm;
423                 dimm->csrow = row;
424                 dimm->cschannel = chn;
425
426                 /* Increment csrow location */
427                 if (layers[0].is_virt_csrow) {
428                         chn++;
429                         if (chn == tot_channels) {
430                                 chn = 0;
431                                 row++;
432                         }
433                 } else {
434                         row++;
435                         if (row == tot_csrows) {
436                                 row = 0;
437                                 chn++;
438                         }
439                 }
440
441                 /* Increment dimm location */
442                 for (j = n_layers - 1; j >= 0; j--) {
443                         pos[j]++;
444                         if (pos[j] < layers[j].size)
445                                 break;
446                         pos[j] = 0;
447                 }
448         }
449
450         mci->op_state = OP_ALLOC;
451
452         return mci;
453
454 error:
455         _edac_mc_free(mci);
456
457         return NULL;
458 }
459 EXPORT_SYMBOL_GPL(edac_mc_alloc);
460
461 /**
462  * edac_mc_free
463  *      'Free' a previously allocated 'mci' structure
464  * @mci: pointer to a struct mem_ctl_info structure
465  */
466 void edac_mc_free(struct mem_ctl_info *mci)
467 {
468         edac_dbg(1, "\n");
469
470         /* If we're not yet registered with sysfs free only what was allocated
471          * in edac_mc_alloc().
472          */
473         if (!device_is_registered(&mci->dev)) {
474                 _edac_mc_free(mci);
475                 return;
476         }
477
478         /* the mci instance is freed here, when the sysfs object is dropped */
479         edac_unregister_sysfs(mci);
480 }
481 EXPORT_SYMBOL_GPL(edac_mc_free);
482
483
484 /**
485  * find_mci_by_dev
486  *
487  *      scan list of controllers looking for the one that manages
488  *      the 'dev' device
489  * @dev: pointer to a struct device related with the MCI
490  */
491 struct mem_ctl_info *find_mci_by_dev(struct device *dev)
492 {
493         struct mem_ctl_info *mci;
494         struct list_head *item;
495
496         edac_dbg(3, "\n");
497
498         list_for_each(item, &mc_devices) {
499                 mci = list_entry(item, struct mem_ctl_info, link);
500
501                 if (mci->pdev == dev)
502                         return mci;
503         }
504
505         return NULL;
506 }
507 EXPORT_SYMBOL_GPL(find_mci_by_dev);
508
509 /*
510  * handler for EDAC to check if NMI type handler has asserted interrupt
511  */
512 static int edac_mc_assert_error_check_and_clear(void)
513 {
514         int old_state;
515
516         if (edac_op_state == EDAC_OPSTATE_POLL)
517                 return 1;
518
519         old_state = edac_err_assert;
520         edac_err_assert = 0;
521
522         return old_state;
523 }
524
525 /*
526  * edac_mc_workq_function
527  *      performs the operation scheduled by a workq request
528  */
529 static void edac_mc_workq_function(struct work_struct *work_req)
530 {
531         struct delayed_work *d_work = to_delayed_work(work_req);
532         struct mem_ctl_info *mci = to_edac_mem_ctl_work(d_work);
533
534         mutex_lock(&mem_ctls_mutex);
535
536         /* if this control struct has movd to offline state, we are done */
537         if (mci->op_state == OP_OFFLINE) {
538                 mutex_unlock(&mem_ctls_mutex);
539                 return;
540         }
541
542         /* Only poll controllers that are running polled and have a check */
543         if (edac_mc_assert_error_check_and_clear() && (mci->edac_check != NULL))
544                 mci->edac_check(mci);
545
546         mutex_unlock(&mem_ctls_mutex);
547
548         /* Reschedule */
549         queue_delayed_work(edac_workqueue, &mci->work,
550                         msecs_to_jiffies(edac_mc_get_poll_msec()));
551 }
552
553 /*
554  * edac_mc_workq_setup
555  *      initialize a workq item for this mci
