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 #include <linux/debugfs.h>
25
26 #include <linux/atomic.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 return ppc_find_vmap_phys(token);
350 }
351
352 /*
353 * On PowerNV platform, we might already have fenced PHB there.
354 * For that case, it's meaningless to recover frozen PE. Intead,
355 * We have to handle fenced PHB firstly.
356 */
eeh_phb_check_failure(struct eeh_pe * pe)357 static int eeh_phb_check_failure(struct eeh_pe *pe)
358 {
359 struct eeh_pe *phb_pe;
360 unsigned long flags;
361 int ret;
362
363 if (!eeh_has_flag(EEH_PROBE_MODE_DEV))
364 return -EPERM;
365
366 /* Find the PHB PE */
367 phb_pe = eeh_phb_pe_get(pe->phb);
368 if (!phb_pe) {
369 pr_warn("%s Can't find PE for PHB#%x\n",
370 __func__, pe->phb->global_number);
371 return -EEXIST;
372 }
373
374 /* If the PHB has been in problematic state */
375 eeh_serialize_lock(&flags);
376 if (phb_pe->state & EEH_PE_ISOLATED) {
377 ret = 0;
378 goto out;
379 }
380
381 /* Check PHB state */
382 ret = eeh_ops->get_state(phb_pe, NULL);
383 if ((ret < 0) ||
384 (ret == EEH_STATE_NOT_SUPPORT) || eeh_state_active(ret)) {
385 ret = 0;
386 goto out;
387 }
388
389 /* Isolate the PHB and send event */
390 eeh_pe_mark_isolated(phb_pe);
391 eeh_serialize_unlock(flags);
392
393 pr_debug("EEH: PHB#%x failure detected, location: %s\n",
394 phb_pe->phb->global_number, eeh_pe_loc_get(phb_pe));
395 eeh_send_failure_event(phb_pe);
396 return 1;
397 out:
398 eeh_serialize_unlock(flags);
399 return ret;
400 }
401
402 /**
403 * eeh_dev_check_failure - Check if all 1's data is due to EEH slot freeze
404 * @edev: eeh device
405 *
406 * Check for an EEH failure for the given device node. Call this
407 * routine if the result of a read was all 0xff's and you want to
408 * find out if this is due to an EEH slot freeze. This routine
409 * will query firmware for the EEH status.
410 *
411 * Returns 0 if there has not been an EEH error; otherwise returns
412 * a non-zero value and queues up a slot isolation event notification.
413 *
414 * It is safe to call this routine in an interrupt context.
415 */
eeh_dev_check_failure(struct eeh_dev * edev)416 int eeh_dev_check_failure(struct eeh_dev *edev)
417 {
418 int ret;
419 unsigned long flags;
420 struct device_node *dn;
421 struct pci_dev *dev;
422 struct eeh_pe *pe, *parent_pe;
423 int rc = 0;
424 const char *location = NULL;
425
426 eeh_stats.total_mmio_ffs++;
427
428 if (!eeh_enabled())
429 return 0;
430
431 if (!edev) {
432 eeh_stats.no_dn++;
433 return 0;
434 }
435 dev = eeh_dev_to_pci_dev(edev);
436 pe = eeh_dev_to_pe(edev);
437
438 /* Access to IO BARs might get this far and still not want checking. */
439 if (!pe) {
440 eeh_stats.ignored_check++;
441 eeh_edev_dbg(edev, "Ignored check\n");
442 return 0;
443 }
444
445 /*
446 * On PowerNV platform, we might already have fenced PHB
447 * there and we need take care of that firstly.
448 */
449 ret = eeh_phb_check_failure(pe);
450 if (ret > 0)
451 return ret;
452
453 /*
454 * If the PE isn't owned by us, we shouldn't check the
455 * state. Instead, let the owner handle it if the PE has
456 * been frozen.
457 */
458 if (eeh_pe_passed(pe))
459 return 0;
460
461 /* If we already have a pending isolation event for this
462 * slot, we know it's bad already, we don't need to check.
463 * Do this checking under a lock; as multiple PCI devices
464 * in one slot might report errors simultaneously, and we
465 * only want one error recovery routine running.
466 */
467 eeh_serialize_lock(&flags);
468 rc = 1;
469 if (pe->state & EEH_PE_ISOLATED) {
470 pe->check_count++;
471 if (pe->check_count == EEH_MAX_FAILS) {
472 dn = pci_device_to_OF_node(dev);
473 if (dn)
474 location = of_get_property(dn, "ibm,loc-code",
475 NULL);
476 eeh_edev_err(edev, "%d reads ignored for recovering device at location=%s driver=%s\n",
477 pe->check_count,
478 location ? location : "unknown",
479 eeh_driver_name(dev));
480 eeh_edev_err(edev, "Might be infinite loop in %s driver\n",
481 eeh_driver_name(dev));
482 dump_stack();
483 }
484 goto dn_unlock;
485 }
486
487 /*
488 * Now test for an EEH failure. This is VERY expensive.
489 * Note that the eeh_config_addr may be a parent device
490 * in the case of a device behind a bridge, or it may be
491 * function zero of a multi-function device.
492 * In any case they must share a common PHB.
493 */
494 ret = eeh_ops->get_state(pe, NULL);
495
496 /* Note that config-io to empty slots may fail;
497 * they are empty when they don't have children.
498 * We will punt with the following conditions: Failure to get
499 * PE's state, EEH not support and Permanently unavailable
500 * state, PE is in good state.
501 */
502 if ((ret < 0) ||
503 (ret == EEH_STATE_NOT_SUPPORT) || eeh_state_active(ret)) {
504 eeh_stats.false_positives++;
505 pe->false_positives++;
506 rc = 0;
507 goto dn_unlock;
508 }
509
510 /*
511 * It should be corner case that the parent PE has been
512 * put into frozen state as well. We should take care
513 * that at first.
514 */
515 parent_pe = pe->parent;
516 while (parent_pe) {
517 /* Hit the ceiling ? */
518 if (parent_pe->type & EEH_PE_PHB)
519 break;
520
521 /* Frozen parent PE ? */
522 ret = eeh_ops->get_state(parent_pe, NULL);
523 if (ret > 0 && !eeh_state_active(ret)) {
524 pe = parent_pe;
525 pr_err("EEH: Failure of PHB#%x-PE#%x will be handled at parent PHB#%x-PE#%x.\n",
526 pe->phb->global_number, pe->addr,
527 pe->phb->global_number, parent_pe->addr);
528 }
529
530 /* Next parent level */
531 parent_pe = parent_pe->parent;
532 }
533
534 eeh_stats.slot_resets++;
535
536 /* Avoid repeated reports of this failure, including problems
537 * with other functions on this device, and functions under
538 * bridges.
539 */
540 eeh_pe_mark_isolated(pe);
541 eeh_serialize_unlock(flags);
542
543 /* Most EEH events are due to device driver bugs. Having
544 * a stack trace will help the device-driver authors figure
545 * out what happened. So print that out.
546 */
547 pr_debug("EEH: %s: Frozen PHB#%x-PE#%x detected\n",
548 __func__, pe->phb->global_number, pe->addr);
549 eeh_send_failure_event(pe);
550
551 return 1;
552
553 dn_unlock:
554 eeh_serialize_unlock(flags);
555 return rc;
556 }
557
558 EXPORT_SYMBOL_GPL(eeh_dev_check_failure);
559
560 /**
561 * eeh_check_failure - Check if all 1's data is due to EEH slot freeze
562 * @token: I/O address
563 *
564 * Check for an EEH failure at the given I/O address. Call this
565 * routine if the result of a read was all 0xff's and you want to
566 * find out if this is due to an EEH slot freeze event. This routine
567 * will query firmware for the EEH status.
