Merge branch 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[cascardo/linux.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 #include <ras/ras_event.h>
37
38 /* lock to memory controller's control array */
39 static DEFINE_MUTEX(mem_ctls_mutex);
40 static LIST_HEAD(mc_devices);
41
42 /*
43  * Used to lock EDAC MC to just one module, avoiding two drivers e. g.
44  *      apei/ghes and i7core_edac to be used at the same time.
45  */
46 static void const *edac_mc_owner;
47
48 static struct bus_type mc_bus[EDAC_MAX_MCS];
49
50 unsigned edac_dimm_info_location(struct dimm_info *dimm, char *buf,
51                                  unsigned len)
52 {
53         struct mem_ctl_info *mci = dimm->mci;
54         int i, n, count = 0;
55         char *p = buf;
56
57         for (i = 0; i < mci->n_layers; i++) {
58                 n = snprintf(p, len, "%s %d ",
59                               edac_layer_name[mci->layers[i].type],
60                               dimm->location[i]);
61                 p += n;
62                 len -= n;
63                 count += n;
64                 if (!len)
65                         break;
66         }
67
68         return count;
69 }
70
71 #ifdef CONFIG_EDAC_DEBUG
72
73 static void edac_mc_dump_channel(struct rank_info *chan)
74 {
75         edac_dbg(4, "  channel->chan_idx = %d\n", chan->chan_idx);
76         edac_dbg(4, "    channel = %p\n", chan);
77         edac_dbg(4, "    channel->csrow = %p\n", chan->csrow);
78         edac_dbg(4, "    channel->dimm = %p\n", chan->dimm);
79 }
80
81 static void edac_mc_dump_dimm(struct dimm_info *dimm, int number)
82 {
83         char location[80];
84
85         edac_dimm_info_location(dimm, location, sizeof(location));
86
87         edac_dbg(4, "%s%i: %smapped as virtual row %d, chan %d\n",
88                  dimm->mci->csbased ? "rank" : "dimm",
89                  number, location, dimm->csrow, dimm->cschannel);
90         edac_dbg(4, "  dimm = %p\n", dimm);
91         edac_dbg(4, "  dimm->label = '%s'\n", dimm->label);
92         edac_dbg(4, "  dimm->nr_pages = 0x%x\n", dimm->nr_pages);
93         edac_dbg(4, "  dimm->grain = %d\n", dimm->grain);
94         edac_dbg(4, "  dimm->nr_pages = 0x%x\n", dimm->nr_pages);
95 }
96
97 static void edac_mc_dump_csrow(struct csrow_info *csrow)
98 {
99         edac_dbg(4, "csrow->csrow_idx = %d\n", csrow->csrow_idx);
100         edac_dbg(4, "  csrow = %p\n", csrow);
101         edac_dbg(4, "  csrow->first_page = 0x%lx\n", csrow->first_page);
102         edac_dbg(4, "  csrow->last_page = 0x%lx\n", csrow->last_page);
103         edac_dbg(4, "  csrow->page_mask = 0x%lx\n", csrow->page_mask);
104         edac_dbg(4, "  csrow->nr_channels = %d\n", csrow->nr_channels);
105         edac_dbg(4, "  csrow->channels = %p\n", csrow->channels);
106         edac_dbg(4, "  csrow->mci = %p\n", csrow->mci);
107 }
108
109 static void edac_mc_dump_mci(struct mem_ctl_info *mci)
110 {
111         edac_dbg(3, "\tmci = %p\n", mci);
112         edac_dbg(3, "\tmci->mtype_cap = %lx\n", mci->mtype_cap);
113         edac_dbg(3, "\tmci->edac_ctl_cap = %lx\n", mci->edac_ctl_cap);
114         edac_dbg(3, "\tmci->edac_cap = %lx\n", mci->edac_cap);
115         edac_dbg(4, "\tmci->edac_check = %p\n", mci->edac_check);
116         edac_dbg(3, "\tmci->nr_csrows = %d, csrows = %p\n",
117                  mci->nr_csrows, mci->csrows);
118         edac_dbg(3, "\tmci->nr_dimms = %d, dimms = %p\n",
119                  mci->tot_dimms, mci->dimms);
120         edac_dbg(3, "\tdev = %p\n", mci->pdev);
121         edac_dbg(3, "\tmod_name:ctl_name = %s:%s\n",
122                  mci->mod_name, mci->ctl_name);
123         edac_dbg(3, "\tpvt_info = %p\n\n", mci->pvt_info);
124 }
125
126 #endif                          /* CONFIG_EDAC_DEBUG */
127
128 const char * const edac_mem_types[] = {
129         [MEM_EMPTY]     = "Empty csrow",
130         [MEM_RESERVED]  = "Reserved csrow type",
131         [MEM_UNKNOWN]   = "Unknown csrow type",
132         [MEM_FPM]       = "Fast page mode RAM",
133         [MEM_EDO]       = "Extended data out RAM",
134         [MEM_BEDO]      = "Burst Extended data out RAM",
135         [MEM_SDR]       = "Single data rate SDRAM",
136         [MEM_RDR]       = "Registered single data rate SDRAM",
137         [MEM_DDR]       = "Double data rate SDRAM",
138         [MEM_RDDR]      = "Registered Double data rate SDRAM",
139         [MEM_RMBS]      = "Rambus DRAM",
140         [MEM_DDR2]      = "Unbuffered DDR2 RAM",
141         [MEM_FB_DDR2]   = "Fully buffered DDR2",
142         [MEM_RDDR2]     = "Registered DDR2 RAM",
143         [MEM_XDR]       = "Rambus XDR",
144         [MEM_DDR3]      = "Unbuffered DDR3 RAM",
145         [MEM_RDDR3]     = "Registered DDR3 RAM",
146         [MEM_LRDDR3]    = "Load-Reduced DDR3 RAM",
147         [MEM_DDR4]      = "Unbuffered DDR4 RAM",
148         [MEM_RDDR4]     = "Registered DDR4 RAM",
149 };
150 EXPORT_SYMBOL_GPL(edac_mem_types);
151
152 /**
153  * edac_align_ptr - Prepares the pointer offsets for a single-shot allocation
154  * @p:          pointer to a pointer with the memory offset to be used. At
155  *              return, this will be incremented to point to the next offset
156  * @size:       Size of the data structure to be reserved
157  * @n_elems:    Number of elements that should be reserved
158  *
159  * If 'size' is a constant, the compiler will optimize this whole function
160  * down to either a no-op or the addition of a constant to the value of '*p'.