556  *      passing in the new delay period in msec
557  *
558  *      locking model:
559  *
560  *              called with the mem_ctls_mutex held
561  */
562 static void edac_mc_workq_setup(struct mem_ctl_info *mci, unsigned msec,
563                                 bool init)
564 {
565         edac_dbg(0, "\n");
566
567         /* if this instance is not in the POLL state, then simply return */
568         if (mci->op_state != OP_RUNNING_POLL)
569                 return;
570
571         if (init)
572                 INIT_DELAYED_WORK(&mci->work, edac_mc_workq_function);
573
574         mod_delayed_work(edac_workqueue, &mci->work, msecs_to_jiffies(msec));
575 }
576
577 /*
578  * edac_mc_workq_teardown
579  *      stop the workq processing on this mci
580  *
581  *      locking model:
582  *
583  *              called WITHOUT lock held
584  */
585 static void edac_mc_workq_teardown(struct mem_ctl_info *mci)
586 {
587         int status;
588
589         if (mci->op_state != OP_RUNNING_POLL)
590                 return;
591
592         status = cancel_delayed_work(&mci->work);
593         if (status == 0) {
594                 edac_dbg(0, "not canceled, flush the queue\n");
595
596                 /* workq instance might be running, wait for it */
597                 flush_workqueue(edac_workqueue);
598         }
599 }
600
601 /*
602  * edac_mc_reset_delay_period(unsigned long value)
603  *
604  *      user space has updated our poll period value, need to
605  *      reset our workq delays
606  */
607 void edac_mc_reset_delay_period(unsigned long value)
608 {
609         struct mem_ctl_info *mci;
610         struct list_head *item;
611
612         mutex_lock(&mem_ctls_mutex);
613
614         list_for_each(item, &mc_devices) {
615                 mci = list_entry(item, struct mem_ctl_info, link);
616
617                 edac_mc_workq_setup(mci, value, false);
618         }
619
620         mutex_unlock(&mem_ctls_mutex);
621 }
622
623
624
625 /* Return 0 on success, 1 on failure.
626  * Before calling this function, caller must
627  * assign a unique value to mci->mc_idx.
628  *
629  *      locking model:
630  *
631  *              called with the mem_ctls_mutex lock held
632  */
633 static int add_mc_to_global_list(struct mem_ctl_info *mci)
634 {
635         struct list_head *item, *insert_before;
636         struct mem_ctl_info *p;
637
638         insert_before = &mc_devices;
639
640         p = find_mci_by_dev(mci->pdev);
641         if (unlikely(p != NULL))
642                 goto fail0;
643
644         list_for_each(item, &mc_devices) {
645                 p = list_entry(item, struct mem_ctl_info, link);
646
647                 if (p->mc_idx >= mci->mc_idx) {
648                         if (unlikely(p->mc_idx == mci->mc_idx))
649                                 goto fail1;
650
651                         insert_before = item;
652                         break;
653                 }
654         }
655
656         list_add_tail_rcu(&mci->link, insert_before);
657         atomic_inc(&edac_handlers);
658         return 0;
659
660 fail0:
661         edac_printk(KERN_WARNING, EDAC_MC,
662                 "%s (%s) %s %s already assigned %d\n", dev_name(p->pdev),
663                 edac_dev_name(mci), p->mod_name, p->ctl_name, p->mc_idx);
664         return 1;
665
666 fail1:
667         edac_printk(KERN_WARNING, EDAC_MC,
668                 "bug in low-level driver: attempt to assign\n"
669                 "    duplicate mc_idx %d in %s()\n", p->mc_idx, __func__);
670         return 1;
671 }
672
673 static int del_mc_from_global_list(struct mem_ctl_info *mci)
674 {
675         int handlers = atomic_dec_return(&edac_handlers);
676         list_del_rcu(&mci->link);
677
678         /* these are for safe removal of devices from global list while
679          * NMI handlers may be traversing list
680          */
681         synchronize_rcu();
682         INIT_LIST_HEAD(&mci->link);
683
684         return handlers;
685 }
686
687 /**
688  * edac_mc_find: Search for a mem_ctl_info structure whose index is 'idx'.