568 *
569 * Note this routine is safe to call in an interrupt context.
570 */
eeh_check_failure(const volatile void __iomem * token)571 int eeh_check_failure(const volatile void __iomem *token)
572 {
573 unsigned long addr;
574 struct eeh_dev *edev;
575
576 /* Finding the phys addr + pci device; this is pretty quick. */
577 addr = eeh_token_to_phys((unsigned long __force) token);
578 edev = eeh_addr_cache_get_dev(addr);
579 if (!edev) {
580 eeh_stats.no_device++;
581 return 0;
582 }
583
584 return eeh_dev_check_failure(edev);
585 }
586 EXPORT_SYMBOL(eeh_check_failure);
587
588
589 /**
590 * eeh_pci_enable - Enable MMIO or DMA transfers for this slot
591 * @pe: EEH PE
592 *
593 * This routine should be called to reenable frozen MMIO or DMA
594 * so that it would work correctly again. It's useful while doing
595 * recovery or log collection on the indicated device.
596 */
eeh_pci_enable(struct eeh_pe * pe,int function)597 int eeh_pci_enable(struct eeh_pe *pe, int function)
598 {
599 int active_flag, rc;
600
601 /*
602 * pHyp doesn't allow to enable IO or DMA on unfrozen PE.
603 * Also, it's pointless to enable them on unfrozen PE. So
604 * we have to check before enabling IO or DMA.
605 */
606 switch (function) {
607 case EEH_OPT_THAW_MMIO:
608 active_flag = EEH_STATE_MMIO_ACTIVE | EEH_STATE_MMIO_ENABLED;
609 break;
610 case EEH_OPT_THAW_DMA:
611 active_flag = EEH_STATE_DMA_ACTIVE;
612 break;
613 case EEH_OPT_DISABLE:
614 case EEH_OPT_ENABLE:
615 case EEH_OPT_FREEZE_PE:
616 active_flag = 0;
617 break;
618 default:
619 pr_warn("%s: Invalid function %d\n",
620 __func__, function);
621 return -EINVAL;
622 }
623
624 /*
625 * Check if IO or DMA has been enabled before
626 * enabling them.
627 */
628 if (active_flag) {
629 rc = eeh_ops->get_state(pe, NULL);
630 if (rc < 0)
631 return rc;
632
633 /* Needn't enable it at all */
634 if (rc == EEH_STATE_NOT_SUPPORT)
635 return 0;
636
637 /* It's already enabled */
638 if (rc & active_flag)
639 return 0;
640 }
641
642
643 /* Issue the request */
644 rc = eeh_ops->set_option(pe, function);
645 if (rc)
646 pr_warn("%s: Unexpected state change %d on "
647 "PHB#%x-PE#%x, err=%d\n",
648 __func__, function, pe->phb->global_number,
649 pe->addr, rc);
650
651 /* Check if the request is finished successfully */
652 if (active_flag) {
653 rc = eeh_wait_state(pe, PCI_BUS_RESET_WAIT_MSEC);
654 if (rc < 0)
655 return rc;
656
657 if (rc & active_flag)
658 return 0;
659
660 return -EIO;
661 }
662
663 return rc;
664 }
665
eeh_disable_and_save_dev_state(struct eeh_dev * edev,void * userdata)666 static void eeh_disable_and_save_dev_state(struct eeh_dev *edev,
667 void *userdata)
668 {
669 struct pci_dev *pdev = eeh_dev_to_pci_dev(edev);
670 struct pci_dev *dev = userdata;
671
672 /*
673 * The caller should have disabled and saved the
674 * state for the specified device
675 */
676 if (!pdev || pdev == dev)
677 return;
678
679 /* Ensure we have D0 power state */
680 pci_set_power_state(pdev, PCI_D0);
681
682 /* Save device state */
683 pci_save_state(pdev);
684
685 /*
686 * Disable device to avoid any DMA traffic and
687 * interrupt from the device
688 */
689 pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
690 }
691
eeh_restore_dev_state(struct eeh_dev * edev,void * userdata)692 static void eeh_restore_dev_state(struct eeh_dev *edev, void *userdata)
693 {
694 struct pci_dev *pdev = eeh_dev_to_pci_dev(edev);
695 struct pci_dev *dev = userdata;
696
697 if (!pdev)
698 return;
699
700 /* Apply customization from firmware */
701 if (eeh_ops->restore_config)
702 eeh_ops->restore_config(edev);
703
704 /* The caller should restore state for the specified device */
705 if (pdev != dev)
706 pci_restore_state(pdev);
707 }
708
709 /**
710 * pcibios_set_pcie_reset_state - Set PCI-E reset state
711 * @dev: pci device struct
712 * @state: reset state to enter
713 *
714 * Return value:
715 * 0 if success
716 */
pcibios_set_pcie_reset_state(struct pci_dev * dev,enum pcie_reset_state state)717 int pcibios_set_pcie_reset_state(struct pci_dev *dev, enum pcie_reset_state state)
718 {
719 struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
720 struct eeh_pe *pe = eeh_dev_to_pe(edev);
721
722 if (!pe) {
723 pr_err("%s: No PE found on PCI device %s\n",
724 __func__, pci_name(dev));
725 return -EINVAL;
726 }
727
728 switch (state) {
729 case pcie_deassert_reset:
730 eeh_ops->reset(pe, EEH_RESET_DEACTIVATE);
731 eeh_unfreeze_pe(pe);
732 if (!(pe->type & EEH_PE_VF))
733 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, true);
734 eeh_pe_dev_traverse(pe, eeh_restore_dev_state, dev);
735 eeh_pe_state_clear(pe, EEH_PE_ISOLATED, true);
736 break;
737 case pcie_hot_reset:
738 eeh_pe_mark_isolated(pe);
739 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, true);
740 eeh_ops->set_option(pe, EEH_OPT_FREEZE_PE);
741 eeh_pe_dev_traverse(pe, eeh_disable_and_save_dev_state, dev);
742 if (!(pe->type & EEH_PE_VF))
743 eeh_pe_state_mark(pe, EEH_PE_CFG_BLOCKED);
744 eeh_ops->reset(pe, EEH_RESET_HOT);
745 break;
746 case pcie_warm_reset:
747 eeh_pe_mark_isolated(pe);
748 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, true);
749 eeh_ops->set_option(pe, EEH_OPT_FREEZE_PE);
750 eeh_pe_dev_traverse(pe, eeh_disable_and_save_dev_state, dev);
751 if (!(pe->type & EEH_PE_VF))
752 eeh_pe_state_mark(pe, EEH_PE_CFG_BLOCKED);
753 eeh_ops->reset(pe, EEH_RESET_FUNDAMENTAL);
754 break;
755 default:
756 eeh_pe_state_clear(pe, EEH_PE_ISOLATED | EEH_PE_CFG_BLOCKED, true);
757 return -EINVAL;
758 }
759
760 return 0;
761 }
762
763 /**
764 * eeh_set_pe_freset - Check the required reset for the indicated device
765 * @data: EEH device
766 * @flag: return value
767 *
768 * Each device might have its preferred reset type: fundamental or
769 * hot reset. The routine is used to collected the information for
770 * the indicated device and its children so that the bunch of the
771 * devices could be reset properly.
