1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Copyright IBM Corporation 2001, 2005, 2006
4 * Copyright Dave Engebretsen & Todd Inglett 2001
5 * Copyright Linas Vepstas 2005, 2006
6 * Copyright 2001-2012 IBM Corporation.
7 *
8 * Please address comments and feedback to Linas Vepstas <linas@austin.ibm.com>
9 */
10
11 #include <linux/delay.h>
12 #include <linux/sched.h>
13 #include <linux/init.h>
14 #include <linux/list.h>
15 #include <linux/pci.h>
16 #include <linux/iommu.h>
17 #include <linux/proc_fs.h>
18 #include <linux/rbtree.h>
19 #include <linux/reboot.h>
20 #include <linux/seq_file.h>
21 #include <linux/spinlock.h>
22 #include <linux/export.h>
23 #include <linux/of.h>
24
25 #include <linux/atomic.h>
26 #include <asm/debugfs.h>
27 #include <asm/eeh.h>
28 #include <asm/eeh_event.h>
29 #include <asm/io.h>
30 #include <asm/iommu.h>
31 #include <asm/machdep.h>
32 #include <asm/ppc-pci.h>
33 #include <asm/rtas.h>
34 #include <asm/pte-walk.h>
35
36
37 /** Overview:
38 * EEH, or "Enhanced Error Handling" is a PCI bridge technology for
39 * dealing with PCI bus errors that can't be dealt with within the
40 * usual PCI framework, except by check-stopping the CPU. Systems
41 * that are designed for high-availability/reliability cannot afford
42 * to crash due to a "mere" PCI error, thus the need for EEH.
43 * An EEH-capable bridge operates by converting a detected error
44 * into a "slot freeze", taking the PCI adapter off-line, making
45 * the slot behave, from the OS'es point of view, as if the slot
46 * were "empty": all reads return 0xff's and all writes are silently
47 * ignored. EEH slot isolation events can be triggered by parity
48 * errors on the address or data busses (e.g. during posted writes),
49 * which in turn might be caused by low voltage on the bus, dust,
50 * vibration, humidity, radioactivity or plain-old failed hardware.
51 *
52 * Note, however, that one of the leading causes of EEH slot
53 * freeze events are buggy device drivers, buggy device microcode,
54 * or buggy device hardware. This is because any attempt by the
55 * device to bus-master data to a memory address that is not
56 * assigned to the device will trigger a slot freeze. (The idea
57 * is to prevent devices-gone-wild from corrupting system memory).
58 * Buggy hardware/drivers will have a miserable time co-existing
59 * with EEH.
60 *
61 * Ideally, a PCI device driver, when suspecting that an isolation
62 * event has occurred (e.g. by reading 0xff's), will then ask EEH
63 * whether this is the case, and then take appropriate steps to
64 * reset the PCI slot, the PCI device, and then resume operations.
65 * However, until that day, the checking is done here, with the
66 * eeh_check_failure() routine embedded in the MMIO macros. If
67 * the slot is found to be isolated, an "EEH Event" is synthesized
68 * and sent out for processing.
69 */
70
71 /* If a device driver keeps reading an MMIO register in an interrupt
72 * handler after a slot isolation event, it might be broken.
73 * This sets the threshold for how many read attempts we allow
74 * before printing an error message.
75 */
76 #define EEH_MAX_FAILS 2100000
77
78 /* Time to wait for a PCI slot to report status, in milliseconds */
79 #define PCI_BUS_RESET_WAIT_MSEC (5*60*1000)
80
81 /*
82 * EEH probe mode support, which is part of the flags,
83 * is to support multiple platforms for EEH. Some platforms
84 * like pSeries do PCI emunation based on device tree.
85 * However, other platforms like powernv probe PCI devices
86 * from hardware. The flag is used to distinguish that.
87 * In addition, struct eeh_ops::probe would be invoked for
88 * particular OF node or PCI device so that the corresponding
89 * PE would be created there.
90 */
91 int eeh_subsystem_flags;
92 EXPORT_SYMBOL(eeh_subsystem_flags);
93
94 /*
95 * EEH allowed maximal frozen times. If one particular PE's
96 * frozen count in last hour exceeds this limit, the PE will
97 * be forced to be offline permanently.
98 */
99 u32 eeh_max_freezes = 5;
100
101 /*
102 * Controls whether a recovery event should be scheduled when an
103 * isolated device is discovered. This is only really useful for
104 * debugging problems with the EEH core.
105 */
106 bool eeh_debugfs_no_recover;
107
108 /* Platform dependent EEH operations */
109 struct eeh_ops *eeh_ops = NULL;
110
111 /* Lock to avoid races due to multiple reports of an error */
112 DEFINE_RAW_SPINLOCK(confirm_error_lock);
113 EXPORT_SYMBOL_GPL(confirm_error_lock);
114
115 /* Lock to protect passed flags */
116 static DEFINE_MUTEX(eeh_dev_mutex);
117
118 /* Buffer for reporting pci register dumps. Its here in BSS, and
119 * not dynamically alloced, so that it ends up in RMO where RTAS
120 * can access it.
121 */
122 #define EEH_PCI_REGS_LOG_LEN 8192
123 static unsigned char pci_regs_buf[EEH_PCI_REGS_LOG_LEN];
124
125 /*
126 * The struct is used to maintain the EEH global statistic
127 * information. Besides, the EEH global statistics will be
128 * exported to user space through procfs
129 */
130 struct eeh_stats {
131 u64 no_device; /* PCI device not found */
132 u64 no_dn; /* OF node not found */
133 u64 no_cfg_addr; /* Config address not found */
134 u64 ignored_check; /* EEH check skipped */
135 u64 total_mmio_ffs; /* Total EEH checks */
136 u64 false_positives; /* Unnecessary EEH checks */
137 u64 slot_resets; /* PE reset */
138 };
139
140 static struct eeh_stats eeh_stats;
141
eeh_setup(char * str)142 static int __init eeh_setup(char *str)
143 {
144 if (!strcmp(str, "off"))
145 eeh_add_flag(EEH_FORCE_DISABLED);
146 else if (!strcmp(str, "early_log"))
147 eeh_add_flag(EEH_EARLY_DUMP_LOG);
148
149 return 1;
150 }
151 __setup("eeh=", eeh_setup);
152
eeh_show_enabled(void)153 void eeh_show_enabled(void)
154 {
155 if (eeh_has_flag(EEH_FORCE_DISABLED))
156 pr_info("EEH: Recovery disabled by kernel parameter.\n");
157 else if (eeh_has_flag(EEH_ENABLED))
158 pr_info("EEH: Capable adapter found: recovery enabled.\n");
159 else
160 pr_info("EEH: No capable adapters found: recovery disabled.\n");
161 }
162
163 /*
164 * This routine captures assorted PCI configuration space data
165 * for the indicated PCI device, and puts them into a buffer
166 * for RTAS error logging.
