Merge git://git.kernel.org/pub/scm/linux/kernel/git/mchehab/linux-edac
[sfrench/cifs-2.6.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 unsigned edac_dimm_info_location(struct dimm_info *dimm, char *buf,
46                                  unsigned len)
47 {
48         struct mem_ctl_info *mci = dimm->mci;
49         int i, n, count = 0;
50         char *p = buf;
51
52         for (i = 0; i < mci->n_layers; i++) {
53                 n = snprintf(p, len, "%s %d ",
54                               edac_layer_name[mci->layers[i].type],
55                               dimm->location[i]);
56                 p += n;
57                 len -= n;
58                 count += n;
59                 if (!len)
60                         break;
61         }
62
63         return count;
64 }
65
66 #ifdef CONFIG_EDAC_DEBUG
67
68 static void edac_mc_dump_channel(struct rank_info *chan)
69 {
70         edac_dbg(4, "  channel->chan_idx = %d\n", chan->chan_idx);
71         edac_dbg(4, "    channel = %p\n", chan);
72         edac_dbg(4, "    channel->csrow = %p\n", chan->csrow);
73         edac_dbg(4, "    channel->dimm = %p\n", chan->dimm);
74 }
75
76 static void edac_mc_dump_dimm(struct dimm_info *dimm, int number)
77 {
78         char location[80];
79
80         edac_dimm_info_location(dimm, location, sizeof(location));
81
82         edac_dbg(4, "%s%i: %smapped as virtual row %d, chan %d\n",
83                  dimm->mci->mem_is_per_rank ? "rank" : "dimm",
84                  number, location, dimm->csrow, dimm->cschannel);
85         edac_dbg(4, "  dimm = %p\n", dimm);
86         edac_dbg(4, "  dimm->label = '%s'\n", dimm->label);
87         edac_dbg(4, "  dimm->nr_pages = 0x%x\n", dimm->nr_pages);
88         edac_dbg(4, "  dimm->grain = %d\n", dimm->grain);
89         edac_dbg(4, "  dimm->nr_pages = 0x%x\n", dimm->nr_pages);
90 }
91
92 static void edac_mc_dump_csrow(struct csrow_info *csrow)
93 {
94         edac_dbg(4, "csrow->csrow_idx = %d\n", csrow->csrow_idx);
95         edac_dbg(4, "  csrow = %p\n", csrow);
96         edac_dbg(4, "  csrow->first_page = 0x%lx\n", csrow->first_page);
97         edac_dbg(4, "  csrow->last_page = 0x%lx\n", csrow->last_page);
98         edac_dbg(4, "  csrow->page_mask = 0x%lx\n", csrow->page_mask);
99         edac_dbg(4, "  csrow->nr_channels = %d\n", csrow->nr_channels);
100         edac_dbg(4, "  csrow->channels = %p\n", csrow->channels);
101         edac_dbg(4, "  csrow->mci = %p\n", csrow->mci);
102 }
103
104 static void edac_mc_dump_mci(struct mem_ctl_info *mci)
105 {
106         edac_dbg(3, "\tmci = %p\n", mci);
107         edac_dbg(3, "\tmci->mtype_cap = %lx\n", mci->mtype_cap);
108         edac_dbg(3, "\tmci->edac_ctl_cap = %lx\n", mci->edac_ctl_cap);
109         edac_dbg(3, "\tmci->edac_cap = %lx\n", mci->edac_cap);
110         edac_dbg(4, "\tmci->edac_check = %p\n", mci->edac_check);
111         edac_dbg(3, "\tmci->nr_csrows = %d, csrows = %p\n",
112                  mci->nr_csrows, mci->csrows);
113         edac_dbg(3, "\tmci->nr_dimms = %d, dimms = %p\n",
114                  mci->tot_dimms, mci->dimms);
115         edac_dbg(3, "\tdev = %p\n", mci->pdev);
116         edac_dbg(3, "\tmod_name:ctl_name = %s:%s\n",
117                  mci->mod_name, mci->ctl_name);
118         edac_dbg(3, "\tpvt_info = %p\n\n", mci->pvt_info);
119 }
120
121 #endif                          /* CONFIG_EDAC_DEBUG */
122
123 /*
124  * keep those in sync with the enum mem_type
125  */
126 const char *edac_mem_types[] = {
127         "Empty csrow",
128         "Reserved csrow type",
129         "Unknown csrow type",
130         "Fast page mode RAM",
131         "Extended data out RAM",
132         "Burst Extended data out RAM",
133         "Single data rate SDRAM",
134         "Registered single data rate SDRAM",
135         "Double data rate SDRAM",
136         "Registered Double data rate SDRAM",
137         "Rambus DRAM",
138         "Unbuffered DDR2 RAM",
139         "Fully buffered DDR2",
140         "Registered DDR2 RAM",
141         "Rambus XDR",
142         "Unbuffered DDR3 RAM",
143         "Registered DDR3 RAM",
144 };
145 EXPORT_SYMBOL_GPL(edac_mem_types);
146
147 /**
148  * edac_align_ptr - Prepares the pointer offsets for a single-shot allocation
149  * @p:          pointer to a pointer with the memory offset to be used. At
150  *              return, this will be incremented to point to the next offset
151  * @size:       Size of the data structure to be reserved
152  * @n_elems:    Number of elements that should be reserved
153  *
154  * If 'size' is a constant, the compiler will optimize this whole function
155  * down to either a no-op or the addition of a constant to the value of '*p'.