161  *
162  * The 'p' pointer is absolutely needed to keep the proper advancing
163  * further in memory to the proper offsets when allocating the struct along
164  * with its embedded structs, as edac_device_alloc_ctl_info() does it
165  * above, for example.
166  *
167  * At return, the pointer 'p' will be incremented to be used on a next call
168  * to this function.
169  */
170 void *edac_align_ptr(void **p, unsigned size, int n_elems)
171 {
172         unsigned align, r;
173         void *ptr = *p;
174
175         *p += size * n_elems;
176
177         /*
178          * 'p' can possibly be an unaligned item X such that sizeof(X) is
179          * 'size'.  Adjust 'p' so that its alignment is at least as
180          * stringent as what the compiler would provide for X and return
181          * the aligned result.
182          * Here we assume that the alignment of a "long long" is the most
183          * stringent alignment that the compiler will ever provide by default.
184          * As far as I know, this is a reasonable assumption.
185          */
186         if (size > sizeof(long))
187                 align = sizeof(long long);
188         else if (size > sizeof(int))
189                 align = sizeof(long);
190         else if (size > sizeof(short))
191                 align = sizeof(int);
192         else if (size > sizeof(char))
193                 align = sizeof(short);
194         else
195                 return (char *)ptr;
196
197         r = (unsigned long)p % align;
198
199         if (r == 0)
200                 return (char *)ptr;
201
202         *p += align - r;
203
204         return (void *)(((unsigned long)ptr) + align - r);
205 }
206
207 static void _edac_mc_free(struct mem_ctl_info *mci)
208 {
209         int i, chn, row;
210         struct csrow_info *csr;
211         const unsigned int tot_dimms = mci->tot_dimms;
212         const unsigned int tot_channels = mci->num_cschannel;
213         const unsigned int tot_csrows = mci->nr_csrows;
214
215         if (mci->dimms) {
216                 for (i = 0; i < tot_dimms; i++)
217                         kfree(mci->dimms[i]);
218                 kfree(mci->dimms);
219         }
220         if (mci->csrows) {
221                 for (row = 0; row < tot_csrows; row++) {
222                         csr = mci->csrows[row];
223                         if (csr) {
224                                 if (csr->channels) {
225                                         for (chn = 0; chn < tot_channels; chn++)
226                                                 kfree(csr->channels[chn]);
227                                         kfree(csr->channels);
228                                 }
229                                 kfree(csr);
230                         }
231                 }
232                 kfree(mci->csrows);
233         }
234         kfree(mci);
235 }
236
237 /**
238  * edac_mc_alloc: Allocate and partially fill a struct mem_ctl_info structure
239  * @mc_num:             Memory controller number
240  * @n_layers:           Number of MC hierarchy layers
241  * layers:              Describes each layer as seen by the Memory Controller
242  * @size_pvt:           size of private storage needed
243  *
244  *
245  * Everything is kmalloc'ed as one big chunk - more efficient.
246  * Only can be used if all structures have the same lifetime - otherwise
247  * you have to allocate and initialize your own structures.
248  *
249  * Use edac_mc_free() to free mc structures allocated by this function.
250  *
251  * NOTE: drivers handle multi-rank memories in different ways: in some
252  * drivers, one multi-rank memory stick is mapped as one entry, while, in
253  * others, a single multi-rank memory stick would be mapped into several
254  * entries. Currently, this function will allocate multiple struct dimm_info
255  * on such scenarios, as grouping the multiple ranks require drivers change.
256  *
257  * Returns:
258  *      On failure: NULL
259  *      On success: struct mem_ctl_info pointer
260  */
261 struct mem_ctl_info *edac_mc_alloc(unsigned mc_num,
262                                    unsigned n_layers,
263                                    struct edac_mc_layer *layers,
264                                    unsigned sz_pvt)
265 {
266         struct mem_ctl_info *mci;
267         struct edac_mc_layer *layer;
268         struct csrow_info *csr;
269         struct rank_info *chan;
270         struct dimm_info *dimm;
271         u32 *ce_per_layer[EDAC_MAX_LAYERS], *ue_per_layer[EDAC_MAX_LAYERS];
272         unsigned pos[EDAC_MAX_LAYERS];
273         unsigned size, tot_dimms = 1, count = 1;
274         unsigned tot_csrows = 1, tot_channels = 1, tot_errcount = 0;
275         void *pvt, *p, *ptr = NULL;
276         int i, j, row, chn, n, len, off;
277         bool per_rank = false;
278
279         BUG_ON(n_layers > EDAC_MAX_LAYERS || n_layers == 0);
280         /*
281          * Calculate the total amount of dimms and csrows/cschannels while
282          * in the old API emulation mode
283          */
284         for (i = 0; i < n_layers; i++) {
285                 tot_dimms *= layers[i].size;
286                 if (layers[i].is_virt_csrow)
287                         tot_csrows *= layers[i].size;
288                 else
289                         tot_channels *= layers[i].size;
290
291                 if (layers[i].type == EDAC_MC_LAYER_CHIP_SELECT)
292                         per_rank = true;
293         }
294
295         /* Figure out the offsets of the various items from the start of an mc
296          * structure.  We want the alignment of each item to be at least as
297          * stringent as what the compiler would provide if we could simply
298          * hardcode everything into a single struct.