689  *
690  * If found, return a pointer to the structure.
691  * Else return NULL.
692  *
693  * Caller must hold mem_ctls_mutex.
694  */
695 struct mem_ctl_info *edac_mc_find(int idx)
696 {
697         struct list_head *item;
698         struct mem_ctl_info *mci;
699
700         list_for_each(item, &mc_devices) {
701                 mci = list_entry(item, struct mem_ctl_info, link);
702
703                 if (mci->mc_idx >= idx) {
704                         if (mci->mc_idx == idx)
705                                 return mci;
706
707                         break;
708                 }
709         }
710
711         return NULL;
712 }
713 EXPORT_SYMBOL(edac_mc_find);
714
715 /**
716  * edac_mc_add_mc: Insert the 'mci' structure into the mci global list and
717  *                 create sysfs entries associated with mci structure
718  * @mci: pointer to the mci structure to be added to the list
719  *
720  * Return:
721  *      0       Success
722  *      !0      Failure
723  */
724
725 /* FIXME - should a warning be printed if no error detection? correction? */
726 int edac_mc_add_mc(struct mem_ctl_info *mci)
727 {
728         int ret = -EINVAL;
729         edac_dbg(0, "\n");
730
731         if (mci->mc_idx >= EDAC_MAX_MCS) {
732                 pr_warn_once("Too many memory controllers: %d\n", mci->mc_idx);
733                 return -ENODEV;
734         }
735
736 #ifdef CONFIG_EDAC_DEBUG
737         if (edac_debug_level >= 3)
738                 edac_mc_dump_mci(mci);
739
740         if (edac_debug_level >= 4) {
741                 int i;
742
743                 for (i = 0; i < mci->nr_csrows; i++) {
744                         struct csrow_info *csrow = mci->csrows[i];
745                         u32 nr_pages = 0;
746                         int j;
747
748                         for (j = 0; j < csrow->nr_channels; j++)
749                                 nr_pages += csrow->channels[j]->dimm->nr_pages;
750                         if (!nr_pages)
751                                 continue;
752                         edac_mc_dump_csrow(csrow);
753                         for (j = 0; j < csrow->nr_channels; j++)
754                                 if (csrow->channels[j]->dimm->nr_pages)
755                                         edac_mc_dump_channel(csrow->channels[j]);
756                 }
757                 for (i = 0; i < mci->tot_dimms; i++)
758                         if (mci->dimms[i]->nr_pages)
759                                 edac_mc_dump_dimm(mci->dimms[i], i);
760         }
761 #endif
762         mutex_lock(&mem_ctls_mutex);
763
764         if (edac_mc_owner && edac_mc_owner != mci->mod_name) {
765                 ret = -EPERM;
766                 goto fail0;
767         }
768
769         if (add_mc_to_global_list(mci))
770                 goto fail0;
771
772         /* set load time so that error rate can be tracked */
773         mci->start_time = jiffies;
774
775         mci->bus = &mc_bus[mci->mc_idx];
776
777         if (edac_create_sysfs_mci_device(mci)) {
778                 edac_mc_printk(mci, KERN_WARNING,
779                         "failed to create sysfs device\n");
780                 goto fail1;
781         }
782
783         /* If there IS a check routine, then we are running POLLED */
784         if (mci->edac_check != NULL) {
785                 /* This instance is NOW RUNNING */
786                 mci->op_state = OP_RUNNING_POLL;
787
788                 edac_mc_workq_setup(mci, edac_mc_get_poll_msec(), true);
789         } else {
790                 mci->op_state = OP_RUNNING_INTERRUPT;
791         }
792
793         /* Report action taken */
794         edac_mc_printk(mci, KERN_INFO, "Giving out device to '%s' '%s':"
795                 " DEV %s\n", mci->mod_name, mci->ctl_name, edac_dev_name(mci));
796
797         edac_mc_owner = mci->mod_name;
798
799         mutex_unlock(&mem_ctls_mutex);
800         return 0;
801
802 fail1:
803         del_mc_from_global_list(mci);
804
805 fail0:
806         mutex_unlock(&mem_ctls_mutex);
807         return ret;
808 }
809 EXPORT_SYMBOL_GPL(edac_mc_add_mc);
810
811 /**
812  * edac_mc_del_mc: Remove sysfs entries for specified mci structure and
813  *                 remove mci structure from global list
814  * @pdev: Pointer to 'struct device' representing mci structure to remove.