156  *
157  * The 'p' pointer is absolutely needed to keep the proper advancing
158  * further in memory to the proper offsets when allocating the struct along
159  * with its embedded structs, as edac_device_alloc_ctl_info() does it
160  * above, for example.
161  *
162  * At return, the pointer 'p' will be incremented to be used on a next call
163  * to this function.
164  */
165 void *edac_align_ptr(void **p, unsigned size, int n_elems)
166 {
167         unsigned align, r;
168         void *ptr = *p;
169
170         *p += size * n_elems;
171
172         /*
173          * 'p' can possibly be an unaligned item X such that sizeof(X) is
174          * 'size'.  Adjust 'p' so that its alignment is at least as
175          * stringent as what the compiler would provide for X and return
176          * the aligned result.
177          * Here we assume that the alignment of a "long long" is the most
178          * stringent alignment that the compiler will ever provide by default.
179          * As far as I know, this is a reasonable assumption.
180          */
181         if (size > sizeof(long))
182                 align = sizeof(long long);
183         else if (size > sizeof(int))
184                 align = sizeof(long);
185         else if (size > sizeof(short))
186                 align = sizeof(int);
187         else if (size > sizeof(char))
188                 align = sizeof(short);
189         else
190                 return (char *)ptr;
191
192         r = (unsigned long)p % align;
193
194         if (r == 0)
195                 return (char *)ptr;
196
197         *p += align - r;
198
199         return (void *)(((unsigned long)ptr) + align - r);
200 }
201
202 static void _edac_mc_free(struct mem_ctl_info *mci)
203 {
204         int i, chn, row;
205         struct csrow_info *csr;
206         const unsigned int tot_dimms = mci->tot_dimms;
207         const unsigned int tot_channels = mci->num_cschannel;
208         const unsigned int tot_csrows = mci->nr_csrows;
209
210         if (mci->dimms) {
211                 for (i = 0; i < tot_dimms; i++)
212                         kfree(mci->dimms[i]);
213                 kfree(mci->dimms);
214         }
215         if (mci->csrows) {
216                 for (row = 0; row < tot_csrows; row++) {
217                         csr = mci->csrows[row];
218                         if (csr) {
219                                 if (csr->channels) {
220                                         for (chn = 0; chn < tot_channels; chn++)
221                                                 kfree(csr->channels[chn]);
222                                         kfree(csr->channels);
223                                 }
224                                 kfree(csr);
225                         }
226                 }
227                 kfree(mci->csrows);
228         }
229         kfree(mci);
230 }
231
232 /**
233  * edac_mc_alloc: Allocate and partially fill a struct mem_ctl_info structure
234  * @mc_num:             Memory controller number
235  * @n_layers:           Number of MC hierarchy layers
236  * layers:              Describes each layer as seen by the Memory Controller
237  * @size_pvt:           size of private storage needed
238  *
239  *
240  * Everything is kmalloc'ed as one big chunk - more efficient.
241  * Only can be used if all structures have the same lifetime - otherwise
242  * you have to allocate and initialize your own structures.
243  *
244  * Use edac_mc_free() to free mc structures allocated by this function.
245  *
246  * NOTE: drivers handle multi-rank memories in different ways: in some
247  * drivers, one multi-rank memory stick is mapped as one entry, while, in
248  * others, a single multi-rank memory stick would be mapped into several
249  * entries. Currently, this function will allocate multiple struct dimm_info
250  * on such scenarios, as grouping the multiple ranks require drivers change.
251  *
252  * Returns:
253  *      On failure: NULL
254  *      On success: struct mem_ctl_info pointer
255  */
256 struct mem_ctl_info *edac_mc_alloc(unsigned mc_num,
257                                    unsigned n_layers,
258                                    struct edac_mc_layer *layers,
259                                    unsigned sz_pvt)
260 {
261         struct mem_ctl_info *mci;
262         struct edac_mc_layer *layer;
263         struct csrow_info *csr;
264         struct rank_info *chan;
265         struct dimm_info *dimm;
266         u32 *ce_per_layer[EDAC_MAX_LAYERS], *ue_per_layer[EDAC_MAX_LAYERS];
267         unsigned pos[EDAC_MAX_LAYERS];
268         unsigned size, tot_dimms = 1, count = 1;
269         unsigned tot_csrows = 1, tot_channels = 1, tot_errcount = 0;
270         void *pvt, *p, *ptr = NULL;
271         int i, j, row, chn, n, len, off;
272         bool per_rank = false;
273
274         BUG_ON(n_layers > EDAC_MAX_LAYERS || n_layers == 0);
275         /*
276          * Calculate the total amount of dimms and csrows/cschannels while
277          * in the old API emulation mode
278          */
279         for (i = 0; i < n_layers; i++) {
280                 tot_dimms *= layers[i].size;
281                 if (layers[i].is_virt_csrow)
282                         tot_csrows *= layers[i].size;
283                 else
284                         tot_channels *= layers[i].size;
285
286                 if (layers[i].type == EDAC_MC_LAYER_CHIP_SELECT)
287                         per_rank = true;
288         }
289
290         /* Figure out the offsets of the various items from the start of an mc
291          * structure.  We want the alignment of each item to be at least as
292          * stringent as what the compiler would provide if we could simply
293          * hardcode everything into a single struct.
