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
eeh_driver_name(struct pci_dev * pdev)402 static inline const char *eeh_driver_name(struct pci_dev *pdev)
403 {
404 if (pdev)
405 return dev_driver_string(&pdev->dev);
406
407 return "<null>";
408 }
409
410 /**
411 * eeh_dev_check_failure - Check if all 1's data is due to EEH slot freeze
412 * @edev: eeh device
413 *
414 * Check for an EEH failure for the given device node. Call this
415 * routine if the result of a read was all 0xff's and you want to
416 * find out if this is due to an EEH slot freeze. This routine
417 * will query firmware for the EEH status.
418 *
419 * Returns 0 if there has not been an EEH error; otherwise returns
420 * a non-zero value and queues up a slot isolation event notification.
421 *
422 * It is safe to call this routine in an interrupt context.
423 */
eeh_dev_check_failure(struct eeh_dev * edev)424 int eeh_dev_check_failure(struct eeh_dev *edev)
425 {
426 int ret;
427 unsigned long flags;
428 struct device_node *dn;
429 struct pci_dev *dev;
430 struct eeh_pe *pe, *parent_pe;
431 int rc = 0;
432 const char *location = NULL;
433
434 eeh_stats.total_mmio_ffs++;
435
436 if (!eeh_enabled())
437 return 0;
438
439 if (!edev) {
440 eeh_stats.no_dn++;
441 return 0;
442 }
443 dev = eeh_dev_to_pci_dev(edev);
444 pe = eeh_dev_to_pe(edev);
445
446 /* Access to IO BARs might get this far and still not want checking. */
447 if (!pe) {
448 eeh_stats.ignored_check++;
449 eeh_edev_dbg(edev, "Ignored check\n");
450 return 0;
451 }
452
453 /*
454 * On PowerNV platform, we might already have fenced PHB
455 * there and we need take care of that firstly.
456 */
457 ret = eeh_phb_check_failure(pe);
458 if (ret > 0)
459 return ret;
460
461 /*
462 * If the PE isn't owned by us, we shouldn't check the
463 * state. Instead, let the owner handle it if the PE has
464 * been frozen.
465 */
466 if (eeh_pe_passed(pe))
467 return 0;
468
469 /* If we already have a pending isolation event for this
470 * slot, we know it's bad already, we don't need to check.
471 * Do this checking under a lock; as multiple PCI devices
472 * in one slot might report errors simultaneously, and we
473 * only want one error recovery routine running.
474 */
475 eeh_serialize_lock(&flags);
476 rc = 1;
477 if (pe->state & EEH_PE_ISOLATED) {
478 pe->check_count++;
479 if (pe->check_count == EEH_MAX_FAILS) {
480 dn = pci_device_to_OF_node(dev);
481 if (dn)
482 location = of_get_property(dn, "ibm,loc-code",
483 NULL);
484 eeh_edev_err(edev, "%d reads ignored for recovering device at location=%s driver=%s\n",
485 pe->check_count,
486 location ? location : "unknown",
487 eeh_driver_name(dev));
488 eeh_edev_err(edev, "Might be infinite loop in %s driver\n",
489 eeh_driver_name(dev));
490 dump_stack();
491 }
492 goto dn_unlock;
493 }
494
495 /*
496 * Now test for an EEH failure. This is VERY expensive.
497 * Note that the eeh_config_addr may be a parent device
498 * in the case of a device behind a bridge, or it may be
499 * function zero of a multi-function device.
500 * In any case they must share a common PHB.
501 */
502 ret = eeh_ops->get_state(pe, NULL);
503
504 /* Note that config-io to empty slots may fail;
505 * they are empty when they don't have children.
506 * We will punt with the following conditions: Failure to get
507 * PE's state, EEH not support and Permanently unavailable
508 * state, PE is in good state.
509 *
510 * On the pSeries, after reaching the threshold, get_state might
511 * return EEH_STATE_NOT_SUPPORT. However, it's possible that the
512 * device state remains uncleared if the device is not marked
513 * pci_channel_io_perm_failure. Therefore, consider logging the
514 * event to let device removal happen.
515 *
516 */
517 if ((ret < 0) ||
518 (ret == EEH_STATE_NOT_SUPPORT &&
519 dev->error_state == pci_channel_io_perm_failure) ||
520 eeh_state_active(ret)) {
521 eeh_stats.false_positives++;
522 pe->false_positives++;
523 rc = 0;
524 goto dn_unlock;
525 }
526
527 /*
528 * It should be corner case that the parent PE has been
529 * put into frozen state as well. We should take care
530 * that at first.
531 */
532 parent_pe = pe->parent;
533 while (parent_pe) {
534 /* Hit the ceiling ? */
535 if (parent_pe->type & EEH_PE_PHB)
536 break;
537
538 /* Frozen parent PE ? */
539 ret = eeh_ops->get_state(parent_pe, NULL);
540 if (ret > 0 && !eeh_state_active(ret)) {
541 pe = parent_pe;
542 pr_err("EEH: Failure of PHB#%x-PE#%x will be handled at parent PHB#%x-PE#%x.\n",
543 pe->phb->global_number, pe->addr,
544 pe->phb->global_number, parent_pe->addr);
545 }
546
547 /* Next parent level */
548 parent_pe = parent_pe->parent;
549 }
550
551 eeh_stats.slot_resets++;
552
553 /* Avoid repeated reports of this failure, including problems
554 * with other functions on this device, and functions under
555 * bridges.