815  *
816  * Return pointer to removed mci structure, or NULL if device not found.
817  */
818 struct mem_ctl_info *edac_mc_del_mc(struct device *dev)
819 {
820         struct mem_ctl_info *mci;
821
822         edac_dbg(0, "\n");
823
824         mutex_lock(&mem_ctls_mutex);
825
826         /* find the requested mci struct in the global list */
827         mci = find_mci_by_dev(dev);
828         if (mci == NULL) {
829                 mutex_unlock(&mem_ctls_mutex);
830                 return NULL;
831         }
832
833         if (!del_mc_from_global_list(mci))
834                 edac_mc_owner = NULL;
835         mutex_unlock(&mem_ctls_mutex);
836
837         /* flush workq processes */
838         edac_mc_workq_teardown(mci);
839
840         /* marking MCI offline */
841         mci->op_state = OP_OFFLINE;
842
843         /* remove from sysfs */
844         edac_remove_sysfs_mci_device(mci);
845
846         edac_printk(KERN_INFO, EDAC_MC,
847                 "Removed device %d for %s %s: DEV %s\n", mci->mc_idx,
848                 mci->mod_name, mci->ctl_name, edac_dev_name(mci));
849
850         return mci;
851 }
852 EXPORT_SYMBOL_GPL(edac_mc_del_mc);
853
854 static void edac_mc_scrub_block(unsigned long page, unsigned long offset,
855                                 u32 size)
856 {
857         struct page *pg;
858         void *virt_addr;
859         unsigned long flags = 0;
860
861         edac_dbg(3, "\n");
862
863         /* ECC error page was not in our memory. Ignore it. */
864         if (!pfn_valid(page))
865                 return;
866
867         /* Find the actual page structure then map it and fix */
868         pg = pfn_to_page(page);
869
870         if (PageHighMem(pg))
871                 local_irq_save(flags);
872
873         virt_addr = kmap_atomic(pg);
874
875         /* Perform architecture specific atomic scrub operation */
876         atomic_scrub(virt_addr + offset, size);
877
878         /* Unmap and complete */
879         kunmap_atomic(virt_addr);
880
881         if (PageHighMem(pg))
882                 local_irq_restore(flags);
883 }
884
885 /* FIXME - should return -1 */
886 int edac_mc_find_csrow_by_page(struct mem_ctl_info *mci, unsigned long page)
887 {
888         struct csrow_info **csrows = mci->csrows;
889         int row, i, j, n;
890
891         edac_dbg(1, "MC%d: 0x%lx\n", mci->mc_idx, page);
892         row = -1;
893
894         for (i = 0; i < mci->nr_csrows; i++) {
895                 struct csrow_info *csrow = csrows[i];
896                 n = 0;
897                 for (j = 0; j < csrow->nr_channels; j++) {
898                         struct dimm_info *dimm = csrow->channels[j]->dimm;
899                         n += dimm->nr_pages;
900                 }
901                 if (n == 0)
902                         continue;
903
904                 edac_dbg(3, "MC%d: first(0x%lx) page(0x%lx) last(0x%lx) mask(0x%lx)\n",
905                          mci->mc_idx,
906                          csrow->first_page, page, csrow->last_page,
907                          csrow->page_mask);
908
909                 if ((page >= csrow->first_page) &&
910                     (page <= csrow->last_page) &&
911                     ((page & csrow->page_mask) ==
912                      (csrow->first_page & csrow->page_mask))) {
913                         row = i;
914                         break;
915                 }
916         }
917
918         if (row == -1)
919                 edac_mc_printk(mci, KERN_ERR,
920                         "could not look up page error address %lx\n",
921                         (unsigned long)page);
922
923         return row;
924 }
925 EXPORT_SYMBOL_GPL(edac_mc_find_csrow_by_page);
926
927 const char *edac_layer_name[] = {
928         [EDAC_MC_LAYER_BRANCH] = "branch",