294          */
295         mci = edac_align_ptr(&ptr, sizeof(*mci), 1);
296         layer = edac_align_ptr(&ptr, sizeof(*layer), n_layers);
297         for (i = 0; i < n_layers; i++) {
298                 count *= layers[i].size;
299                 edac_dbg(4, "errcount layer %d size %d\n", i, count);
300                 ce_per_layer[i] = edac_align_ptr(&ptr, sizeof(u32), count);
301                 ue_per_layer[i] = edac_align_ptr(&ptr, sizeof(u32), count);
302                 tot_errcount += 2 * count;
303         }
304
305         edac_dbg(4, "allocating %d error counters\n", tot_errcount);
306         pvt = edac_align_ptr(&ptr, sz_pvt, 1);
307         size = ((unsigned long)pvt) + sz_pvt;
308
309         edac_dbg(1, "allocating %u bytes for mci data (%d %s, %d csrows/channels)\n",
310                  size,
311                  tot_dimms,
312                  per_rank ? "ranks" : "dimms",
313                  tot_csrows * tot_channels);
314
315         mci = kzalloc(size, GFP_KERNEL);
316         if (mci == NULL)
317                 return NULL;
318
319         /* Adjust pointers so they point within the memory we just allocated
320          * rather than an imaginary chunk of memory located at address 0.
321          */
322         layer = (struct edac_mc_layer *)(((char *)mci) + ((unsigned long)layer));
323         for (i = 0; i < n_layers; i++) {
324                 mci->ce_per_layer[i] = (u32 *)((char *)mci + ((unsigned long)ce_per_layer[i]));
325                 mci->ue_per_layer[i] = (u32 *)((char *)mci + ((unsigned long)ue_per_layer[i]));
326         }
327         pvt = sz_pvt ? (((char *)mci) + ((unsigned long)pvt)) : NULL;
328
329         /* setup index and various internal pointers */
330         mci->mc_idx = mc_num;
331         mci->tot_dimms = tot_dimms;
332         mci->pvt_info = pvt;
333         mci->n_layers = n_layers;
334         mci->layers = layer;
335         memcpy(mci->layers, layers, sizeof(*layer) * n_layers);
336         mci->nr_csrows = tot_csrows;
337         mci->num_cschannel = tot_channels;
338         mci->mem_is_per_rank = per_rank;
339
340         /*
341          * Alocate and fill the csrow/channels structs
342          */
343         mci->csrows = kcalloc(sizeof(*mci->csrows), tot_csrows, GFP_KERNEL);
344         if (!mci->csrows)
345                 goto error;
346         for (row = 0; row < tot_csrows; row++) {
347                 csr = kzalloc(sizeof(**mci->csrows), GFP_KERNEL);
348                 if (!csr)
349                         goto error;
350                 mci->csrows[row] = csr;
351                 csr->csrow_idx = row;
352                 csr->mci = mci;
353                 csr->nr_channels = tot_channels;
354                 csr->channels = kcalloc(sizeof(*csr->channels), tot_channels,
355                                         GFP_KERNEL);
356                 if (!csr->channels)
357                         goto error;
358
359                 for (chn = 0; chn < tot_channels; chn++) {
360                         chan = kzalloc(sizeof(**csr->channels), GFP_KERNEL);
361                         if (!chan)
362                                 goto error;
363                         csr->channels[chn] = chan;
364                         chan->chan_idx = chn;
365                         chan->csrow = csr;
366                 }
367         }
368
369         /*
370          * Allocate and fill the dimm structs
371          */
372         mci->dimms  = kcalloc(sizeof(*mci->dimms), tot_dimms, GFP_KERNEL);
373         if (!mci->dimms)
374                 goto error;
375
376         memset(&pos, 0, sizeof(pos));
377         row = 0;
378         chn = 0;
379         for (i = 0; i < tot_dimms; i++) {
380                 chan = mci->csrows[row]->channels[chn];
381                 off = EDAC_DIMM_OFF(layer, n_layers, pos[0], pos[1], pos[2]);
382                 if (off < 0 || off >= tot_dimms) {
383                         edac_mc_printk(mci, KERN_ERR, "EDAC core bug: EDAC_DIMM_OFF is trying to do an illegal data access\n");
384                         goto error;
385                 }
386
387                 dimm = kzalloc(sizeof(**mci->dimms), GFP_KERNEL);
388                 if (!dimm)
389                         goto error;
390                 mci->dimms[off] = dimm;
391                 dimm->mci = mci;
392
393                 /*
394                  * Copy DIMM location and initialize it.
395                  */
396                 len = sizeof(dimm->label);
397                 p = dimm->label;
398                 n = snprintf(p, len, "mc#%u", mc_num);
399                 p += n;
400                 len -= n;
401                 for (j = 0; j < n_layers; j++) {
402                         n = snprintf(p, len, "%s#%u",
403                                      edac_layer_name[layers[j].type],
404                                      pos[j]);
405                         p += n;
406                         len -= n;
407                         dimm->location[j] = pos[j];
408
409                         if (len <= 0)
410                                 break;
411                 }
412
413                 /* Link it to the csrows old API data */
414                 chan->dimm = dimm;
415                 dimm->csrow = row;
416                 dimm->cschannel = chn;
417
418                 /* Increment csrow location */
419                 if (layers[0].is_virt_csrow) {
420                         chn++;
421                         if (chn == tot_channels) {
422                                 chn = 0;
423                                 row++;
424                         }
425                 } else {
426                         row++;
427                         if (row == tot_csrows) {
428                                 row = 0;
429                                 chn++;
430                         }
431                 }
432
433                 /* Increment dimm location */
434                 for (j = n_layers - 1; j >= 0; j--) {
435                         pos[j]++;
436                         if (pos[j] < layers[j].size)
437                                 break;
438                         pos[j] = 0;
439                 }
440         }
441
442         mci->op_state = OP_ALLOC;
443
444         /* at this point, the root kobj is valid, and in order to
445          * 'free' the object, then the function:
446          *      edac_mc_unregister_sysfs_main_kobj() must be called
447          * which will perform kobj unregistration and the actual free
448          * will occur during the kobject callback operation
449          */
450
451         return mci;
452
453 error:
454         _edac_mc_free(mci);
455
456         return NULL;
457 }
458 EXPORT_SYMBOL_GPL(edac_mc_alloc);
459
460 /**
461  * edac_mc_free
462  *      'Free' a previously allocated 'mci' structure
463  * @mci: pointer to a struct mem_ctl_info structure
464  */
465 void edac_mc_free(struct mem_ctl_info *mci)
466 {
467         edac_dbg(1, "\n");
468
469         /* If we're not yet registered with sysfs free only what was allocated
470          * in edac_mc_alloc().