772 */
eeh_set_dev_freset(struct eeh_dev * edev,void * flag)773 static void eeh_set_dev_freset(struct eeh_dev *edev, void *flag)
774 {
775 struct pci_dev *dev;
776 unsigned int *freset = (unsigned int *)flag;
777
778 dev = eeh_dev_to_pci_dev(edev);
779 if (dev)
780 *freset |= dev->needs_freset;
781 }
782
eeh_pe_refreeze_passed(struct eeh_pe * root)783 static void eeh_pe_refreeze_passed(struct eeh_pe *root)
784 {
785 struct eeh_pe *pe;
786 int state;
787
788 eeh_for_each_pe(root, pe) {
789 if (eeh_pe_passed(pe)) {
790 state = eeh_ops->get_state(pe, NULL);
791 if (state &
792 (EEH_STATE_MMIO_ACTIVE | EEH_STATE_MMIO_ENABLED)) {
793 pr_info("EEH: Passed-through PE PHB#%x-PE#%x was thawed by reset, re-freezing for safety.\n",
794 pe->phb->global_number, pe->addr);
795 eeh_pe_set_option(pe, EEH_OPT_FREEZE_PE);
796 }
797 }
798 }
799 }
800
801 /**
802 * eeh_pe_reset_full - Complete a full reset process on the indicated PE
803 * @pe: EEH PE
804 *
805 * This function executes a full reset procedure on a PE, including setting
806 * the appropriate flags, performing a fundamental or hot reset, and then
807 * deactivating the reset status. It is designed to be used within the EEH
808 * subsystem, as opposed to eeh_pe_reset which is exported to drivers and
809 * only performs a single operation at a time.
810 *
811 * This function will attempt to reset a PE three times before failing.
812 */
eeh_pe_reset_full(struct eeh_pe * pe,bool include_passed)813 int eeh_pe_reset_full(struct eeh_pe *pe, bool include_passed)
814 {
815 int reset_state = (EEH_PE_RESET | EEH_PE_CFG_BLOCKED);
816 int type = EEH_RESET_HOT;
817 unsigned int freset = 0;
818 int i, state = 0, ret;
819
820 /*
821 * Determine the type of reset to perform - hot or fundamental.
822 * Hot reset is the default operation, unless any device under the
823 * PE requires a fundamental reset.
824 */
825 eeh_pe_dev_traverse(pe, eeh_set_dev_freset, &freset);
826
827 if (freset)
828 type = EEH_RESET_FUNDAMENTAL;
829
830 /* Mark the PE as in reset state and block config space accesses */
831 eeh_pe_state_mark(pe, reset_state);
832
833 /* Make three attempts at resetting the bus */
834 for (i = 0; i < 3; i++) {
835 ret = eeh_pe_reset(pe, type, include_passed);
836 if (!ret)
837 ret = eeh_pe_reset(pe, EEH_RESET_DEACTIVATE,
838 include_passed);
839 if (ret) {
840 ret = -EIO;
841 pr_warn("EEH: Failure %d resetting PHB#%x-PE#%x (attempt %d)\n\n",
842 state, pe->phb->global_number, pe->addr, i + 1);
843 continue;
844 }
845 if (i)
846 pr_warn("EEH: PHB#%x-PE#%x: Successful reset (attempt %d)\n",
847 pe->phb->global_number, pe->addr, i + 1);
848
849 /* Wait until the PE is in a functioning state */
850 state = eeh_wait_state(pe, PCI_BUS_RESET_WAIT_MSEC);
851 if (state < 0) {
852 pr_warn("EEH: Unrecoverable slot failure on PHB#%x-PE#%x",
853 pe->phb->global_number, pe->addr);
854 ret = -ENOTRECOVERABLE;
855 break;
856 }
857 if (eeh_state_active(state))
858 break;
859 else
860 pr_warn("EEH: PHB#%x-PE#%x: Slot inactive after reset: 0x%x (attempt %d)\n",
861 pe->phb->global_number, pe->addr, state, i + 1);
862 }
863
864 /* Resetting the PE may have unfrozen child PEs. If those PEs have been
865 * (potentially) passed through to a guest, re-freeze them:
866 */
867 if (!include_passed)
868 eeh_pe_refreeze_passed(pe);
869
870 eeh_pe_state_clear(pe, reset_state, true);
871 return ret;
872 }
873
874 /**
875 * eeh_save_bars - Save device bars
876 * @edev: PCI device associated EEH device
877 *
878 * Save the values of the device bars. Unlike the restore
879 * routine, this routine is *not* recursive. This is because
880 * PCI devices are added individually; but, for the restore,
881 * an entire slot is reset at a time.
882 */
eeh_save_bars(struct eeh_dev * edev)883 void eeh_save_bars(struct eeh_dev *edev)
884 {
885 int i;
886
887 if (!edev)
888 return;
889
890 for (i = 0; i < 16; i++)
891 eeh_ops->read_config(edev, i * 4, 4, &edev->config_space[i]);
892
893 /*
894 * For PCI bridges including root port, we need enable bus
895 * master explicitly. Otherwise, it can't fetch IODA table
896 * entries correctly. So we cache the bit in advance so that
897 * we can restore it after reset, either PHB range or PE range.
898 */
899 if (edev->mode & EEH_DEV_BRIDGE)
900 edev->config_space[1] |= PCI_COMMAND_MASTER;
901 }
902
eeh_reboot_notifier(struct notifier_block * nb,unsigned long action,void * unused)903 static int eeh_reboot_notifier(struct notifier_block *nb,
904 unsigned long action, void *unused)
905 {
906 eeh_clear_flag(EEH_ENABLED);
907 return NOTIFY_DONE;
908 }
909
910 static struct notifier_block eeh_reboot_nb = {
911 .notifier_call = eeh_reboot_notifier,
912 };
913
eeh_device_notifier(struct notifier_block * nb,unsigned long action,void * data)914 static int eeh_device_notifier(struct notifier_block *nb,
915 unsigned long action, void *data)
916 {
917 struct device *dev = data;
918
919 switch (action) {
920 /*
921 * Note: It's not possible to perform EEH device addition (i.e.
922 * {pseries,pnv}_pcibios_bus_add_device()) here because it depends on
923 * the device's resources, which have not yet been set up.
924 */
925 case BUS_NOTIFY_DEL_DEVICE:
926 eeh_remove_device(to_pci_dev(dev));
927 break;
928 default:
929 break;
930 }
931 return NOTIFY_DONE;
932 }
933
934 static struct notifier_block eeh_device_nb = {
935 .notifier_call = eeh_device_notifier,
936 };
937
938 /**
939 * eeh_init - System wide EEH initialization
940 *
941 * It's the platform's job to call this from an arch_initcall().