929         [EDAC_MC_LAYER_CHANNEL] = "channel",
930         [EDAC_MC_LAYER_SLOT] = "slot",
931         [EDAC_MC_LAYER_CHIP_SELECT] = "csrow",
932         [EDAC_MC_LAYER_ALL_MEM] = "memory",
933 };
934 EXPORT_SYMBOL_GPL(edac_layer_name);
935
936 static void edac_inc_ce_error(struct mem_ctl_info *mci,
937                               bool enable_per_layer_report,
938                               const int pos[EDAC_MAX_LAYERS],
939                               const u16 count)
940 {
941         int i, index = 0;
942
943         mci->ce_mc += count;
944
945         if (!enable_per_layer_report) {
946                 mci->ce_noinfo_count += count;
947                 return;
948         }
949
950         for (i = 0; i < mci->n_layers; i++) {
951                 if (pos[i] < 0)
952                         break;
953                 index += pos[i];
954                 mci->ce_per_layer[i][index] += count;
955
956                 if (i < mci->n_layers - 1)
957                         index *= mci->layers[i + 1].size;
958         }
959 }
960
961 static void edac_inc_ue_error(struct mem_ctl_info *mci,
962                                     bool enable_per_layer_report,
963                                     const int pos[EDAC_MAX_LAYERS],
964                                     const u16 count)
965 {
966         int i, index = 0;
967
968         mci->ue_mc += count;
969
970         if (!enable_per_layer_report) {
971                 mci->ce_noinfo_count += count;
972                 return;
973         }
974
975         for (i = 0; i < mci->n_layers; i++) {
976                 if (pos[i] < 0)
977                         break;
978                 index += pos[i];
979                 mci->ue_per_layer[i][index] += count;
980
981                 if (i < mci->n_layers - 1)
982                         index *= mci->layers[i + 1].size;
983         }
984 }
985
986 static void edac_ce_error(struct mem_ctl_info *mci,
987                           const u16 error_count,
988                           const int pos[EDAC_MAX_LAYERS],
989                           const char *msg,
990                           const char *location,
991                           const char *label,
992                           const char *detail,
993                           const char *other_detail,
994                           const bool enable_per_layer_report,
995                           const unsigned long page_frame_number,
996                           const unsigned long offset_in_page,
997                           long grain)
998 {
999         unsigned long remapped_page;
1000         char *msg_aux = "";
1001
1002         if (*msg)
1003                 msg_aux = " ";
1004
1005         if (edac_mc_get_log_ce()) {
1006                 if (other_detail && *other_detail)
1007                         edac_mc_printk(mci, KERN_WARNING,
1008                                        "%d CE %s%son %s (%s %s - %s)\n",
1009                                        error_count, msg, msg_aux, label,
1010                                        location, detail, other_detail);
1011                 else
1012                         edac_mc_printk(mci, KERN_WARNING,
1013                                        "%d CE %s%son %s (%s %s)\n",
1014                                        error_count, msg, msg_aux, label,
1015                                        location, detail);
1016         }
1017         edac_inc_ce_error(mci, enable_per_layer_report, pos, error_count);
1018
1019         if (mci->scrub_mode & SCRUB_SW_SRC) {
1020                 /*
1021                         * Some memory controllers (called MCs below) can remap
1022                         * memory so that it is still available at a different
1023                         * address when PCI devices map into memory.