167 */
eeh_dump_dev_log(struct eeh_dev * edev,char * buf,size_t len)168 static size_t eeh_dump_dev_log(struct eeh_dev *edev, char *buf, size_t len)
169 {
170 u32 cfg;
171 int cap, i;
172 int n = 0, l = 0;
173 char buffer[128];
174
175 n += scnprintf(buf+n, len-n, "%04x:%02x:%02x.%01x\n",
176 edev->pe->phb->global_number, edev->bdfn >> 8,
177 PCI_SLOT(edev->bdfn), PCI_FUNC(edev->bdfn));
178 pr_warn("EEH: of node=%04x:%02x:%02x.%01x\n",
179 edev->pe->phb->global_number, edev->bdfn >> 8,
180 PCI_SLOT(edev->bdfn), PCI_FUNC(edev->bdfn));
181
182 eeh_ops->read_config(edev, PCI_VENDOR_ID, 4, &cfg);
183 n += scnprintf(buf+n, len-n, "dev/vend:%08x\n", cfg);
184 pr_warn("EEH: PCI device/vendor: %08x\n", cfg);
185
186 eeh_ops->read_config(edev, PCI_COMMAND, 4, &cfg);
187 n += scnprintf(buf+n, len-n, "cmd/stat:%x\n", cfg);
188 pr_warn("EEH: PCI cmd/status register: %08x\n", cfg);
189
190 /* Gather bridge-specific registers */
191 if (edev->mode & EEH_DEV_BRIDGE) {
192 eeh_ops->read_config(edev, PCI_SEC_STATUS, 2, &cfg);
193 n += scnprintf(buf+n, len-n, "sec stat:%x\n", cfg);
194 pr_warn("EEH: Bridge secondary status: %04x\n", cfg);
195
196 eeh_ops->read_config(edev, PCI_BRIDGE_CONTROL, 2, &cfg);
197 n += scnprintf(buf+n, len-n, "brdg ctl:%x\n", cfg);
198 pr_warn("EEH: Bridge control: %04x\n", cfg);
199 }
200
201 /* Dump out the PCI-X command and status regs */
202 cap = edev->pcix_cap;
203 if (cap) {
204 eeh_ops->read_config(edev, cap, 4, &cfg);
205 n += scnprintf(buf+n, len-n, "pcix-cmd:%x\n", cfg);
206 pr_warn("EEH: PCI-X cmd: %08x\n", cfg);
207
208 eeh_ops->read_config(edev, cap+4, 4, &cfg);
209 n += scnprintf(buf+n, len-n, "pcix-stat:%x\n", cfg);
210 pr_warn("EEH: PCI-X status: %08x\n", cfg);
211 }
212
213 /* If PCI-E capable, dump PCI-E cap 10 */
214 cap = edev->pcie_cap;
215 if (cap) {
216 n += scnprintf(buf+n, len-n, "pci-e cap10:\n");
217 pr_warn("EEH: PCI-E capabilities and status follow:\n");
218
219 for (i=0; i<=8; i++) {
220 eeh_ops->read_config(edev, cap+4*i, 4, &cfg);
221 n += scnprintf(buf+n, len-n, "%02x:%x\n", 4*i, cfg);
222
223 if ((i % 4) == 0) {
224 if (i != 0)
225 pr_warn("%s\n", buffer);
226
227 l = scnprintf(buffer, sizeof(buffer),
228 "EEH: PCI-E %02x: %08x ",
229 4*i, cfg);
230 } else {
231 l += scnprintf(buffer+l, sizeof(buffer)-l,
232 "%08x ", cfg);
233 }
234
235 }
236
237 pr_warn("%s\n", buffer);
238 }
239
240 /* If AER capable, dump it */
241 cap = edev->aer_cap;
242 if (cap) {
243 n += scnprintf(buf+n, len-n, "pci-e AER:\n");
244 pr_warn("EEH: PCI-E AER capability register set follows:\n");
245
246 for (i=0; i<=13; i++) {
247 eeh_ops->read_config(edev, cap+4*i, 4, &cfg);
248 n += scnprintf(buf+n, len-n, "%02x:%x\n", 4*i, cfg);
249
250 if ((i % 4) == 0) {
251 if (i != 0)
252 pr_warn("%s\n", buffer);
253
254 l = scnprintf(buffer, sizeof(buffer),
255 "EEH: PCI-E AER %02x: %08x ",
256 4*i, cfg);
257 } else {
258 l += scnprintf(buffer+l, sizeof(buffer)-l,
259 "%08x ", cfg);
260 }
261 }
262
263 pr_warn("%s\n", buffer);
264 }
265
266 return n;
267 }
268
eeh_dump_pe_log(struct eeh_pe * pe,void * flag)269 static void *eeh_dump_pe_log(struct eeh_pe *pe, void *flag)
270 {
271 struct eeh_dev *edev, *tmp;
272 size_t *plen = flag;
273
274 eeh_pe_for_each_dev(pe, edev, tmp)
275 *plen += eeh_dump_dev_log(edev, pci_regs_buf + *plen,
276 EEH_PCI_REGS_LOG_LEN - *plen);
277
278 return NULL;
279 }
280
281 /**
282 * eeh_slot_error_detail - Generate combined log including driver log and error log
283 * @pe: EEH PE
284 * @severity: temporary or permanent error log
285 *
286 * This routine should be called to generate the combined log, which
287 * is comprised of driver log and error log. The driver log is figured
288 * out from the config space of the corresponding PCI device, while
289 * the error log is fetched through platform dependent function call.
290 */
eeh_slot_error_detail(struct eeh_pe * pe,int severity)291 void eeh_slot_error_detail(struct eeh_pe *pe, int severity)
292 {
293 size_t loglen = 0;
294
295 /*
296 * When the PHB is fenced or dead, it's pointless to collect
297 * the data from PCI config space because it should return
298 * 0xFF's. For ER, we still retrieve the data from the PCI
299 * config space.
300 *
301 * For pHyp, we have to enable IO for log retrieval. Otherwise,
302 * 0xFF's is always returned from PCI config space.
303 *
304 * When the @severity is EEH_LOG_PERM, the PE is going to be
305 * removed. Prior to that, the drivers for devices included in
306 * the PE will be closed. The drivers rely on working IO path
307 * to bring the devices to quiet state. Otherwise, PCI traffic
308 * from those devices after they are removed is like to cause
309 * another unexpected EEH error.
310 */
311 if (!(pe->type & EEH_PE_PHB)) {
312 if (eeh_has_flag(EEH_ENABLE_IO_FOR_LOG) ||
313 severity == EEH_LOG_PERM)
314 eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
315
316 /*
317 * The config space of some PCI devices can't be accessed
318 * when their PEs are in frozen state. Otherwise, fenced
319 * PHB might be seen. Those PEs are identified with flag
320 * EEH_PE_CFG_RESTRICTED, indicating EEH_PE_CFG_BLOCKED
321 * is set automatically when the PE is put to EEH_PE_ISOLATED.
322 *
323 * Restoring BARs possibly triggers PCI config access in
324 * (OPAL) firmware and then causes fenced PHB. If the
325 * PCI config is blocked with flag EEH_PE_CFG_BLOCKED, it's
326 * pointless to restore BARs and dump config space.
327 */
328 eeh_ops->configure_bridge(pe);
329 if (!(pe->state & EEH_PE_CFG_BLOCKED)) {
330 eeh_pe_restore_bars(pe);
331
332 pci_regs_buf[0] = 0;
333 eeh_pe_traverse(pe, eeh_dump_pe_log, &loglen);
334 }
335 }
336
337 eeh_ops->get_log(pe, severity, pci_regs_buf, loglen);
338 }
339
340 /**
341 * eeh_token_to_phys - Convert EEH address token to phys address
342 * @token: I/O token, should be address in the form 0xA....
343 *
344 * This routine should be called to convert virtual I/O address
345 * to physical one.
346 */
eeh_token_to_phys(unsigned long token)347 static inline unsigned long eeh_token_to_phys(unsigned long token)
348 {
349 pte_t *ptep;
350 unsigned long pa;
351 int hugepage_shift;
352
353 /*
354 * We won't find hugepages here(this is iomem). Hence we are not
355 * worried about _PAGE_SPLITTING/collapse. Also we will not hit
356 * page table free, because of init_mm.
357 */
358 ptep = find_init_mm_pte(token, &hugepage_shift);
359 if (!ptep)
360 return token;
361
362 pa = pte_pfn(*ptep);
363
364 /* On radix we can do hugepage mappings for io, so handle that */
365 if (hugepage_shift) {
366 pa <<= hugepage_shift;
367 pa |= token & ((1ul << hugepage_shift) - 1);
368 } else {
369 pa <<= PAGE_SHIFT;
370 pa |= token & (PAGE_SIZE - 1);
371 }
372
373 return pa;
374 }
375
376 /*
377 * On PowerNV platform, we might already have fenced PHB there.
378 * For that case, it's meaningless to recover frozen PE. Intead,
379 * We have to handle fenced PHB firstly.
380 */
eeh_phb_check_failure(struct eeh_pe * pe)381 static int eeh_phb_check_failure(struct eeh_pe *pe)
382 {
383 struct eeh_pe *phb_pe;
384 unsigned long flags;
385 int ret;
386
387 if (!eeh_has_flag(EEH_PROBE_MODE_DEV))
388 return -EPERM;
389
390 /* Find the PHB PE */
391 phb_pe = eeh_phb_pe_get(pe->phb);
392 if (!phb_pe) {
393 pr_warn("%s Can't find PE for PHB#%x\n",
394 __func__, pe->phb->global_number);
395 return -EEXIST;
396 }
397
398 /* If the PHB has been in problematic state */
399 eeh_serialize_lock(&flags);
400 if (phb_pe->state & EEH_PE_ISOLATED) {
401 ret = 0;
402 goto out;
403 }
404
405 /* Check PHB state */
406 ret = eeh_ops->get_state(phb_pe, NULL);
407 if ((ret < 0) ||
408 (ret == EEH_STATE_NOT_SUPPORT) || eeh_state_active(ret)) {
409 ret = 0;
410 goto out;
411 }
412
413 /* Isolate the PHB and send event */
414 eeh_pe_mark_isolated(phb_pe);
415 eeh_serialize_unlock(flags);
416
417 pr_debug("EEH: PHB#%x failure detected, location: %s\n",
418 phb_pe->phb->global_number, eeh_pe_loc_get(phb_pe));
419 eeh_send_failure_event(phb_pe);
420 return 1;
421 out:
422 eeh_serialize_unlock(flags);
423 return ret;
424 }
425
426 /**
427 * eeh_dev_check_failure - Check if all 1's data is due to EEH slot freeze
428 * @edev: eeh device
429 *
430 * Check for an EEH failure for the given device node. Call this
431 * routine if the result of a read was all 0xff's and you want to
432 * find out if this is due to an EEH slot freeze. This routine
433 * will query firmware for the EEH status.
434 *
435 * Returns 0 if there has not been an EEH error; otherwise returns
436 * a non-zero value and queues up a slot isolation event notification.
437 *
438 * It is safe to call this routine in an interrupt context.
