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.
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/
11 * Modified by Dave Peterson and Doug Thompson
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>
33 #include "edac_core.h"
34 #include "edac_module.h"
35 #include <ras/ras_event.h>
37 #ifdef CONFIG_EDAC_ATOMIC_SCRUB
40 #define edac_atomic_scrub(va, size) do { } while (0)
43 /* lock to memory controller's control array */
44 static DEFINE_MUTEX(mem_ctls_mutex);
45 static LIST_HEAD(mc_devices);
48 * Used to lock EDAC MC to just one module, avoiding two drivers e. g.
49 * apei/ghes and i7core_edac to be used at the same time.
51 static void const *edac_mc_owner;
53 static struct bus_type mc_bus[EDAC_MAX_MCS];
55 unsigned edac_dimm_info_location(struct dimm_info *dimm, char *buf,
58 struct mem_ctl_info *mci = dimm->mci;
62 for (i = 0; i < mci->n_layers; i++) {
63 n = snprintf(p, len, "%s %d ",
64 edac_layer_name[mci->layers[i].type],
76 #ifdef CONFIG_EDAC_DEBUG
78 static void edac_mc_dump_channel(struct rank_info *chan)
80 edac_dbg(4, " channel->chan_idx = %d\n", chan->chan_idx);
81 edac_dbg(4, " channel = %p\n", chan);
82 edac_dbg(4, " channel->csrow = %p\n", chan->csrow);
83 edac_dbg(4, " channel->dimm = %p\n", chan->dimm);
86 static void edac_mc_dump_dimm(struct dimm_info *dimm, int number)
90 edac_dimm_info_location(dimm, location, sizeof(location));
92 edac_dbg(4, "%s%i: %smapped as virtual row %d, chan %d\n",
93 dimm->mci->csbased ? "rank" : "dimm",
94 number, location, dimm->csrow, dimm->cschannel);
95 edac_dbg(4, " dimm = %p\n", dimm);
96 edac_dbg(4, " dimm->label = '%s'\n", dimm->label);
97 edac_dbg(4, " dimm->nr_pages = 0x%x\n", dimm->nr_pages);
98 edac_dbg(4, " dimm->grain = %d\n", dimm->grain);
99 edac_dbg(4, " dimm->nr_pages = 0x%x\n", dimm->nr_pages);
102 static void edac_mc_dump_csrow(struct csrow_info *csrow)
104 edac_dbg(4, "csrow->csrow_idx = %d\n", csrow->csrow_idx);
105 edac_dbg(4, " csrow = %p\n", csrow);
106 edac_dbg(4, " csrow->first_page = 0x%lx\n", csrow->first_page);
107 edac_dbg(4, " csrow->last_page = 0x%lx\n", csrow->last_page);
108 edac_dbg(4, " csrow->page_mask = 0x%lx\n", csrow->page_mask);
109 edac_dbg(4, " csrow->nr_channels = %d\n", csrow->nr_channels);
110 edac_dbg(4, " csrow->channels = %p\n", csrow->channels);
111 edac_dbg(4, " csrow->mci = %p\n", csrow->mci);
114 static void edac_mc_dump_mci(struct mem_ctl_info *mci)
116 edac_dbg(3, "\tmci = %p\n", mci);
117 edac_dbg(3, "\tmci->mtype_cap = %lx\n", mci->mtype_cap);
118 edac_dbg(3, "\tmci->edac_ctl_cap = %lx\n", mci->edac_ctl_cap);
119 edac_dbg(3, "\tmci->edac_cap = %lx\n", mci->edac_cap);
120 edac_dbg(4, "\tmci->edac_check = %p\n", mci->edac_check);
121 edac_dbg(3, "\tmci->nr_csrows = %d, csrows = %p\n",
122 mci->nr_csrows, mci->csrows);
123 edac_dbg(3, "\tmci->nr_dimms = %d, dimms = %p\n",
124 mci->tot_dimms, mci->dimms);
125 edac_dbg(3, "\tdev = %p\n", mci->pdev);
126 edac_dbg(3, "\tmod_name:ctl_name = %s:%s\n",
127 mci->mod_name, mci->ctl_name);
128 edac_dbg(3, "\tpvt_info = %p\n\n", mci->pvt_info);
131 #endif /* CONFIG_EDAC_DEBUG */
133 const char * const edac_mem_types[] = {
134 [MEM_EMPTY] = "Empty csrow",
135 [MEM_RESERVED] = "Reserved csrow type",
136 [MEM_UNKNOWN] = "Unknown csrow type",
137 [MEM_FPM] = "Fast page mode RAM",
138 [MEM_EDO] = "Extended data out RAM",
139 [MEM_BEDO] = "Burst Extended data out RAM",
140 [MEM_SDR] = "Single data rate SDRAM",
141 [MEM_RDR] = "Registered single data rate SDRAM",
142 [MEM_DDR] = "Double data rate SDRAM",
143 [MEM_RDDR] = "Registered Double data rate SDRAM",
144 [MEM_RMBS] = "Rambus DRAM",
145 [MEM_DDR2] = "Unbuffered DDR2 RAM",
146 [MEM_FB_DDR2] = "Fully buffered DDR2",
147 [MEM_RDDR2] = "Registered DDR2 RAM",
148 [MEM_XDR] = "Rambus XDR",
149 [MEM_DDR3] = "Unbuffered DDR3 RAM",
150 [MEM_RDDR3] = "Registered DDR3 RAM",
151 [MEM_LRDDR3] = "Load-Reduced DDR3 RAM",
152 [MEM_DDR4] = "Unbuffered DDR4 RAM",
153 [MEM_RDDR4] = "Registered DDR4 RAM",
155 EXPORT_SYMBOL_GPL(edac_mem_types);
158 * edac_align_ptr - Prepares the pointer offsets for a single-shot allocation
159 * @p: pointer to a pointer with the memory offset to be used. At
160 * return, this will be incremented to point to the next offset
161 * @size: Size of the data structure to be reserved
162 * @n_elems: Number of elements that should be reserved
164 * If 'size' is a constant, the compiler will optimize this whole function
165 * down to either a no-op or the addition of a constant to the value of '*p'.