556 */
557 eeh_pe_mark_isolated(pe);
558 eeh_serialize_unlock(flags);
559
560 /* Most EEH events are due to device driver bugs. Having
561 * a stack trace will help the device-driver authors figure
562 * out what happened. So print that out.
563 */
564 pr_debug("EEH: %s: Frozen PHB#%x-PE#%x detected\n",
565 __func__, pe->phb->global_number, pe->addr);
566 eeh_send_failure_event(pe);
567
568 return 1;
569
570 dn_unlock:
571 eeh_serialize_unlock(flags);
572 return rc;
573 }
574
575 EXPORT_SYMBOL_GPL(eeh_dev_check_failure);
576
577 /**
578 * eeh_check_failure - Check if all 1's data is due to EEH slot freeze
579 * @token: I/O address
580 *
581 * Check for an EEH failure at the given I/O address. Call this
582 * routine if the result of a read was all 0xff's and you want to
583 * find out if this is due to an EEH slot freeze event. This routine
584 * will query firmware for the EEH status.
585 *
586 * Note this routine is safe to call in an interrupt context.
587 */
eeh_check_failure(const volatile void __iomem * token)588 int eeh_check_failure(const volatile void __iomem *token)
589 {
590 unsigned long addr;
591 struct eeh_dev *edev;
592
593 /* Finding the phys addr + pci device; this is pretty quick. */
594 addr = eeh_token_to_phys((unsigned long __force) token);
595 edev = eeh_addr_cache_get_dev(addr);
596 if (!edev) {
597 eeh_stats.no_device++;
598 return 0;
599 }
600
601 return eeh_dev_check_failure(edev);
602 }
603 EXPORT_SYMBOL(eeh_check_failure);
604
605
606 /**
607 * eeh_pci_enable - Enable MMIO or DMA transfers for this slot
608 * @pe: EEH PE
609 * @function: EEH option
610 *
611 * This routine should be called to reenable frozen MMIO or DMA
612 * so that it would work correctly again. It's useful while doing
613 * recovery or log collection on the indicated device.
614 */
eeh_pci_enable(struct eeh_pe * pe,int function)615 int eeh_pci_enable(struct eeh_pe *pe, int function)
616 {
617 int active_flag, rc;
618
619 /*
620 * pHyp doesn't allow to enable IO or DMA on unfrozen PE.
621 * Also, it's pointless to enable them on unfrozen PE. So
622 * we have to check before enabling IO or DMA.
623 */
624 switch (function) {
625 case EEH_OPT_THAW_MMIO:
626 active_flag = EEH_STATE_MMIO_ACTIVE | EEH_STATE_MMIO_ENABLED;
627 break;
628 case EEH_OPT_THAW_DMA:
629 active_flag = EEH_STATE_DMA_ACTIVE;
630 break;
631 case EEH_OPT_DISABLE:
632 case EEH_OPT_ENABLE:
633 case EEH_OPT_FREEZE_PE:
634 active_flag = 0;
635 break;
636 default:
637 pr_warn("%s: Invalid function %d\n",
638 __func__, function);
639 return -EINVAL;
640 }
641
642 /*
643 * Check if IO or DMA has been enabled before
644 * enabling them.
645 */
646 if (active_flag) {
647 rc = eeh_ops->get_state(pe, NULL);
648 if (rc < 0)
649 return rc;
650
651 /* Needn't enable it at all */
652 if (rc == EEH_STATE_NOT_SUPPORT)
653 return 0;
654
655 /* It's already enabled */
656 if (rc & active_flag)
657 return 0;
658 }
659
660
661 /* Issue the request */
662 rc = eeh_ops->set_option(pe, function);
663 if (rc)
664 pr_warn("%s: Unexpected state change %d on "
665 "PHB#%x-PE#%x, err=%d\n",
666 __func__, function, pe->phb->global_number,
667 pe->addr, rc);
668
669 /* Check if the request is finished successfully */
670 if (active_flag) {
671 rc = eeh_wait_state(pe, PCI_BUS_RESET_WAIT_MSEC);
672 if (rc < 0)
673 return rc;
674
675 if (rc & active_flag)
676 return 0;
677
678 return -EIO;
679 }
680
681 return rc;
682 }
683
eeh_disable_and_save_dev_state(struct eeh_dev * edev,void * userdata)684 static void eeh_disable_and_save_dev_state(struct eeh_dev *edev,
685 void *userdata)
686 {
687 struct pci_dev *pdev = eeh_dev_to_pci_dev(edev);
688 struct pci_dev *dev = userdata;
689
690 /*
691 * The caller should have disabled and saved the
692 * state for the specified device
693 */
694 if (!pdev || pdev == dev)
695 return;
696
697 /* Ensure we have D0 power state */
698 pci_set_power_state(pdev, PCI_D0);
699
700 /* Save device state */
701 pci_save_state(pdev);
702
703 /*
704 * Disable device to avoid any DMA traffic and
705 * interrupt from the device
706 */
707 pci_write_config_word(pdev, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE);
708 }
709
eeh_restore_dev_state(struct eeh_dev * edev,void * userdata)710 static void eeh_restore_dev_state(struct eeh_dev *edev, void *userdata)
711 {
712 struct pci_dev *pdev = eeh_dev_to_pci_dev(edev);