942 */
eeh_init(struct eeh_ops * ops)943 int eeh_init(struct eeh_ops *ops)
944 {
945 struct pci_controller *hose, *tmp;
946 int ret = 0;
947
948 /* the platform should only initialise EEH once */
949 if (WARN_ON(eeh_ops))
950 return -EEXIST;
951 if (WARN_ON(!ops))
952 return -ENOENT;
953 eeh_ops = ops;
954
955 /* Register reboot notifier */
956 ret = register_reboot_notifier(&eeh_reboot_nb);
957 if (ret) {
958 pr_warn("%s: Failed to register reboot notifier (%d)\n",
959 __func__, ret);
960 return ret;
961 }
962
963 ret = bus_register_notifier(&pci_bus_type, &eeh_device_nb);
964 if (ret) {
965 pr_warn("%s: Failed to register bus notifier (%d)\n",
966 __func__, ret);
967 return ret;
968 }
969
970 /* Initialize PHB PEs */
971 list_for_each_entry_safe(hose, tmp, &hose_list, list_node)
972 eeh_phb_pe_create(hose);
973
974 eeh_addr_cache_init();
975
976 /* Initialize EEH event */
977 return eeh_event_init();
978 }
979
980 /**
981 * eeh_probe_device() - Perform EEH initialization for the indicated pci device
982 * @dev: pci device for which to set up EEH
983 *
984 * This routine must be used to complete EEH initialization for PCI
985 * devices that were added after system boot (e.g. hotplug, dlpar).
986 */
eeh_probe_device(struct pci_dev * dev)987 void eeh_probe_device(struct pci_dev *dev)
988 {
989 struct eeh_dev *edev;
990
991 pr_debug("EEH: Adding device %s\n", pci_name(dev));
992
993 /*
994 * pci_dev_to_eeh_dev() can only work if eeh_probe_dev() was
995 * already called for this device.
996 */
997 if (WARN_ON_ONCE(pci_dev_to_eeh_dev(dev))) {
998 pci_dbg(dev, "Already bound to an eeh_dev!\n");
999 return;
1000 }
1001
1002 edev = eeh_ops->probe(dev);
1003 if (!edev) {
1004 pr_debug("EEH: Adding device failed\n");
1005 return;
1006 }
1007
1008 /*
1009 * FIXME: We rely on pcibios_release_device() to remove the
1010 * existing EEH state. The release function is only called if
1011 * the pci_dev's refcount drops to zero so if something is
1012 * keeping a ref to a device (e.g. a filesystem) we need to
1013 * remove the old EEH state.
1014 *
1015 * FIXME: HEY MA, LOOK AT ME, NO LOCKING!
1016 */
1017 if (edev->pdev && edev->pdev != dev) {
1018 eeh_pe_tree_remove(edev);
1019 eeh_addr_cache_rmv_dev(edev->pdev);
1020 eeh_sysfs_remove_device(edev->pdev);
1021
1022 /*
1023 * We definitely should have the PCI device removed
1024 * though it wasn't correctly. So we needn't call
1025 * into error handler afterwards.
1026 */
1027 edev->mode |= EEH_DEV_NO_HANDLER;
1028 }
1029
1030 /* bind the pdev and the edev together */
1031 edev->pdev = dev;
1032 dev->dev.archdata.edev = edev;
1033 eeh_addr_cache_insert_dev(dev);
1034 eeh_sysfs_add_device(dev);
1035 }
1036
1037 /**
1038 * eeh_remove_device - Undo EEH setup for the indicated pci device
1039 * @dev: pci device to be removed
1040 *
1041 * This routine should be called when a device is removed from
1042 * a running system (e.g. by hotplug or dlpar). It unregisters
1043 * the PCI device from the EEH subsystem. I/O errors affecting
1044 * this device will no longer be detected after this call; thus,
1045 * i/o errors affecting this slot may leave this device unusable.
1046 */
eeh_remove_device(struct pci_dev * dev)1047 void eeh_remove_device(struct pci_dev *dev)
1048 {
1049 struct eeh_dev *edev;
1050
1051 if (!dev || !eeh_enabled())
1052 return;
1053 edev = pci_dev_to_eeh_dev(dev);
1054
1055 /* Unregister the device with the EEH/PCI address search system */
1056 dev_dbg(&dev->dev, "EEH: Removing device\n");
1057
1058 if (!edev || !edev->pdev || !edev->pe) {
1059 dev_dbg(&dev->dev, "EEH: Device not referenced!\n");
1060 return;
1061 }
1062
1063 /*
1064 * During the hotplug for EEH error recovery, we need the EEH
1065 * device attached to the parent PE in order for BAR restore
1066 * a bit later. So we keep it for BAR restore and remove it
1067 * from the parent PE during the BAR resotre.
1068 */
1069 edev->pdev = NULL;
1070
1071 /*
1072 * eeh_sysfs_remove_device() uses pci_dev_to_eeh_dev() so we need to
1073 * remove the sysfs files before clearing dev.archdata.edev
1074 */
1075 if (edev->mode & EEH_DEV_SYSFS)
1076 eeh_sysfs_remove_device(dev);
1077
1078 /*
1079 * We're removing from the PCI subsystem, that means
1080 * the PCI device driver can't support EEH or not
1081 * well. So we rely on hotplug completely to do recovery
1082 * for the specific PCI device.
1083 */
1084 edev->mode |= EEH_DEV_NO_HANDLER;
1085
1086 eeh_addr_cache_rmv_dev(dev);
1087
1088 /*
1089 * The flag "in_error" is used to trace EEH devices for VFs
1090 * in error state or not. It's set in eeh_report_error(). If
1091 * it's not set, eeh_report_{reset,resume}() won't be called
1092 * for the VF EEH device.
1093 */
1094 edev->in_error = false;
1095 dev->dev.archdata.edev = NULL;
1096 if (!(edev->pe->state & EEH_PE_KEEP))
1097 eeh_pe_tree_remove(edev);
1098 else
1099 edev->mode |= EEH_DEV_DISCONNECTED;
1100 }
1101
eeh_unfreeze_pe(struct eeh_pe * pe)1102 int eeh_unfreeze_pe(struct eeh_pe *pe)
1103 {
1104 int ret;
1105
1106 ret = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
1107 if (ret) {
1108 pr_warn("%s: Failure %d enabling IO on PHB#%x-PE#%x\n",
1109 __func__, ret, pe->phb->global_number, pe->addr);
1110 return ret;
1111 }
1112
1113 ret = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
1114 if (ret) {
1115 pr_warn("%s: Failure %d enabling DMA on PHB#%x-PE#%x\n",
1116 __func__, ret, pe->phb->global_number, pe->addr);
1117 return ret;
1118 }
1119
1120 return ret;
1121 }
1122
1123
1124 static struct pci_device_id eeh_reset_ids[] = {
1125 { PCI_DEVICE(0x19a2, 0x0710) }, /* Emulex, BE */
1126 { PCI_DEVICE(0x10df, 0xe220) }, /* Emulex, Lancer */
1127 { PCI_DEVICE(0x14e4, 0x1657) }, /* Broadcom BCM5719 */
1128 { 0 }
1129 };
1130
eeh_pe_change_owner(struct eeh_pe * pe)1131 static int eeh_pe_change_owner(struct eeh_pe *pe)
1132 {
1133 struct eeh_dev *edev, *tmp;
1134 struct pci_dev *pdev;
1135 struct pci_device_id *id;
1136 int ret;
1137
1138 /* Check PE state */
1139 ret = eeh_ops->get_state(pe, NULL);
1140 if (ret < 0 || ret == EEH_STATE_NOT_SUPPORT)
1141 return 0;
1142
1143 /* Unfrozen PE, nothing to do */
1144 if (eeh_state_active(ret))
1145 return 0;
1146
1147 /* Frozen PE, check if it needs PE level reset */
1148 eeh_pe_for_each_dev(pe, edev, tmp) {
1149 pdev = eeh_dev_to_pci_dev(edev);
1150 if (!pdev)
1151 continue;
1152
1153 for (id = &eeh_reset_ids[0]; id->vendor != 0; id++) {
1154 if (id->vendor != PCI_ANY_ID &&
1155 id->vendor != pdev->vendor)
1156 continue;
1157 if (id->device != PCI_ANY_ID &&
1158 id->device != pdev->device)
1159 continue;
1160 if (id->subvendor != PCI_ANY_ID &&
1161 id->subvendor != pdev->subsystem_vendor)
1162 continue;
1163 if (id->subdevice != PCI_ANY_ID &&
1164 id->subdevice != pdev->subsystem_device)
1165 continue;
1166
1167 return eeh_pe_reset_and_recover(pe);
1168 }
1169 }
1170
1171 ret = eeh_unfreeze_pe(pe);
1172 if (!ret)
1173 eeh_pe_state_clear(pe, EEH_PE_ISOLATED, true);
1174 return ret;
1175 }
1176
1177 /**
1178 * eeh_dev_open - Increase count of pass through devices for PE
1179 * @pdev: PCI device
1180 *
1181 * Increase count of passed through devices for the indicated
1182 * PE. In the result, the EEH errors detected on the PE won't be
1183 * reported. The PE owner will be responsible for detection
1184 * and recovery.