167 * The 'p' pointer is absolutely needed to keep the proper advancing
168 * further in memory to the proper offsets when allocating the struct along
169 * with its embedded structs, as edac_device_alloc_ctl_info() does it
170 * above, for example.
172 * At return, the pointer 'p' will be incremented to be used on a next call
175 void *edac_align_ptr(void **p, unsigned size, int n_elems)
180 *p += size * n_elems;
183 * 'p' can possibly be an unaligned item X such that sizeof(X) is
184 * 'size'. Adjust 'p' so that its alignment is at least as
185 * stringent as what the compiler would provide for X and return
186 * the aligned result.
187 * Here we assume that the alignment of a "long long" is the most
188 * stringent alignment that the compiler will ever provide by default.
189 * As far as I know, this is a reasonable assumption.
191 if (size > sizeof(long))
192 align = sizeof(long long);
193 else if (size > sizeof(int))
194 align = sizeof(long);
195 else if (size > sizeof(short))
197 else if (size > sizeof(char))
198 align = sizeof(short);
202 r = (unsigned long)p % align;
209 return (void *)(((unsigned long)ptr) + align - r);
212 static void _edac_mc_free(struct mem_ctl_info *mci)
215 struct csrow_info *csr;
216 const unsigned int tot_dimms = mci->tot_dimms;
217 const unsigned int tot_channels = mci->num_cschannel;
218 const unsigned int tot_csrows = mci->nr_csrows;
221 for (i = 0; i < tot_dimms; i++)
222 kfree(mci->dimms[i]);
226 for (row = 0; row < tot_csrows; row++) {
227 csr = mci->csrows[row];
230 for (chn = 0; chn < tot_channels; chn++)
231 kfree(csr->channels[chn]);
232 kfree(csr->channels);
243 * edac_mc_alloc: Allocate and partially fill a struct mem_ctl_info structure
244 * @mc_num: Memory controller number
245 * @n_layers: Number of MC hierarchy layers
246 * layers: Describes each layer as seen by the Memory Controller
247 * @size_pvt: size of private storage needed
250 * Everything is kmalloc'ed as one big chunk - more efficient.
251 * Only can be used if all structures have the same lifetime - otherwise
252 * you have to allocate and initialize your own structures.
254 * Use edac_mc_free() to free mc structures allocated by this function.
256 * NOTE: drivers handle multi-rank memories in different ways: in some
257 * drivers, one multi-rank memory stick is mapped as one entry, while, in
258 * others, a single multi-rank memory stick would be mapped into several
259 * entries. Currently, this function will allocate multiple struct dimm_info
260 * on such scenarios, as grouping the multiple ranks require drivers change.
264 * On success: struct mem_ctl_info pointer
266 struct mem_ctl_info *edac_mc_alloc(unsigned mc_num,
268 struct edac_mc_layer *layers,
271 struct mem_ctl_info *mci;
272 struct edac_mc_layer *layer;
273 struct csrow_info *csr;
274 struct rank_info *chan;
275 struct dimm_info *dimm;
276 u32 *ce_per_layer[EDAC_MAX_LAYERS], *ue_per_layer[EDAC_MAX_LAYERS];
277 unsigned pos[EDAC_MAX_LAYERS];
278 unsigned size, tot_dimms = 1, count = 1;
279 unsigned tot_csrows = 1, tot_channels = 1, tot_errcount = 0;
280 void *pvt, *p, *ptr = NULL;
281 int i, j, row, chn, n, len, off;
282 bool per_rank = false;
284 BUG_ON(n_layers > EDAC_MAX_LAYERS || n_layers == 0);
286 * Calculate the total amount of dimms and csrows/cschannels while
287 * in the old API emulation mode
289 for (i = 0; i < n_layers; i++) {
290 tot_dimms *= layers[i].size;
291 if (layers[i].is_virt_csrow)
292 tot_csrows *= layers[i].size;
294 tot_channels *= layers[i].size;
296 if (layers[i].type == EDAC_MC_LAYER_CHIP_SELECT)
300 /* Figure out the offsets of the various items from the start of an mc
301 * structure. We want the alignment of each item to be at least as
302 * stringent as what the compiler would provide if we could simply
303 * hardcode everything into a single struct.
