1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2017-2019 Borislav Petkov, SUSE Labs.
8 #include <linux/kernel.h>
9 #include <linux/workqueue.h>
16 * RAS Correctable Errors Collector
18 * This is a simple gadget which collects correctable errors and counts their
19 * occurrence per physical page address.
21 * We've opted for possibly the simplest data structure to collect those - an
22 * array of the size of a memory page. It stores 512 u64's with the following
25 * [63 ... PFN ... 12 | 11 ... generation ... 10 | 9 ... count ... 0]
27 * The generation in the two highest order bits is two bits which are set to 11b
28 * on every insertion. During the course of each entry's existence, the
29 * generation field gets decremented during spring cleaning to 10b, then 01b and
32 * This way we're employing the natural numeric ordering to make sure that newly
33 * inserted/touched elements have higher 12-bit counts (which we've manufactured)
34 * and thus iterating over the array initially won't kick out those elements
35 * which were inserted last.
37 * Spring cleaning is what we do when we reach a certain number CLEAN_ELEMS of
38 * elements entered into the array, during which, we're decaying all elements.
39 * If, after decay, an element gets inserted again, its generation is set to 11b
40 * to make sure it has higher numerical count than other, older elements and
41 * thus emulate an LRU-like behavior when deleting elements to free up space
44 * When an element reaches it's max count of action_threshold, we try to poison
45 * it by assuming that errors triggered action_threshold times in a single page
46 * are excessive and that page shouldn't be used anymore. action_threshold is
47 * initialized to COUNT_MASK which is the maximum.
49 * That error event entry causes cec_add_elem() to return !0 value and thus
50 * signal to its callers to log the error.
52 * To the question why we've chosen a page and moving elements around with
53 * memmove(), it is because it is a very simple structure to handle and max data
54 * movement is 4K which on highly optimized modern CPUs is almost unnoticeable.
55 * We wanted to avoid the pointer traversal of more complex structures like a
56 * linked list or some sort of a balancing search tree.
58 * Deleting an element takes O(n) but since it is only a single page, it should
59 * be fast enough and it shouldn't happen all too often depending on error
64 #define pr_fmt(fmt) "RAS: " fmt
67 * We use DECAY_BITS bits of PAGE_SHIFT bits for counting decay, i.e., how long
68 * elements have stayed in the array without having been accessed again.
71 #define DECAY_MASK ((1ULL << DECAY_BITS) - 1)
72 #define MAX_ELEMS (PAGE_SIZE / sizeof(u64))
75 * Threshold amount of inserted elements after which we start spring
78 #define CLEAN_ELEMS (MAX_ELEMS >> DECAY_BITS)
80 /* Bits which count the number of errors happened in this 4K page. */
81 #define COUNT_BITS (PAGE_SHIFT - DECAY_BITS)
82 #define COUNT_MASK ((1ULL << COUNT_BITS) - 1)
83 #define FULL_COUNT_MASK (PAGE_SIZE - 1)
86 * u64: [ 63 ... 12 | DECAY_BITS | COUNT_BITS ]
89 #define PFN(e) ((e) >> PAGE_SHIFT)
90 #define DECAY(e) (((e) >> COUNT_BITS) & DECAY_MASK)
91 #define COUNT(e) ((unsigned int)(e) & COUNT_MASK)
92 #define FULL_COUNT(e) ((e) & (PAGE_SIZE - 1))
94 static struct ce_array {
95 u64 *array; /* container page */
96 unsigned int n; /* number of elements in the array */
98 unsigned int decay_count; /*
99 * number of element insertions/increments
100 * since the last spring cleaning.
103 u64 pfns_poisoned; /*
104 * number of PFNs which got poisoned.
108 * The number of correctable errors
109 * entered into the collector.
113 * Times we did spring cleaning.
118 __u32 disabled : 1, /* cmdline disabled */
125 static DEFINE_MUTEX(ce_mutex);
128 /* Amount of errors after which we offline */
129 static u64 action_threshold = COUNT_MASK;
131 /* Each element "decays" each decay_interval which is 24hrs by default. */
132 #define CEC_DECAY_DEFAULT_INTERVAL 24 * 60 * 60 /* 24 hrs */
133 #define CEC_DECAY_MIN_INTERVAL 1 * 60 * 60 /* 1h */
134 #define CEC_DECAY_MAX_INTERVAL 30 * 24 * 60 * 60 /* one month */
135 static struct delayed_work cec_work;
136 static u64 decay_interval = CEC_DECAY_DEFAULT_INTERVAL;
139 * Decrement decay value. We're using DECAY_BITS bits to denote decay of an
140 * element in the array. On insertion and any access, it gets reset to max.