1185 */
eeh_dev_open(struct pci_dev * pdev)1186 int eeh_dev_open(struct pci_dev *pdev)
1187 {
1188 struct eeh_dev *edev;
1189 int ret = -ENODEV;
1190
1191 mutex_lock(&eeh_dev_mutex);
1192
1193 /* No PCI device ? */
1194 if (!pdev)
1195 goto out;
1196
1197 /* No EEH device or PE ? */
1198 edev = pci_dev_to_eeh_dev(pdev);
1199 if (!edev || !edev->pe)
1200 goto out;
1201
1202 /*
1203 * The PE might have been put into frozen state, but we
1204 * didn't detect that yet. The passed through PCI devices
1205 * in frozen PE won't work properly. Clear the frozen state
1206 * in advance.
1207 */
1208 ret = eeh_pe_change_owner(edev->pe);
1209 if (ret)
1210 goto out;
1211
1212 /* Increase PE's pass through count */
1213 atomic_inc(&edev->pe->pass_dev_cnt);
1214 mutex_unlock(&eeh_dev_mutex);
1215
1216 return 0;
1217 out:
1218 mutex_unlock(&eeh_dev_mutex);
1219 return ret;
1220 }
1221 EXPORT_SYMBOL_GPL(eeh_dev_open);
1222
1223 /**
1224 * eeh_dev_release - Decrease count of pass through devices for PE
1225 * @pdev: PCI device
1226 *
1227 * Decrease count of pass through devices for the indicated PE. If
1228 * there is no passed through device in PE, the EEH errors detected
1229 * on the PE will be reported and handled as usual.
1230 */
eeh_dev_release(struct pci_dev * pdev)1231 void eeh_dev_release(struct pci_dev *pdev)
1232 {
1233 struct eeh_dev *edev;
1234
1235 mutex_lock(&eeh_dev_mutex);
1236
1237 /* No PCI device ? */
1238 if (!pdev)
1239 goto out;
1240
1241 /* No EEH device ? */
1242 edev = pci_dev_to_eeh_dev(pdev);
1243 if (!edev || !edev->pe || !eeh_pe_passed(edev->pe))
1244 goto out;
1245
1246 /* Decrease PE's pass through count */
1247 WARN_ON(atomic_dec_if_positive(&edev->pe->pass_dev_cnt) < 0);
1248 eeh_pe_change_owner(edev->pe);
1249 out:
1250 mutex_unlock(&eeh_dev_mutex);
1251 }
1252 EXPORT_SYMBOL(eeh_dev_release);
1253
1254 #ifdef CONFIG_IOMMU_API
1255
dev_has_iommu_table(struct device * dev,void * data)1256 static int dev_has_iommu_table(struct device *dev, void *data)
1257 {
1258 struct pci_dev *pdev = to_pci_dev(dev);
1259 struct pci_dev **ppdev = data;
1260
1261 if (!dev)
1262 return 0;
1263
1264 if (device_iommu_mapped(dev)) {
1265 *ppdev = pdev;
1266 return 1;
1267 }
1268
1269 return 0;
1270 }
1271
1272 /**
1273 * eeh_iommu_group_to_pe - Convert IOMMU group to EEH PE
1274 * @group: IOMMU group
1275 *
1276 * The routine is called to convert IOMMU group to EEH PE.
1277 */
eeh_iommu_group_to_pe(struct iommu_group * group)1278 struct eeh_pe *eeh_iommu_group_to_pe(struct iommu_group *group)
1279 {
1280 struct pci_dev *pdev = NULL;
1281 struct eeh_dev *edev;
1282 int ret;
1283
1284 /* No IOMMU group ? */
1285 if (!group)
1286 return NULL;
1287
1288 ret = iommu_group_for_each_dev(group, &pdev, dev_has_iommu_table);
1289 if (!ret || !pdev)
1290 return NULL;
1291
1292 /* No EEH device or PE ? */
1293 edev = pci_dev_to_eeh_dev(pdev);
1294 if (!edev || !edev->pe)
1295 return NULL;
1296
1297 return edev->pe;
1298 }
1299 EXPORT_SYMBOL_GPL(eeh_iommu_group_to_pe);
1300
1301 #endif /* CONFIG_IOMMU_API */
1302
1303 /**
1304 * eeh_pe_set_option - Set options for the indicated PE
1305 * @pe: EEH PE
1306 * @option: requested option
1307 *
1308 * The routine is called to enable or disable EEH functionality
1309 * on the indicated PE, to enable IO or DMA for the frozen PE.
1310 */
eeh_pe_set_option(struct eeh_pe * pe,int option)1311 int eeh_pe_set_option(struct eeh_pe *pe, int option)
1312 {
1313 int ret = 0;
1314
1315 /* Invalid PE ? */
1316 if (!pe)
1317 return -ENODEV;
1318
1319 /*
1320 * EEH functionality could possibly be disabled, just
1321 * return error for the case. And the EEH functinality
1322 * isn't expected to be disabled on one specific PE.
1323 */
1324 switch (option) {
1325 case EEH_OPT_ENABLE:
1326 if (eeh_enabled()) {
1327 ret = eeh_pe_change_owner(pe);
1328 break;
1329 }
1330 ret = -EIO;
1331 break;
1332 case EEH_OPT_DISABLE:
1333 break;
1334 case EEH_OPT_THAW_MMIO:
1335 case EEH_OPT_THAW_DMA:
1336 case EEH_OPT_FREEZE_PE:
1337 if (!eeh_ops || !eeh_ops->set_option) {
1338 ret = -ENOENT;
1339 break;
1340 }
1341
1342 ret = eeh_pci_enable(pe, option);
1343 break;
1344 default:
1345 pr_debug("%s: Option %d out of range (%d, %d)\n",
1346 __func__, option, EEH_OPT_DISABLE, EEH_OPT_THAW_DMA);
1347 ret = -EINVAL;
1348 }
1349
1350 return ret;
1351 }
1352 EXPORT_SYMBOL_GPL(eeh_pe_set_option);
1353
1354 /**
1355 * eeh_pe_get_state - Retrieve PE's state
1356 * @pe: EEH PE
1357 *
1358 * Retrieve the PE's state, which includes 3 aspects: enabled
1359 * DMA, enabled IO and asserted reset.