305 mci = edac_align_ptr(&ptr, sizeof(*mci), 1);
306 layer = edac_align_ptr(&ptr, sizeof(*layer), n_layers);
307 for (i = 0; i < n_layers; i++) {
308 count *= layers[i].size;
309 edac_dbg(4, "errcount layer %d size %d\n", i, count);
310 ce_per_layer[i] = edac_align_ptr(&ptr, sizeof(u32), count);
311 ue_per_layer[i] = edac_align_ptr(&ptr, sizeof(u32), count);
312 tot_errcount += 2 * count;
315 edac_dbg(4, "allocating %d error counters\n", tot_errcount);
316 pvt = edac_align_ptr(&ptr, sz_pvt, 1);
317 size = ((unsigned long)pvt) + sz_pvt;
319 edac_dbg(1, "allocating %u bytes for mci data (%d %s, %d csrows/channels)\n",
322 per_rank ? "ranks" : "dimms",
323 tot_csrows * tot_channels);
325 mci = kzalloc(size, GFP_KERNEL);
329 /* Adjust pointers so they point within the memory we just allocated
330 * rather than an imaginary chunk of memory located at address 0.
332 layer = (struct edac_mc_layer *)(((char *)mci) + ((unsigned long)layer));
333 for (i = 0; i < n_layers; i++) {
334 mci->ce_per_layer[i] = (u32 *)((char *)mci + ((unsigned long)ce_per_layer[i]));
335 mci->ue_per_layer[i] = (u32 *)((char *)mci + ((unsigned long)ue_per_layer[i]));
337 pvt = sz_pvt ? (((char *)mci) + ((unsigned long)pvt)) : NULL;
339 /* setup index and various internal pointers */
340 mci->mc_idx = mc_num;
341 mci->tot_dimms = tot_dimms;
343 mci->n_layers = n_layers;
345 memcpy(mci->layers, layers, sizeof(*layer) * n_layers);
346 mci->nr_csrows = tot_csrows;
347 mci->num_cschannel = tot_channels;
348 mci->csbased = per_rank;
351 * Alocate and fill the csrow/channels structs
353 mci->csrows = kcalloc(tot_csrows, sizeof(*mci->csrows), GFP_KERNEL);
356 for (row = 0; row < tot_csrows; row++) {
357 csr = kzalloc(sizeof(**mci->csrows), GFP_KERNEL);
360 mci->csrows[row] = csr;
361 csr->csrow_idx = row;
363 csr->nr_channels = tot_channels;
364 csr->channels = kcalloc(tot_channels, sizeof(*csr->channels),
369 for (chn = 0; chn < tot_channels; chn++) {
370 chan = kzalloc(sizeof(**csr->channels), GFP_KERNEL);
373 csr->channels[chn] = chan;
374 chan->chan_idx = chn;
380 * Allocate and fill the dimm structs
382 mci->dimms = kcalloc(tot_dimms, sizeof(*mci->dimms), GFP_KERNEL);
386 memset(&pos, 0, sizeof(pos));
389 for (i = 0; i < tot_dimms; i++) {
390 chan = mci->csrows[row]->channels[chn];
391 off = EDAC_DIMM_OFF(layer, n_layers, pos[0], pos[1], pos[2]);
392 if (off < 0 || off >= tot_dimms) {
393 edac_mc_printk(mci, KERN_ERR, "EDAC core bug: EDAC_DIMM_OFF is trying to do an illegal data access\n");
397 dimm = kzalloc(sizeof(**mci->dimms), GFP_KERNEL);
400 mci->dimms[off] = dimm;
404 * Copy DIMM location and initialize it.
