]> Pileus Git - ~andy/linux/blob - drivers/edac/edac_mc.c
Merge branch 'nfsd-next' of git://linux-nfs.org/~bfields/linux
[~andy/linux] / 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,
795                 "Giving out device to module %s controller %s: DEV %s (%s)\n",
796                 mci->mod_name, mci->ctl_name, mci->dev_name,
797                 edac_op_state_to_string(mci->op_state));
798
799         edac_mc_owner = mci->mod_name;
800
801         mutex_unlock(&mem_ctls_mutex);
802         return 0;
803
804 fail1:
805         del_mc_from_global_list(mci);
806
807 fail0:
808         mutex_unlock(&mem_ctls_mutex);
809         return ret;
810 }
811 EXPORT_SYMBOL_GPL(edac_mc_add_mc);
812
813 /**
814  * edac_mc_del_mc: Remove sysfs entries for specified mci structure and
815  *                 remove mci structure from global list
816  * @pdev: Pointer to 'struct device' representing mci structure to remove.
817  *
818  * Return pointer to removed mci structure, or NULL if device not found.
819  */
820 struct mem_ctl_info *edac_mc_del_mc(struct device *dev)
821 {
822         struct mem_ctl_info *mci;
823
824         edac_dbg(0, "\n");
825
826         mutex_lock(&mem_ctls_mutex);
827
828         /* find the requested mci struct in the global list */
829         mci = find_mci_by_dev(dev);
830         if (mci == NULL) {
831                 mutex_unlock(&mem_ctls_mutex);
832                 return NULL;
833         }
834
835         if (!del_mc_from_global_list(mci))
836                 edac_mc_owner = NULL;
837         mutex_unlock(&mem_ctls_mutex);
838
839         /* flush workq processes */
840         edac_mc_workq_teardown(mci);
841
842         /* marking MCI offline */
843         mci->op_state = OP_OFFLINE;
844
845         /* remove from sysfs */
846         edac_remove_sysfs_mci_device(mci);
847
848         edac_printk(KERN_INFO, EDAC_MC,
849                 "Removed device %d for %s %s: DEV %s\n", mci->mc_idx,
850                 mci->mod_name, mci->ctl_name, edac_dev_name(mci));
851
852         return mci;
853 }
854 EXPORT_SYMBOL_GPL(edac_mc_del_mc);
855
856 static void edac_mc_scrub_block(unsigned long page, unsigned long offset,
857                                 u32 size)
858 {
859         struct page *pg;
860         void *virt_addr;
861         unsigned long flags = 0;
862
863         edac_dbg(3, "\n");
864
865         /* ECC error page was not in our memory. Ignore it. */
866         if (!pfn_valid(page))
867                 return;
868
869         /* Find the actual page structure then map it and fix */
870         pg = pfn_to_page(page);
871
872         if (PageHighMem(pg))
873                 local_irq_save(flags);
874
875         virt_addr = kmap_atomic(pg);
876
877         /* Perform architecture specific atomic scrub operation */
878         atomic_scrub(virt_addr + offset, size);
879
880         /* Unmap and complete */
881         kunmap_atomic(virt_addr);
882
883         if (PageHighMem(pg))
884                 local_irq_restore(flags);
885 }
886
887 /* FIXME - should return -1 */
888 int edac_mc_find_csrow_by_page(struct mem_ctl_info *mci, unsigned long page)
889 {
890         struct csrow_info **csrows = mci->csrows;
891         int row, i, j, n;
892
893         edac_dbg(1, "MC%d: 0x%lx\n", mci->mc_idx, page);
894         row = -1;
895
896         for (i = 0; i < mci->nr_csrows; i++) {
897                 struct csrow_info *csrow = csrows[i];
898                 n = 0;
899                 for (j = 0; j < csrow->nr_channels; j++) {
900                         struct dimm_info *dimm = csrow->channels[j]->dimm;
901                         n += dimm->nr_pages;
902                 }
903                 if (n == 0)
904                         continue;
905
906                 edac_dbg(3, "MC%d: first(0x%lx) page(0x%lx) last(0x%lx) mask(0x%lx)\n",
907                          mci->mc_idx,
908                          csrow->first_page, page, csrow->last_page,
909                          csrow->page_mask);
910
911                 if ((page >= csrow->first_page) &&
912                     (page <= csrow->last_page) &&
913                     ((page & csrow->page_mask) ==
914                      (csrow->first_page & csrow->page_mask))) {
915                         row = i;