299          */
300         mci = edac_align_ptr(&ptr, sizeof(*mci), 1);
301         layer = edac_align_ptr(&ptr, sizeof(*layer), n_layers);
302         for (i = 0; i < n_layers; i++) {
303                 count *= layers[i].size;
304                 edac_dbg(4, "errcount layer %d size %d\n", i, count);
305                 ce_per_layer[i] = edac_align_ptr(&ptr, sizeof(u32), count);
306                 ue_per_layer[i] = edac_align_ptr(&ptr, sizeof(u32), count);
307                 tot_errcount += 2 * count;
308         }
309
310         edac_dbg(4, "allocating %d error counters\n", tot_errcount);
311         pvt = edac_align_ptr(&ptr, sz_pvt, 1);
312         size = ((unsigned long)pvt) + sz_pvt;
313
314         edac_dbg(1, "allocating %u bytes for mci data (%d %s, %d csrows/channels)\n",
315                  size,
316                  tot_dimms,
317                  per_rank ? "ranks" : "dimms",
318                  tot_csrows * tot_channels);
319
320         mci = kzalloc(size, GFP_KERNEL);
321         if (mci == NULL)
322                 return NULL;
323
324         /* Adjust pointers so they point within the memory we just allocated
325          * rather than an imaginary chunk of memory located at address 0.
326          */
327         layer = (struct edac_mc_layer *)(((char *)mci) + ((unsigned long)layer));
328         for (i = 0; i < n_layers; i++) {
329                 mci->ce_per_layer[i] = (u32 *)((char *)mci + ((unsigned long)ce_per_layer[i]));
330                 mci->ue_per_layer[i] = (u32 *)((char *)mci + ((unsigned long)ue_per_layer[i]));
331         }
332         pvt = sz_pvt ? (((char *)mci) + ((unsigned long)pvt)) : NULL;
333
334         /* setup index and various internal pointers */
335         mci->mc_idx = mc_num;
336         mci->tot_dimms = tot_dimms;
337         mci->pvt_info = pvt;
338         mci->n_layers = n_layers;
339         mci->layers = layer;
340         memcpy(mci->layers, layers, sizeof(*layer) * n_layers);
341         mci->nr_csrows = tot_csrows;
342         mci->num_cschannel = tot_channels;
343         mci->csbased = per_rank;
344
345         /*
346          * Alocate and fill the csrow/channels structs
347          */
348         mci->csrows = kcalloc(tot_csrows, sizeof(*mci->csrows), GFP_KERNEL);
349         if (!mci->csrows)
350                 goto error;
351         for (row = 0; row < tot_csrows; row++) {
352                 csr = kzalloc(sizeof(**mci->csrows), GFP_KERNEL);
353                 if (!csr)
354                         goto error;
355                 mci->csrows[row] = csr;
356                 csr->csrow_idx = row;
357                 csr->mci = mci;
358                 csr->nr_channels = tot_channels;
359                 csr->channels = kcalloc(tot_channels, sizeof(*csr->channels),
360                                         GFP_KERNEL);
361                 if (!csr->channels)
362                         goto error;
363
364                 for (chn = 0; chn < tot_channels; chn++) {
365                         chan = kzalloc(sizeof(**csr->channels), GFP_KERNEL);
366                         if (!chan)
367                                 goto error;
368                         csr->channels[chn] = chan;
369                         chan->chan_idx = chn;
370                         chan->csrow = csr;
371                 }
372         }
373
374         /*
375          * Allocate and fill the dimm structs
376          */
377         mci->dimms  = kcalloc(tot_dimms, sizeof(*mci->dimms), GFP_KERNEL);
378         if (!mci->dimms)
379                 goto error;
380
381         memset(&pos, 0, sizeof(pos));
382         row = 0;
383         chn = 0;
384         for (i = 0; i < tot_dimms; i++) {
385                 chan = mci->csrows[row]->channels[chn];
386                 off = EDAC_DIMM_OFF(layer, n_layers, pos[0], pos[1], pos[2]);
387                 if (off < 0 || off >= tot_dimms) {
388                         edac_mc_printk(mci, KERN_ERR, "EDAC core bug: EDAC_DIMM_OFF is trying to do an illegal data access\n");
389                         goto error;
390                 }
391
392                 dimm = kzalloc(sizeof(**mci->dimms), GFP_KERNEL);
393                 if (!dimm)
394                         goto error;
395                 mci->dimms[off] = dimm;
396                 dimm->mci = mci;
397
398                 /*
399                  * Copy DIMM location and initialize it.
400                  */
401                 len = sizeof(dimm->label);
402                 p = dimm->label;
403                 n = snprintf(p, len, "mc#%u", mc_num);
404                 p += n;
405                 len -= n;
406                 for (j = 0; j < n_layers; j++) {
407                         n = snprintf(p, len, "%s#%u",
408                                      edac_layer_name[layers[j].type],
409                                      pos[j]);
410                         p += n;
411                         len -= n;
412                         dimm->location[j] = pos[j];
413
414                         if (len <= 0)
415                                 break;
416                 }
417
418                 /* Link it to the csrows old API data */
419                 chan->dimm = dimm;
420                 dimm->csrow = row;
421                 dimm->cschannel = chn;
422
423                 /* Increment csrow location */
424                 if (layers[0].is_virt_csrow) {
425                         chn++;
426                         if (chn == tot_channels) {
427                                 chn = 0;
428                                 row++;
429                         }
430                 } else {
431                         row++;
432                         if (row == tot_csrows) {
433                                 row = 0;
434                                 chn++;
435                         }
436                 }
437
438                 /* Increment dimm location */
439                 for (j = n_layers - 1; j >= 0; j--) {
440                         pos[j]++;
441                         if (pos[j] < layers[j].size)
442                                 break;
443                         pos[j] = 0;
444                 }
445         }
446
447         mci->op_state = OP_ALLOC;
448
449         return mci;
450
451 error:
452         _edac_mc_free(mci);
453
454         return NULL;
455 }
456 EXPORT_SYMBOL_GPL(edac_mc_alloc);
457
458 /**
459  * edac_mc_free
460  *      'Free' a previously allocated 'mci' structure
461  * @mci: pointer to a struct mem_ctl_info structure
462  */
463 void edac_mc_free(struct mem_ctl_info *mci)
464 {
465         edac_dbg(1, "\n");
466
467         /* If we're not yet registered with sysfs free only what was allocated
468          * in edac_mc_alloc().