439 */
eeh_dev_check_failure(struct eeh_dev * edev)440 int eeh_dev_check_failure(struct eeh_dev *edev)
441 {
442 int ret;
443 unsigned long flags;
444 struct device_node *dn;
445 struct pci_dev *dev;
446 struct eeh_pe *pe, *parent_pe;
447 int rc = 0;
448 const char *location = NULL;
449
450 eeh_stats.total_mmio_ffs++;
451
452 if (!eeh_enabled())
453 return 0;
454
455 if (!edev) {
456 eeh_stats.no_dn++;
457 return 0;
458 }
459 dev = eeh_dev_to_pci_dev(edev);
460 pe = eeh_dev_to_pe(edev);
461
462 /* Access to IO BARs might get this far and still not want checking. */
463 if (!pe) {
464 eeh_stats.ignored_check++;
465 eeh_edev_dbg(edev, "Ignored check\n");
466 return 0;
467 }
468
469 /*
470 * On PowerNV platform, we might already have fenced PHB
471 * there and we need take care of that firstly.
472 */
473 ret = eeh_phb_check_failure(pe);
474 if (ret > 0)
475 return ret;
476
477 /*
478 * If the PE isn't owned by us, we shouldn't check the
479 * state. Instead, let the owner handle it if the PE has
480 * been frozen.
481 */
482 if (eeh_pe_passed(pe))
483 return 0;
484
485 /* If we already have a pending isolation event for this
486 * slot, we know it's bad already, we don't need to check.
487 * Do this checking under a lock; as multiple PCI devices
488 * in one slot might report errors simultaneously, and we
489 * only want one error recovery routine running.
490 */
491 eeh_serialize_lock(&flags);
492 rc = 1;
493 if (pe->state & EEH_PE_ISOLATED) {
494 pe->check_count++;
495 if (pe->check_count == EEH_MAX_FAILS) {
496 dn = pci_device_to_OF_node(dev);
497 if (dn)
498 location = of_get_property(dn, "ibm,loc-code",
499 NULL);
500 eeh_edev_err(edev, "%d reads ignored for recovering device at location=%s driver=%s\n",
501 pe->check_count,
502 location ? location : "unknown",
503 eeh_driver_name(dev));
504 eeh_edev_err(edev, "Might be infinite loop in %s driver\n",
505 eeh_driver_name(dev));
506 dump_stack();
507 }
508 goto dn_unlock;
509 }
510
511 /*
512 * Now test for an EEH failure. This is VERY expensive.
513 * Note that the eeh_config_addr may be a parent device
514 * in the case of a device behind a bridge, or it may be
515 * function zero of a multi-function device.
516 * In any case they must share a common PHB.
517 */
518 ret = eeh_ops->get_state(pe, NULL);
519
520 /* Note that config-io to empty slots may fail;
521 * they are empty when they don't have children.
522 * We will punt with the following conditions: Failure to get
523 * PE's state, EEH not support and Permanently unavailable
524 * state, PE is in good state.
525 */
526 if ((ret < 0) ||
527 (ret == EEH_STATE_NOT_SUPPORT) || eeh_state_active(ret)) {
528 eeh_stats.false_positives++;
529 pe->false_positives++;
530 rc = 0;
531 goto dn_unlock;
532 }
533
534 /*
535 * It should be corner case that the parent PE has been
536 * put into frozen state as well. We should take care
537 * that at first.
538 */
539 parent_pe = pe->parent;
540 while (parent_pe) {
541 /* Hit the ceiling ? */
542 if (parent_pe->type & EEH_PE_PHB)
543 break;
544
545 /* Frozen parent PE ? */
546 ret = eeh_ops->get_state(parent_pe, NULL);
547 if (ret > 0 && !eeh_state_active(ret)) {
548 pe = parent_pe;
549 pr_err("EEH: Failure of PHB#%x-PE#%x will be handled at parent PHB#%x-PE#%x.\n",
550 pe->phb->global_number, pe->addr,
551 pe->phb->global_number, parent_pe->addr);
552 }
553
554 /* Next parent level */
555 parent_pe = parent_pe->parent;
556 }
557
558 eeh_stats.slot_resets++;
559
560 /* Avoid repeated reports of this failure, including problems
561 * with other functions on this device, and functions under
562 * bridges.
563 */
564 eeh_pe_mark_isolated(pe);
565 eeh_serialize_unlock(flags);
566
567 /* Most EEH events are due to device driver bugs. Having
568 * a stack trace will help the device-driver authors figure
569 * out what happened. So print that out.
570 */
571 pr_debug("EEH: %s: Frozen PHB#%x-PE#%x detected\n",
572 __func__, pe->phb->global_number, pe->addr);
573 eeh_send_failure_event(pe);
574
575 return 1;
576
577 dn_unlock:
578 eeh_serialize_unlock(flags);
579 return rc;
580 }
581
582 EXPORT_SYMBOL_GPL(eeh_dev_check_failure);
583
584 /**
585 * eeh_check_failure - Check if all 1's data is due to EEH slot freeze
586 * @token: I/O address
587 *
588 * Check for an EEH failure at the given I/O address. Call this
589 * routine if the result of a read was all 0xff's and you want to
590 * find out if this is due to an EEH slot freeze event. This routine
591 * will query firmware for the EEH status.
592 *
593 * Note this routine is safe to call in an interrupt context.
594 */
eeh_check_failure(const volatile void __iomem * token)595 int eeh_check_failure(const volatile void __iomem *token)
596 {
597 unsigned long addr;
598 struct eeh_dev *edev;
599
600 /* Finding the phys addr + pci device; this is pretty quick. */
601 addr = eeh_token_to_phys((unsigned long __force) token);
602 edev = eeh_addr_cache_get_dev(addr);
603 if (!edev) {
604 eeh_stats.no_device++;
605 return 0;
606 }
607
608 return eeh_dev_check_failure(edev);
609 }
610 EXPORT_SYMBOL(eeh_check_failure);
611
612
613 /**
614 * eeh_pci_enable - Enable MMIO or DMA transfers for this slot
615 * @pe: EEH PE
616 *
617 * This routine should be called to reenable frozen MMIO or DMA
618 * so that it would work correctly again. It's useful while doing
619 * recovery or log collection on the indicated device.
620 */
eeh_pci_enable(struct eeh_pe * pe,int function)621 int eeh_pci_enable(struct eeh_pe *pe, int function)
622 {
623 int active_flag, rc;
624
625 /*
626 * pHyp doesn't allow to enable IO or DMA on unfrozen PE.
627 * Also, it's pointless to enable them on unfrozen PE. So
628 * we have to check before enabling IO or DMA.
629 */
630 switch (function) {
631 case EEH_OPT_THAW_MMIO:
632 active_flag = EEH_STATE_MMIO_ACTIVE | EEH_STATE_MMIO_ENABLED;
633 break;
634 case EEH_OPT_THAW_DMA:
635 active_flag = EEH_STATE_DMA_ACTIVE;
636 break;
637 case EEH_OPT_DISABLE:
638 case EEH_OPT_ENABLE:
639 case EEH_OPT_FREEZE_PE:
640 active_flag = 0;
641 break;
642 default:
643 pr_warn("%s: Invalid function %d\n",
644 __func__, function);
645 return -EINVAL;
646 }
647
648 /*
649 * Check if IO or DMA has been enabled before
650 * enabling them.