471          */
472         if (!device_is_registered(&mci->dev)) {
473                 _edac_mc_free(mci);
474                 return;
475         }
476
477         /* the mci instance is freed here, when the sysfs object is dropped */
478         edac_unregister_sysfs(mci);
479 }
480 EXPORT_SYMBOL_GPL(edac_mc_free);
481
482
483 /**
484  * find_mci_by_dev
485  *
486  *      scan list of controllers looking for the one that manages
487  *      the 'dev' device
488  * @dev: pointer to a struct device related with the MCI
489  */
490 struct mem_ctl_info *find_mci_by_dev(struct device *dev)
491 {
492         struct mem_ctl_info *mci;
493         struct list_head *item;
494
495         edac_dbg(3, "\n");
496
497         list_for_each(item, &mc_devices) {
498                 mci = list_entry(item, struct mem_ctl_info, link);
499
500                 if (mci->pdev == dev)
501                         return mci;
502         }
503
504         return NULL;
505 }
506 EXPORT_SYMBOL_GPL(find_mci_by_dev);
507
508 /*
509  * handler for EDAC to check if NMI type handler has asserted interrupt
510  */
511 static int edac_mc_assert_error_check_and_clear(void)
512 {
513         int old_state;
514
515         if (edac_op_state == EDAC_OPSTATE_POLL)
516                 return 1;
517
518         old_state = edac_err_assert;
519         edac_err_assert = 0;
520
521         return old_state;
522 }
523
524 /*
525  * edac_mc_workq_function
526  *      performs the operation scheduled by a workq request
527  */
528 static void edac_mc_workq_function(struct work_struct *work_req)
529 {
530         struct delayed_work *d_work = to_delayed_work(work_req);
531         struct mem_ctl_info *mci = to_edac_mem_ctl_work(d_work);
532
533         mutex_lock(&mem_ctls_mutex);
534
535         /* if this control struct has movd to offline state, we are done */
536         if (mci->op_state == OP_OFFLINE) {
537                 mutex_unlock(&mem_ctls_mutex);
538                 return;
539         }
540
541         /* Only poll controllers that are running polled and have a check */
542         if (edac_mc_assert_error_check_and_clear() && (mci->edac_check != NULL))
543                 mci->edac_check(mci);
544
545         mutex_unlock(&mem_ctls_mutex);
546
547         /* Reschedule */
548         queue_delayed_work(edac_workqueue, &mci->work,
549                         msecs_to_jiffies(edac_mc_get_poll_msec()));
550 }
551
552 /*
553  * edac_mc_workq_setup
554  *      initialize a workq item for this mci
555  *      passing in the new delay period in msec
556  *
557  *      locking model:
558  *
559  *              called with the mem_ctls_mutex held
560  */
561 static void edac_mc_workq_setup(struct mem_ctl_info *mci, unsigned msec)
562 {
563         edac_dbg(0, "\n");
564
565         /* if this instance is not in the POLL state, then simply return */
566         if (mci->op_state != OP_RUNNING_POLL)
567                 return;
568
569         INIT_DELAYED_WORK(&mci->work, edac_mc_workq_function);
570         mod_delayed_work(edac_workqueue, &mci->work, msecs_to_jiffies(msec));
571 }
572
573 /*
574  * edac_mc_workq_teardown
575  *      stop the workq processing on this mci
576  *
577  *      locking model:
578  *
579  *              called WITHOUT lock held
580  */
581 static void edac_mc_workq_teardown(struct mem_ctl_info *mci)
582 {
583         int status;
584
585         if (mci->op_state != OP_RUNNING_POLL)
586                 return;
587
588         status = cancel_delayed_work(&mci->work);
589         if (status == 0) {
590                 edac_dbg(0, "not canceled, flush the queue\n");
591
592                 /* workq instance might be running, wait for it */
593                 flush_workqueue(edac_workqueue);
594         }
595 }
596
597 /*
598  * edac_mc_reset_delay_period(unsigned long value)
599  *
600  *      user space has updated our poll period value, need to
601  *      reset our workq delays
602  */
603 void edac_mc_reset_delay_period(int value)
604 {
605         struct mem_ctl_info *mci;
606         struct list_head *item;
607
608         mutex_lock(&mem_ctls_mutex);
609
610         list_for_each(item, &mc_devices) {
611                 mci = list_entry(item, struct mem_ctl_info, link);
612
613                 edac_mc_workq_setup(mci, (unsigned long) value);
614         }
615
616         mutex_unlock(&mem_ctls_mutex);
617 }
618
619
620
621 /* Return 0 on success, 1 on failure.
622  * Before calling this function, caller must
623  * assign a unique value to mci->mc_idx.