1360 */
eeh_pe_get_state(struct eeh_pe * pe)1361 int eeh_pe_get_state(struct eeh_pe *pe)
1362 {
1363 int result, ret = 0;
1364 bool rst_active, dma_en, mmio_en;
1365
1366 /* Existing PE ? */
1367 if (!pe)
1368 return -ENODEV;
1369
1370 if (!eeh_ops || !eeh_ops->get_state)
1371 return -ENOENT;
1372
1373 /*
1374 * If the parent PE is owned by the host kernel and is undergoing
1375 * error recovery, we should return the PE state as temporarily
1376 * unavailable so that the error recovery on the guest is suspended
1377 * until the recovery completes on the host.
1378 */
1379 if (pe->parent &&
1380 !(pe->state & EEH_PE_REMOVED) &&
1381 (pe->parent->state & (EEH_PE_ISOLATED | EEH_PE_RECOVERING)))
1382 return EEH_PE_STATE_UNAVAIL;
1383
1384 result = eeh_ops->get_state(pe, NULL);
1385 rst_active = !!(result & EEH_STATE_RESET_ACTIVE);
1386 dma_en = !!(result & EEH_STATE_DMA_ENABLED);
1387 mmio_en = !!(result & EEH_STATE_MMIO_ENABLED);
1388
1389 if (rst_active)
1390 ret = EEH_PE_STATE_RESET;
1391 else if (dma_en && mmio_en)
1392 ret = EEH_PE_STATE_NORMAL;
1393 else if (!dma_en && !mmio_en)
1394 ret = EEH_PE_STATE_STOPPED_IO_DMA;
1395 else if (!dma_en && mmio_en)
1396 ret = EEH_PE_STATE_STOPPED_DMA;
1397 else
1398 ret = EEH_PE_STATE_UNAVAIL;
1399
1400 return ret;
1401 }
1402 EXPORT_SYMBOL_GPL(eeh_pe_get_state);
1403
eeh_pe_reenable_devices(struct eeh_pe * pe,bool include_passed)1404 static int eeh_pe_reenable_devices(struct eeh_pe *pe, bool include_passed)
1405 {
1406 struct eeh_dev *edev, *tmp;
1407 struct pci_dev *pdev;
1408 int ret = 0;
1409
1410 eeh_pe_restore_bars(pe);
1411
1412 /*
1413 * Reenable PCI devices as the devices passed
1414 * through are always enabled before the reset.
1415 */
1416 eeh_pe_for_each_dev(pe, edev, tmp) {
1417 pdev = eeh_dev_to_pci_dev(edev);
1418 if (!pdev)
1419 continue;
1420
1421 ret = pci_reenable_device(pdev);
1422 if (ret) {
1423 pr_warn("%s: Failure %d reenabling %s\n",
1424 __func__, ret, pci_name(pdev));
1425 return ret;
1426 }
1427 }
1428
1429 /* The PE is still in frozen state */
1430 if (include_passed || !eeh_pe_passed(pe)) {
1431 ret = eeh_unfreeze_pe(pe);
1432 } else
1433 pr_info("EEH: Note: Leaving passthrough PHB#%x-PE#%x frozen.\n",
1434 pe->phb->global_number, pe->addr);
1435 if (!ret)
1436 eeh_pe_state_clear(pe, EEH_PE_ISOLATED, include_passed);
1437 return ret;
1438 }
1439
1440
1441 /**
1442 * eeh_pe_reset - Issue PE reset according to specified type
1443 * @pe: EEH PE
1444 * @option: reset type
1445 *
1446 * The routine is called to reset the specified PE with the
1447 * indicated type, either fundamental reset or hot reset.
1448 * PE reset is the most important part for error recovery.
1449 */
eeh_pe_reset(struct eeh_pe * pe,int option,bool include_passed)1450 int eeh_pe_reset(struct eeh_pe *pe, int option, bool include_passed)
1451 {
1452 int ret = 0;
1453
1454 /* Invalid PE ? */
1455 if (!pe)
1456 return -ENODEV;
1457
1458 if (!eeh_ops || !eeh_ops->set_option || !eeh_ops->reset)
1459 return -ENOENT;
1460
1461 switch (option) {
1462 case EEH_RESET_DEACTIVATE:
1463 ret = eeh_ops->reset(pe, option);
1464 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, include_passed);
1465 if (ret)
1466 break;
1467
1468 ret = eeh_pe_reenable_devices(pe, include_passed);
1469 break;
1470 case EEH_RESET_HOT:
1471 case EEH_RESET_FUNDAMENTAL:
1472 /*
1473 * Proactively freeze the PE to drop all MMIO access
1474 * during reset, which should be banned as it's always
1475 * cause recursive EEH error.
1476 */
1477 eeh_ops->set_option(pe, EEH_OPT_FREEZE_PE);
1478
1479 eeh_pe_state_mark(pe, EEH_PE_CFG_BLOCKED);
1480 ret = eeh_ops->reset(pe, option);
1481 break;
1482 default:
1483 pr_debug("%s: Unsupported option %d\n",
1484 __func__, option);
1485 ret = -EINVAL;
1486 }
1487
1488 return ret;
1489 }
1490 EXPORT_SYMBOL_GPL(eeh_pe_reset);
1491
1492 /**
1493 * eeh_pe_configure - Configure PCI bridges after PE reset
1494 * @pe: EEH PE
1495 *
1496 * The routine is called to restore the PCI config space for
1497 * those PCI devices, especially PCI bridges affected by PE
1498 * reset issued previously.
1499 */
eeh_pe_configure(struct eeh_pe * pe)1500 int eeh_pe_configure(struct eeh_pe *pe)
1501 {
1502 int ret = 0;
1503
1504 /* Invalid PE ? */
1505 if (!pe)
1506 return -ENODEV;
1507
1508 return ret;
1509 }
1510 EXPORT_SYMBOL_GPL(eeh_pe_configure);
1511
1512 /**
1513 * eeh_pe_inject_err - Injecting the specified PCI error to the indicated PE
1514 * @pe: the indicated PE
1515 * @type: error type
1516 * @function: error function
1517 * @addr: address
1518 * @mask: address mask
1519 *
1520 * The routine is called to inject the specified PCI error, which
1521 * is determined by @type and @function, to the indicated PE for
1522 * testing purpose.