406 len = sizeof(dimm->label);
408 n = snprintf(p, len, "mc#%u", mc_num);
411 for (j = 0; j < n_layers; j++) {
412 n = snprintf(p, len, "%s#%u",
413 edac_layer_name[layers[j].type],
417 dimm->location[j] = pos[j];
423 /* Link it to the csrows old API data */
426 dimm->cschannel = chn;
428 /* Increment csrow location */
429 if (layers[0].is_virt_csrow) {
431 if (chn == tot_channels) {
437 if (row == tot_csrows) {
443 /* Increment dimm location */
444 for (j = n_layers - 1; j >= 0; j--) {
446 if (pos[j] < layers[j].size)
452 mci->op_state = OP_ALLOC;
461 EXPORT_SYMBOL_GPL(edac_mc_alloc);
465 * 'Free' a previously allocated 'mci' structure
466 * @mci: pointer to a struct mem_ctl_info structure
468 void edac_mc_free(struct mem_ctl_info *mci)
472 /* If we're not yet registered with sysfs free only what was allocated
473 * in edac_mc_alloc().
475 if (!device_is_registered(&mci->dev)) {
480 /* the mci instance is freed here, when the sysfs object is dropped */
481 edac_unregister_sysfs(mci);
483 EXPORT_SYMBOL_GPL(edac_mc_free);
489 * scan list of controllers looking for the one that manages
491 * @dev: pointer to a struct device related with the MCI
493 struct mem_ctl_info *find_mci_by_dev(struct device *dev)
495 struct mem_ctl_info *mci;
496 struct list_head *item;
500 list_for_each(item, &mc_devices) {
501 mci = list_entry(item, struct mem_ctl_info, link);
503 if (mci->pdev == dev)
509 EXPORT_SYMBOL_GPL(find_mci_by_dev);
512 * handler for EDAC to check if NMI type handler has asserted interrupt
514 static int edac_mc_assert_error_check_and_clear(void)
518 if (edac_op_state == EDAC_OPSTATE_POLL)
521 old_state = edac_err_assert;
528 * edac_mc_workq_function
529 * performs the operation scheduled by a workq request
531 static void edac_mc_workq_function(struct work_struct *work_req)
533 struct delayed_work *d_work = to_delayed_work(work_req);
534 struct mem_ctl_info *mci = to_edac_mem_ctl_work(d_work);
536 mutex_lock(&mem_ctls_mutex);
538 /* if this control struct has movd to offline state, we are done */
539 if (mci->op_state == OP_OFFLINE) {
540 mutex_unlock(&mem_ctls_mutex);
544 /* Only poll controllers that are running polled and have a check */
545 if (edac_mc_assert_error_check_and_clear() && (mci->edac_check != NULL))
546 mci->edac_check(mci);
548 mutex_unlock(&mem_ctls_mutex);
551 queue_delayed_work(edac_workqueue, &mci->work,
552 msecs_to_jiffies(edac_mc_get_poll_msec()));
556 * edac_mc_workq_setup
557 * initialize a workq item for this mci
558 * passing in the new delay period in msec
562 * called with the mem_ctls_mutex held
564 static void edac_mc_workq_setup(struct mem_ctl_info *mci, unsigned msec,
569 /* if this instance is not in the POLL state, then simply return */
570 if (mci->op_state != OP_RUNNING_POLL)
574 INIT_DELAYED_WORK(&mci->work, edac_mc_workq_function);
576 mod_delayed_work(edac_workqueue, &mci->work, msecs_to_jiffies(msec));
580 * edac_mc_workq_teardown
581 * stop the workq processing on this mci
585 * called WITHOUT lock held
587 static void edac_mc_workq_teardown(struct mem_ctl_info *mci)
589 mci->op_state = OP_OFFLINE;
591 cancel_delayed_work_sync(&mci->work);
592 flush_workqueue(edac_workqueue);
596 * edac_mc_reset_delay_period(unsigned long value)
598 * user space has updated our poll period value, need to
599 * reset our workq delays
601 void edac_mc_reset_delay_period(unsigned long value)
603 struct mem_ctl_info *mci;
604 struct list_head *item;
606 mutex_lock(&mem_ctls_mutex);
608 list_for_each(item, &mc_devices) {
609 mci = list_entry(item, struct mem_ctl_info, link);
611 edac_mc_workq_setup(mci, value, false);
614 mutex_unlock(&mem_ctls_mutex);
619 /* Return 0 on success, 1 on failure.
620 * Before calling this function, caller must
621 * assign a unique value to mci->mc_idx.