469          */
470         if (!device_is_registered(&mci->dev)) {
471                 _edac_mc_free(mci);
472                 return;
473         }
474
475         /* the mci instance is freed here, when the sysfs object is dropped */
476         edac_unregister_sysfs(mci);
477 }
478 EXPORT_SYMBOL_GPL(edac_mc_free);
479
480
481 /**
482  * find_mci_by_dev
483  *
484  *      scan list of controllers looking for the one that manages
485  *      the 'dev' device
486  * @dev: pointer to a struct device related with the MCI
487  */
488 struct mem_ctl_info *find_mci_by_dev(struct device *dev)
489 {
490         struct mem_ctl_info *mci;
491         struct list_head *item;
492
493         edac_dbg(3, "\n");
494
495         list_for_each(item, &mc_devices) {
496                 mci = list_entry(item, struct mem_ctl_info, link);
497
498                 if (mci->pdev == dev)
499                         return mci;
500         }
501
502         return NULL;
503 }
504 EXPORT_SYMBOL_GPL(find_mci_by_dev);
505
506 /*
507  * handler for EDAC to check if NMI type handler has asserted interrupt
508  */
509 static int edac_mc_assert_error_check_and_clear(void)
510 {
511         int old_state;
512
513         if (edac_op_state == EDAC_OPSTATE_POLL)
514                 return 1;
515
516         old_state = edac_err_assert;
517         edac_err_assert = 0;
518
519         return old_state;
520 }
521
522 /*
523  * edac_mc_workq_function
524  *      performs the operation scheduled by a workq request
525  */
526 static void edac_mc_workq_function(struct work_struct *work_req)
527 {
528         struct delayed_work *d_work = to_delayed_work(work_req);
529         struct mem_ctl_info *mci = to_edac_mem_ctl_work(d_work);
530
531         mutex_lock(&mem_ctls_mutex);
532
533         /* if this control struct has movd to offline state, we are done */
534         if (mci->op_state == OP_OFFLINE) {
535                 mutex_unlock(&mem_ctls_mutex);
536                 return;
537         }
538
539         /* Only poll controllers that are running polled and have a check */
540         if (edac_mc_assert_error_check_and_clear() && (mci->edac_check != NULL))
541                 mci->edac_check(mci);
542
543         mutex_unlock(&mem_ctls_mutex);
544
545         /* Reschedule */
546         queue_delayed_work(edac_workqueue, &mci->work,
547                         msecs_to_jiffies(edac_mc_get_poll_msec()));
548 }
549
550 /*
551  * edac_mc_workq_setup
552  *      initialize a workq item for this mci
553  *      passing in the new delay period in msec
554  *
555  *      locking model:
556  *
557  *              called with the mem_ctls_mutex held
558  */
559 static void edac_mc_workq_setup(struct mem_ctl_info *mci, unsigned msec,
560                                 bool init)
561 {
562         edac_dbg(0, "\n");
563
564         /* if this instance is not in the POLL state, then simply return */
565         if (mci->op_state != OP_RUNNING_POLL)
566                 return;
567
568         if (init)
569                 INIT_DELAYED_WORK(&mci->work, edac_mc_workq_function);
570
571         mod_delayed_work(edac_workqueue, &mci->work, msecs_to_jiffies(msec));
572 }
573
574 /*
575  * edac_mc_workq_teardown
576  *      stop the workq processing on this mci
577  *
578  *      locking model:
579  *
580  *              called WITHOUT lock held
581  */
582 static void edac_mc_workq_teardown(struct mem_ctl_info *mci)
583 {
584         int status;
585
586         if (mci->op_state != OP_RUNNING_POLL)
587                 return;
588
589         status = cancel_delayed_work(&mci->work);
590         if (status == 0) {
591                 edac_dbg(0, "not canceled, flush the queue\n");
592
593                 /* workq instance might be running, wait for it */
594                 flush_workqueue(edac_workqueue);
595         }
596 }
597
598 /*
599  * edac_mc_reset_delay_period(unsigned long value)
600  *
601  *      user space has updated our poll period value, need to
602  *      reset our workq delays
603  */
604 void edac_mc_reset_delay_period(unsigned long value)
605 {
606         struct mem_ctl_info *mci;
607         struct list_head *item;
608
609         mutex_lock(&mem_ctls_mutex);
610
611         list_for_each(item, &mc_devices) {
612                 mci = list_entry(item, struct mem_ctl_info, link);
613
614                 edac_mc_workq_setup(mci, value, false);
615         }
616
617         mutex_unlock(&mem_ctls_mutex);
618 }
619
620
621
622 /* Return 0 on success, 1 on failure.
623  * Before calling this function, caller must
624  * assign a unique value to mci->mc_idx.
625  *
626  *      locking model:
627  *
628  *              called with the mem_ctls_mutex lock held
629  */
630 static int add_mc_to_global_list(struct mem_ctl_info *mci)
631 {
632         struct list_head *item, *insert_before;
633         struct mem_ctl_info *p;
634
635         insert_before = &mc_devices;
636
637         p = find_mci_by_dev(mci->pdev);
638         if (unlikely(p != NULL))
639                 goto fail0;
640
641         list_for_each(item, &mc_devices) {
642                 p = list_entry(item, struct mem_ctl_info, link);
643
644                 if (p->mc_idx >= mci->mc_idx) {
645                         if (unlikely(p->mc_idx == mci->mc_idx))
646                                 goto fail1;
647
648                         insert_before = item;
649                         break;
650                 }
651         }
652
653         list_add_tail_rcu(&mci->link, insert_before);
654         atomic_inc(&edac_handlers);
655         return 0;
656
657 fail0:
658         edac_printk(KERN_WARNING, EDAC_MC,
659                 "%s (%s) %s %s already assigned %d\n", dev_name(p->pdev),
660                 edac_dev_name(mci), p->mod_name, p->ctl_name, p->mc_idx);
661         return 1;
662
663 fail1:
664         edac_printk(KERN_WARNING, EDAC_MC,
665                 "bug in low-level driver: attempt to assign\n"
666                 "    duplicate mc_idx %d in %s()\n", p->mc_idx, __func__);
667         return 1;
668 }
669
670 static int del_mc_from_global_list(struct mem_ctl_info *mci)
671 {
672         int handlers = atomic_dec_return(&edac_handlers);
673         list_del_rcu(&mci->link);
674
675         /* these are for safe removal of devices from global list while
676          * NMI handlers may be traversing list
677          */
678         synchronize_rcu();
679         INIT_LIST_HEAD(&mci->link);
680
681         return handlers;
682 }
683
684 /**
685  * edac_mc_find: Search for a mem_ctl_info structure whose index is 'idx'.