651 */
652 if (active_flag) {
653 rc = eeh_ops->get_state(pe, NULL);
654 if (rc < 0)
655 return rc;
656
657 /* Needn't enable it at all */
658 if (rc == EEH_STATE_NOT_SUPPORT)
659 return 0;
660
661 /* It's already enabled */
662 if (rc & active_flag)
663 return 0;
664 }
665
666
667 /* Issue the request */
668 rc = eeh_ops->set_option(pe, function);
669 if (rc)
670 pr_warn("%s: Unexpected state change %d on "
671 "PHB#%x-PE#%x, err=%d\n",
672 __func__, function, pe->phb->global_number,
673 pe->addr, rc);
674
675 /* Check if the request is finished successfully */
676 if (active_flag) {
677 rc = eeh_wait_state(pe, PCI_BUS_RESET_WAIT_MSEC);
678 if (rc < 0)
679 return rc;
680
681 if (rc & active_flag)
682 return 0;
683
684 return -EIO;
685 }
686
687 return rc;
688 }
689
eeh_disable_and_save_dev_state(struct eeh_dev * edev,void * userdata)690 static void eeh_disable_and_save_dev_state(struct eeh_dev *edev,
691 void *userdata)
692 {
693 struct pci_dev *pdev = eeh_dev_to_pci_dev(edev);
694 struct pci_dev *dev = userdata;
695
696 /*
697 * The caller should have disabled and saved the
698 * state for the specified device
699 */
700 if (!pdev || pdev == dev)
701 return;
702
703 /* Ensure we have D0 power state */
704 pci_set_power_state(pdev, PCI_D0);
705
706 /* Save device state */
707 pci_save_state(pdev);
708
709 /*
710 * Disable device to avoid any DMA traffic and
711 * interrupt from the device
712 */
713 pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
714 }
715
eeh_restore_dev_state(struct eeh_dev * edev,void * userdata)716 static void eeh_restore_dev_state(struct eeh_dev *edev, void *userdata)
717 {
718 struct pci_dev *pdev = eeh_dev_to_pci_dev(edev);
719 struct pci_dev *dev = userdata;
720
721 if (!pdev)
722 return;
723
724 /* Apply customization from firmware */
725 if (eeh_ops->restore_config)
726 eeh_ops->restore_config(edev);
727
728 /* The caller should restore state for the specified device */
729 if (pdev != dev)
730 pci_restore_state(pdev);
731 }
732
733 /**
734 * pcibios_set_pcie_reset_state - Set PCI-E reset state
735 * @dev: pci device struct
736 * @state: reset state to enter
737 *
738 * Return value:
739 * 0 if success
740 */
pcibios_set_pcie_reset_state(struct pci_dev * dev,enum pcie_reset_state state)741 int pcibios_set_pcie_reset_state(struct pci_dev *dev, enum pcie_reset_state state)
742 {
743 struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
744 struct eeh_pe *pe = eeh_dev_to_pe(edev);
745
746 if (!pe) {
747 pr_err("%s: No PE found on PCI device %s\n",
748 __func__, pci_name(dev));
749 return -EINVAL;
750 }
751
752 switch (state) {
753 case pcie_deassert_reset:
754 eeh_ops->reset(pe, EEH_RESET_DEACTIVATE);
755 eeh_unfreeze_pe(pe);
756 if (!(pe->type & EEH_PE_VF))
757 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, true);
758 eeh_pe_dev_traverse(pe, eeh_restore_dev_state, dev);
759 eeh_pe_state_clear(pe, EEH_PE_ISOLATED, true);
760 break;
761 case pcie_hot_reset:
762 eeh_pe_mark_isolated(pe);
763 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, true);
764 eeh_ops->set_option(pe, EEH_OPT_FREEZE_PE);
765 eeh_pe_dev_traverse(pe, eeh_disable_and_save_dev_state, dev);
766 if (!(pe->type & EEH_PE_VF))
767 eeh_pe_state_mark(pe, EEH_PE_CFG_BLOCKED);
768 eeh_ops->reset(pe, EEH_RESET_HOT);
769 break;
770 case pcie_warm_reset:
771 eeh_pe_mark_isolated(pe);
772 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, true);
773 eeh_ops->set_option(pe, EEH_OPT_FREEZE_PE);
774 eeh_pe_dev_traverse(pe, eeh_disable_and_save_dev_state, dev);
775 if (!(pe->type & EEH_PE_VF))
776 eeh_pe_state_mark(pe, EEH_PE_CFG_BLOCKED);
777 eeh_ops->reset(pe, EEH_RESET_FUNDAMENTAL);
778 break;
779 default:
780 eeh_pe_state_clear(pe, EEH_PE_ISOLATED | EEH_PE_CFG_BLOCKED, true);
781 return -EINVAL;
782 };
783
784 return 0;
785 }
786
787 /**
788 * eeh_set_pe_freset - Check the required reset for the indicated device
789 * @data: EEH device
790 * @flag: return value
791 *
792 * Each device might have its preferred reset type: fundamental or
793 * hot reset. The routine is used to collected the information for
794 * the indicated device and its children so that the bunch of the
795 * devices could be reset properly.
796 */
eeh_set_dev_freset(struct eeh_dev * edev,void * flag)797 static void eeh_set_dev_freset(struct eeh_dev *edev, void *flag)
798 {
799 struct pci_dev *dev;
800 unsigned int *freset = (unsigned int *)flag;
801
802 dev = eeh_dev_to_pci_dev(edev);
803 if (dev)
804 *freset |= dev->needs_freset;
805 }
806
eeh_pe_refreeze_passed(struct eeh_pe * root)807 static void eeh_pe_refreeze_passed(struct eeh_pe *root)
808 {
809 struct eeh_pe *pe;
810 int state;
811
812 eeh_for_each_pe(root, pe) {
813 if (eeh_pe_passed(pe)) {
814 state = eeh_ops->get_state(pe, NULL);
815 if (state &
816 (EEH_STATE_MMIO_ACTIVE | EEH_STATE_MMIO_ENABLED)) {
817 pr_info("EEH: Passed-through PE PHB#%x-PE#%x was thawed by reset, re-freezing for safety.\n",
818 pe->phb->global_number, pe->addr);
819 eeh_pe_set_option(pe, EEH_OPT_FREEZE_PE);
820 }
821 }
822 }
823 }
824
825 /**
826 * eeh_pe_reset_full - Complete a full reset process on the indicated PE
827 * @pe: EEH PE
828 *
829 * This function executes a full reset procedure on a PE, including setting
830 * the appropriate flags, performing a fundamental or hot reset, and then
831 * deactivating the reset status. It is designed to be used within the EEH
832 * subsystem, as opposed to eeh_pe_reset which is exported to drivers and
833 * only performs a single operation at a time.
834 *
835 * This function will attempt to reset a PE three times before failing.
836 */
eeh_pe_reset_full(struct eeh_pe * pe,bool include_passed)837 int eeh_pe_reset_full(struct eeh_pe *pe, bool include_passed)
838 {
839 int reset_state = (EEH_PE_RESET | EEH_PE_CFG_BLOCKED);
840 int type = EEH_RESET_HOT;
841 unsigned int freset = 0;
842 int i, state = 0, ret;
843
844 /*
845 * Determine the type of reset to perform - hot or fundamental.
846 * Hot reset is the default operation, unless any device under the
847 * PE requires a fundamental reset.
848 */
849 eeh_pe_dev_traverse(pe, eeh_set_dev_freset, &freset);
850
851 if (freset)
852 type = EEH_RESET_FUNDAMENTAL;
853
854 /* Mark the PE as in reset state and block config space accesses */
855 eeh_pe_state_mark(pe, reset_state);
856
857 /* Make three attempts at resetting the bus */
858 for (i = 0; i < 3; i++) {
859 ret = eeh_pe_reset(pe, type, include_passed);
860 if (!ret)
861 ret = eeh_pe_reset(pe, EEH_RESET_DEACTIVATE,
862 include_passed);
863 if (ret) {
864 ret = -EIO;
865 pr_warn("EEH: Failure %d resetting PHB#%x-PE#%x (attempt %d)\n\n",
866 state, pe->phb->global_number, pe->addr, i + 1);
867 continue;
868 }
869 if (i)
870 pr_warn("EEH: PHB#%x-PE#%x: Successful reset (attempt %d)\n",
871 pe->phb->global_number, pe->addr, i + 1);
872
873 /* Wait until the PE is in a functioning state */
874 state = eeh_wait_state(pe, PCI_BUS_RESET_WAIT_MSEC);
875 if (state < 0) {
876 pr_warn("EEH: Unrecoverable slot failure on PHB#%x-PE#%x",
877 pe->phb->global_number, pe->addr);
878 ret = -ENOTRECOVERABLE;
879 break;
880 }
881 if (eeh_state_active(state))
882 break;
883 else
884 pr_warn("EEH: PHB#%x-PE#%x: Slot inactive after reset: 0x%x (attempt %d)\n",
885 pe->phb->global_number, pe->addr, state, i + 1);
886 }
887
888 /* Resetting the PE may have unfrozen child PEs. If those PEs have been
889 * (potentially) passed through to a guest, re-freeze them:
890 */
891 if (!include_passed)
892 eeh_pe_refreeze_passed(pe);
893
894 eeh_pe_state_clear(pe, reset_state, true);
895 return ret;
896 }
897
898 /**
899 * eeh_save_bars - Save device bars
900 * @edev: PCI device associated EEH device
901 *
902 * Save the values of the device bars. Unlike the restore
903 * routine, this routine is *not* recursive. This is because
904 * PCI devices are added individually; but, for the restore,
905 * an entire slot is reset at a time.
906 */
eeh_save_bars(struct eeh_dev * edev)907 void eeh_save_bars(struct eeh_dev *edev)
908 {
909 int i;
910
911 if (!edev)
912 return;
913
914 for (i = 0; i < 16; i++)
915 eeh_ops->read_config(edev, i * 4, 4, &edev->config_space[i]);
916
917 /*
918 * For PCI bridges including root port, we need enable bus
919 * master explicitly. Otherwise, it can't fetch IODA table
920 * entries correctly. So we cache the bit in advance so that
921 * we can restore it after reset, either PHB range or PE range.
922 */
923 if (edev->mode & EEH_DEV_BRIDGE)
924 edev->config_space[1] |= PCI_COMMAND_MASTER;
925 }
926
eeh_reboot_notifier(struct notifier_block * nb,unsigned long action,void * unused)927 static int eeh_reboot_notifier(struct notifier_block *nb,
928 unsigned long action, void *unused)
929 {
930 eeh_clear_flag(EEH_ENABLED);
931 return NOTIFY_DONE;
932 }
933
934 static struct notifier_block eeh_reboot_nb = {
935 .notifier_call = eeh_reboot_notifier,
936 };
937
eeh_device_notifier(struct notifier_block * nb,unsigned long action,void * data)938 static int eeh_device_notifier(struct notifier_block *nb,
939 unsigned long action, void *data)
940 {
941 struct device *dev = data;
942
943 switch (action) {
944 /*
945 * Note: It's not possible to perform EEH device addition (i.e.