1024                         * MC's that can't do this, lose the memory where PCI
1025                         * devices are mapped. This mapping is MC-dependent
1026                         * and so we call back into the MC driver for it to
1027                         * map the MC page to a physical (CPU) page which can
1028                         * then be mapped to a virtual page - which can then
1029                         * be scrubbed.
1030                         */
1031                 remapped_page = mci->ctl_page_to_phys ?
1032                         mci->ctl_page_to_phys(mci, page_frame_number) :
1033                         page_frame_number;
1034
1035                 edac_mc_scrub_block(remapped_page,
1036                                         offset_in_page, grain);
1037         }
1038 }
1039
1040 static void edac_ue_error(struct mem_ctl_info *mci,
1041                           const u16 error_count,
1042                           const int pos[EDAC_MAX_LAYERS],
1043                           const char *msg,
1044                           const char *location,
1045                           const char *label,
1046                           const char *detail,
1047                           const char *other_detail,
1048                           const bool enable_per_layer_report)
1049 {
1050         char *msg_aux = "";
1051
1052         if (*msg)
1053                 msg_aux = " ";
1054
1055         if (edac_mc_get_log_ue()) {
1056                 if (other_detail && *other_detail)
1057                         edac_mc_printk(mci, KERN_WARNING,
1058                                        "%d UE %s%son %s (%s %s - %s)\n",
1059                                        error_count, msg, msg_aux, label,
1060                                        location, detail, other_detail);
1061                 else
1062                         edac_mc_printk(mci, KERN_WARNING,
1063                                        "%d UE %s%son %s (%s %s)\n",
1064                                        error_count, msg, msg_aux, label,
1065                                        location, detail);
1066         }
1067
1068         if (edac_mc_get_panic_on_ue()) {
1069                 if (other_detail && *other_detail)
1070                         panic("UE %s%son %s (%s%s - %s)\n",
1071                               msg, msg_aux, label, location, detail, other_detail);
1072                 else
1073                         panic("UE %s%son %s (%s%s)\n",
1074                               msg, msg_aux, label, location, detail);
1075         }
1076
1077         edac_inc_ue_error(mci, enable_per_layer_report, pos, error_count);
1078 }
1079
1080 /**
1081  * edac_raw_mc_handle_error - reports a memory event to userspace without doing
1082  *                            anything to discover the error location
1083  *
1084  * @type:               severity of the error (CE/UE/Fatal)
1085  * @mci:                a struct mem_ctl_info pointer
1086  * @e:                  error description
1087  *
1088  * This raw function is used internally by edac_mc_handle_error(). It should
1089  * only be called directly when the hardware error come directly from BIOS,
1090  * like in the case of APEI GHES driver.
1091  */
1092 void edac_raw_mc_handle_error(const enum hw_event_mc_err_type type,
1093                               struct mem_ctl_info *mci,
1094                               struct edac_raw_error_desc *e)
1095 {
1096         char detail[80];
1097         int pos[EDAC_MAX_LAYERS] = { e->top_layer, e->mid_layer, e->low_layer };
1098
1099         /* Memory type dependent details about the error */
1100         if (type == HW_EVENT_ERR_CORRECTED) {
1101                 snprintf(detail, sizeof(detail),
1102                         "page:0x%lx offset:0x%lx grain:%ld syndrome:0x%lx",
1103                         e->page_frame_number, e->offset_in_page,
1104                         e->grain, e->syndrome);
1105                 edac_ce_error(mci, e->error_count, pos, e->msg, e->location, e->label,
1106                               detail, e->other_detail, e->enable_per_layer_report,
1107                               e->page_frame_number, e->offset_in_page, e->grain);
1108         } else {
1109                 snprintf(detail, sizeof(detail),
1110                         "page:0x%lx offset:0x%lx grain:%ld",
1111                         e->page_frame_number, e->offset_in_page, e->grain);
1112
1113                 edac_ue_error(mci, e->error_count, pos, e->msg, e->location, e->label,