916                         break;
917                 }
918         }
919
920         if (row == -1)
921                 edac_mc_printk(mci, KERN_ERR,
922                         "could not look up page error address %lx\n",
923                         (unsigned long)page);
924
925         return row;
926 }
927 EXPORT_SYMBOL_GPL(edac_mc_find_csrow_by_page);
928
929 const char *edac_layer_name[] = {
930         [EDAC_MC_LAYER_BRANCH] = "branch",
931         [EDAC_MC_LAYER_CHANNEL] = "channel",
932         [EDAC_MC_LAYER_SLOT] = "slot",
933         [EDAC_MC_LAYER_CHIP_SELECT] = "csrow",
934         [EDAC_MC_LAYER_ALL_MEM] = "memory",
935 };
936 EXPORT_SYMBOL_GPL(edac_layer_name);
937
938 static void edac_inc_ce_error(struct mem_ctl_info *mci,
939                               bool enable_per_layer_report,
940                               const int pos[EDAC_MAX_LAYERS],
941                               const u16 count)
942 {
943         int i, index = 0;
944
945         mci->ce_mc += count;
946
947         if (!enable_per_layer_report) {
948                 mci->ce_noinfo_count += count;
949                 return;
950         }
951
952         for (i = 0; i < mci->n_layers; i++) {
953                 if (pos[i] < 0)
954                         break;
955                 index += pos[i];
956                 mci->ce_per_layer[i][index] += count;
957
958                 if (i < mci->n_layers - 1)
959                         index *= mci->layers[i + 1].size;
960         }
961 }
962
963 static void edac_inc_ue_error(struct mem_ctl_info *mci,
964                                     bool enable_per_layer_report,
965                                     const int pos[EDAC_MAX_LAYERS],
966                                     const u16 count)
967 {
968         int i, index = 0;
969
970         mci->ue_mc += count;
971
972         if (!enable_per_layer_report) {
973                 mci->ce_noinfo_count += count;
974                 return;
975         }
976
977         for (i = 0; i < mci->n_layers; i++) {
978                 if (pos[i] < 0)
979                         break;
980                 index += pos[i];
981                 mci->ue_per_layer[i][index] += count;
982
983                 if (i < mci->n_layers - 1)
984                         index *= mci->layers[i + 1].size;
985         }
986 }
987
988 static void edac_ce_error(struct mem_ctl_info *mci,
989                           const u16 error_count,
990                           const int pos[EDAC_MAX_LAYERS],
991                           const char *msg,
992                           const char *location,
993                           const char *label,
994                           const char *detail,
995                           const char *other_detail,
996                           const bool enable_per_layer_report,
997                           const unsigned long page_frame_number,
998                           const unsigned long offset_in_page,
999                           long grain)
1000 {
1001         unsigned long remapped_page;
1002         char *msg_aux = "";
1003
1004         if (*msg)
1005                 msg_aux = " ";
1006
1007         if (edac_mc_get_log_ce()) {
1008                 if (other_detail && *other_detail)
1009                         edac_mc_printk(mci, KERN_WARNING,
1010                                        "%d CE %s%son %s (%s %s - %s)\n",
1011                                        error_count, msg, msg_aux, label,
1012                                        location, detail, other_detail);
1013                 else
1014                         edac_mc_printk(mci, KERN_WARNING,
1015                                        "%d CE %s%son %s (%s %s)\n",
1016                                        error_count, msg, msg_aux, label,
1017                                        location, detail);
1018         }
1019         edac_inc_ce_error(mci, enable_per_layer_report, pos, error_count);
1020
1021         if (mci->scrub_mode & SCRUB_SW_SRC) {
1022                 /*
1023                         * Some memory controllers (called MCs below) can remap
1024                         * memory so that it is still available at a different
1025                         * address when PCI devices map into memory.