946 * {pseries,pnv}_pcibios_bus_add_device()) here because it depends on
947 * the device's resources, which have not yet been set up.
948 */
949 case BUS_NOTIFY_DEL_DEVICE:
950 eeh_remove_device(to_pci_dev(dev));
951 break;
952 default:
953 break;
954 }
955 return NOTIFY_DONE;
956 }
957
958 static struct notifier_block eeh_device_nb = {
959 .notifier_call = eeh_device_notifier,
960 };
961
962 /**
963 * eeh_init - System wide EEH initialization
964 *
965 * It's the platform's job to call this from an arch_initcall().
966 */
eeh_init(struct eeh_ops * ops)967 int eeh_init(struct eeh_ops *ops)
968 {
969 struct pci_controller *hose, *tmp;
970 int ret = 0;
971
972 /* the platform should only initialise EEH once */
973 if (WARN_ON(eeh_ops))
974 return -EEXIST;
975 if (WARN_ON(!ops))
976 return -ENOENT;
977 eeh_ops = ops;
978
979 /* Register reboot notifier */
980 ret = register_reboot_notifier(&eeh_reboot_nb);
981 if (ret) {
982 pr_warn("%s: Failed to register reboot notifier (%d)\n",
983 __func__, ret);
984 return ret;
985 }
986
987 ret = bus_register_notifier(&pci_bus_type, &eeh_device_nb);
988 if (ret) {
989 pr_warn("%s: Failed to register bus notifier (%d)\n",
990 __func__, ret);
991 return ret;
992 }
993
994 /* Initialize PHB PEs */
995 list_for_each_entry_safe(hose, tmp, &hose_list, list_node)
996 eeh_phb_pe_create(hose);
997
998 eeh_addr_cache_init();
999
1000 /* Initialize EEH event */
1001 return eeh_event_init();
1002 }
1003
1004 /**
1005 * eeh_probe_device() - Perform EEH initialization for the indicated pci device
1006 * @dev: pci device for which to set up EEH
1007 *
1008 * This routine must be used to complete EEH initialization for PCI
1009 * devices that were added after system boot (e.g. hotplug, dlpar).
1010 */
eeh_probe_device(struct pci_dev * dev)1011 void eeh_probe_device(struct pci_dev *dev)
1012 {
1013 struct eeh_dev *edev;
1014
1015 pr_debug("EEH: Adding device %s\n", pci_name(dev));
1016
1017 /*
1018 * pci_dev_to_eeh_dev() can only work if eeh_probe_dev() was
1019 * already called for this device.
1020 */
1021 if (WARN_ON_ONCE(pci_dev_to_eeh_dev(dev))) {
1022 pci_dbg(dev, "Already bound to an eeh_dev!\n");
1023 return;
1024 }
1025
1026 edev = eeh_ops->probe(dev);
1027 if (!edev) {
1028 pr_debug("EEH: Adding device failed\n");
1029 return;
1030 }
1031
1032 /*
1033 * FIXME: We rely on pcibios_release_device() to remove the
1034 * existing EEH state. The release function is only called if
1035 * the pci_dev's refcount drops to zero so if something is
1036 * keeping a ref to a device (e.g. a filesystem) we need to
1037 * remove the old EEH state.
1038 *
1039 * FIXME: HEY MA, LOOK AT ME, NO LOCKING!
1040 */
1041 if (edev->pdev && edev->pdev != dev) {
1042 eeh_pe_tree_remove(edev);
1043 eeh_addr_cache_rmv_dev(edev->pdev);
1044 eeh_sysfs_remove_device(edev->pdev);
1045
1046 /*
1047 * We definitely should have the PCI device removed
1048 * though it wasn't correctly. So we needn't call
1049 * into error handler afterwards.
1050 */
1051 edev->mode |= EEH_DEV_NO_HANDLER;
1052 }
1053
1054 /* bind the pdev and the edev together */
1055 edev->pdev = dev;
1056 dev->dev.archdata.edev = edev;
1057 eeh_addr_cache_insert_dev(dev);
1058 eeh_sysfs_add_device(dev);
1059 }
1060
1061 /**
1062 * eeh_remove_device - Undo EEH setup for the indicated pci device
1063 * @dev: pci device to be removed
1064 *
1065 * This routine should be called when a device is removed from
1066 * a running system (e.g. by hotplug or dlpar). It unregisters
1067 * the PCI device from the EEH subsystem. I/O errors affecting
1068 * this device will no longer be detected after this call; thus,
1069 * i/o errors affecting this slot may leave this device unusable.
1070 */
eeh_remove_device(struct pci_dev * dev)1071 void eeh_remove_device(struct pci_dev *dev)
1072 {
1073 struct eeh_dev *edev;
1074
1075 if (!dev || !eeh_enabled())
1076 return;
1077 edev = pci_dev_to_eeh_dev(dev);
1078
1079 /* Unregister the device with the EEH/PCI address search system */
1080 dev_dbg(&dev->dev, "EEH: Removing device\n");
1081
1082 if (!edev || !edev->pdev || !edev->pe) {
1083 dev_dbg(&dev->dev, "EEH: Device not referenced!\n");
1084 return;
1085 }
1086
1087 /*
1088 * During the hotplug for EEH error recovery, we need the EEH
1089 * device attached to the parent PE in order for BAR restore
1090 * a bit later. So we keep it for BAR restore and remove it
1091 * from the parent PE during the BAR resotre.
1092 */
1093 edev->pdev = NULL;
1094
1095 /*
1096 * eeh_sysfs_remove_device() uses pci_dev_to_eeh_dev() so we need to
1097 * remove the sysfs files before clearing dev.archdata.edev
1098 */
1099 if (edev->mode & EEH_DEV_SYSFS)
1100 eeh_sysfs_remove_device(dev);
1101
1102 /*
1103 * We're removing from the PCI subsystem, that means
1104 * the PCI device driver can't support EEH or not
1105 * well. So we rely on hotplug completely to do recovery
1106 * for the specific PCI device.
1107 */
1108 edev->mode |= EEH_DEV_NO_HANDLER;
1109
1110 eeh_addr_cache_rmv_dev(dev);
1111
1112 /*
1113 * The flag "in_error" is used to trace EEH devices for VFs
1114 * in error state or not. It's set in eeh_report_error(). If
1115 * it's not set, eeh_report_{reset,resume}() won't be called
1116 * for the VF EEH device.
1117 */
1118 edev->in_error = false;
1119 dev->dev.archdata.edev = NULL;
1120 if (!(edev->pe->state & EEH_PE_KEEP))
1121 eeh_pe_tree_remove(edev);
1122 else
1123 edev->mode |= EEH_DEV_DISCONNECTED;
1124 }
1125
eeh_unfreeze_pe(struct eeh_pe * pe)1126 int eeh_unfreeze_pe(struct eeh_pe *pe)
1127 {
1128 int ret;
1129
1130 ret = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
1131 if (ret) {
1132 pr_warn("%s: Failure %d enabling IO on PHB#%x-PE#%x\n",
1133 __func__, ret, pe->phb->global_number, pe->addr);
1134 return ret;
1135 }
1136
1137 ret = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
1138 if (ret) {
1139 pr_warn("%s: Failure %d enabling DMA on PHB#%x-PE#%x\n",
1140 __func__, ret, pe->phb->global_number, pe->addr);
1141 return ret;
1142 }
1143
1144 return ret;
1145 }
1146
1147
1148 static struct pci_device_id eeh_reset_ids[] = {
1149 { PCI_DEVICE(0x19a2, 0x0710) }, /* Emulex, BE */
1150 { PCI_DEVICE(0x10df, 0xe220) }, /* Emulex, Lancer */
1151 { PCI_DEVICE(0x14e4, 0x1657) }, /* Broadcom BCM5719 */
1152 { 0 }
1153 };
1154
eeh_pe_change_owner(struct eeh_pe * pe)1155 static int eeh_pe_change_owner(struct eeh_pe *pe)
1156 {
1157 struct eeh_dev *edev, *tmp;
1158 struct pci_dev *pdev;
1159 struct pci_device_id *id;
1160 int ret;
1161
1162 /* Check PE state */
1163 ret = eeh_ops->get_state(pe, NULL);
1164 if (ret < 0 || ret == EEH_STATE_NOT_SUPPORT)
1165 return 0;
1166
1167 /* Unfrozen PE, nothing to do */
1168 if (eeh_state_active(ret))
1169 return 0;
1170
1171 /* Frozen PE, check if it needs PE level reset */
1172 eeh_pe_for_each_dev(pe, edev, tmp) {
1173 pdev = eeh_dev_to_pci_dev(edev);
1174 if (!pdev)
1175 continue;
1176
1177 for (id = &eeh_reset_ids[0]; id->vendor != 0; id++) {
1178 if (id->vendor != PCI_ANY_ID &&
1179 id->vendor != pdev->vendor)
1180 continue;
1181 if (id->device != PCI_ANY_ID &&
1182 id->device != pdev->device)
1183 continue;
1184 if (id->subvendor != PCI_ANY_ID &&
1185 id->subvendor != pdev->subsystem_vendor)
1186 continue;
1187 if (id->subdevice != PCI_ANY_ID &&
1188 id->subdevice != pdev->subsystem_device)
1189 continue;
1190
1191 return eeh_pe_reset_and_recover(pe);
1192 }
1193 }
1194
1195 ret = eeh_unfreeze_pe(pe);
1196 if (!ret)
1197 eeh_pe_state_clear(pe, EEH_PE_ISOLATED, true);
1198 return ret;
1199 }
1200
1201 /**
1202 * eeh_dev_open - Increase count of pass through devices for PE
1203 * @pdev: PCI device
1204 *
1205 * Increase count of passed through devices for the indicated
1206 * PE. In the result, the EEH errors detected on the PE won't be
1207 * reported. The PE owner will be responsible for detection
1208 * and recovery.