1114                               detail, e->other_detail, e->enable_per_layer_report);
1115         }
1116
1117
1118 }
1119 EXPORT_SYMBOL_GPL(edac_raw_mc_handle_error);
1120
1121 /**
1122  * edac_mc_handle_error - reports a memory event to userspace
1123  *
1124  * @type:               severity of the error (CE/UE/Fatal)
1125  * @mci:                a struct mem_ctl_info pointer
1126  * @error_count:        Number of errors of the same type
1127  * @page_frame_number:  mem page where the error occurred
1128  * @offset_in_page:     offset of the error inside the page
1129  * @syndrome:           ECC syndrome
1130  * @top_layer:          Memory layer[0] position
1131  * @mid_layer:          Memory layer[1] position
1132  * @low_layer:          Memory layer[2] position
1133  * @msg:                Message meaningful to the end users that
1134  *                      explains the event
1135  * @other_detail:       Technical details about the event that
1136  *                      may help hardware manufacturers and
1137  *                      EDAC developers to analyse the event
1138  */
1139 void edac_mc_handle_error(const enum hw_event_mc_err_type type,
1140                           struct mem_ctl_info *mci,
1141                           const u16 error_count,
1142                           const unsigned long page_frame_number,
1143                           const unsigned long offset_in_page,
1144                           const unsigned long syndrome,
1145                           const int top_layer,
1146                           const int mid_layer,
1147                           const int low_layer,
1148                           const char *msg,
1149                           const char *other_detail)
1150 {
1151         char *p;
1152         int row = -1, chan = -1;
1153         int pos[EDAC_MAX_LAYERS] = { top_layer, mid_layer, low_layer };
1154         int i, n_labels = 0;
1155         u8 grain_bits;
1156         struct edac_raw_error_desc *e = &mci->error_desc;
1157
1158         edac_dbg(3, "MC%d\n", mci->mc_idx);
1159
1160         /* Fills the error report buffer */
1161         memset(e, 0, sizeof (*e));
1162         e->error_count = error_count;
1163         e->top_layer = top_layer;
1164         e->mid_layer = mid_layer;
1165         e->low_layer = low_layer;
1166         e->page_frame_number = page_frame_number;
1167         e->offset_in_page = offset_in_page;
1168         e->syndrome = syndrome;
1169         e->msg = msg;
1170         e->other_detail = other_detail;
1171
1172         /*
1173          * Check if the event report is consistent and if the memory
1174          * location is known. If it is known, enable_per_layer_report will be
1175          * true, the DIMM(s) label info will be filled and the per-layer
1176          * error counters will be incremented.
1177          */
1178         for (i = 0; i < mci->n_layers; i++) {
1179                 if (pos[i] >= (int)mci->layers[i].size) {
1180
1181                         edac_mc_printk(mci, KERN_ERR,
1182                                        "INTERNAL ERROR: %s value is out of range (%d >= %d)\n",
1183                                        edac_layer_name[mci->layers[i].type],
1184                                        pos[i], mci->layers[i].size);
1185                         /*
1186                          * Instead of just returning it, let's use what's
1187                          * known about the error. The increment routines and
1188                          * the DIMM filter logic will do the right thing by
1189                          * pointing the likely damaged DIMMs.
1190                          */
1191                         pos[i] = -1;
1192                 }
1193                 if (pos[i] >= 0)
1194                         e->enable_per_layer_report = true;
1195         }
1196
1197         /*
1198          * Get the dimm label/grain that applies to the match criteria.
1199          * As the error algorithm may not be able to point to just one memory
1200          * stick, the logic here will get all possible labels that could
1201          * pottentially be affected by the error.
1202          * On FB-DIMM memory controllers, for uncorrected errors, it is common
1203          * to have only the MC channel and the MC dimm (also called "branch")
1204          * but the channel is not known, as the memory is arranged in pairs,
1205          * where each memory belongs to a separate channel within the same
1206          * branch.