686  *
687  * If found, return a pointer to the structure.
688  * Else return NULL.
689  *
690  * Caller must hold mem_ctls_mutex.
691  */
692 struct mem_ctl_info *edac_mc_find(int idx)
693 {
694         struct list_head *item;
695         struct mem_ctl_info *mci;
696
697         list_for_each(item, &mc_devices) {
698                 mci = list_entry(item, struct mem_ctl_info, link);
699
700                 if (mci->mc_idx >= idx) {
701                         if (mci->mc_idx == idx)
702                                 return mci;
703
704                         break;
705                 }
706         }
707
708         return NULL;
709 }
710 EXPORT_SYMBOL(edac_mc_find);
711
712 /**
713  * edac_mc_add_mc: Insert the 'mci' structure into the mci global list and
714  *                 create sysfs entries associated with mci structure
715  * @mci: pointer to the mci structure to be added to the list
716  *
717  * Return:
718  *      0       Success
719  *      !0      Failure
720  */
721
722 /* FIXME - should a warning be printed if no error detection? correction? */
723 int edac_mc_add_mc(struct mem_ctl_info *mci)
724 {
725         int ret = -EINVAL;
726         edac_dbg(0, "\n");
727
728         if (mci->mc_idx >= EDAC_MAX_MCS) {
729                 pr_warn_once("Too many memory controllers: %d\n", mci->mc_idx);
730                 return -ENODEV;
731         }
732
733 #ifdef CONFIG_EDAC_DEBUG
734         if (edac_debug_level >= 3)
735                 edac_mc_dump_mci(mci);
736
737         if (edac_debug_level >= 4) {
738                 int i;
739
740                 for (i = 0; i < mci->nr_csrows; i++) {
741                         struct csrow_info *csrow = mci->csrows[i];
742                         u32 nr_pages = 0;
743                         int j;
744
745                         for (j = 0; j < csrow->nr_channels; j++)
746                                 nr_pages += csrow->channels[j]->dimm->nr_pages;
747                         if (!nr_pages)
748                                 continue;
749                         edac_mc_dump_csrow(csrow);
750                         for (j = 0; j < csrow->nr_channels; j++)
751                                 if (csrow->channels[j]->dimm->nr_pages)
752                                         edac_mc_dump_channel(csrow->channels[j]);
753                 }
754                 for (i = 0; i < mci->tot_dimms; i++)
755                         if (mci->dimms[i]->nr_pages)
756                                 edac_mc_dump_dimm(mci->dimms[i], i);
757         }
758 #endif
759         mutex_lock(&mem_ctls_mutex);
760
761         if (edac_mc_owner && edac_mc_owner != mci->mod_name) {
762                 ret = -EPERM;
763                 goto fail0;
764         }
765
766         if (add_mc_to_global_list(mci))
767                 goto fail0;
768
769         /* set load time so that error rate can be tracked */
770         mci->start_time = jiffies;
771
772         mci->bus = &mc_bus[mci->mc_idx];
773
774         if (edac_create_sysfs_mci_device(mci)) {
775                 edac_mc_printk(mci, KERN_WARNING,
776                         "failed to create sysfs device\n");
777                 goto fail1;
778         }
779
780         /* If there IS a check routine, then we are running POLLED */
781         if (mci->edac_check != NULL) {
782                 /* This instance is NOW RUNNING */
783                 mci->op_state = OP_RUNNING_POLL;
784
785                 edac_mc_workq_setup(mci, edac_mc_get_poll_msec(), true);
786         } else {
787                 mci->op_state = OP_RUNNING_INTERRUPT;
788         }
789
790         /* Report action taken */
791         edac_mc_printk(mci, KERN_INFO,
792                 "Giving out device to module %s controller %s: DEV %s (%s)\n",
793                 mci->mod_name, mci->ctl_name, mci->dev_name,
794                 edac_op_state_to_string(mci->op_state));
795
796         edac_mc_owner = mci->mod_name;
797
798         mutex_unlock(&mem_ctls_mutex);
799         return 0;
800
801 fail1:
802         del_mc_from_global_list(mci);
803
804 fail0:
805         mutex_unlock(&mem_ctls_mutex);
806         return ret;
807 }
808 EXPORT_SYMBOL_GPL(edac_mc_add_mc);
809
810 /**
811  * edac_mc_del_mc: Remove sysfs entries for specified mci structure and
812  *                 remove mci structure from global list
813  * @pdev: Pointer to 'struct device' representing mci structure to remove.
814  *
815  * Return pointer to removed mci structure, or NULL if device not found.