1209 */
eeh_dev_open(struct pci_dev * pdev)1210 int eeh_dev_open(struct pci_dev *pdev)
1211 {
1212 struct eeh_dev *edev;
1213 int ret = -ENODEV;
1214
1215 mutex_lock(&eeh_dev_mutex);
1216
1217 /* No PCI device ? */
1218 if (!pdev)
1219 goto out;
1220
1221 /* No EEH device or PE ? */
1222 edev = pci_dev_to_eeh_dev(pdev);
1223 if (!edev || !edev->pe)
1224 goto out;
1225
1226 /*
1227 * The PE might have been put into frozen state, but we
1228 * didn't detect that yet. The passed through PCI devices
1229 * in frozen PE won't work properly. Clear the frozen state
1230 * in advance.
1231 */
1232 ret = eeh_pe_change_owner(edev->pe);
1233 if (ret)
1234 goto out;
1235
1236 /* Increase PE's pass through count */
1237 atomic_inc(&edev->pe->pass_dev_cnt);
1238 mutex_unlock(&eeh_dev_mutex);
1239
1240 return 0;
1241 out:
1242 mutex_unlock(&eeh_dev_mutex);
1243 return ret;
1244 }
1245 EXPORT_SYMBOL_GPL(eeh_dev_open);
1246
1247 /**
1248 * eeh_dev_release - Decrease count of pass through devices for PE
1249 * @pdev: PCI device
1250 *
1251 * Decrease count of pass through devices for the indicated PE. If
1252 * there is no passed through device in PE, the EEH errors detected
1253 * on the PE will be reported and handled as usual.
1254 */
eeh_dev_release(struct pci_dev * pdev)1255 void eeh_dev_release(struct pci_dev *pdev)
1256 {
1257 struct eeh_dev *edev;
1258
1259 mutex_lock(&eeh_dev_mutex);
1260
1261 /* No PCI device ? */
1262 if (!pdev)
1263 goto out;
1264
1265 /* No EEH device ? */
1266 edev = pci_dev_to_eeh_dev(pdev);
1267 if (!edev || !edev->pe || !eeh_pe_passed(edev->pe))
1268 goto out;
1269
1270 /* Decrease PE's pass through count */
1271 WARN_ON(atomic_dec_if_positive(&edev->pe->pass_dev_cnt) < 0);
1272 eeh_pe_change_owner(edev->pe);
1273 out:
1274 mutex_unlock(&eeh_dev_mutex);
1275 }
1276 EXPORT_SYMBOL(eeh_dev_release);
1277
1278 #ifdef CONFIG_IOMMU_API
1279
dev_has_iommu_table(struct device * dev,void * data)1280 static int dev_has_iommu_table(struct device *dev, void *data)
1281 {
1282 struct pci_dev *pdev = to_pci_dev(dev);
1283 struct pci_dev **ppdev = data;
1284
1285 if (!dev)
1286 return 0;
1287
1288 if (device_iommu_mapped(dev)) {
1289 *ppdev = pdev;
1290 return 1;
1291 }
1292
1293 return 0;
1294 }
1295
1296 /**
1297 * eeh_iommu_group_to_pe - Convert IOMMU group to EEH PE
1298 * @group: IOMMU group
1299 *
1300 * The routine is called to convert IOMMU group to EEH PE.
1301 */
eeh_iommu_group_to_pe(struct iommu_group * group)1302 struct eeh_pe *eeh_iommu_group_to_pe(struct iommu_group *group)
1303 {
1304 struct pci_dev *pdev = NULL;
1305 struct eeh_dev *edev;
1306 int ret;
1307
1308 /* No IOMMU group ? */
1309 if (!group)
1310 return NULL;
1311
1312 ret = iommu_group_for_each_dev(group, &pdev, dev_has_iommu_table);
1313 if (!ret || !pdev)
1314 return NULL;
1315
1316 /* No EEH device or PE ? */
1317 edev = pci_dev_to_eeh_dev(pdev);
1318 if (!edev || !edev->pe)
1319 return NULL;
1320
1321 return edev->pe;
1322 }
1323 EXPORT_SYMBOL_GPL(eeh_iommu_group_to_pe);
1324
1325 #endif /* CONFIG_IOMMU_API */
1326
1327 /**
1328 * eeh_pe_set_option - Set options for the indicated PE
1329 * @pe: EEH PE
1330 * @option: requested option
1331 *
1332 * The routine is called to enable or disable EEH functionality
1333 * on the indicated PE, to enable IO or DMA for the frozen PE.
1334 */
eeh_pe_set_option(struct eeh_pe * pe,int option)1335 int eeh_pe_set_option(struct eeh_pe *pe, int option)
1336 {
1337 int ret = 0;
1338
1339 /* Invalid PE ? */
1340 if (!pe)
1341 return -ENODEV;
1342
1343 /*
1344 * EEH functionality could possibly be disabled, just
1345 * return error for the case. And the EEH functinality
1346 * isn't expected to be disabled on one specific PE.
1347 */
1348 switch (option) {
1349 case EEH_OPT_ENABLE:
1350 if (eeh_enabled()) {
1351 ret = eeh_pe_change_owner(pe);
1352 break;
1353 }
1354 ret = -EIO;
1355 break;
1356 case EEH_OPT_DISABLE:
1357 break;
1358 case EEH_OPT_THAW_MMIO:
1359 case EEH_OPT_THAW_DMA:
1360 case EEH_OPT_FREEZE_PE:
1361 if (!eeh_ops || !eeh_ops->set_option) {
1362 ret = -ENOENT;
1363 break;
1364 }
1365
1366 ret = eeh_pci_enable(pe, option);
1367 break;
1368 default:
1369 pr_debug("%s: Option %d out of range (%d, %d)\n",
1370 __func__, option, EEH_OPT_DISABLE, EEH_OPT_THAW_DMA);
1371 ret = -EINVAL;
1372 }
1373
1374 return ret;
1375 }
1376 EXPORT_SYMBOL_GPL(eeh_pe_set_option);
1377
1378 /**
1379 * eeh_pe_get_state - Retrieve PE's state
1380 * @pe: EEH PE
1381 *
1382 * Retrieve the PE's state, which includes 3 aspects: enabled
1383 * DMA, enabled IO and asserted reset.
1384 */
eeh_pe_get_state(struct eeh_pe * pe)1385 int eeh_pe_get_state(struct eeh_pe *pe)
1386 {
1387 int result, ret = 0;
1388 bool rst_active, dma_en, mmio_en;
1389
1390 /* Existing PE ? */
1391 if (!pe)
1392 return -ENODEV;
1393
1394 if (!eeh_ops || !eeh_ops->get_state)
1395 return -ENOENT;
1396
1397 /*
1398 * If the parent PE is owned by the host kernel and is undergoing
1399 * error recovery, we should return the PE state as temporarily
1400 * unavailable so that the error recovery on the guest is suspended
1401 * until the recovery completes on the host.
1402 */
1403 if (pe->parent &&
1404 !(pe->state & EEH_PE_REMOVED) &&
1405 (pe->parent->state & (EEH_PE_ISOLATED | EEH_PE_RECOVERING)))
1406 return EEH_PE_STATE_UNAVAIL;
1407
1408 result = eeh_ops->get_state(pe, NULL);
1409 rst_active = !!(result & EEH_STATE_RESET_ACTIVE);
1410 dma_en = !!(result & EEH_STATE_DMA_ENABLED);
1411 mmio_en = !!(result & EEH_STATE_MMIO_ENABLED);
1412
1413 if (rst_active)
1414 ret = EEH_PE_STATE_RESET;
1415 else if (dma_en && mmio_en)
1416 ret = EEH_PE_STATE_NORMAL;
1417 else if (!dma_en && !mmio_en)
1418 ret = EEH_PE_STATE_STOPPED_IO_DMA;
1419 else if (!dma_en && mmio_en)
1420 ret = EEH_PE_STATE_STOPPED_DMA;
1421 else
1422 ret = EEH_PE_STATE_UNAVAIL;
1423
1424 return ret;
1425 }
1426 EXPORT_SYMBOL_GPL(eeh_pe_get_state);
1427
eeh_pe_reenable_devices(struct eeh_pe * pe,bool include_passed)1428 static int eeh_pe_reenable_devices(struct eeh_pe *pe, bool include_passed)
1429 {
1430 struct eeh_dev *edev, *tmp;
1431 struct pci_dev *pdev;
1432 int ret = 0;
1433
1434 eeh_pe_restore_bars(pe);
1435
1436 /*
1437 * Reenable PCI devices as the devices passed
1438 * through are always enabled before the reset.