624  *
625  *      locking model:
626  *
627  *              called with the mem_ctls_mutex lock held
628  */
629 static int add_mc_to_global_list(struct mem_ctl_info *mci)
630 {
631         struct list_head *item, *insert_before;
632         struct mem_ctl_info *p;
633
634         insert_before = &mc_devices;
635
636         p = find_mci_by_dev(mci->pdev);
637         if (unlikely(p != NULL))
638                 goto fail0;
639
640         list_for_each(item, &mc_devices) {
641                 p = list_entry(item, struct mem_ctl_info, link);
642
643                 if (p->mc_idx >= mci->mc_idx) {
644                         if (unlikely(p->mc_idx == mci->mc_idx))
645                                 goto fail1;
646
647                         insert_before = item;
648                         break;
649                 }
650         }
651
652         list_add_tail_rcu(&mci->link, insert_before);
653         atomic_inc(&edac_handlers);
654         return 0;
655
656 fail0:
657         edac_printk(KERN_WARNING, EDAC_MC,
658                 "%s (%s) %s %s already assigned %d\n", dev_name(p->pdev),
659                 edac_dev_name(mci), p->mod_name, p->ctl_name, p->mc_idx);
660         return 1;
661
662 fail1:
663         edac_printk(KERN_WARNING, EDAC_MC,
664                 "bug in low-level driver: attempt to assign\n"
665                 "    duplicate mc_idx %d in %s()\n", p->mc_idx, __func__);
666         return 1;
667 }
668
669 static void del_mc_from_global_list(struct mem_ctl_info *mci)
670 {
671         atomic_dec(&edac_handlers);
672         list_del_rcu(&mci->link);
673
674         /* these are for safe removal of devices from global list while
675          * NMI handlers may be traversing list
676          */
677         synchronize_rcu();
678         INIT_LIST_HEAD(&mci->link);
679 }
680
681 /**
682  * edac_mc_find: Search for a mem_ctl_info structure whose index is 'idx'.
683  *
684  * If found, return a pointer to the structure.
685  * Else return NULL.
686  *
687  * Caller must hold mem_ctls_mutex.
688  */
689 struct mem_ctl_info *edac_mc_find(int idx)
690 {
691         struct list_head *item;
692         struct mem_ctl_info *mci;
693
694         list_for_each(item, &mc_devices) {
695                 mci = list_entry(item, struct mem_ctl_info, link);
696
697                 if (mci->mc_idx >= idx) {
698                         if (mci->mc_idx == idx)
699                                 return mci;
700
701                         break;
702                 }
703         }
704
705         return NULL;
706 }
707 EXPORT_SYMBOL(edac_mc_find);
708
709 /**
710  * edac_mc_add_mc: Insert the 'mci' structure into the mci global list and
711  *                 create sysfs entries associated with mci structure
712  * @mci: pointer to the mci structure to be added to the list
713  *
714  * Return:
715  *      0       Success
716  *      !0      Failure
717  */
718
719 /* FIXME - should a warning be printed if no error detection? correction? */
720 int edac_mc_add_mc(struct mem_ctl_info *mci)
721 {
722         edac_dbg(0, "\n");
723
724 #ifdef CONFIG_EDAC_DEBUG
725         if (edac_debug_level >= 3)
726                 edac_mc_dump_mci(mci);
727
728         if (edac_debug_level >= 4) {
729                 int i;
730
731                 for (i = 0; i < mci->nr_csrows; i++) {
732                         struct csrow_info *csrow = mci->csrows[i];
733                         u32 nr_pages = 0;
734                         int j;
735
736                         for (j = 0; j < csrow->nr_channels; j++)
737                                 nr_pages += csrow->channels[j]->dimm->nr_pages;
738                         if (!nr_pages)
739                                 continue;
740                         edac_mc_dump_csrow(csrow);
741                         for (j = 0; j < csrow->nr_channels; j++)
742                                 if (csrow->channels[j]->dimm->nr_pages)
743                                         edac_mc_dump_channel(csrow->channels[j]);
744                 }
745                 for (i = 0; i < mci->tot_dimms; i++)
746                         if (mci->dimms[i]->nr_pages)
747                                 edac_mc_dump_dimm(mci->dimms[i], i);
748         }
749 #endif
750         mutex_lock(&mem_ctls_mutex);
751
752         if (add_mc_to_global_list(mci))
753                 goto fail0;
754
755         /* set load time so that error rate can be tracked */
756         mci->start_time = jiffies;
757
758         if (edac_create_sysfs_mci_device(mci)) {
759                 edac_mc_printk(mci, KERN_WARNING,
760                         "failed to create sysfs device\n");
761                 goto fail1;
762         }
763
764         /* If there IS a check routine, then we are running POLLED */
765         if (mci->edac_check != NULL) {
766                 /* This instance is NOW RUNNING */
767                 mci->op_state = OP_RUNNING_POLL;
768
769                 edac_mc_workq_setup(mci, edac_mc_get_poll_msec());
770         } else {
771                 mci->op_state = OP_RUNNING_INTERRUPT;
772         }
773
774         /* Report action taken */
775         edac_mc_printk(mci, KERN_INFO, "Giving out device to '%s' '%s':"
776                 " DEV %s\n", mci->mod_name, mci->ctl_name, edac_dev_name(mci));
777
778         mutex_unlock(&mem_ctls_mutex);
779         return 0;
780
781 fail1:
782         del_mc_from_global_list(mci);
783
784 fail0:
785         mutex_unlock(&mem_ctls_mutex);
786         return 1;
787 }
788 EXPORT_SYMBOL_GPL(edac_mc_add_mc);
789
790 /**
791  * edac_mc_del_mc: Remove sysfs entries for specified mci structure and
792  *                 remove mci structure from global list
793  * @pdev: Pointer to 'struct device' representing mci structure to remove.