713 struct pci_dev *dev = userdata;
714
715 if (!pdev)
716 return;
717
718 /* Apply customization from firmware */
719 if (eeh_ops->restore_config)
720 eeh_ops->restore_config(edev);
721
722 /* The caller should restore state for the specified device */
723 if (pdev != dev)
724 pci_restore_state(pdev);
725 }
726
727 /**
728 * pcibios_set_pcie_reset_state - Set PCI-E reset state
729 * @dev: pci device struct
730 * @state: reset state to enter
731 *
732 * Return value:
733 * 0 if success
734 */
pcibios_set_pcie_reset_state(struct pci_dev * dev,enum pcie_reset_state state)735 int pcibios_set_pcie_reset_state(struct pci_dev *dev, enum pcie_reset_state state)
736 {
737 struct eeh_dev *edev = pci_dev_to_eeh_dev(dev);
738 struct eeh_pe *pe = eeh_dev_to_pe(edev);
739
740 if (!pe) {
741 pr_err("%s: No PE found on PCI device %s\n",
742 __func__, pci_name(dev));
743 return -EINVAL;
744 }
745
746 switch (state) {
747 case pcie_deassert_reset:
748 eeh_ops->reset(pe, EEH_RESET_DEACTIVATE);
749 eeh_unfreeze_pe(pe);
750 if (!(pe->type & EEH_PE_VF))
751 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, true);
752 eeh_pe_dev_traverse(pe, eeh_restore_dev_state, dev);
753 eeh_pe_state_clear(pe, EEH_PE_ISOLATED, true);
754 break;
755 case pcie_hot_reset:
756 eeh_pe_mark_isolated(pe);
757 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, true);
758 eeh_ops->set_option(pe, EEH_OPT_FREEZE_PE);
759 eeh_pe_dev_traverse(pe, eeh_disable_and_save_dev_state, dev);
760 if (!(pe->type & EEH_PE_VF))
761 eeh_pe_state_mark(pe, EEH_PE_CFG_BLOCKED);
762 eeh_ops->reset(pe, EEH_RESET_HOT);
763 break;
764 case pcie_warm_reset:
765 eeh_pe_mark_isolated(pe);
766 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, true);
767 eeh_ops->set_option(pe, EEH_OPT_FREEZE_PE);
768 eeh_pe_dev_traverse(pe, eeh_disable_and_save_dev_state, dev);
769 if (!(pe->type & EEH_PE_VF))
770 eeh_pe_state_mark(pe, EEH_PE_CFG_BLOCKED);
771 eeh_ops->reset(pe, EEH_RESET_FUNDAMENTAL);
772 break;
773 default:
774 eeh_pe_state_clear(pe, EEH_PE_ISOLATED | EEH_PE_CFG_BLOCKED, true);
775 return -EINVAL;
776 }
777
778 return 0;
779 }
780
781 /**
782 * eeh_set_dev_freset - Check the required reset for the indicated device
783 * @edev: EEH device
784 * @flag: return value
785 *
786 * Each device might have its preferred reset type: fundamental or
787 * hot reset. The routine is used to collected the information for
788 * the indicated device and its children so that the bunch of the
789 * devices could be reset properly.
790 */
eeh_set_dev_freset(struct eeh_dev * edev,void * flag)791 static void eeh_set_dev_freset(struct eeh_dev *edev, void *flag)
792 {
793 struct pci_dev *dev;
794 unsigned int *freset = (unsigned int *)flag;
795
796 dev = eeh_dev_to_pci_dev(edev);
797 if (dev)
798 *freset |= dev->needs_freset;
799 }
800
eeh_pe_refreeze_passed(struct eeh_pe * root)801 static void eeh_pe_refreeze_passed(struct eeh_pe *root)
802 {
803 struct eeh_pe *pe;
804 int state;
805
806 eeh_for_each_pe(root, pe) {
807 if (eeh_pe_passed(pe)) {
808 state = eeh_ops->get_state(pe, NULL);
809 if (state &
810 (EEH_STATE_MMIO_ACTIVE | EEH_STATE_MMIO_ENABLED)) {
811 pr_info("EEH: Passed-through PE PHB#%x-PE#%x was thawed by reset, re-freezing for safety.\n",
812 pe->phb->global_number, pe->addr);
813 eeh_pe_set_option(pe, EEH_OPT_FREEZE_PE);
814 }
815 }
816 }
817 }
818
819 /**
820 * eeh_pe_reset_full - Complete a full reset process on the indicated PE
821 * @pe: EEH PE
822 * @include_passed: include passed-through devices?
823 *
824 * This function executes a full reset procedure on a PE, including setting
825 * the appropriate flags, performing a fundamental or hot reset, and then
826 * deactivating the reset status. It is designed to be used within the EEH
827 * subsystem, as opposed to eeh_pe_reset which is exported to drivers and
828 * only performs a single operation at a time.
829 *
830 * This function will attempt to reset a PE three times before failing.
831 */
eeh_pe_reset_full(struct eeh_pe * pe,bool include_passed)832 int eeh_pe_reset_full(struct eeh_pe *pe, bool include_passed)
833 {
834 int reset_state = (EEH_PE_RESET | EEH_PE_CFG_BLOCKED);
835 int type = EEH_RESET_HOT;
836 unsigned int freset = 0;
837 int i, state = 0, ret;
838
839 /*
840 * Determine the type of reset to perform - hot or fundamental.