1439 */
1440 eeh_pe_for_each_dev(pe, edev, tmp) {
1441 pdev = eeh_dev_to_pci_dev(edev);
1442 if (!pdev)
1443 continue;
1444
1445 ret = pci_reenable_device(pdev);
1446 if (ret) {
1447 pr_warn("%s: Failure %d reenabling %s\n",
1448 __func__, ret, pci_name(pdev));
1449 return ret;
1450 }
1451 }
1452
1453 /* The PE is still in frozen state */
1454 if (include_passed || !eeh_pe_passed(pe)) {
1455 ret = eeh_unfreeze_pe(pe);
1456 } else
1457 pr_info("EEH: Note: Leaving passthrough PHB#%x-PE#%x frozen.\n",
1458 pe->phb->global_number, pe->addr);
1459 if (!ret)
1460 eeh_pe_state_clear(pe, EEH_PE_ISOLATED, include_passed);
1461 return ret;
1462 }
1463
1464
1465 /**
1466 * eeh_pe_reset - Issue PE reset according to specified type
1467 * @pe: EEH PE
1468 * @option: reset type
1469 *
1470 * The routine is called to reset the specified PE with the
1471 * indicated type, either fundamental reset or hot reset.
1472 * PE reset is the most important part for error recovery.
1473 */
eeh_pe_reset(struct eeh_pe * pe,int option,bool include_passed)1474 int eeh_pe_reset(struct eeh_pe *pe, int option, bool include_passed)
1475 {
1476 int ret = 0;
1477
1478 /* Invalid PE ? */
1479 if (!pe)
1480 return -ENODEV;
1481
1482 if (!eeh_ops || !eeh_ops->set_option || !eeh_ops->reset)
1483 return -ENOENT;
1484
1485 switch (option) {
1486 case EEH_RESET_DEACTIVATE:
1487 ret = eeh_ops->reset(pe, option);
1488 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, include_passed);
1489 if (ret)
1490 break;
1491
1492 ret = eeh_pe_reenable_devices(pe, include_passed);
1493 break;
1494 case EEH_RESET_HOT:
1495 case EEH_RESET_FUNDAMENTAL:
1496 /*
1497 * Proactively freeze the PE to drop all MMIO access
1498 * during reset, which should be banned as it's always
1499 * cause recursive EEH error.
1500 */
1501 eeh_ops->set_option(pe, EEH_OPT_FREEZE_PE);
1502
1503 eeh_pe_state_mark(pe, EEH_PE_CFG_BLOCKED);
1504 ret = eeh_ops->reset(pe, option);
1505 break;
1506 default:
1507 pr_debug("%s: Unsupported option %d\n",
1508 __func__, option);
1509 ret = -EINVAL;
1510 }
1511
1512 return ret;
1513 }
1514 EXPORT_SYMBOL_GPL(eeh_pe_reset);
1515
1516 /**
1517 * eeh_pe_configure - Configure PCI bridges after PE reset
1518 * @pe: EEH PE
1519 *
1520 * The routine is called to restore the PCI config space for
1521 * those PCI devices, especially PCI bridges affected by PE
1522 * reset issued previously.
1523 */
eeh_pe_configure(struct eeh_pe * pe)1524 int eeh_pe_configure(struct eeh_pe *pe)
1525 {
1526 int ret = 0;
1527
1528 /* Invalid PE ? */
1529 if (!pe)
1530 return -ENODEV;
1531
1532 return ret;
1533 }
1534 EXPORT_SYMBOL_GPL(eeh_pe_configure);
1535
1536 /**
1537 * eeh_pe_inject_err - Injecting the specified PCI error to the indicated PE
1538 * @pe: the indicated PE
1539 * @type: error type
1540 * @function: error function
1541 * @addr: address
1542 * @mask: address mask
1543 *
1544 * The routine is called to inject the specified PCI error, which
1545 * is determined by @type and @function, to the indicated PE for
1546 * testing purpose.
1547 */
eeh_pe_inject_err(struct eeh_pe * pe,int type,int func,unsigned long addr,unsigned long mask)1548 int eeh_pe_inject_err(struct eeh_pe *pe, int type, int func,
1549 unsigned long addr, unsigned long mask)
1550 {
1551 /* Invalid PE ? */
1552 if (!pe)
1553 return -ENODEV;
1554
1555 /* Unsupported operation ? */
1556 if (!eeh_ops || !eeh_ops->err_inject)
1557 return -ENOENT;
1558
1559 /* Check on PCI error type */
1560 if (type != EEH_ERR_TYPE_32 && type != EEH_ERR_TYPE_64)
1561 return -EINVAL;
1562
1563 /* Check on PCI error function */
1564 if (func < EEH_ERR_FUNC_MIN || func > EEH_ERR_FUNC_MAX)
1565 return -EINVAL;
1566
1567 return eeh_ops->err_inject(pe, type, func, addr, mask);
1568 }
1569 EXPORT_SYMBOL_GPL(eeh_pe_inject_err);
1570
proc_eeh_show(struct seq_file * m,void * v)1571 static int proc_eeh_show(struct seq_file *m, void *v)
1572 {
1573 if (!eeh_enabled()) {
1574 seq_printf(m, "EEH Subsystem is globally disabled\n");
1575 seq_printf(m, "eeh_total_mmio_ffs=%llu\n", eeh_stats.total_mmio_ffs);
1576 } else {
1577 seq_printf(m, "EEH Subsystem is enabled\n");
1578 seq_printf(m,
1579 "no device=%llu\n"
1580 "no device node=%llu\n"
1581 "no config address=%llu\n"
1582 "check not wanted=%llu\n"
1583 "eeh_total_mmio_ffs=%llu\n"
1584 "eeh_false_positives=%llu\n"
1585 "eeh_slot_resets=%llu\n",
1586 eeh_stats.no_device,
1587 eeh_stats.no_dn,
1588 eeh_stats.no_cfg_addr,
1589 eeh_stats.ignored_check,
1590 eeh_stats.total_mmio_ffs,
1591 eeh_stats.false_positives,
1592 eeh_stats.slot_resets);
1593 }
1594
1595 return 0;
1596 }
1597
1598 #ifdef CONFIG_DEBUG_FS
eeh_enable_dbgfs_set(void * data,u64 val)1599 static int eeh_enable_dbgfs_set(void *data, u64 val)
1600 {
1601 if (val)
1602 eeh_clear_flag(EEH_FORCE_DISABLED);
1603 else
1604 eeh_add_flag(EEH_FORCE_DISABLED);
1605
1606 return 0;
1607 }
1608
eeh_enable_dbgfs_get(void * data,u64 * val)1609 static int eeh_enable_dbgfs_get(void *data, u64 *val)
1610 {
1611 if (eeh_enabled())
1612 *val = 0x1ul;
1613 else
1614 *val = 0x0ul;
1615 return 0;
1616 }
1617
1618 DEFINE_DEBUGFS_ATTRIBUTE(eeh_enable_dbgfs_ops, eeh_enable_dbgfs_get,
1619 eeh_enable_dbgfs_set, "0x%llx\n");
1620
eeh_force_recover_write(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1621 static ssize_t eeh_force_recover_write(struct file *filp,
1622 const char __user *user_buf,
1623 size_t count, loff_t *ppos)
1624 {
1625 struct pci_controller *hose;
1626 uint32_t phbid, pe_no;
1627 struct eeh_pe *pe;
1628 char buf[20];
1629 int ret;
1630
1631 ret = simple_write_to_buffer(buf, sizeof(buf), ppos, user_buf, count);
1632 if (!ret)
1633 return -EFAULT;
1634
1635 /*
1636 * When PE is NULL the event is a "special" event. Rather than
1637 * recovering a specific PE it forces the EEH core to scan for failed
1638 * PHBs and recovers each. This needs to be done before any device
1639 * recoveries can occur.