816  */
817 struct mem_ctl_info *edac_mc_del_mc(struct device *dev)
818 {
819         struct mem_ctl_info *mci;
820
821         edac_dbg(0, "\n");
822
823         mutex_lock(&mem_ctls_mutex);
824
825         /* find the requested mci struct in the global list */
826         mci = find_mci_by_dev(dev);
827         if (mci == NULL) {
828                 mutex_unlock(&mem_ctls_mutex);
829                 return NULL;
830         }
831
832         if (!del_mc_from_global_list(mci))
833                 edac_mc_owner = NULL;
834         mutex_unlock(&mem_ctls_mutex);
835
836         /* flush workq processes */
837         edac_mc_workq_teardown(mci);
838
839         /* marking MCI offline */
840         mci->op_state = OP_OFFLINE;
841
842         /* remove from sysfs */
843         edac_remove_sysfs_mci_device(mci);
844
845         edac_printk(KERN_INFO, EDAC_MC,
846                 "Removed device %d for %s %s: DEV %s\n", mci->mc_idx,
847                 mci->mod_name, mci->ctl_name, edac_dev_name(mci));
848
849         return mci;
850 }
851 EXPORT_SYMBOL_GPL(edac_mc_del_mc);
852
853 static void edac_mc_scrub_block(unsigned long page, unsigned long offset,
854                                 u32 size)
855 {
856         struct page *pg;
857         void *virt_addr;
858         unsigned long flags = 0;
859
860         edac_dbg(3, "\n");
861
862         /* ECC error page was not in our memory. Ignore it. */
863         if (!pfn_valid(page))
864                 return;
865
866         /* Find the actual page structure then map it and fix */
867         pg = pfn_to_page(page);
868
869         if (PageHighMem(pg))
870                 local_irq_save(flags);
871
872         virt_addr = kmap_atomic(pg);
873
874         /* Perform architecture specific atomic scrub operation */
875         atomic_scrub(virt_addr + offset, size);
876
877         /* Unmap and complete */
878         kunmap_atomic(virt_addr);
879
880         if (PageHighMem(pg))
881                 local_irq_restore(flags);
882 }
883
884 /* FIXME - should return -1 */
885 int edac_mc_find_csrow_by_page(struct mem_ctl_info *mci, unsigned long page)
886 {
887         struct csrow_info **csrows = mci->csrows;
888         int row, i, j, n;
889
890         edac_dbg(1, "MC%d: 0x%lx\n", mci->mc_idx, page);
891         row = -1;
892
893         for (i = 0; i < mci->nr_csrows; i++) {
894                 struct csrow_info *csrow = csrows[i];
895                 n = 0;
896                 for (j = 0; j < csrow->nr_channels; j++) {
897                         struct dimm_info *dimm = csrow->channels[j]->dimm;
898                         n += dimm->nr_pages;
899                 }
900                 if (n == 0)
901                         continue;
902
903                 edac_dbg(3, "MC%d: first(0x%lx) page(0x%lx) last(0x%lx) mask(0x%lx)\n",
904                          mci->mc_idx,
905                          csrow->first_page, page, csrow->last_page,
906                          csrow->page_mask);
907
908                 if ((page >= csrow->first_page) &&
909                     (page <= csrow->last_page) &&
910                     ((page & csrow->page_mask) ==
911                      (csrow->first_page & csrow->page_mask))) {
912                         row = i;
913                         break;
914                 }
915         }
916
917         if (row == -1)
918                 edac_mc_printk(mci, KERN_ERR,
919                         "could not look up page error address %lx\n",
920                         (unsigned long)page);
921
922         return row;
923 }
924 EXPORT_SYMBOL_GPL(edac_mc_find_csrow_by_page);
925
926 const char *edac_layer_name[] = {
927         [EDAC_MC_LAYER_BRANCH] = "branch",
928         [EDAC_MC_LAYER_CHANNEL] = "channel",
929         [EDAC_MC_LAYER_SLOT] = "slot",
930         [EDAC_MC_LAYER_CHIP_SELECT] = "csrow",
931         [EDAC_MC_LAYER_ALL_MEM] = "memory",
932 };
933 EXPORT_SYMBOL_GPL(edac_layer_name);
934
935 static void edac_inc_ce_error(struct mem_ctl_info *mci,
936                               bool enable_per_layer_report,
937                               const int pos[EDAC_MAX_LAYERS],
938                               const u16 count)
939 {
940         int i, index = 0;
941
942         mci->ce_mc += count;
943
944         if (!enable_per_layer_report) {
945                 mci->ce_noinfo_count += count;
946                 return;
947         }
948
949         for (i = 0; i < mci->n_layers; i++) {
950                 if (pos[i] < 0)
951                         break;
952                 index += pos[i];
953                 mci->ce_per_layer[i][index] += count;
954
955                 if (i < mci->n_layers - 1)
956                         index *= mci->layers[i + 1].size;
957         }
958 }
959
960 static void edac_inc_ue_error(struct mem_ctl_info *mci,
961                                     bool enable_per_layer_report,
962                                     const int pos[EDAC_MAX_LAYERS],
963                                     const u16 count)
964 {
965         int i, index = 0;
966
967         mci->ue_mc += count;
968
969         if (!enable_per_layer_report) {
970                 mci->ce_noinfo_count += count;
971                 return;
972         }
973
974         for (i = 0; i < mci->n_layers; i++) {
975                 if (pos[i] < 0)
976                         break;
977                 index += pos[i];
978                 mci->ue_per_layer[i][index] += count;
979
980                 if (i < mci->n_layers - 1)
981                         index *= mci->layers[i + 1].size;
982         }
983 }
984
985 static void edac_ce_error(struct mem_ctl_info *mci,
986                           const u16 error_count,
987                           const int pos[EDAC_MAX_LAYERS],
988                           const char *msg,
989                           const char *location,
990                           const char *label,
991                           const char *detail,
992                           const char *other_detail,
993                           const bool enable_per_layer_report,
994                           const unsigned long page_frame_number,
995                           const unsigned long offset_in_page,
996                           long grain)
997 {
998         unsigned long remapped_page;
999         char *msg_aux = "";
1000
1001         if (*msg)
1002                 msg_aux = " ";
1003
1004         if (edac_mc_get_log_ce()) {
1005                 if (other_detail && *other_detail)
1006                         edac_mc_printk(mci, KERN_WARNING,
1007                                        "%d CE %s%son %s (%s %s - %s)\n",
1008                                        error_count, msg, msg_aux, label,
1009                                        location, detail, other_detail);
1010                 else
1011                         edac_mc_printk(mci, KERN_WARNING,
1012                                        "%d CE %s%son %s (%s %s)\n",
1013                                        error_count, msg, msg_aux, label,
1014                                        location, detail);
1015         }
1016         edac_inc_ce_error(mci, enable_per_layer_report, pos, error_count);
1017
1018         if (mci->scrub_mode == SCRUB_SW_SRC) {
1019                 /*
1020                         * Some memory controllers (called MCs below) can remap
1021                         * memory so that it is still available at a different
1022                         * address when PCI devices map into memory.