794  *
795  * Return pointer to removed mci structure, or NULL if device not found.
796  */
797 struct mem_ctl_info *edac_mc_del_mc(struct device *dev)
798 {
799         struct mem_ctl_info *mci;
800
801         edac_dbg(0, "\n");
802
803         mutex_lock(&mem_ctls_mutex);
804
805         /* find the requested mci struct in the global list */
806         mci = find_mci_by_dev(dev);
807         if (mci == NULL) {
808                 mutex_unlock(&mem_ctls_mutex);
809                 return NULL;
810         }
811
812         del_mc_from_global_list(mci);
813         mutex_unlock(&mem_ctls_mutex);
814
815         /* flush workq processes */
816         edac_mc_workq_teardown(mci);
817
818         /* marking MCI offline */
819         mci->op_state = OP_OFFLINE;
820
821         /* remove from sysfs */
822         edac_remove_sysfs_mci_device(mci);
823
824         edac_printk(KERN_INFO, EDAC_MC,
825                 "Removed device %d for %s %s: DEV %s\n", mci->mc_idx,
826                 mci->mod_name, mci->ctl_name, edac_dev_name(mci));
827
828         return mci;
829 }
830 EXPORT_SYMBOL_GPL(edac_mc_del_mc);
831
832 static void edac_mc_scrub_block(unsigned long page, unsigned long offset,
833                                 u32 size)
834 {
835         struct page *pg;
836         void *virt_addr;
837         unsigned long flags = 0;
838
839         edac_dbg(3, "\n");
840
841         /* ECC error page was not in our memory. Ignore it. */
842         if (!pfn_valid(page))
843                 return;
844
845         /* Find the actual page structure then map it and fix */
846         pg = pfn_to_page(page);
847
848         if (PageHighMem(pg))
849                 local_irq_save(flags);
850
851         virt_addr = kmap_atomic(pg);
852
853         /* Perform architecture specific atomic scrub operation */
854         atomic_scrub(virt_addr + offset, size);
855
856         /* Unmap and complete */
857         kunmap_atomic(virt_addr);
858
859         if (PageHighMem(pg))
860                 local_irq_restore(flags);
861 }
862
863 /* FIXME - should return -1 */
864 int edac_mc_find_csrow_by_page(struct mem_ctl_info *mci, unsigned long page)
865 {
866         struct csrow_info **csrows = mci->csrows;
867         int row, i, j, n;
868
869         edac_dbg(1, "MC%d: 0x%lx\n", mci->mc_idx, page);
870         row = -1;
871
872         for (i = 0; i < mci->nr_csrows; i++) {
873                 struct csrow_info *csrow = csrows[i];
874                 n = 0;
875                 for (j = 0; j < csrow->nr_channels; j++) {
876                         struct dimm_info *dimm = csrow->channels[j]->dimm;
877                         n += dimm->nr_pages;
878                 }
879                 if (n == 0)
880                         continue;
881
882                 edac_dbg(3, "MC%d: first(0x%lx) page(0x%lx) last(0x%lx) mask(0x%lx)\n",
883                          mci->mc_idx,
884                          csrow->first_page, page, csrow->last_page,
885                          csrow->page_mask);
886
887                 if ((page >= csrow->first_page) &&
888                     (page <= csrow->last_page) &&
889                     ((page & csrow->page_mask) ==
890                      (csrow->first_page & csrow->page_mask))) {
891                         row = i;
892                         break;
893                 }
894         }
895
896         if (row == -1)
897                 edac_mc_printk(mci, KERN_ERR,
898                         "could not look up page error address %lx\n",
899                         (unsigned long)page);
900
901         return row;
902 }
903 EXPORT_SYMBOL_GPL(edac_mc_find_csrow_by_page);
904
905 const char *edac_layer_name[] = {
906         [EDAC_MC_LAYER_BRANCH] = "branch",
907         [EDAC_MC_LAYER_CHANNEL] = "channel",
908         [EDAC_MC_LAYER_SLOT] = "slot",
909         [EDAC_MC_LAYER_CHIP_SELECT] = "csrow",
910 };
911 EXPORT_SYMBOL_GPL(edac_layer_name);
912
913 static void edac_inc_ce_error(struct mem_ctl_info *mci,
914                               bool enable_per_layer_report,
915                               const int pos[EDAC_MAX_LAYERS],
916                               const u16 count)
917 {
918         int i, index = 0;
919
920         mci->ce_mc += count;
921
922         if (!enable_per_layer_report) {
923                 mci->ce_noinfo_count += count;
924                 return;
925         }
926
927         for (i = 0; i < mci->n_layers; i++) {
928                 if (pos[i] < 0)
929                         break;
930                 index += pos[i];
931                 mci->ce_per_layer[i][index] += count;
932
933                 if (i < mci->n_layers - 1)
934                         index *= mci->layers[i + 1].size;
935         }
936 }
937
938 static void edac_inc_ue_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->ue_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->ue_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_ce_error(struct mem_ctl_info *mci,
964                           const u16 error_count,
965                           const int pos[EDAC_MAX_LAYERS],
966                           const char *msg,
967                           const char *location,
968                           const char *label,
969                           const char *detail,
970                           const char *other_detail,
971                           const bool enable_per_layer_report,
972                           const unsigned long page_frame_number,
973                           const unsigned long offset_in_page,
974                           long grain)
975 {
976         unsigned long remapped_page;
977
978         if (edac_mc_get_log_ce()) {
979                 if (other_detail && *other_detail)
980                         edac_mc_printk(mci, KERN_WARNING,
981                                        "%d CE %s on %s (%s %s - %s)\n",
982                                        error_count,
983                                        msg, label, location,
984                                        detail, other_detail);
985                 else
986                         edac_mc_printk(mci, KERN_WARNING,
987                                        "%d CE %s on %s (%s %s)\n",
988                                        error_count,
989                                        msg, label, location,
990                                        detail);
991         }
992         edac_inc_ce_error(mci, enable_per_layer_report, pos, error_count);
993
994         if (mci->scrub_mode & SCRUB_SW_SRC) {
995                 /*
996                         * Some memory controllers (called MCs below) can remap
997                         * memory so that it is still available at a different
998                         * address when PCI devices map into memory.