1026                         * MC's that can't do this, lose the memory where PCI
1027                         * devices are mapped. This mapping is MC-dependent
1028                         * and so we call back into the MC driver for it to
1029                         * map the MC page to a physical (CPU) page which can
1030                         * then be mapped to a virtual page - which can then
1031                         * be scrubbed.
1032                         */
1033                 remapped_page = mci->ctl_page_to_phys ?
1034                         mci->ctl_page_to_phys(mci, page_frame_number) :
1035                         page_frame_number;
1036
1037                 edac_mc_scrub_block(remapped_page,
1038                                         offset_in_page, grain);
1039         }
1040 }
1041
1042 static void edac_ue_error(struct mem_ctl_info *mci,
1043                           const u16 error_count,
1044                           const int pos[EDAC_MAX_LAYERS],
1045                           const char *msg,
1046                           const char *location,
1047                           const char *label,
1048                           const char *detail,
1049                           const char *other_detail,
1050                           const bool enable_per_layer_report)
1051 {
1052         char *msg_aux = "";
1053
1054         if (*msg)
1055                 msg_aux = " ";
1056
1057         if (edac_mc_get_log_ue()) {
1058                 if (other_detail && *other_detail)
1059                         edac_mc_printk(mci, KERN_WARNING,
1060                                        "%d UE %s%son %s (%s %s - %s)\n",
1061                                        error_count, msg, msg_aux, label,
1062                                        location, detail, other_detail);
1063                 else
1064                         edac_mc_printk(mci, KERN_WARNING,
1065                                        "%d UE %s%son %s (%s %s)\n",
1066                                        error_count, msg, msg_aux, label,
1067                                        location, detail);
1068         }
1069
1070         if (edac_mc_get_panic_on_ue()) {
1071                 if (other_detail && *other_detail)
1072                         panic("UE %s%son %s (%s%s - %s)\n",
1073                               msg, msg_aux, label, location, detail, other_detail);
1074                 else
1075                         panic("UE %s%son %s (%s%s)\n",
1076                               msg, msg_aux, label, location, detail);
1077         }
1078
1079         edac_inc_ue_error(mci, enable_per_layer_report, pos, error_count);
1080 }
1081
1082 /**
1083  * edac_raw_mc_handle_error - reports a memory event to userspace without doing
1084  *                            anything to discover the error location
1085  *
1086  * @type:               severity of the error (CE/UE/Fatal)
1087  * @mci:                a struct mem_ctl_info pointer
1088  * @e:                  error description
1089  *
1090  * This raw function is used internally by edac_mc_handle_error(). It should
1091  * only be called directly when the hardware error come directly from BIOS,
1092  * like in the case of APEI GHES driver.