1023                         * MC's that can't do this, lose the memory where PCI
1024                         * devices are mapped. This mapping is MC-dependent
1025                         * and so we call back into the MC driver for it to
1026                         * map the MC page to a physical (CPU) page which can
1027                         * then be mapped to a virtual page - which can then
1028                         * be scrubbed.
1029                         */
1030                 remapped_page = mci->ctl_page_to_phys ?
1031                         mci->ctl_page_to_phys(mci, page_frame_number) :
1032                         page_frame_number;
1033
1034                 edac_mc_scrub_block(remapped_page,
1035                                         offset_in_page, grain);
1036         }
1037 }
1038
1039 static void edac_ue_error(struct mem_ctl_info *mci,
1040                           const u16 error_count,
1041                           const int pos[EDAC_MAX_LAYERS],
1042                           const char *msg,
1043                           const char *location,
1044                           const char *label,
1045                           const char *detail,
1046                           const char *other_detail,
1047                           const bool enable_per_layer_report)
1048 {
1049         char *msg_aux = "";
1050
1051         if (*msg)
1052                 msg_aux = " ";
1053
1054         if (edac_mc_get_log_ue()) {
1055                 if (other_detail && *other_detail)
1056                         edac_mc_printk(mci, KERN_WARNING,
1057                                        "%d UE %s%son %s (%s %s - %s)\n",
1058                                        error_count, msg, msg_aux, label,
1059                                        location, detail, other_detail);
1060                 else
1061                         edac_mc_printk(mci, KERN_WARNING,
1062                                        "%d UE %s%son %s (%s %s)\n",
1063                                        error_count, msg, msg_aux, label,
1064                                        location, detail);
1065         }
1066
1067         if (edac_mc_get_panic_on_ue()) {
1068                 if (other_detail && *other_detail)
1069                         panic("UE %s%son %s (%s%s - %s)\n",
1070                               msg, msg_aux, label, location, detail, other_detail);
1071                 else
1072                         panic("UE %s%son %s (%s%s)\n",
1073                               msg, msg_aux, label, location, detail);
1074         }
1075
1076         edac_inc_ue_error(mci, enable_per_layer_report, pos, error_count);
1077 }
1078
1079 /**
1080  * edac_raw_mc_handle_error - reports a memory event to userspace without doing
1081  *                            anything to discover the error location
1082  *
1083  * @type:               severity of the error (CE/UE/Fatal)
1084  * @mci:                a struct mem_ctl_info pointer
1085  * @e:                  error description
1086  *
1087  * This raw function is used internally by edac_mc_handle_error(). It should
1088  * only be called directly when the hardware error come directly from BIOS,
1089  * like in the case of APEI GHES driver.
1090  */
1091 void edac_raw_mc_handle_error(const enum hw_event_mc_err_type type,
1092                               struct mem_ctl_info *mci,
1093                               struct edac_raw_error_desc *e)
1094 {
1095         char detail[80];
1096         int pos[EDAC_MAX_LAYERS] = { e->top_layer, e->mid_layer, e->low_layer };
1097
1098         /* Memory type dependent details about the error */
1099         if (type == HW_EVENT_ERR_CORRECTED) {
1100                 snprintf(detail, sizeof(detail),
1101                         "page:0x%lx offset:0x%lx grain:%ld syndrome:0x%lx",
1102                         e->page_frame_number, e->offset_in_page,
1103                         e->grain, e->syndrome);
1104                 edac_ce_error(mci, e->error_count, pos, e->msg, e->location, e->label,
1105                               detail, e->other_detail, e->enable_per_layer_report,
1106                               e->page_frame_number, e->offset_in_page, e->grain);
1107         } else {
1108                 snprintf(detail, sizeof(detail),
1109                         "page:0x%lx offset:0x%lx grain:%ld",
1110                         e->page_frame_number, e->offset_in_page, e->grain);
1111
1112                 edac_ue_error(mci, e->error_count, pos, e->msg, e->location, e->label,
1113                               detail, e->other_detail, e->enable_per_layer_report);
1114         }
1115
1116
1117 }
1118 EXPORT_SYMBOL_GPL(edac_raw_mc_handle_error);
1119
1120 /**
1121  * edac_mc_handle_error - reports a memory event to userspace
1122  *
1123  * @type:               severity of the error (CE/UE/Fatal)
1124  * @mci:                a struct mem_ctl_info pointer
1125  * @error_count:        Number of errors of the same type
1126  * @page_frame_number:  mem page where the error occurred
1127  * @offset_in_page:     offset of the error inside the page
1128  * @syndrome:           ECC syndrome
1129  * @top_layer:          Memory layer[0] position
1130  * @mid_layer:          Memory layer[1] position
1131  * @low_layer:          Memory layer[2] position
1132  * @msg:                Message meaningful to the end users that
1133  *                      explains the event
1134  * @other_detail:       Technical details about the event that
1135  *                      may help hardware manufacturers and
1136  *                      EDAC developers to analyse the event
1137  */
1138 void edac_mc_handle_error(const enum hw_event_mc_err_type type,
1139                           struct mem_ctl_info *mci,
1140                           const u16 error_count,
1141                           const unsigned long page_frame_number,
1142                           const unsigned long offset_in_page,
1143                           const unsigned long syndrome,
1144                           const int top_layer,
1145                           const int mid_layer,
1146                           const int low_layer,
1147                           const char *msg,
1148                           const char *other_detail)
1149 {
1150         char *p;
1151         int row = -1, chan = -1;
1152         int pos[EDAC_MAX_LAYERS] = { top_layer, mid_layer, low_layer };
1153         int i, n_labels = 0;
1154         u8 grain_bits;
1155         struct edac_raw_error_desc *e = &mci->error_desc;
1156
1157         edac_dbg(3, "MC%d\n", mci->mc_idx);
1158
1159         /* Fills the error report buffer */
1160         memset(e, 0, sizeof (*e));
1161         e->error_count = error_count;
1162         e->top_layer = top_layer;
1163         e->mid_layer = mid_layer;
1164         e->low_layer = low_layer;
1165         e->page_frame_number = page_frame_number;
1166         e->offset_in_page = offset_in_page;
1167         e->syndrome = syndrome;
1168         e->msg = msg;
1169         e->other_detail = other_detail;
1170
1171         /*
1172          * Check if the event report is consistent and if the memory
1173          * location is known. If it is known, enable_per_layer_report will be
1174          * true, the DIMM(s) label info will be filled and the per-layer
1175          * error counters will be incremented.