142 static void do_spring_cleaning(struct ce_array *ca)
146 for (i = 0; i < ca->n; i++) {
147 u8 decay = DECAY(ca->array[i]);
154 ca->array[i] &= ~(DECAY_MASK << COUNT_BITS);
155 ca->array[i] |= (decay << COUNT_BITS);
162 * @interval in seconds
164 static void cec_mod_work(unsigned long interval)
169 mod_delayed_work(system_wq, &cec_work, round_jiffies(iv));
172 static void cec_work_fn(struct work_struct *work)
174 mutex_lock(&ce_mutex);
175 do_spring_cleaning(&ce_arr);
176 mutex_unlock(&ce_mutex);
178 cec_mod_work(decay_interval);
182 * @to: index of the smallest element which is >= then @pfn.
184 * Return the index of the pfn if found, otherwise negative value.
186 static int __find_elem(struct ce_array *ca, u64 pfn, unsigned int *to)
188 int min = 0, max = ca->n - 1;
192 int i = (min + max) >> 1;
194 this_pfn = PFN(ca->array[i]);
198 else if (this_pfn > pfn)
200 else if (this_pfn == pfn) {
209 * When the loop terminates without finding @pfn, min has the index of
210 * the element slot where the new @pfn should be inserted. The loop
211 * terminates when min > max, which means the min index points to the
212 * bigger element while the max index to the smaller element, in-between
213 * which the new @pfn belongs to.
215 * For more details, see exercise 1, Section 6.2.1 in TAOCP, vol. 3.
223 static int find_elem(struct ce_array *ca, u64 pfn, unsigned int *to)
231 return __find_elem(ca, pfn, to);
234 static void del_elem(struct ce_array *ca, int idx)
236 /* Save us a function call when deleting the last element. */
237 if (ca->n - (idx + 1))
238 memmove((void *)&ca->array[idx],
239 (void *)&ca->array[idx + 1],
240 (ca->n - (idx + 1)) * sizeof(u64));
245 static u64 del_lru_elem_unlocked(struct ce_array *ca)
247 unsigned int min = FULL_COUNT_MASK;
250 for (i = 0; i < ca->n; i++) {
251 unsigned int this = FULL_COUNT(ca->array[i]);
259 del_elem(ca, min_idx);
261 return PFN(ca->array[min_idx]);
265 * We return the 0th pfn in the error case under the assumption that it cannot
266 * be poisoned and excessive CEs in there are a serious deal anyway.
268 static u64 __maybe_unused del_lru_elem(void)
270 struct ce_array *ca = &ce_arr;
276 mutex_lock(&ce_mutex);
277 pfn = del_lru_elem_unlocked(ca);
278 mutex_unlock(&ce_mutex);
283 static bool sanity_check(struct ce_array *ca)
289 for (i = 0; i < ca->n; i++) {
290 u64 this = PFN(ca->array[i]);
292 if (WARN(prev > this, "prev: 0x%016llx <-> this: 0x%016llx\n", prev, this))
301 pr_info("Sanity check dump:\n{ n: %d\n", ca->n);
302 for (i = 0; i < ca->n; i++) {
303 u64 this = PFN(ca->array[i]);
305 pr_info(" %03d: [%016llx|%03llx]\n", i, this, FULL_COUNT(ca->array[i]));
313 * cec_add_elem - Add an element to the CEC array.
314 * @pfn: page frame number to insert
319 * - >0: when the inserted pfn was offlined
321 static int cec_add_elem(u64 pfn)
323 struct ce_array *ca = &ce_arr;
324 int count, err, ret = 0;
328 * We can be called very early on the identify_cpu() path where we are
329 * not initialized yet. We ignore the error for simplicity.
331 if (!ce_arr.array || ce_arr.disabled)
334 mutex_lock(&ce_mutex);
338 /* Array full, free the LRU slot. */
339 if (ca->n == MAX_ELEMS)
340 WARN_ON(!del_lru_elem_unlocked(ca));
342 err = find_elem(ca, pfn, &to);
345 * Shift range [to-end] to make room for one more element.
347 memmove((void *)&ca->array[to + 1],
348 (void *)&ca->array[to],
349 (ca->n - to) * sizeof(u64));
351 ca->array[to] = pfn << PAGE_SHIFT;
355 /* Add/refresh element generation and increment count */
356 ca->array[to] |= DECAY_MASK << COUNT_BITS;
359 /* Check action threshold and soft-offline, if reached. */
360 count = COUNT(ca->array[to]);
361 if (count >= action_threshold) {
362 u64 pfn = ca->array[to] >> PAGE_SHIFT;
364 if (!pfn_valid(pfn)) {
365 pr_warn("CEC: Invalid pfn: 0x%llx\n", pfn);
367 /* We have reached max count for this page, soft-offline it. */
368 pr_err("Soft-offlining pfn: 0x%llx\n", pfn);
369 memory_failure_queue(pfn, MF_SOFT_OFFLINE);
376 * Return a >0 value to callers, to denote that we've reached
377 * the offlining threshold.