841 * Hot reset is the default operation, unless any device under the
842 * PE requires a fundamental reset.
843 */
844 eeh_pe_dev_traverse(pe, eeh_set_dev_freset, &freset);
845
846 if (freset)
847 type = EEH_RESET_FUNDAMENTAL;
848
849 /* Mark the PE as in reset state and block config space accesses */
850 eeh_pe_state_mark(pe, reset_state);
851
852 /* Make three attempts at resetting the bus */
853 for (i = 0; i < 3; i++) {
854 ret = eeh_pe_reset(pe, type, include_passed);
855 if (!ret)
856 ret = eeh_pe_reset(pe, EEH_RESET_DEACTIVATE,
857 include_passed);
858 if (ret) {
859 ret = -EIO;
860 pr_warn("EEH: Failure %d resetting PHB#%x-PE#%x (attempt %d)\n\n",
861 state, pe->phb->global_number, pe->addr, i + 1);
862 continue;
863 }
864 if (i)
865 pr_warn("EEH: PHB#%x-PE#%x: Successful reset (attempt %d)\n",
866 pe->phb->global_number, pe->addr, i + 1);
867
868 /* Wait until the PE is in a functioning state */
869 state = eeh_wait_state(pe, PCI_BUS_RESET_WAIT_MSEC);
870 if (state < 0) {
871 pr_warn("EEH: Unrecoverable slot failure on PHB#%x-PE#%x",
872 pe->phb->global_number, pe->addr);
873 ret = -ENOTRECOVERABLE;
874 break;
875 }
876 if (eeh_state_active(state))
877 break;
878 else
879 pr_warn("EEH: PHB#%x-PE#%x: Slot inactive after reset: 0x%x (attempt %d)\n",
880 pe->phb->global_number, pe->addr, state, i + 1);
881 }
882
883 /* Resetting the PE may have unfrozen child PEs. If those PEs have been
884 * (potentially) passed through to a guest, re-freeze them:
885 */
886 if (!include_passed)
887 eeh_pe_refreeze_passed(pe);
888
889 eeh_pe_state_clear(pe, reset_state, true);
890 return ret;
891 }
892
893 /**
894 * eeh_save_bars - Save device bars
895 * @edev: PCI device associated EEH device
896 *
897 * Save the values of the device bars. Unlike the restore
898 * routine, this routine is *not* recursive. This is because
899 * PCI devices are added individually; but, for the restore,
900 * an entire slot is reset at a time.
901 */
eeh_save_bars(struct eeh_dev * edev)902 void eeh_save_bars(struct eeh_dev *edev)
903 {
904 int i;
905
906 if (!edev)
907 return;
908
909 for (i = 0; i < 16; i++)
910 eeh_ops->read_config(edev, i * 4, 4, &edev->config_space[i]);
911
912 /*
913 * For PCI bridges including root port, we need enable bus
914 * master explicitly. Otherwise, it can't fetch IODA table
915 * entries correctly. So we cache the bit in advance so that
916 * we can restore it after reset, either PHB range or PE range.
917 */
918 if (edev->mode & EEH_DEV_BRIDGE)
919 edev->config_space[1] |= PCI_COMMAND_MASTER;
920 }
921
eeh_reboot_notifier(struct notifier_block * nb,unsigned long action,void * unused)922 static int eeh_reboot_notifier(struct notifier_block *nb,
923 unsigned long action, void *unused)
924 {
925 eeh_clear_flag(EEH_ENABLED);
926 return NOTIFY_DONE;
927 }
928
929 static struct notifier_block eeh_reboot_nb = {
930 .notifier_call = eeh_reboot_notifier,
931 };
932
eeh_device_notifier(struct notifier_block * nb,unsigned long action,void * data)933 static int eeh_device_notifier(struct notifier_block *nb,
934 unsigned long action, void *data)
935 {
936 struct device *dev = data;
937
938 switch (action) {
939 /*
940 * Note: It's not possible to perform EEH device addition (i.e.
941 * {pseries,pnv}_pcibios_bus_add_device()) here because it depends on
942 * the device's resources, which have not yet been set up.
943 */
944 case BUS_NOTIFY_DEL_DEVICE:
945 eeh_remove_device(to_pci_dev(dev));
946 break;
947 default:
948 break;
949 }
950 return NOTIFY_DONE;
951 }
952
953 static struct notifier_block eeh_device_nb = {
954 .notifier_call = eeh_device_notifier,
955 };
956
957 /**
958 * eeh_init - System wide EEH initialization
959 * @ops: struct to trace EEH operation callback functions
960 *
961 * It's the platform's job to call this from an arch_initcall().