1176          */
1177         for (i = 0; i < mci->n_layers; i++) {
1178                 if (pos[i] >= (int)mci->layers[i].size) {
1179
1180                         edac_mc_printk(mci, KERN_ERR,
1181                                        "INTERNAL ERROR: %s value is out of range (%d >= %d)\n",
1182                                        edac_layer_name[mci->layers[i].type],
1183                                        pos[i], mci->layers[i].size);
1184                         /*
1185                          * Instead of just returning it, let's use what's
1186                          * known about the error. The increment routines and
1187                          * the DIMM filter logic will do the right thing by
1188                          * pointing the likely damaged DIMMs.
1189                          */
1190                         pos[i] = -1;
1191                 }
1192                 if (pos[i] >= 0)
1193                         e->enable_per_layer_report = true;
1194         }
1195
1196         /*
1197          * Get the dimm label/grain that applies to the match criteria.
1198          * As the error algorithm may not be able to point to just one memory
1199          * stick, the logic here will get all possible labels that could
1200          * pottentially be affected by the error.
1201          * On FB-DIMM memory controllers, for uncorrected errors, it is common
1202          * to have only the MC channel and the MC dimm (also called "branch")
1203          * but the channel is not known, as the memory is arranged in pairs,
1204          * where each memory belongs to a separate channel within the same
1205          * branch.
1206          */
1207         p = e->label;
1208         *p = '\0';
1209
1210         for (i = 0; i < mci->tot_dimms; i++) {
1211                 struct dimm_info *dimm = mci->dimms[i];
1212
1213                 if (top_layer >= 0 && top_layer != dimm->location[0])
1214                         continue;
1215                 if (mid_layer >= 0 && mid_layer != dimm->location[1])
1216                         continue;
1217                 if (low_layer >= 0 && low_layer != dimm->location[2])
1218                         continue;
1219
1220                 /* get the max grain, over the error match range */
1221                 if (dimm->grain > e->grain)
1222                         e->grain = dimm->grain;
1223
1224                 /*
1225                  * If the error is memory-controller wide, there's no need to
1226                  * seek for the affected DIMMs because the whole
1227                  * channel/memory controller/...  may be affected.
1228                  * Also, don't show errors for empty DIMM slots.
1229                  */
1230                 if (e->enable_per_layer_report && dimm->nr_pages) {
1231                         if (n_labels >= EDAC_MAX_LABELS) {
1232                                 e->enable_per_layer_report = false;
1233                                 break;
1234                         }
1235                         n_labels++;
1236                         if (p != e->label) {
1237                                 strcpy(p, OTHER_LABEL);
1238                                 p += strlen(OTHER_LABEL);
1239                         }
1240                         strcpy(p, dimm->label);
1241                         p += strlen(p);
1242                         *p = '\0';
1243
1244                         /*
1245                          * get csrow/channel of the DIMM, in order to allow
1246                          * incrementing the compat API counters
1247                          */
1248                         edac_dbg(4, "%s csrows map: (%d,%d)\n",
1249                                  mci->csbased ? "rank" : "dimm",
1250                                  dimm->csrow, dimm->cschannel);
1251                         if (row == -1)
1252                                 row = dimm->csrow;
1253                         else if (row >= 0 && row != dimm->csrow)
1254                                 row = -2;
1255
1256                         if (chan == -1)
1257                                 chan = dimm->cschannel;
1258                         else if (chan >= 0 && chan != dimm->cschannel)
1259                                 chan = -2;
1260                 }
1261         }
1262
1263         if (!e->enable_per_layer_report) {
1264                 strcpy(e->label, "any memory");
1265         } else {
1266                 edac_dbg(4, "csrow/channel to increment: (%d,%d)\n", row, chan);
1267                 if (p == e->label)
1268                         strcpy(e->label, "unknown memory");
1269                 if (type == HW_EVENT_ERR_CORRECTED) {
1270                         if (row >= 0) {
1271                                 mci->csrows[row]->ce_count += error_count;
1272                                 if (chan >= 0)
1273                                         mci->csrows[row]->channels[chan]->ce_count += error_count;
1274                         }
1275                 } else
1276                         if (row >= 0)
1277                                 mci->csrows[row]->ue_count += error_count;
1278         }
1279
1280         /* Fill the RAM location data */
1281         p = e->location;
1282
1283         for (i = 0; i < mci->n_layers; i++) {
1284                 if (pos[i] < 0)
1285                         continue;
1286
1287                 p += sprintf(p, "%s:%d ",
1288                              edac_layer_name[mci->layers[i].type],
1289                              pos[i]);
1290         }
1291         if (p > e->location)
1292                 *(p - 1) = '\0';
1293
1294         /* Report the error via the trace interface */
1295         grain_bits = fls_long(e->grain) + 1;
1296         trace_mc_event(type, e->msg, e->label, e->error_count,
1297                        mci->mc_idx, e->top_layer, e->mid_layer, e->low_layer,
1298                        PAGES_TO_MiB(e->page_frame_number) | e->offset_in_page,
1299                        grain_bits, e->syndrome, e->other_detail);
1300
1301         edac_raw_mc_handle_error(type, mci, e);
1302 }
1303 EXPORT_SYMBOL_GPL(edac_mc_handle_error);