625 * called with the mem_ctls_mutex lock held
627 static int add_mc_to_global_list(struct mem_ctl_info *mci)
629 struct list_head *item, *insert_before;
630 struct mem_ctl_info *p;
632 insert_before = &mc_devices;
634 p = find_mci_by_dev(mci->pdev);
635 if (unlikely(p != NULL))
638 list_for_each(item, &mc_devices) {
639 p = list_entry(item, struct mem_ctl_info, link);
641 if (p->mc_idx >= mci->mc_idx) {
642 if (unlikely(p->mc_idx == mci->mc_idx))
645 insert_before = item;
650 list_add_tail_rcu(&mci->link, insert_before);
651 atomic_inc(&edac_handlers);
655 edac_printk(KERN_WARNING, EDAC_MC,
656 "%s (%s) %s %s already assigned %d\n", dev_name(p->pdev),
657 edac_dev_name(mci), p->mod_name, p->ctl_name, p->mc_idx);
661 edac_printk(KERN_WARNING, EDAC_MC,
662 "bug in low-level driver: attempt to assign\n"
663 " duplicate mc_idx %d in %s()\n", p->mc_idx, __func__);
667 static int del_mc_from_global_list(struct mem_ctl_info *mci)
669 int handlers = atomic_dec_return(&edac_handlers);
670 list_del_rcu(&mci->link);
672 /* these are for safe removal of devices from global list while
673 * NMI handlers may be traversing list
676 INIT_LIST_HEAD(&mci->link);
682 * edac_mc_find: Search for a mem_ctl_info structure whose index is 'idx'.
684 * If found, return a pointer to the structure.
687 * Caller must hold mem_ctls_mutex.
689 struct mem_ctl_info *edac_mc_find(int idx)
691 struct list_head *item;
692 struct mem_ctl_info *mci;
694 list_for_each(item, &mc_devices) {
695 mci = list_entry(item, struct mem_ctl_info, link);
697 if (mci->mc_idx >= idx) {
698 if (mci->mc_idx == idx)
707 EXPORT_SYMBOL(edac_mc_find);
710 * edac_mc_add_mc_with_groups: Insert the 'mci' structure into the mci
711 * global list and create sysfs entries associated with mci structure
712 * @mci: pointer to the mci structure to be added to the list
713 * @groups: optional attribute groups for the driver-specific sysfs entries
720 /* FIXME - should a warning be printed if no error detection? correction? */
721 int edac_mc_add_mc_with_groups(struct mem_ctl_info *mci,
722 const struct attribute_group **groups)
727 if (mci->mc_idx >= EDAC_MAX_MCS) {
728 pr_warn_once("Too many memory controllers: %d\n", mci->mc_idx);
732 #ifdef CONFIG_EDAC_DEBUG
733 if (edac_debug_level >= 3)
734 edac_mc_dump_mci(mci);
736 if (edac_debug_level >= 4) {
739 for (i = 0; i < mci->nr_csrows; i++) {
740 struct csrow_info *csrow = mci->csrows[i];
744 for (j = 0; j < csrow->nr_channels; j++)
745 nr_pages += csrow->channels[j]->dimm->nr_pages;
748 edac_mc_dump_csrow(csrow);
749 for (j = 0; j < csrow->nr_channels; j++)
750 if (csrow->channels[j]->dimm->nr_pages)
751 edac_mc_dump_channel(csrow->channels[j]);
753 for (i = 0; i < mci->tot_dimms; i++)
754 if (mci->dimms[i]->nr_pages)
755 edac_mc_dump_dimm(mci->dimms[i], i);
758 mutex_lock(&mem_ctls_mutex);
760 if (edac_mc_owner && edac_mc_owner != mci->mod_name) {
765 if (add_mc_to_global_list(mci))
768 /* set load time so that error rate can be tracked */
769 mci->start_time = jiffies;
771 mci->bus = &mc_bus[mci->mc_idx];
773 if (edac_create_sysfs_mci_device(mci, groups)) {
774 edac_mc_printk(mci, KERN_WARNING,
775 "failed to create sysfs device\n");
779 /* If there IS a check routine, then we are running POLLED */
780 if (mci->edac_check != NULL) {
781 /* This instance is NOW RUNNING */
782 mci->op_state = OP_RUNNING_POLL;
784 edac_mc_workq_setup(mci, edac_mc_get_poll_msec(), true);
786 mci->op_state = OP_RUNNING_INTERRUPT;
789 /* Report action taken */
790 edac_mc_printk(mci, KERN_INFO,
791 "Giving out device to module %s controller %s: DEV %s (%s)\n",
792 mci->mod_name, mci->ctl_name, mci->dev_name,
793 edac_op_state_to_string(mci->op_state));
795 edac_mc_owner = mci->mod_name;
797 mutex_unlock(&mem_ctls_mutex);
801 del_mc_from_global_list(mci);
804 mutex_unlock(&mem_ctls_mutex);
807 EXPORT_SYMBOL_GPL(edac_mc_add_mc_with_groups);
810 * edac_mc_del_mc: Remove sysfs entries for specified mci structure and
811 * remove mci structure from global list
812 * @pdev: Pointer to 'struct device' representing mci structure to remove.
814 * Return pointer to removed mci structure, or NULL if device not found.