962 */
eeh_init(struct eeh_ops * ops)963 int eeh_init(struct eeh_ops *ops)
964 {
965 struct pci_controller *hose, *tmp;
966 int ret = 0;
967
968 /* the platform should only initialise EEH once */
969 if (WARN_ON(eeh_ops))
970 return -EEXIST;
971 if (WARN_ON(!ops))
972 return -ENOENT;
973 eeh_ops = ops;
974
975 /* Register reboot notifier */
976 ret = register_reboot_notifier(&eeh_reboot_nb);
977 if (ret) {
978 pr_warn("%s: Failed to register reboot notifier (%d)\n",
979 __func__, ret);
980 return ret;
981 }
982
983 ret = bus_register_notifier(&pci_bus_type, &eeh_device_nb);
984 if (ret) {
985 pr_warn("%s: Failed to register bus notifier (%d)\n",
986 __func__, ret);
987 return ret;
988 }
989
990 /* Initialize PHB PEs */
991 list_for_each_entry_safe(hose, tmp, &hose_list, list_node)
992 eeh_phb_pe_create(hose);
993
994 eeh_addr_cache_init();
995
996 /* Initialize EEH event */
997 return eeh_event_init();
998 }
999
1000 /**
1001 * eeh_probe_device() - Perform EEH initialization for the indicated pci device
1002 * @dev: pci device for which to set up EEH
1003 *
1004 * This routine must be used to complete EEH initialization for PCI
1005 * devices that were added after system boot (e.g. hotplug, dlpar).
1006 */
eeh_probe_device(struct pci_dev * dev)1007 void eeh_probe_device(struct pci_dev *dev)
1008 {
1009 struct eeh_dev *edev;
1010
1011 pr_debug("EEH: Adding device %s\n", pci_name(dev));
1012
1013 /*
1014 * pci_dev_to_eeh_dev() can only work if eeh_probe_dev() was
1015 * already called for this device.
1016 */
1017 if (WARN_ON_ONCE(pci_dev_to_eeh_dev(dev))) {
1018 pci_dbg(dev, "Already bound to an eeh_dev!\n");
1019 return;
1020 }
1021
1022 edev = eeh_ops->probe(dev);
1023 if (!edev) {
1024 pr_debug("EEH: Adding device failed\n");
1025 return;
1026 }
1027
1028 /*
1029 * FIXME: We rely on pcibios_release_device() to remove the
1030 * existing EEH state. The release function is only called if
1031 * the pci_dev's refcount drops to zero so if something is
1032 * keeping a ref to a device (e.g. a filesystem) we need to
1033 * remove the old EEH state.
1034 *
1035 * FIXME: HEY MA, LOOK AT ME, NO LOCKING!
1036 */
1037 if (edev->pdev && edev->pdev != dev) {
1038 eeh_pe_tree_remove(edev);
1039 eeh_addr_cache_rmv_dev(edev->pdev);
1040 eeh_sysfs_remove_device(edev->pdev);
1041
1042 /*
1043 * We definitely should have the PCI device removed
1044 * though it wasn't correctly. So we needn't call
1045 * into error handler afterwards.
1046 */
1047 edev->mode |= EEH_DEV_NO_HANDLER;
1048 }
1049
1050 /* bind the pdev and the edev together */
1051 edev->pdev = dev;
1052 dev->dev.archdata.edev = edev;
1053 eeh_addr_cache_insert_dev(dev);
1054 eeh_sysfs_add_device(dev);
1055 }
1056
1057 /**
1058 * eeh_remove_device - Undo EEH setup for the indicated pci device
1059 * @dev: pci device to be removed
1060 *
1061 * This routine should be called when a device is removed from
1062 * a running system (e.g. by hotplug or dlpar). It unregisters
1063 * the PCI device from the EEH subsystem. I/O errors affecting
1064 * this device will no longer be detected after this call; thus,
1065 * i/o errors affecting this slot may leave this device unusable.
1066 */
eeh_remove_device(struct pci_dev * dev)1067 void eeh_remove_device(struct pci_dev *dev)
1068 {
1069 struct eeh_dev *edev;
1070
1071 if (!dev || !eeh_enabled())
1072 return;
1073 edev = pci_dev_to_eeh_dev(dev);
1074
1075 /* Unregister the device with the EEH/PCI address search system */
1076 dev_dbg(&dev->dev, "EEH: Removing device\n");
1077
1078 if (!edev || !edev->pdev || !edev->pe) {
1079 dev_dbg(&dev->dev, "EEH: Device not referenced!\n");
1080 return;
1081 }
1082
1083 /*
1084 * During the hotplug for EEH error recovery, we need the EEH
1085 * device attached to the parent PE in order for BAR restore
1086 * a bit later. So we keep it for BAR restore and remove it
1087 * from the parent PE during the BAR resotre.
1088 */
1089 edev->pdev = NULL;
1090
1091 /*
1092 * eeh_sysfs_remove_device() uses pci_dev_to_eeh_dev() so we need to
1093 * remove the sysfs files before clearing dev.archdata.edev
1094 */
1095 if (edev->mode & EEH_DEV_SYSFS)
1096 eeh_sysfs_remove_device(dev);
1097
1098 /*
1099 * We're removing from the PCI subsystem, that means
1100 * the PCI device driver can't support EEH or not
1101 * well. So we rely on hotplug completely to do recovery
1102 * for the specific PCI device.
1103 */
1104 edev->mode |= EEH_DEV_NO_HANDLER;
1105
1106 eeh_addr_cache_rmv_dev(dev);
1107
1108 /*
1109 * The flag "in_error" is used to trace EEH devices for VFs
1110 * in error state or not. It's set in eeh_report_error(). If
1111 * it's not set, eeh_report_{reset,resume}() won't be called
1112 * for the VF EEH device.