1207          */
1208         p = e->label;
1209         *p = '\0';
1210
1211         for (i = 0; i < mci->tot_dimms; i++) {
1212                 struct dimm_info *dimm = mci->dimms[i];
1213
1214                 if (top_layer >= 0 && top_layer != dimm->location[0])
1215                         continue;
1216                 if (mid_layer >= 0 && mid_layer != dimm->location[1])
1217                         continue;
1218                 if (low_layer >= 0 && low_layer != dimm->location[2])
1219                         continue;
1220
1221                 /* get the max grain, over the error match range */
1222                 if (dimm->grain > e->grain)
1223                         e->grain = dimm->grain;
1224
1225                 /*
1226                  * If the error is memory-controller wide, there's no need to
1227                  * seek for the affected DIMMs because the whole
1228                  * channel/memory controller/...  may be affected.
1229                  * Also, don't show errors for empty DIMM slots.
1230                  */
1231                 if (e->enable_per_layer_report && dimm->nr_pages) {
1232                         if (n_labels >= EDAC_MAX_LABELS) {
1233                                 e->enable_per_layer_report = false;
1234                                 break;
1235                         }
1236                         n_labels++;
1237                         if (p != e->label) {
1238                                 strcpy(p, OTHER_LABEL);
1239                                 p += strlen(OTHER_LABEL);
1240                         }
1241                         strcpy(p, dimm->label);
1242                         p += strlen(p);
1243                         *p = '\0';
1244
1245                         /*
1246                          * get csrow/channel of the DIMM, in order to allow
1247                          * incrementing the compat API counters
1248                          */
1249                         edac_dbg(4, "%s csrows map: (%d,%d)\n",
1250                                  mci->csbased ? "rank" : "dimm",
1251                                  dimm->csrow, dimm->cschannel);
1252                         if (row == -1)
1253                                 row = dimm->csrow;
1254                         else if (row >= 0 && row != dimm->csrow)
1255                                 row = -2;
1256
1257                         if (chan == -1)
1258                                 chan = dimm->cschannel;
1259                         else if (chan >= 0 && chan != dimm->cschannel)
1260                                 chan = -2;
1261                 }
1262         }
1263
1264         if (!e->enable_per_layer_report) {
1265                 strcpy(e->label, "any memory");
1266         } else {
1267                 edac_dbg(4, "csrow/channel to increment: (%d,%d)\n", row, chan);
1268                 if (p == e->label)
1269                         strcpy(e->label, "unknown memory");
1270                 if (type == HW_EVENT_ERR_CORRECTED) {
1271                         if (row >= 0) {
1272                                 mci->csrows[row]->ce_count += error_count;
1273                                 if (chan >= 0)
1274                                         mci->csrows[row]->channels[chan]->ce_count += error_count;
1275                         }
1276                 } else
1277                         if (row >= 0)
1278                                 mci->csrows[row]->ue_count += error_count;
1279         }
1280
1281         /* Fill the RAM location data */
1282         p = e->location;
1283
1284         for (i = 0; i < mci->n_layers; i++) {
1285                 if (pos[i] < 0)
1286                         continue;
1287
1288                 p += sprintf(p, "%s:%d ",
1289                              edac_layer_name[mci->layers[i].type],
1290                              pos[i]);
1291         }
1292         if (p > e->location)
1293                 *(p - 1) = '\0';
1294
1295         /* Report the error via the trace interface */
1296         grain_bits = fls_long(e->grain) + 1;
1297         trace_mc_event(type, e->msg, e->label, e->error_count,
1298                        mci->mc_idx, e->top_layer, e->mid_layer, e->low_layer,
1299                        PAGES_TO_MiB(e->page_frame_number) | e->offset_in_page,
1300                        grain_bits, e->syndrome, e->other_detail);
1301
1302         edac_raw_mc_handle_error(type, mci, e);
1303 }
1304 EXPORT_SYMBOL_GPL(edac_mc_handle_error);