386 if (ca->decay_count >= CLEAN_ELEMS)
387 do_spring_cleaning(ca);
389 WARN_ON_ONCE(sanity_check(ca));
392 mutex_unlock(&ce_mutex);
397 static int u64_get(void *data, u64 *val)
404 static int pfn_set(void *data, u64 val)
413 DEFINE_DEBUGFS_ATTRIBUTE(pfn_ops, u64_get, pfn_set, "0x%llx\n");
415 static int decay_interval_set(void *data, u64 val)
417 if (val < CEC_DECAY_MIN_INTERVAL)
420 if (val > CEC_DECAY_MAX_INTERVAL)
424 decay_interval = val;
426 cec_mod_work(decay_interval);
430 DEFINE_DEBUGFS_ATTRIBUTE(decay_interval_ops, u64_get, decay_interval_set, "%lld\n");
432 static int action_threshold_set(void *data, u64 val)
436 if (val > COUNT_MASK)
439 action_threshold = val;
443 DEFINE_DEBUGFS_ATTRIBUTE(action_threshold_ops, u64_get, action_threshold_set, "%lld\n");
445 static const char * const bins[] = { "00", "01", "10", "11" };
447 static int array_show(struct seq_file *m, void *v)
449 struct ce_array *ca = &ce_arr;
452 mutex_lock(&ce_mutex);
454 seq_printf(m, "{ n: %d\n", ca->n);
455 for (i = 0; i < ca->n; i++) {
456 u64 this = PFN(ca->array[i]);
458 seq_printf(m, " %3d: [%016llx|%s|%03llx]\n",
459 i, this, bins[DECAY(ca->array[i])], COUNT(ca->array[i]));
462 seq_printf(m, "}\n");
464 seq_printf(m, "Stats:\nCEs: %llu\nofflined pages: %llu\n",
465 ca->ces_entered, ca->pfns_poisoned);
467 seq_printf(m, "Flags: 0x%x\n", ca->flags);
469 seq_printf(m, "Decay interval: %lld seconds\n", decay_interval);
470 seq_printf(m, "Decays: %lld\n", ca->decays_done);
472 seq_printf(m, "Action threshold: %lld\n", action_threshold);
474 mutex_unlock(&ce_mutex);
479 DEFINE_SHOW_ATTRIBUTE(array);
481 static int __init create_debugfs_nodes(void)
483 struct dentry *d, *pfn, *decay, *count, *array;
485 d = debugfs_create_dir("cec", ras_debugfs_dir);
487 pr_warn("Error creating cec debugfs node!\n");
491 decay = debugfs_create_file("decay_interval", S_IRUSR | S_IWUSR, d,
492 &decay_interval, &decay_interval_ops);
494 pr_warn("Error creating decay_interval debugfs node!\n");
498 count = debugfs_create_file("action_threshold", S_IRUSR | S_IWUSR, d,
499 &action_threshold, &action_threshold_ops);
501 pr_warn("Error creating action_threshold debugfs node!\n");
505 if (!IS_ENABLED(CONFIG_RAS_CEC_DEBUG))
508 pfn = debugfs_create_file("pfn", S_IRUSR | S_IWUSR, d, &dfs_pfn, &pfn_ops);
510 pr_warn("Error creating pfn debugfs node!\n");
514 array = debugfs_create_file("array", S_IRUSR, d, NULL, &array_fops);
516 pr_warn("Error creating array debugfs node!\n");
523 debugfs_remove_recursive(d);
528 static int cec_notifier(struct notifier_block *nb, unsigned long val,
531 struct mce *m = (struct mce *)data;
536 /* We eat only correctable DRAM errors with usable addresses. */
537 if (mce_is_memory_error(m) &&
538 mce_is_correctable(m) &&
539 mce_usable_address(m)) {
540 if (!cec_add_elem(m->addr >> PAGE_SHIFT)) {
541 m->kflags |= MCE_HANDLED_CEC;
549 static struct notifier_block cec_nb = {
550 .notifier_call = cec_notifier,
551 .priority = MCE_PRIO_CEC,
554 static int __init cec_init(void)
560 * Intel systems may avoid uncorrectable errors
561 * if pages with corrected errors are aggressively
564 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
565 action_threshold = 2;
567 ce_arr.array = (void *)get_zeroed_page(GFP_KERNEL);
569 pr_err("Error allocating CE array page!\n");
573 if (create_debugfs_nodes()) {
574 free_page((unsigned long)ce_arr.array);
578 INIT_DELAYED_WORK(&cec_work, cec_work_fn);
579 schedule_delayed_work(&cec_work, CEC_DECAY_DEFAULT_INTERVAL);
581 mce_register_decode_chain(&cec_nb);
583 pr_info("Correctable Errors collector initialized.\n");
586 late_initcall(cec_init);
588 int __init parse_cec_param(char *str)
596 if (!strcmp(str, "cec_disable"))