1113 */
1114 edev->in_error = false;
1115 dev->dev.archdata.edev = NULL;
1116 if (!(edev->pe->state & EEH_PE_KEEP))
1117 eeh_pe_tree_remove(edev);
1118 else
1119 edev->mode |= EEH_DEV_DISCONNECTED;
1120 }
1121
eeh_unfreeze_pe(struct eeh_pe * pe)1122 int eeh_unfreeze_pe(struct eeh_pe *pe)
1123 {
1124 int ret;
1125
1126 ret = eeh_pci_enable(pe, EEH_OPT_THAW_MMIO);
1127 if (ret) {
1128 pr_warn("%s: Failure %d enabling IO on PHB#%x-PE#%x\n",
1129 __func__, ret, pe->phb->global_number, pe->addr);
1130 return ret;
1131 }
1132
1133 ret = eeh_pci_enable(pe, EEH_OPT_THAW_DMA);
1134 if (ret) {
1135 pr_warn("%s: Failure %d enabling DMA on PHB#%x-PE#%x\n",
1136 __func__, ret, pe->phb->global_number, pe->addr);
1137 return ret;
1138 }
1139
1140 return ret;
1141 }
1142
1143
1144 static struct pci_device_id eeh_reset_ids[] = {
1145 { PCI_DEVICE(0x19a2, 0x0710) }, /* Emulex, BE */
1146 { PCI_DEVICE(0x10df, 0xe220) }, /* Emulex, Lancer */
1147 { PCI_DEVICE(0x14e4, 0x1657) }, /* Broadcom BCM5719 */
1148 { 0 }
1149 };
1150
eeh_pe_change_owner(struct eeh_pe * pe)1151 static int eeh_pe_change_owner(struct eeh_pe *pe)
1152 {
1153 struct eeh_dev *edev, *tmp;
1154 struct pci_dev *pdev;
1155 struct pci_device_id *id;
1156 int ret;
1157
1158 /* Check PE state */
1159 ret = eeh_ops->get_state(pe, NULL);
1160 if (ret < 0 || ret == EEH_STATE_NOT_SUPPORT)
1161 return 0;
1162
1163 /* Unfrozen PE, nothing to do */
1164 if (eeh_state_active(ret))
1165 return 0;
1166
1167 /* Frozen PE, check if it needs PE level reset */
1168 eeh_pe_for_each_dev(pe, edev, tmp) {
1169 pdev = eeh_dev_to_pci_dev(edev);
1170 if (!pdev)
1171 continue;
1172
1173 for (id = &eeh_reset_ids[0]; id->vendor != 0; id++) {
1174 if (id->vendor != PCI_ANY_ID &&
1175 id->vendor != pdev->vendor)
1176 continue;
1177 if (id->device != PCI_ANY_ID &&
1178 id->device != pdev->device)
1179 continue;
1180 if (id->subvendor != PCI_ANY_ID &&
1181 id->subvendor != pdev->subsystem_vendor)
1182 continue;
1183 if (id->subdevice != PCI_ANY_ID &&
1184 id->subdevice != pdev->subsystem_device)
1185 continue;
1186
1187 return eeh_pe_reset_and_recover(pe);
1188 }
1189 }
1190
1191 ret = eeh_unfreeze_pe(pe);
1192 if (!ret)
1193 eeh_pe_state_clear(pe, EEH_PE_ISOLATED, true);
1194 return ret;
1195 }
1196
1197 /**
1198 * eeh_dev_open - Increase count of pass through devices for PE
1199 * @pdev: PCI device
1200 *
1201 * Increase count of passed through devices for the indicated
1202 * PE. In the result, the EEH errors detected on the PE won't be
1203 * reported. The PE owner will be responsible for detection
1204 * and recovery.
1205 */
eeh_dev_open(struct pci_dev * pdev)1206 int eeh_dev_open(struct pci_dev *pdev)
1207 {
1208 struct eeh_dev *edev;
1209 int ret = -ENODEV;
1210
1211 mutex_lock(&eeh_dev_mutex);
1212
1213 /* No PCI device ? */
1214 if (!pdev)
1215 goto out;
1216
1217 /* No EEH device or PE ? */
1218 edev = pci_dev_to_eeh_dev(pdev);
1219 if (!edev || !edev->pe)
1220 goto out;
1221
1222 /*
1223 * The PE might have been put into frozen state, but we
1224 * didn't detect that yet. The passed through PCI devices
1225 * in frozen PE won't work properly. Clear the frozen state
1226 * in advance.
1227 */
1228 ret = eeh_pe_change_owner(edev->pe);
1229 if (ret)
1230 goto out;
1231
1232 /* Increase PE's pass through count */
1233 atomic_inc(&edev->pe->pass_dev_cnt);
1234 mutex_unlock(&eeh_dev_mutex);
1235
1236 return 0;
1237 out:
1238 mutex_unlock(&eeh_dev_mutex);
1239 return ret;
1240 }
1241 EXPORT_SYMBOL_GPL(eeh_dev_open);
1242
1243 /**
1244 * eeh_dev_release - Decrease count of pass through devices for PE
1245 * @pdev: PCI device
1246 *
1247 * Decrease count of pass through devices for the indicated PE. If
1248 * there is no passed through device in PE, the EEH errors detected
1249 * on the PE will be reported and handled as usual.