999                         * MC's that can't do this, lose the memory where PCI
1000                         * devices are mapped. This mapping is MC-dependent
1001                         * and so we call back into the MC driver for it to
1002                         * map the MC page to a physical (CPU) page which can
1003                         * then be mapped to a virtual page - which can then
1004                         * be scrubbed.
1005                         */
1006                 remapped_page = mci->ctl_page_to_phys ?
1007                         mci->ctl_page_to_phys(mci, page_frame_number) :
1008                         page_frame_number;
1009
1010                 edac_mc_scrub_block(remapped_page,
1011                                         offset_in_page, grain);
1012         }
1013 }
1014
1015 static void edac_ue_error(struct mem_ctl_info *mci,
1016                           const u16 error_count,
1017                           const int pos[EDAC_MAX_LAYERS],
1018                           const char *msg,
1019                           const char *location,
1020                           const char *label,
1021                           const char *detail,
1022                           const char *other_detail,
1023                           const bool enable_per_layer_report)
1024 {
1025         if (edac_mc_get_log_ue()) {
1026                 if (other_detail && *other_detail)
1027                         edac_mc_printk(mci, KERN_WARNING,
1028                                        "%d UE %s on %s (%s %s - %s)\n",
1029                                        error_count,
1030                                        msg, label, location, detail,
1031                                        other_detail);
1032                 else
1033                         edac_mc_printk(mci, KERN_WARNING,
1034                                        "%d UE %s on %s (%s %s)\n",
1035                                        error_count,
1036                                        msg, label, location, detail);
1037         }
1038
1039         if (edac_mc_get_panic_on_ue()) {
1040                 if (other_detail && *other_detail)
1041                         panic("UE %s on %s (%s%s - %s)\n",
1042                               msg, label, location, detail, other_detail);
1043                 else
1044                         panic("UE %s on %s (%s%s)\n",
1045                               msg, label, location, detail);
1046         }
1047
1048         edac_inc_ue_error(mci, enable_per_layer_report, pos, error_count);
1049 }
1050
1051 #define OTHER_LABEL " or "
1052
1053 /**
1054  * edac_mc_handle_error - reports a memory event to userspace
1055  *
1056  * @type:               severity of the error (CE/UE/Fatal)
1057  * @mci:                a struct mem_ctl_info pointer
1058  * @error_count:        Number of errors of the same type
1059  * @page_frame_number:  mem page where the error occurred
1060  * @offset_in_page:     offset of the error inside the page
1061  * @syndrome:           ECC syndrome
1062  * @top_layer:          Memory layer[0] position
1063  * @mid_layer:          Memory layer[1] position
1064  * @low_layer:          Memory layer[2] position
1065  * @msg:                Message meaningful to the end users that
1066  *                      explains the event
1067  * @other_detail:       Technical details about the event that
1068  *                      may help hardware manufacturers and
1069  *                      EDAC developers to analyse the event
1070  */
1071 void edac_mc_handle_error(const enum hw_event_mc_err_type type,
1072                           struct mem_ctl_info *mci,
1073                           const u16 error_count,
1074                           const unsigned long page_frame_number,
1075                           const unsigned long offset_in_page,
1076                           const unsigned long syndrome,
1077                           const int top_layer,
1078                           const int mid_layer,
1079                           const int low_layer,
1080                           const char *msg,
1081                           const char *other_detail)
1082 {
1083         /* FIXME: too much for stack: move it to some pre-alocated area */
1084         char detail[80], location[80];
1085         char label[(EDAC_MC_LABEL_LEN + 1 + sizeof(OTHER_LABEL)) * mci->tot_dimms];
1086         char *p;
1087         int row = -1, chan = -1;
1088         int pos[EDAC_MAX_LAYERS] = { top_layer, mid_layer, low_layer };
1089         int i;
1090         long grain;
1091         bool enable_per_layer_report = false;
1092         u8 grain_bits;
1093
1094         edac_dbg(3, "MC%d\n", mci->mc_idx);
1095
1096         /*
1097          * Check if the event report is consistent and if the memory
1098          * location is known. If it is known, enable_per_layer_report will be
1099          * true, the DIMM(s) label info will be filled and the per-layer
1100          * error counters will be incremented.