816 struct mem_ctl_info *edac_mc_del_mc(struct device *dev)
818 struct mem_ctl_info *mci;
822 mutex_lock(&mem_ctls_mutex);
824 /* find the requested mci struct in the global list */
825 mci = find_mci_by_dev(dev);
827 mutex_unlock(&mem_ctls_mutex);
831 if (!del_mc_from_global_list(mci))
832 edac_mc_owner = NULL;
833 mutex_unlock(&mem_ctls_mutex);
835 /* flush workq processes */
836 edac_mc_workq_teardown(mci);
838 /* marking MCI offline */
839 mci->op_state = OP_OFFLINE;
841 /* remove from sysfs */
842 edac_remove_sysfs_mci_device(mci);
844 edac_printk(KERN_INFO, EDAC_MC,
845 "Removed device %d for %s %s: DEV %s\n", mci->mc_idx,
846 mci->mod_name, mci->ctl_name, edac_dev_name(mci));
850 EXPORT_SYMBOL_GPL(edac_mc_del_mc);
852 static void edac_mc_scrub_block(unsigned long page, unsigned long offset,
857 unsigned long flags = 0;
861 /* ECC error page was not in our memory. Ignore it. */
862 if (!pfn_valid(page))
865 /* Find the actual page structure then map it and fix */
866 pg = pfn_to_page(page);
869 local_irq_save(flags);
871 virt_addr = kmap_atomic(pg);
873 /* Perform architecture specific atomic scrub operation */
874 edac_atomic_scrub(virt_addr + offset, size);
876 /* Unmap and complete */
877 kunmap_atomic(virt_addr);
880 local_irq_restore(flags);
883 /* FIXME - should return -1 */
884 int edac_mc_find_csrow_by_page(struct mem_ctl_info *mci, unsigned long page)
886 struct csrow_info **csrows = mci->csrows;
889 edac_dbg(1, "MC%d: 0x%lx\n", mci->mc_idx, page);
892 for (i = 0; i < mci->nr_csrows; i++) {
893 struct csrow_info *csrow = csrows[i];
895 for (j = 0; j < csrow->nr_channels; j++) {
896 struct dimm_info *dimm = csrow->channels[j]->dimm;
902 edac_dbg(3, "MC%d: first(0x%lx) page(0x%lx) last(0x%lx) mask(0x%lx)\n",
904 csrow->first_page, page, csrow->last_page,
907 if ((page >= csrow->first_page) &&
908 (page <= csrow->last_page) &&
909 ((page & csrow->page_mask) ==
910 (csrow->first_page & csrow->page_mask))) {
917 edac_mc_printk(mci, KERN_ERR,
918 "could not look up page error address %lx\n",
919 (unsigned long)page);
923 EXPORT_SYMBOL_GPL(edac_mc_find_csrow_by_page);
925 const char *edac_layer_name[] = {
926 [EDAC_MC_LAYER_BRANCH] = "branch",
927 [EDAC_MC_LAYER_CHANNEL] = "channel",
928 [EDAC_MC_LAYER_SLOT] = "slot",
929 [EDAC_MC_LAYER_CHIP_SELECT] = "csrow",
930 [EDAC_MC_LAYER_ALL_MEM] = "memory",
932 EXPORT_SYMBOL_GPL(edac_layer_name);
934 static void edac_inc_ce_error(struct mem_ctl_info *mci,
935 bool enable_per_layer_report,
936 const int pos[EDAC_MAX_LAYERS],
943 if (!enable_per_layer_report) {
944 mci->ce_noinfo_count += count;
948 for (i = 0; i < mci->n_layers; i++) {
952 mci->ce_per_layer[i][index] += count;
954 if (i < mci->n_layers - 1)
955 index *= mci->layers[i + 1].size;
959 static void edac_inc_ue_error(struct mem_ctl_info *mci,
960 bool enable_per_layer_report,
961 const int pos[EDAC_MAX_LAYERS],
968 if (!enable_per_layer_report) {
969 mci->ue_noinfo_count += count;
973 for (i = 0; i < mci->n_layers; i++) {
977 mci->ue_per_layer[i][index] += count;
979 if (i < mci->n_layers - 1)
980 index *= mci->layers[i + 1].size;
984 static void edac_ce_error(struct mem_ctl_info *mci,
985 const u16 error_count,
986 const int pos[EDAC_MAX_LAYERS],
988 const char *location,
991 const char *other_detail,
992 const bool enable_per_layer_report,
993 const unsigned long page_frame_number,
994 const unsigned long offset_in_page,
997 unsigned long remapped_page;
1003 if (edac_mc_get_log_ce()) {
1004 if (other_detail && *other_detail)
1005 edac_mc_printk(mci, KERN_WARNING,
1006 "%d CE %s%son %s (%s %s - %s)\n",
1007 error_count, msg, msg_aux, label,
1008 location, detail, other_detail);
1010 edac_mc_printk(mci, KERN_WARNING,
1011 "%d CE %s%son %s (%s %s)\n",
1012 error_count, msg, msg_aux, label,
1015 edac_inc_ce_error(mci, enable_per_layer_report, pos, error_count);
1017 if (mci->scrub_mode == SCRUB_SW_SRC) {
1019 * Some memory controllers (called MCs below) can remap
1020 * memory so that it is still available at a different
1021 * address when PCI devices map into memory.