1250 */
eeh_dev_release(struct pci_dev * pdev)1251 void eeh_dev_release(struct pci_dev *pdev)
1252 {
1253 struct eeh_dev *edev;
1254
1255 mutex_lock(&eeh_dev_mutex);
1256
1257 /* No PCI device ? */
1258 if (!pdev)
1259 goto out;
1260
1261 /* No EEH device ? */
1262 edev = pci_dev_to_eeh_dev(pdev);
1263 if (!edev || !edev->pe || !eeh_pe_passed(edev->pe))
1264 goto out;
1265
1266 /* Decrease PE's pass through count */
1267 WARN_ON(atomic_dec_if_positive(&edev->pe->pass_dev_cnt) < 0);
1268 eeh_pe_change_owner(edev->pe);
1269 out:
1270 mutex_unlock(&eeh_dev_mutex);
1271 }
1272 EXPORT_SYMBOL(eeh_dev_release);
1273
1274 #ifdef CONFIG_IOMMU_API
1275
1276 /**
1277 * eeh_iommu_group_to_pe - Convert IOMMU group to EEH PE
1278 * @group: IOMMU group
1279 *
1280 * The routine is called to convert IOMMU group to EEH PE.
1281 */
eeh_iommu_group_to_pe(struct iommu_group * group)1282 struct eeh_pe *eeh_iommu_group_to_pe(struct iommu_group *group)
1283 {
1284 struct pci_dev *pdev = NULL;
1285 struct eeh_dev *edev;
1286 int ret;
1287
1288 /* No IOMMU group ? */
1289 if (!group)
1290 return NULL;
1291
1292 ret = iommu_group_for_each_dev(group, &pdev, dev_has_iommu_table);
1293 if (!ret || !pdev)
1294 return NULL;
1295
1296 /* No EEH device or PE ? */
1297 edev = pci_dev_to_eeh_dev(pdev);
1298 if (!edev || !edev->pe)
1299 return NULL;
1300
1301 return edev->pe;
1302 }
1303 EXPORT_SYMBOL_GPL(eeh_iommu_group_to_pe);
1304
1305 #endif /* CONFIG_IOMMU_API */
1306
1307 /**
1308 * eeh_pe_set_option - Set options for the indicated PE
1309 * @pe: EEH PE
1310 * @option: requested option
1311 *
1312 * The routine is called to enable or disable EEH functionality
1313 * on the indicated PE, to enable IO or DMA for the frozen PE.
1314 */
eeh_pe_set_option(struct eeh_pe * pe,int option)1315 int eeh_pe_set_option(struct eeh_pe *pe, int option)
1316 {
1317 int ret = 0;
1318
1319 /* Invalid PE ? */
1320 if (!pe)
1321 return -ENODEV;
1322
1323 /*
1324 * EEH functionality could possibly be disabled, just
1325 * return error for the case. And the EEH functionality
1326 * isn't expected to be disabled on one specific PE.
1327 */
1328 switch (option) {
1329 case EEH_OPT_ENABLE:
1330 if (eeh_enabled()) {
1331 ret = eeh_pe_change_owner(pe);
1332 break;
1333 }
1334 ret = -EIO;
1335 break;
1336 case EEH_OPT_DISABLE:
1337 break;
1338 case EEH_OPT_THAW_MMIO:
1339 case EEH_OPT_THAW_DMA:
1340 case EEH_OPT_FREEZE_PE:
1341 if (!eeh_ops || !eeh_ops->set_option) {
1342 ret = -ENOENT;
1343 break;
1344 }
1345
1346 ret = eeh_pci_enable(pe, option);
1347 break;
1348 default:
1349 pr_debug("%s: Option %d out of range (%d, %d)\n",
1350 __func__, option, EEH_OPT_DISABLE, EEH_OPT_THAW_DMA);
1351 ret = -EINVAL;
1352 }
1353
1354 return ret;
1355 }
1356 EXPORT_SYMBOL_GPL(eeh_pe_set_option);
1357
1358 /**
1359 * eeh_pe_get_state - Retrieve PE's state
1360 * @pe: EEH PE
1361 *
1362 * Retrieve the PE's state, which includes 3 aspects: enabled
1363 * DMA, enabled IO and asserted reset.
1364 */
eeh_pe_get_state(struct eeh_pe * pe)1365 int eeh_pe_get_state(struct eeh_pe *pe)
1366 {
1367 int result, ret = 0;
1368 bool rst_active, dma_en, mmio_en;
1369
1370 /* Existing PE ? */
1371 if (!pe)
1372 return -ENODEV;
1373
1374 if (!eeh_ops || !eeh_ops->get_state)
1375 return -ENOENT;
1376
1377 /*
1378 * If the parent PE is owned by the host kernel and is undergoing
1379 * error recovery, we should return the PE state as temporarily
1380 * unavailable so that the error recovery on the guest is suspended
1381 * until the recovery completes on the host.