1523 */
eeh_pe_inject_err(struct eeh_pe * pe,int type,int func,unsigned long addr,unsigned long mask)1524 int eeh_pe_inject_err(struct eeh_pe *pe, int type, int func,
1525 unsigned long addr, unsigned long mask)
1526 {
1527 /* Invalid PE ? */
1528 if (!pe)
1529 return -ENODEV;
1530
1531 /* Unsupported operation ? */
1532 if (!eeh_ops || !eeh_ops->err_inject)
1533 return -ENOENT;
1534
1535 /* Check on PCI error type */
1536 if (type != EEH_ERR_TYPE_32 && type != EEH_ERR_TYPE_64)
1537 return -EINVAL;
1538
1539 /* Check on PCI error function */
1540 if (func < EEH_ERR_FUNC_MIN || func > EEH_ERR_FUNC_MAX)
1541 return -EINVAL;
1542
1543 return eeh_ops->err_inject(pe, type, func, addr, mask);
1544 }
1545 EXPORT_SYMBOL_GPL(eeh_pe_inject_err);
1546
1547 #ifdef CONFIG_PROC_FS
proc_eeh_show(struct seq_file * m,void * v)1548 static int proc_eeh_show(struct seq_file *m, void *v)
1549 {
1550 if (!eeh_enabled()) {
1551 seq_printf(m, "EEH Subsystem is globally disabled\n");
1552 seq_printf(m, "eeh_total_mmio_ffs=%llu\n", eeh_stats.total_mmio_ffs);
1553 } else {
1554 seq_printf(m, "EEH Subsystem is enabled\n");
1555 seq_printf(m,
1556 "no device=%llu\n"
1557 "no device node=%llu\n"
1558 "no config address=%llu\n"
1559 "check not wanted=%llu\n"
1560 "eeh_total_mmio_ffs=%llu\n"
1561 "eeh_false_positives=%llu\n"
1562 "eeh_slot_resets=%llu\n",
1563 eeh_stats.no_device,
1564 eeh_stats.no_dn,
1565 eeh_stats.no_cfg_addr,
1566 eeh_stats.ignored_check,
1567 eeh_stats.total_mmio_ffs,
1568 eeh_stats.false_positives,
1569 eeh_stats.slot_resets);
1570 }
1571
1572 return 0;
1573 }
1574 #endif /* CONFIG_PROC_FS */
1575
1576 #ifdef CONFIG_DEBUG_FS
1577
1578
eeh_debug_lookup_pdev(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1579 static struct pci_dev *eeh_debug_lookup_pdev(struct file *filp,
1580 const char __user *user_buf,
1581 size_t count, loff_t *ppos)
1582 {
1583 uint32_t domain, bus, dev, fn;
1584 struct pci_dev *pdev;
1585 char buf[20];
1586 int ret;
1587
1588 memset(buf, 0, sizeof(buf));
1589 ret = simple_write_to_buffer(buf, sizeof(buf)-1, ppos, user_buf, count);
1590 if (!ret)
1591 return ERR_PTR(-EFAULT);
1592
1593 ret = sscanf(buf, "%x:%x:%x.%x", &domain, &bus, &dev, &fn);
1594 if (ret != 4) {
1595 pr_err("%s: expected 4 args, got %d\n", __func__, ret);
1596 return ERR_PTR(-EINVAL);
1597 }
1598
1599 pdev = pci_get_domain_bus_and_slot(domain, bus, (dev << 3) | fn);
1600 if (!pdev)
1601 return ERR_PTR(-ENODEV);
1602
1603 return pdev;
1604 }
1605
eeh_enable_dbgfs_set(void * data,u64 val)1606 static int eeh_enable_dbgfs_set(void *data, u64 val)
1607 {
1608 if (val)
1609 eeh_clear_flag(EEH_FORCE_DISABLED);
1610 else
1611 eeh_add_flag(EEH_FORCE_DISABLED);
1612
1613 return 0;
1614 }
1615
eeh_enable_dbgfs_get(void * data,u64 * val)1616 static int eeh_enable_dbgfs_get(void *data, u64 *val)
1617 {
1618 if (eeh_enabled())
1619 *val = 0x1ul;
1620 else
1621 *val = 0x0ul;
1622 return 0;
1623 }
1624
1625 DEFINE_DEBUGFS_ATTRIBUTE(eeh_enable_dbgfs_ops, eeh_enable_dbgfs_get,
1626 eeh_enable_dbgfs_set, "0x%llx\n");
1627
eeh_force_recover_write(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1628 static ssize_t eeh_force_recover_write(struct file *filp,
1629 const char __user *user_buf,
1630 size_t count, loff_t *ppos)
1631 {
1632 struct pci_controller *hose;
1633 uint32_t phbid, pe_no;
1634 struct eeh_pe *pe;
1635 char buf[20];
1636 int ret;
1637
1638 ret = simple_write_to_buffer(buf, sizeof(buf), ppos, user_buf, count);
1639 if (!ret)
1640 return -EFAULT;
1641
1642 /*
1643 * When PE is NULL the event is a "special" event. Rather than
1644 * recovering a specific PE it forces the EEH core to scan for failed
1645 * PHBs and recovers each. This needs to be done before any device
1646 * recoveries can occur.
1647 */
1648 if (!strncmp(buf, "hwcheck", 7)) {
1649 __eeh_send_failure_event(NULL);
1650 return count;
1651 }
1652
1653 ret = sscanf(buf, "%x:%x", &phbid, &pe_no);
1654 if (ret != 2)
1655 return -EINVAL;
1656
1657 hose = pci_find_controller_for_domain(phbid);
1658 if (!hose)
1659 return -ENODEV;
1660
1661 /* Retrieve PE */
1662 pe = eeh_pe_get(hose, pe_no);
1663 if (!pe)
1664 return -ENODEV;
1665
1666 /*
1667 * We don't do any state checking here since the detection
1668 * process is async to the recovery process. The recovery
1669 * thread *should* not break even if we schedule a recovery
1670 * from an odd state (e.g. PE removed, or recovery of a
1671 * non-isolated PE)
1672 */
1673 __eeh_send_failure_event(pe);
1674
1675 return ret < 0 ? ret : count;
1676 }
1677
1678 static const struct file_operations eeh_force_recover_fops = {
1679 .open = simple_open,
1680 .llseek = no_llseek,
1681 .write = eeh_force_recover_write,
1682 };
1683
eeh_debugfs_dev_usage(struct file * filp,char __user * user_buf,size_t count,loff_t * ppos)1684 static ssize_t eeh_debugfs_dev_usage(struct file *filp,
1685 char __user *user_buf,
1686 size_t count, loff_t *ppos)
1687 {
1688 static const char usage[] = "input format: <domain>:<bus>:<dev>.<fn>\n";
1689
1690 return simple_read_from_buffer(user_buf, count, ppos,
1691 usage, sizeof(usage) - 1);
1692 }
1693
eeh_dev_check_write(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1694 static ssize_t eeh_dev_check_write(struct file *filp,
1695 const char __user *user_buf,
1696 size_t count, loff_t *ppos)
1697 {
1698 struct pci_dev *pdev;
1699 struct eeh_dev *edev;
1700 int ret;
1701
1702 pdev = eeh_debug_lookup_pdev(filp, user_buf, count, ppos);
1703 if (IS_ERR(pdev))
1704 return PTR_ERR(pdev);
1705
1706 edev = pci_dev_to_eeh_dev(pdev);
1707 if (!edev) {
1708 pci_err(pdev, "No eeh_dev for this device!\n");
1709 pci_dev_put(pdev);
1710 return -ENODEV;
1711 }
1712
1713 ret = eeh_dev_check_failure(edev);
1714 pci_info(pdev, "eeh_dev_check_failure(%s) = %d\n",
1715 pci_name(pdev), ret);
1716
1717 pci_dev_put(pdev);
1718
1719 return count;
1720 }
1721
1722 static const struct file_operations eeh_dev_check_fops = {
1723 .open = simple_open,
1724 .llseek = no_llseek,
1725 .write = eeh_dev_check_write,
1726 .read = eeh_debugfs_dev_usage,
1727 };
1728
eeh_debugfs_break_device(struct pci_dev * pdev)1729 static int eeh_debugfs_break_device(struct pci_dev *pdev)
1730 {
1731 struct resource *bar = NULL;
1732 void __iomem *mapped;
1733 u16 old, bit;
1734 int i, pos;
1735
1736 /* Do we have an MMIO BAR to disable? */
1737 for (i = 0; i <= PCI_STD_RESOURCE_END; i++) {
1738 struct resource *r = &pdev->resource[i];
1739
1740 if (!r->flags || !r->start)
1741 continue;
1742 if (r->flags & IORESOURCE_IO)
1743 continue;
1744 if (r->flags & IORESOURCE_UNSET)
1745 continue;
1746
1747 bar = r;
1748 break;
1749 }
1750
1751 if (!bar) {
1752 pci_err(pdev, "Unable to find Memory BAR to cause EEH with\n");
1753 return -ENXIO;
1754 }
1755
1756 pci_err(pdev, "Going to break: %pR\n", bar);
1757
1758 if (pdev->is_virtfn) {
1759 #ifndef CONFIG_PCI_IOV
1760 return -ENXIO;
1761 #else
1762 /*
1763 * VFs don't have a per-function COMMAND register, so the best
1764 * we can do is clear the Memory Space Enable bit in the PF's
1765 * SRIOV control reg.