1093  */
1094 void edac_raw_mc_handle_error(const enum hw_event_mc_err_type type,
1095                               struct mem_ctl_info *mci,
1096                               struct edac_raw_error_desc *e)
1097 {
1098         char detail[80];
1099         int pos[EDAC_MAX_LAYERS] = { e->top_layer, e->mid_layer, e->low_layer };
1100
1101         /* Memory type dependent details about the error */
1102         if (type == HW_EVENT_ERR_CORRECTED) {
1103                 snprintf(detail, sizeof(detail),
1104                         "page:0x%lx offset:0x%lx grain:%ld syndrome:0x%lx",
1105                         e->page_frame_number, e->offset_in_page,
1106                         e->grain, e->syndrome);
1107                 edac_ce_error(mci, e->error_count, pos, e->msg, e->location, e->label,
1108                               detail, e->other_detail, e->enable_per_layer_report,
1109                               e->page_frame_number, e->offset_in_page, e->grain);
1110         } else {
1111                 snprintf(detail, sizeof(detail),
1112                         "page:0x%lx offset:0x%lx grain:%ld",
1113                         e->page_frame_number, e->offset_in_page, e->grain);
1114
1115                 edac_ue_error(mci, e->error_count, pos, e->msg, e->location, e->label,
1116                               detail, e->other_detail, e->enable_per_layer_report);
1117         }
1118
1119
1120 }
1121 EXPORT_SYMBOL_GPL(edac_raw_mc_handle_error);
1122
1123 /**
1124  * edac_mc_handle_error - reports a memory event to userspace
1125  *
1126  * @type:               severity of the error (CE/UE/Fatal)
1127  * @mci:                a struct mem_ctl_info pointer
1128  * @error_count:        Number of errors of the same type
1129  * @page_frame_number:  mem page where the error occurred
1130  * @offset_in_page:     offset of the error inside the page
1131  * @syndrome:           ECC syndrome
1132  * @top_layer:          Memory layer[0] position
1133  * @mid_layer:          Memory layer[1] position
1134  * @low_layer:          Memory layer[2] position
1135  * @msg:                Message meaningful to the end users that
1136  *                      explains the event
1137  * @other_detail:       Technical details about the event that
1138  *                      may help hardware manufacturers and
1139  *                      EDAC developers to analyse the event
1140  */
1141 void edac_mc_handle_error(const enum hw_event_mc_err_type type,
1142                           struct mem_ctl_info *mci,
1143                           const u16 error_count,
1144                           const unsigned long page_frame_number,
1145                           const unsigned long offset_in_page,
1146                           const unsigned long syndrome,
1147                           const int top_layer,
1148                           const int mid_layer,
1149                           const int low_layer,
1150                           const char *msg,
1151                           const char *other_detail)
1152 {
1153         char *p;
1154         int row = -1, chan = -1;
1155         int pos[EDAC_MAX_LAYERS] = { top_layer, mid_layer, low_layer };
1156         int i, n_labels = 0;
1157         u8 grain_bits;
1158         struct edac_raw_error_desc *e = &mci->error_desc;
1159
1160         edac_dbg(3, "MC%d\n", mci->mc_idx);
1161
1162         /* Fills the error report buffer */
1163         memset(e, 0, sizeof (*e));
1164         e->error_count = error_count;
1165         e->top_layer = top_layer;
1166         e->mid_layer = mid_layer;
1167         e->low_layer = low_layer;
1168         e->page_frame_number = page_frame_number;
1169         e->offset_in_page = offset_in_page;
1170         e->syndrome = syndrome;
1171         e->msg = msg;
1172         e->other_detail = other_detail;
1173
1174         /*
1175          * Check if the event report is consistent and if the memory
1176          * location is known. If it is known, enable_per_layer_report will be
1177          * true, the DIMM(s) label info will be filled and the per-layer
1178          * error counters will be incremented.
1179          */
1180         for (i = 0; i < mci->n_layers; i++) {
1181                 if (pos[i] >= (int)mci->layers[i].size) {
1182
1183                         edac_mc_printk(mci, KERN_ERR,
1184                                        "INTERNAL ERROR: %s value is out of range (%d >= %d)\n",
1185                                        edac_layer_name[mci->layers[i].type],
1186                                        pos[i], mci->layers[i].size);
1187                         /*
1188                          * Instead of just returning it, let's use what's
1189                          * known about the error. The increment routines and
1190                          * the DIMM filter logic will do the right thing by
1191                          * pointing the likely damaged DIMMs.
1192                          */
1193                         pos[i] = -1;
1194                 }
1195                 if (pos[i] >= 0)
1196                         e->enable_per_layer_report = true;
1197         }
1198
1199         /*
1200          * Get the dimm label/grain that applies to the match criteria.
1201          * As the error algorithm may not be able to point to just one memory
1202          * stick, the logic here will get all possible labels that could
1203          * pottentially be affected by the error.
1204          * On FB-DIMM memory controllers, for uncorrected errors, it is common
1205          * to have only the MC channel and the MC dimm (also called "branch")
1206          * but the channel is not known, as the memory is arranged in pairs,
1207          * where each memory belongs to a separate channel within the same
1208          * branch.