1022 * MC's that can't do this, lose the memory where PCI
1023 * devices are mapped. This mapping is MC-dependent
1024 * and so we call back into the MC driver for it to
1025 * map the MC page to a physical (CPU) page which can
1026 * then be mapped to a virtual page - which can then
1029 remapped_page = mci->ctl_page_to_phys ?
1030 mci->ctl_page_to_phys(mci, page_frame_number) :
1033 edac_mc_scrub_block(remapped_page,
1034 offset_in_page, grain);
1038 static void edac_ue_error(struct mem_ctl_info *mci,
1039 const u16 error_count,
1040 const int pos[EDAC_MAX_LAYERS],
1042 const char *location,
1045 const char *other_detail,
1046 const bool enable_per_layer_report)
1053 if (edac_mc_get_log_ue()) {
1054 if (other_detail && *other_detail)
1055 edac_mc_printk(mci, KERN_WARNING,
1056 "%d UE %s%son %s (%s %s - %s)\n",
1057 error_count, msg, msg_aux, label,
1058 location, detail, other_detail);
1060 edac_mc_printk(mci, KERN_WARNING,
1061 "%d UE %s%son %s (%s %s)\n",
1062 error_count, msg, msg_aux, label,
1066 if (edac_mc_get_panic_on_ue()) {
1067 if (other_detail && *other_detail)
1068 panic("UE %s%son %s (%s%s - %s)\n",
1069 msg, msg_aux, label, location, detail, other_detail);
1071 panic("UE %s%son %s (%s%s)\n",
1072 msg, msg_aux, label, location, detail);
1075 edac_inc_ue_error(mci, enable_per_layer_report, pos, error_count);
1079 * edac_raw_mc_handle_error - reports a memory event to userspace without doing
1080 * anything to discover the error location
1082 * @type: severity of the error (CE/UE/Fatal)
1083 * @mci: a struct mem_ctl_info pointer
1084 * @e: error description
1086 * This raw function is used internally by edac_mc_handle_error(). It should
1087 * only be called directly when the hardware error come directly from BIOS,
1088 * like in the case of APEI GHES driver.
1090 void edac_raw_mc_handle_error(const enum hw_event_mc_err_type type,
1091 struct mem_ctl_info *mci,
1092 struct edac_raw_error_desc *e)
1095 int pos[EDAC_MAX_LAYERS] = { e->top_layer, e->mid_layer, e->low_layer };
1097 /* Memory type dependent details about the error */
1098 if (type == HW_EVENT_ERR_CORRECTED) {
1099 snprintf(detail, sizeof(detail),
1100 "page:0x%lx offset:0x%lx grain:%ld syndrome:0x%lx",
1101 e->page_frame_number, e->offset_in_page,
1102 e->grain, e->syndrome);
1103 edac_ce_error(mci, e->error_count, pos, e->msg, e->location, e->label,
1104 detail, e->other_detail, e->enable_per_layer_report,
1105 e->page_frame_number, e->offset_in_page, e->grain);
1107 snprintf(detail, sizeof(detail),
1108 "page:0x%lx offset:0x%lx grain:%ld",
1109 e->page_frame_number, e->offset_in_page, e->grain);
1111 edac_ue_error(mci, e->error_count, pos, e->msg, e->location, e->label,
1112 detail, e->other_detail, e->enable_per_layer_report);
1117 EXPORT_SYMBOL_GPL(edac_raw_mc_handle_error);
1120 * edac_mc_handle_error - reports a memory event to userspace
1122 * @type: severity of the error (CE/UE/Fatal)
1123 * @mci: a struct mem_ctl_info pointer
1124 * @error_count: Number of errors of the same type
1125 * @page_frame_number: mem page where the error occurred
1126 * @offset_in_page: offset of the error inside the page
1127 * @syndrome: ECC syndrome
1128 * @top_layer: Memory layer[0] position
1129 * @mid_layer: Memory layer[1] position
1130 * @low_layer: Memory layer[2] position
1131 * @msg: Message meaningful to the end users that
1132 * explains the event
1133 * @other_detail: Technical details about the event that
1134 * may help hardware manufacturers and
1135 * EDAC developers to analyse the event
1137 void edac_mc_handle_error(const enum hw_event_mc_err_type type,
1138 struct mem_ctl_info *mci,
1139 const u16 error_count,
1140 const unsigned long page_frame_number,
1141 const unsigned long offset_in_page,
1142 const unsigned long syndrome,
1143 const int top_layer,
1144 const int mid_layer,
1145 const int low_layer,
1147 const char *other_detail)
1150 int row = -1, chan = -1;
1151 int pos[EDAC_MAX_LAYERS] = { top_layer, mid_layer, low_layer };
1152 int i, n_labels = 0;
1154 struct edac_raw_error_desc *e = &mci->error_desc;
1156 edac_dbg(3, "MC%d\n", mci->mc_idx);
1158 /* Fills the error report buffer */
1159 memset(e, 0, sizeof (*e));
1160 e->error_count = error_count;
1161 e->top_layer = top_layer;
1162 e->mid_layer = mid_layer;
1163 e->low_layer = low_layer;
1164 e->page_frame_number = page_frame_number;
1165 e->offset_in_page = offset_in_page;
1166 e->syndrome = syndrome;
1168 e->other_detail = other_detail;
1171 * Check if the event report is consistent and if the memory
1172 * location is known. If it is known, enable_per_layer_report will be
1173 * true, the DIMM(s) label info will be filled and the per-layer
1174 * error counters will be incremented.