1766 *
1767 * Unfortunately, this requires that we have a PF (i.e doesn't
1768 * work for a passed-through VF) and it has the potential side
1769 * effect of also causing an EEH on every other VF under the
1770 * PF. Oh well.
1771 */
1772 pdev = pdev->physfn;
1773 if (!pdev)
1774 return -ENXIO; /* passed through VFs have no PF */
1775
1776 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_SRIOV);
1777 pos += PCI_SRIOV_CTRL;
1778 bit = PCI_SRIOV_CTRL_MSE;
1779 #endif /* !CONFIG_PCI_IOV */
1780 } else {
1781 bit = PCI_COMMAND_MEMORY;
1782 pos = PCI_COMMAND;
1783 }
1784
1785 /*
1786 * Process here is:
1787 *
1788 * 1. Disable Memory space.
1789 *
1790 * 2. Perform an MMIO to the device. This should result in an error
1791 * (CA / UR) being raised by the device which results in an EEH
1792 * PE freeze. Using the in_8() accessor skips the eeh detection hook
1793 * so the freeze hook so the EEH Detection machinery won't be
1794 * triggered here. This is to match the usual behaviour of EEH
1795 * where the HW will asyncronously freeze a PE and it's up to
1796 * the kernel to notice and deal with it.
1797 *
1798 * 3. Turn Memory space back on. This is more important for VFs
1799 * since recovery will probably fail if we don't. For normal
1800 * the COMMAND register is reset as a part of re-initialising
1801 * the device.
1802 *
1803 * Breaking stuff is the point so who cares if it's racy ;)
1804 */
1805 pci_read_config_word(pdev, pos, &old);
1806
1807 mapped = ioremap(bar->start, PAGE_SIZE);
1808 if (!mapped) {
1809 pci_err(pdev, "Unable to map MMIO BAR %pR\n", bar);
1810 return -ENXIO;
1811 }
1812
1813 pci_write_config_word(pdev, pos, old & ~bit);
1814 in_8(mapped);
1815 pci_write_config_word(pdev, pos, old);
1816
1817 iounmap(mapped);
1818
1819 return 0;
1820 }
1821
eeh_dev_break_write(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1822 static ssize_t eeh_dev_break_write(struct file *filp,
1823 const char __user *user_buf,
1824 size_t count, loff_t *ppos)
1825 {
1826 struct pci_dev *pdev;
1827 int ret;
1828
1829 pdev = eeh_debug_lookup_pdev(filp, user_buf, count, ppos);
1830 if (IS_ERR(pdev))
1831 return PTR_ERR(pdev);
1832
1833 ret = eeh_debugfs_break_device(pdev);
1834 pci_dev_put(pdev);
1835
1836 if (ret < 0)
1837 return ret;
1838
1839 return count;
1840 }
1841
1842 static const struct file_operations eeh_dev_break_fops = {
1843 .open = simple_open,
1844 .llseek = no_llseek,
1845 .write = eeh_dev_break_write,
1846 .read = eeh_debugfs_dev_usage,
1847 };
1848
eeh_dev_can_recover(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1849 static ssize_t eeh_dev_can_recover(struct file *filp,
1850 const char __user *user_buf,
1851 size_t count, loff_t *ppos)
1852 {
1853 struct pci_driver *drv;
1854 struct pci_dev *pdev;
1855 size_t ret;
1856
1857 pdev = eeh_debug_lookup_pdev(filp, user_buf, count, ppos);
1858 if (IS_ERR(pdev))
1859 return PTR_ERR(pdev);
1860
1861 /*
1862 * In order for error recovery to work the driver needs to implement
1863 * .error_detected(), so it can quiesce IO to the device, and
1864 * .slot_reset() so it can re-initialise the device after a reset.
1865 *
1866 * Ideally they'd implement .resume() too, but some drivers which
1867 * we need to support (notably IPR) don't so I guess we can tolerate
1868 * that.
1869 *
1870 * .mmio_enabled() is mostly there as a work-around for devices which
1871 * take forever to re-init after a hot reset. Implementing that is
1872 * strictly optional.
1873 */
1874 drv = pci_dev_driver(pdev);
1875 if (drv &&
1876 drv->err_handler &&
1877 drv->err_handler->error_detected &&
1878 drv->err_handler->slot_reset) {
1879 ret = count;
1880 } else {
1881 ret = -EOPNOTSUPP;
1882 }
1883
1884 pci_dev_put(pdev);
1885
1886 return ret;
1887 }
1888
1889 static const struct file_operations eeh_dev_can_recover_fops = {
1890 .open = simple_open,
1891 .llseek = no_llseek,
1892 .write = eeh_dev_can_recover,
1893 .read = eeh_debugfs_dev_usage,
1894 };
1895
1896 #endif
1897
eeh_init_proc(void)1898 static int __init eeh_init_proc(void)
1899 {
1900 if (machine_is(pseries) || machine_is(powernv)) {
1901 proc_create_single("powerpc/eeh", 0, NULL, proc_eeh_show);
1902 #ifdef CONFIG_DEBUG_FS
1903 debugfs_create_file_unsafe("eeh_enable", 0600,
1904 arch_debugfs_dir, NULL,
1905 &eeh_enable_dbgfs_ops);
1906 debugfs_create_u32("eeh_max_freezes", 0600,
1907 arch_debugfs_dir, &eeh_max_freezes);
1908 debugfs_create_bool("eeh_disable_recovery", 0600,
1909 arch_debugfs_dir,
1910 &eeh_debugfs_no_recover);
1911 debugfs_create_file_unsafe("eeh_dev_check", 0600,
1912 arch_debugfs_dir, NULL,
1913 &eeh_dev_check_fops);
1914 debugfs_create_file_unsafe("eeh_dev_break", 0600,
1915 arch_debugfs_dir, NULL,
1916 &eeh_dev_break_fops);
1917 debugfs_create_file_unsafe("eeh_force_recover", 0600,
1918 arch_debugfs_dir, NULL,
1919 &eeh_force_recover_fops);
1920 debugfs_create_file_unsafe("eeh_dev_can_recover", 0600,
1921 arch_debugfs_dir, NULL,
1922 &eeh_dev_can_recover_fops);
1923 eeh_cache_debugfs_init();
1924 #endif
1925 }
1926
1927 return 0;
1928 }
1929 __initcall(eeh_init_proc);
1930