1101          */
1102         for (i = 0; i < mci->n_layers; i++) {
1103                 if (pos[i] >= (int)mci->layers[i].size) {
1104                         if (type == HW_EVENT_ERR_CORRECTED)
1105                                 p = "CE";
1106                         else
1107                                 p = "UE";
1108
1109                         edac_mc_printk(mci, KERN_ERR,
1110                                        "INTERNAL ERROR: %s value is out of range (%d >= %d)\n",
1111                                        edac_layer_name[mci->layers[i].type],
1112                                        pos[i], mci->layers[i].size);
1113                         /*
1114                          * Instead of just returning it, let's use what's
1115                          * known about the error. The increment routines and
1116                          * the DIMM filter logic will do the right thing by
1117                          * pointing the likely damaged DIMMs.
1118                          */
1119                         pos[i] = -1;
1120                 }
1121                 if (pos[i] >= 0)
1122                         enable_per_layer_report = true;
1123         }
1124
1125         /*
1126          * Get the dimm label/grain that applies to the match criteria.
1127          * As the error algorithm may not be able to point to just one memory
1128          * stick, the logic here will get all possible labels that could
1129          * pottentially be affected by the error.
1130          * On FB-DIMM memory controllers, for uncorrected errors, it is common
1131          * to have only the MC channel and the MC dimm (also called "branch")
1132          * but the channel is not known, as the memory is arranged in pairs,
1133          * where each memory belongs to a separate channel within the same
1134          * branch.
1135          */
1136         grain = 0;
1137         p = label;
1138         *p = '\0';
1139         for (i = 0; i < mci->tot_dimms; i++) {
1140                 struct dimm_info *dimm = mci->dimms[i];
1141
1142                 if (top_layer >= 0 && top_layer != dimm->location[0])
1143                         continue;
1144                 if (mid_layer >= 0 && mid_layer != dimm->location[1])
1145                         continue;
1146                 if (low_layer >= 0 && low_layer != dimm->location[2])
1147                         continue;
1148
1149                 /* get the max grain, over the error match range */
1150                 if (dimm->grain > grain)
1151                         grain = dimm->grain;
1152
1153                 /*
1154                  * If the error is memory-controller wide, there's no need to
1155                  * seek for the affected DIMMs because the whole
1156                  * channel/memory controller/...  may be affected.
1157                  * Also, don't show errors for empty DIMM slots.
1158                  */
1159                 if (enable_per_layer_report && dimm->nr_pages) {
1160                         if (p != label) {
1161                                 strcpy(p, OTHER_LABEL);
1162                                 p += strlen(OTHER_LABEL);
1163                         }
1164                         strcpy(p, dimm->label);
1165                         p += strlen(p);
1166                         *p = '\0';
1167
1168                         /*
1169                          * get csrow/channel of the DIMM, in order to allow
1170                          * incrementing the compat API counters
1171                          */
1172                         edac_dbg(4, "%s csrows map: (%d,%d)\n",
1173                                  mci->mem_is_per_rank ? "rank" : "dimm",
1174                                  dimm->csrow, dimm->cschannel);
1175                         if (row == -1)
1176                                 row = dimm->csrow;
1177                         else if (row >= 0 && row != dimm->csrow)
1178                                 row = -2;
1179
1180                         if (chan == -1)
1181                                 chan = dimm->cschannel;
1182                         else if (chan >= 0 && chan != dimm->cschannel)
1183                                 chan = -2;
1184                 }
1185         }
1186
1187         if (!enable_per_layer_report) {
1188                 strcpy(label, "any memory");
1189         } else {
1190                 edac_dbg(4, "csrow/channel to increment: (%d,%d)\n", row, chan);
1191                 if (p == label)
1192                         strcpy(label, "unknown memory");
1193                 if (type == HW_EVENT_ERR_CORRECTED) {
1194                         if (row >= 0) {
1195                                 mci->csrows[row]->ce_count += error_count;
1196                                 if (chan >= 0)
1197                                         mci->csrows[row]->channels[chan]->ce_count += error_count;
1198                         }
1199                 } else
1200                         if (row >= 0)
1201                                 mci->csrows[row]->ue_count += error_count;
1202         }
1203
1204         /* Fill the RAM location data */
1205         p = location;
1206         for (i = 0; i < mci->n_layers; i++) {
1207                 if (pos[i] < 0)
1208                         continue;
1209
1210                 p += sprintf(p, "%s:%d ",
1211                              edac_layer_name[mci->layers[i].type],
1212                              pos[i]);
1213         }
1214         if (p > location)
1215                 *(p - 1) = '\0';
1216
1217         /* Report the error via the trace interface */
1218
1219         grain_bits = fls_long(grain) + 1;
1220         trace_mc_event(type, msg, label, error_count,
1221                        mci->mc_idx, top_layer, mid_layer, low_layer,
1222                        PAGES_TO_MiB(page_frame_number) | offset_in_page,
1223                        grain_bits, syndrome, other_detail);
1224
1225         /* Memory type dependent details about the error */
1226         if (type == HW_EVENT_ERR_CORRECTED) {
1227                 snprintf(detail, sizeof(detail),
1228                         "page:0x%lx offset:0x%lx grain:%ld syndrome:0x%lx",
1229                         page_frame_number, offset_in_page,
1230                         grain, syndrome);
1231                 edac_ce_error(mci, error_count, pos, msg, location, label,
1232                               detail, other_detail, enable_per_layer_report,
1233                               page_frame_number, offset_in_page, grain);
1234         } else {
1235                 snprintf(detail, sizeof(detail),
1236                         "page:0x%lx offset:0x%lx grain:%ld",
1237                         page_frame_number, offset_in_page, grain);
1238
1239                 edac_ue_error(mci, error_count, pos, msg, location, label,
1240                               detail, other_detail, enable_per_layer_report);
1241         }
1242 }
1243 EXPORT_SYMBOL_GPL(edac_mc_handle_error);