1640 */
1641 if (!strncmp(buf, "hwcheck", 7)) {
1642 __eeh_send_failure_event(NULL);
1643 return count;
1644 }
1645
1646 ret = sscanf(buf, "%x:%x", &phbid, &pe_no);
1647 if (ret != 2)
1648 return -EINVAL;
1649
1650 hose = pci_find_controller_for_domain(phbid);
1651 if (!hose)
1652 return -ENODEV;
1653
1654 /* Retrieve PE */
1655 pe = eeh_pe_get(hose, pe_no);
1656 if (!pe)
1657 return -ENODEV;
1658
1659 /*
1660 * We don't do any state checking here since the detection
1661 * process is async to the recovery process. The recovery
1662 * thread *should* not break even if we schedule a recovery
1663 * from an odd state (e.g. PE removed, or recovery of a
1664 * non-isolated PE)
1665 */
1666 __eeh_send_failure_event(pe);
1667
1668 return ret < 0 ? ret : count;
1669 }
1670
1671 static const struct file_operations eeh_force_recover_fops = {
1672 .open = simple_open,
1673 .llseek = no_llseek,
1674 .write = eeh_force_recover_write,
1675 };
1676
eeh_debugfs_dev_usage(struct file * filp,char __user * user_buf,size_t count,loff_t * ppos)1677 static ssize_t eeh_debugfs_dev_usage(struct file *filp,
1678 char __user *user_buf,
1679 size_t count, loff_t *ppos)
1680 {
1681 static const char usage[] = "input format: <domain>:<bus>:<dev>.<fn>\n";
1682
1683 return simple_read_from_buffer(user_buf, count, ppos,
1684 usage, sizeof(usage) - 1);
1685 }
1686
eeh_dev_check_write(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1687 static ssize_t eeh_dev_check_write(struct file *filp,
1688 const char __user *user_buf,
1689 size_t count, loff_t *ppos)
1690 {
1691 uint32_t domain, bus, dev, fn;
1692 struct pci_dev *pdev;
1693 struct eeh_dev *edev;
1694 char buf[20];
1695 int ret;
1696
1697 memset(buf, 0, sizeof(buf));
1698 ret = simple_write_to_buffer(buf, sizeof(buf)-1, ppos, user_buf, count);
1699 if (!ret)
1700 return -EFAULT;
1701
1702 ret = sscanf(buf, "%x:%x:%x.%x", &domain, &bus, &dev, &fn);
1703 if (ret != 4) {
1704 pr_err("%s: expected 4 args, got %d\n", __func__, ret);
1705 return -EINVAL;
1706 }
1707
1708 pdev = pci_get_domain_bus_and_slot(domain, bus, (dev << 3) | fn);
1709 if (!pdev)
1710 return -ENODEV;
1711
1712 edev = pci_dev_to_eeh_dev(pdev);
1713 if (!edev) {
1714 pci_err(pdev, "No eeh_dev for this device!\n");
1715 pci_dev_put(pdev);
1716 return -ENODEV;
1717 }
1718
1719 ret = eeh_dev_check_failure(edev);
1720 pci_info(pdev, "eeh_dev_check_failure(%04x:%02x:%02x.%01x) = %d\n",
1721 domain, bus, dev, fn, ret);
1722
1723 pci_dev_put(pdev);
1724
1725 return count;
1726 }
1727
1728 static const struct file_operations eeh_dev_check_fops = {
1729 .open = simple_open,
1730 .llseek = no_llseek,
1731 .write = eeh_dev_check_write,
1732 .read = eeh_debugfs_dev_usage,
1733 };
1734
eeh_debugfs_break_device(struct pci_dev * pdev)1735 static int eeh_debugfs_break_device(struct pci_dev *pdev)
1736 {
1737 struct resource *bar = NULL;
1738 void __iomem *mapped;
1739 u16 old, bit;
1740 int i, pos;
1741
1742 /* Do we have an MMIO BAR to disable? */
1743 for (i = 0; i <= PCI_STD_RESOURCE_END; i++) {
1744 struct resource *r = &pdev->resource[i];
1745
1746 if (!r->flags || !r->start)
1747 continue;
1748 if (r->flags & IORESOURCE_IO)
1749 continue;
1750 if (r->flags & IORESOURCE_UNSET)
1751 continue;
1752
1753 bar = r;
1754 break;
1755 }
1756
1757 if (!bar) {
1758 pci_err(pdev, "Unable to find Memory BAR to cause EEH with\n");
1759 return -ENXIO;
1760 }
1761
1762 pci_err(pdev, "Going to break: %pR\n", bar);
1763
1764 if (pdev->is_virtfn) {
1765 #ifndef CONFIG_PCI_IOV
1766 return -ENXIO;
1767 #else
1768 /*
1769 * VFs don't have a per-function COMMAND register, so the best
1770 * we can do is clear the Memory Space Enable bit in the PF's
1771 * SRIOV control reg.
1772 *
1773 * Unfortunately, this requires that we have a PF (i.e doesn't
1774 * work for a passed-through VF) and it has the potential side
1775 * effect of also causing an EEH on every other VF under the
1776 * PF. Oh well.
1777 */
1778 pdev = pdev->physfn;
1779 if (!pdev)
1780 return -ENXIO; /* passed through VFs have no PF */
1781
1782 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_SRIOV);
1783 pos += PCI_SRIOV_CTRL;
1784 bit = PCI_SRIOV_CTRL_MSE;
1785 #endif /* !CONFIG_PCI_IOV */
1786 } else {
1787 bit = PCI_COMMAND_MEMORY;
1788 pos = PCI_COMMAND;
1789 }
1790
1791 /*
1792 * Process here is:
1793 *
1794 * 1. Disable Memory space.
1795 *
1796 * 2. Perform an MMIO to the device. This should result in an error
1797 * (CA / UR) being raised by the device which results in an EEH
1798 * PE freeze. Using the in_8() accessor skips the eeh detection hook
1799 * so the freeze hook so the EEH Detection machinery won't be
1800 * triggered here. This is to match the usual behaviour of EEH
1801 * where the HW will asyncronously freeze a PE and it's up to
1802 * the kernel to notice and deal with it.
1803 *
1804 * 3. Turn Memory space back on. This is more important for VFs
1805 * since recovery will probably fail if we don't. For normal
1806 * the COMMAND register is reset as a part of re-initialising
1807 * the device.
1808 *
1809 * Breaking stuff is the point so who cares if it's racy ;)
1810 */
1811 pci_read_config_word(pdev, pos, &old);
1812
1813 mapped = ioremap(bar->start, PAGE_SIZE);
1814 if (!mapped) {
1815 pci_err(pdev, "Unable to map MMIO BAR %pR\n", bar);
1816 return -ENXIO;
1817 }
1818
1819 pci_write_config_word(pdev, pos, old & ~bit);
1820 in_8(mapped);
1821 pci_write_config_word(pdev, pos, old);
1822
1823 iounmap(mapped);
1824
1825 return 0;
1826 }
1827
eeh_dev_break_write(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1828 static ssize_t eeh_dev_break_write(struct file *filp,
1829 const char __user *user_buf,
1830 size_t count, loff_t *ppos)
1831 {
1832 uint32_t domain, bus, dev, fn;
1833 struct pci_dev *pdev;
1834 char buf[20];
1835 int ret;
1836
1837 memset(buf, 0, sizeof(buf));
1838 ret = simple_write_to_buffer(buf, sizeof(buf)-1, ppos, user_buf, count);
1839 if (!ret)
1840 return -EFAULT;
1841
1842 ret = sscanf(buf, "%x:%x:%x.%x", &domain, &bus, &dev, &fn);
1843 if (ret != 4) {
1844 pr_err("%s: expected 4 args, got %d\n", __func__, ret);
1845 return -EINVAL;
1846 }
1847
1848 pdev = pci_get_domain_bus_and_slot(domain, bus, (dev << 3) | fn);
1849 if (!pdev)
1850 return -ENODEV;
1851
1852 ret = eeh_debugfs_break_device(pdev);
1853 pci_dev_put(pdev);
1854
1855 if (ret < 0)
1856 return ret;
1857
1858 return count;
1859 }
1860
1861 static const struct file_operations eeh_dev_break_fops = {
1862 .open = simple_open,
1863 .llseek = no_llseek,
1864 .write = eeh_dev_break_write,
1865 .read = eeh_debugfs_dev_usage,
1866 };
1867
1868 #endif
1869
eeh_init_proc(void)1870 static int __init eeh_init_proc(void)
1871 {
1872 if (machine_is(pseries) || machine_is(powernv)) {
1873 proc_create_single("powerpc/eeh", 0, NULL, proc_eeh_show);
1874 #ifdef CONFIG_DEBUG_FS
1875 debugfs_create_file_unsafe("eeh_enable", 0600,
1876 powerpc_debugfs_root, NULL,
1877 &eeh_enable_dbgfs_ops);
1878 debugfs_create_u32("eeh_max_freezes", 0600,
1879 powerpc_debugfs_root, &eeh_max_freezes);
1880 debugfs_create_bool("eeh_disable_recovery", 0600,
1881 powerpc_debugfs_root,
1882 &eeh_debugfs_no_recover);
1883 debugfs_create_file_unsafe("eeh_dev_check", 0600,
1884 powerpc_debugfs_root, NULL,
1885 &eeh_dev_check_fops);
1886 debugfs_create_file_unsafe("eeh_dev_break", 0600,
1887 powerpc_debugfs_root, NULL,
1888 &eeh_dev_break_fops);
1889 debugfs_create_file_unsafe("eeh_force_recover", 0600,
1890 powerpc_debugfs_root, NULL,
1891 &eeh_force_recover_fops);
1892 eeh_cache_debugfs_init();
1893 #endif
1894 }
1895
1896 return 0;
1897 }
1898 __initcall(eeh_init_proc);
1899