1176 for (i = 0; i < mci->n_layers; i++) {
1177 if (pos[i] >= (int)mci->layers[i].size) {
1179 edac_mc_printk(mci, KERN_ERR,
1180 "INTERNAL ERROR: %s value is out of range (%d >= %d)\n",
1181 edac_layer_name[mci->layers[i].type],
1182 pos[i], mci->layers[i].size);
1184 * Instead of just returning it, let's use what's
1185 * known about the error. The increment routines and
1186 * the DIMM filter logic will do the right thing by
1187 * pointing the likely damaged DIMMs.
1192 e->enable_per_layer_report = true;
1196 * Get the dimm label/grain that applies to the match criteria.
1197 * As the error algorithm may not be able to point to just one memory
1198 * stick, the logic here will get all possible labels that could
1199 * pottentially be affected by the error.
1200 * On FB-DIMM memory controllers, for uncorrected errors, it is common
1201 * to have only the MC channel and the MC dimm (also called "branch")
1202 * but the channel is not known, as the memory is arranged in pairs,
1203 * where each memory belongs to a separate channel within the same
1209 for (i = 0; i < mci->tot_dimms; i++) {
1210 struct dimm_info *dimm = mci->dimms[i];
1212 if (top_layer >= 0 && top_layer != dimm->location[0])
1214 if (mid_layer >= 0 && mid_layer != dimm->location[1])
1216 if (low_layer >= 0 && low_layer != dimm->location[2])
1219 /* get the max grain, over the error match range */
1220 if (dimm->grain > e->grain)
1221 e->grain = dimm->grain;
1224 * If the error is memory-controller wide, there's no need to
1225 * seek for the affected DIMMs because the whole
1226 * channel/memory controller/... may be affected.
1227 * Also, don't show errors for empty DIMM slots.
1229 if (e->enable_per_layer_report && dimm->nr_pages) {
1230 if (n_labels >= EDAC_MAX_LABELS) {
1231 e->enable_per_layer_report = false;
1235 if (p != e->label) {
1236 strcpy(p, OTHER_LABEL);
1237 p += strlen(OTHER_LABEL);
1239 strcpy(p, dimm->label);
1244 * get csrow/channel of the DIMM, in order to allow
1245 * incrementing the compat API counters
1247 edac_dbg(4, "%s csrows map: (%d,%d)\n",
1248 mci->csbased ? "rank" : "dimm",
1249 dimm->csrow, dimm->cschannel);
1252 else if (row >= 0 && row != dimm->csrow)
1256 chan = dimm->cschannel;
1257 else if (chan >= 0 && chan != dimm->cschannel)
1262 if (!e->enable_per_layer_report) {
1263 strcpy(e->label, "any memory");
1265 edac_dbg(4, "csrow/channel to increment: (%d,%d)\n", row, chan);
1267 strcpy(e->label, "unknown memory");
1268 if (type == HW_EVENT_ERR_CORRECTED) {
1270 mci->csrows[row]->ce_count += error_count;
1272 mci->csrows[row]->channels[chan]->ce_count += error_count;
1276 mci->csrows[row]->ue_count += error_count;
1279 /* Fill the RAM location data */
1282 for (i = 0; i < mci->n_layers; i++) {
1286 p += sprintf(p, "%s:%d ",
1287 edac_layer_name[mci->layers[i].type],
1290 if (p > e->location)
1293 /* Report the error via the trace interface */
1294 grain_bits = fls_long(e->grain) + 1;
1295 trace_mc_event(type, e->msg, e->label, e->error_count,
1296 mci->mc_idx, e->top_layer, e->mid_layer, e->low_layer,
1297 (e->page_frame_number << PAGE_SHIFT) | e->offset_in_page,
1298 grain_bits, e->syndrome, e->other_detail);
1300 edac_raw_mc_handle_error(type, mci, e);
1302 EXPORT_SYMBOL_GPL(edac_mc_handle_error);