1209          */
1210         p = e->label;
1211         *p = '\0';
1212
1213         for (i = 0; i < mci->tot_dimms; i++) {
1214                 struct dimm_info *dimm = mci->dimms[i];
1215
1216                 if (top_layer >= 0 && top_layer != dimm->location[0])
1217                         continue;
1218                 if (mid_layer >= 0 && mid_layer != dimm->location[1])
1219                         continue;
1220                 if (low_layer >= 0 && low_layer != dimm->location[2])
1221                         continue;
1222
1223                 /* get the max grain, over the error match range */
1224                 if (dimm->grain > e->grain)
1225                         e->grain = dimm->grain;
1226
1227                 /*
1228                  * If the error is memory-controller wide, there's no need to
1229                  * seek for the affected DIMMs because the whole
1230                  * channel/memory controller/...  may be affected.
1231                  * Also, don't show errors for empty DIMM slots.
1232                  */
1233                 if (e->enable_per_layer_report && dimm->nr_pages) {
1234                         if (n_labels >= EDAC_MAX_LABELS) {
1235                                 e->enable_per_layer_report = false;
1236                                 break;
1237                         }
1238                         n_labels++;
1239                         if (p != e->label) {
1240                                 strcpy(p, OTHER_LABEL);
1241                                 p += strlen(OTHER_LABEL);
1242                         }
1243                         strcpy(p, dimm->label);
1244                         p += strlen(p);
1245                         *p = '\0';
1246
1247                         /*
1248                          * get csrow/channel of the DIMM, in order to allow
1249                          * incrementing the compat API counters
1250                          */
1251                         edac_dbg(4, "%s csrows map: (%d,%d)\n",
1252                                  mci->csbased ? "rank" : "dimm",
1253                                  dimm->csrow, dimm->cschannel);
1254                         if (row == -1)
1255                                 row = dimm->csrow;
1256                         else if (row >= 0 && row != dimm->csrow)
1257                                 row = -2;
1258
1259                         if (chan == -1)
1260                                 chan = dimm->cschannel;
1261                         else if (chan >= 0 && chan != dimm->cschannel)
1262                                 chan = -2;
1263                 }
1264         }
1265
1266         if (!e->enable_per_layer_report) {
1267                 strcpy(e->label, "any memory");
1268         } else {
1269                 edac_dbg(4, "csrow/channel to increment: (%d,%d)\n", row, chan);
1270                 if (p == e->label)
1271                         strcpy(e->label, "unknown memory");
1272                 if (type == HW_EVENT_ERR_CORRECTED) {
1273                         if (row >= 0) {
1274                                 mci->csrows[row]->ce_count += error_count;
1275                                 if (chan >= 0)
1276                                         mci->csrows[row]->channels[chan]->ce_count += error_count;
1277                         }
1278                 } else
1279                         if (row >= 0)
1280                                 mci->csrows[row]->ue_count += error_count;
1281         }
1282
1283         /* Fill the RAM location data */
1284         p = e->location;
1285
1286         for (i = 0; i < mci->n_layers; i++) {
1287                 if (pos[i] < 0)
1288                         continue;
1289
1290                 p += sprintf(p, "%s:%d ",
1291                              edac_layer_name[mci->layers[i].type],
1292                              pos[i]);
1293         }
1294         if (p > e->location)
1295                 *(p - 1) = '\0';
1296
1297         /* Report the error via the trace interface */
1298         grain_bits = fls_long(e->grain) + 1;
1299         trace_mc_event(type, e->msg, e->label, e->error_count,
1300                        mci->mc_idx, e->top_layer, e->mid_layer, e->low_layer,
1301                        PAGES_TO_MiB(e->page_frame_number) | e->offset_in_page,
1302                        grain_bits, e->syndrome, e->other_detail);
1303
1304         edac_raw_mc_handle_error(type, mci, e);
1305 }
1306 EXPORT_SYMBOL_GPL(edac_mc_handle_error);