1382 */
1383 if (pe->parent &&
1384 !(pe->state & EEH_PE_REMOVED) &&
1385 (pe->parent->state & (EEH_PE_ISOLATED | EEH_PE_RECOVERING)))
1386 return EEH_PE_STATE_UNAVAIL;
1387
1388 result = eeh_ops->get_state(pe, NULL);
1389 rst_active = !!(result & EEH_STATE_RESET_ACTIVE);
1390 dma_en = !!(result & EEH_STATE_DMA_ENABLED);
1391 mmio_en = !!(result & EEH_STATE_MMIO_ENABLED);
1392
1393 if (rst_active)
1394 ret = EEH_PE_STATE_RESET;
1395 else if (dma_en && mmio_en)
1396 ret = EEH_PE_STATE_NORMAL;
1397 else if (!dma_en && !mmio_en)
1398 ret = EEH_PE_STATE_STOPPED_IO_DMA;
1399 else if (!dma_en && mmio_en)
1400 ret = EEH_PE_STATE_STOPPED_DMA;
1401 else
1402 ret = EEH_PE_STATE_UNAVAIL;
1403
1404 return ret;
1405 }
1406 EXPORT_SYMBOL_GPL(eeh_pe_get_state);
1407
eeh_pe_reenable_devices(struct eeh_pe * pe,bool include_passed)1408 static int eeh_pe_reenable_devices(struct eeh_pe *pe, bool include_passed)
1409 {
1410 struct eeh_dev *edev, *tmp;
1411 struct pci_dev *pdev;
1412 int ret = 0;
1413
1414 eeh_pe_restore_bars(pe);
1415
1416 /*
1417 * Reenable PCI devices as the devices passed
1418 * through are always enabled before the reset.
1419 */
1420 eeh_pe_for_each_dev(pe, edev, tmp) {
1421 pdev = eeh_dev_to_pci_dev(edev);
1422 if (!pdev)
1423 continue;
1424
1425 ret = pci_reenable_device(pdev);
1426 if (ret) {
1427 pr_warn("%s: Failure %d reenabling %s\n",
1428 __func__, ret, pci_name(pdev));
1429 return ret;
1430 }
1431 }
1432
1433 /* The PE is still in frozen state */
1434 if (include_passed || !eeh_pe_passed(pe)) {
1435 ret = eeh_unfreeze_pe(pe);
1436 } else
1437 pr_info("EEH: Note: Leaving passthrough PHB#%x-PE#%x frozen.\n",
1438 pe->phb->global_number, pe->addr);
1439 if (!ret)
1440 eeh_pe_state_clear(pe, EEH_PE_ISOLATED, include_passed);
1441 return ret;
1442 }
1443
1444
1445 /**
1446 * eeh_pe_reset - Issue PE reset according to specified type
1447 * @pe: EEH PE
1448 * @option: reset type
1449 * @include_passed: include passed-through devices?
1450 *
1451 * The routine is called to reset the specified PE with the
1452 * indicated type, either fundamental reset or hot reset.
1453 * PE reset is the most important part for error recovery.
1454 */
eeh_pe_reset(struct eeh_pe * pe,int option,bool include_passed)1455 int eeh_pe_reset(struct eeh_pe *pe, int option, bool include_passed)
1456 {
1457 int ret = 0;
1458
1459 /* Invalid PE ? */
1460 if (!pe)
1461 return -ENODEV;
1462
1463 if (!eeh_ops || !eeh_ops->set_option || !eeh_ops->reset)
1464 return -ENOENT;
1465
1466 switch (option) {
1467 case EEH_RESET_DEACTIVATE:
1468 ret = eeh_ops->reset(pe, option);
1469 eeh_pe_state_clear(pe, EEH_PE_CFG_BLOCKED, include_passed);
1470 if (ret)
1471 break;
1472
1473 ret = eeh_pe_reenable_devices(pe, include_passed);
1474 break;
1475 case EEH_RESET_HOT:
1476 case EEH_RESET_FUNDAMENTAL:
1477 /*
1478 * Proactively freeze the PE to drop all MMIO access
1479 * during reset, which should be banned as it's always
1480 * cause recursive EEH error.
1481 */
1482 eeh_ops->set_option(pe, EEH_OPT_FREEZE_PE);
1483
1484 eeh_pe_state_mark(pe, EEH_PE_CFG_BLOCKED);
1485 ret = eeh_ops->reset(pe, option);
1486 break;
1487 default:
1488 pr_debug("%s: Unsupported option %d\n",
1489 __func__, option);
1490 ret = -EINVAL;
1491 }
1492
1493 return ret;
1494 }
1495 EXPORT_SYMBOL_GPL(eeh_pe_reset);
1496
1497 /**
1498 * eeh_pe_configure - Configure PCI bridges after PE reset
1499 * @pe: EEH PE
1500 *
1501 * The routine is called to restore the PCI config space for
1502 * those PCI devices, especially PCI bridges affected by PE
1503 * reset issued previously.
1504 */
eeh_pe_configure(struct eeh_pe * pe)1505 int eeh_pe_configure(struct eeh_pe *pe)
1506 {
1507 int ret = 0;
1508
1509 /* Invalid PE ? */
1510 if (!pe)
1511 return -ENODEV;
1512
1513 return ret;
1514 }
1515 EXPORT_SYMBOL_GPL(eeh_pe_configure);
1516
1517 /**
1518 * eeh_pe_inject_err - Injecting the specified PCI error to the indicated PE
1519 * @pe: the indicated PE
1520 * @type: error type
1521 * @func: error function
1522 * @addr: address
1523 * @mask: address mask
1524 *
1525 * The routine is called to inject the specified PCI error, which
1526 * is determined by @type and @func, to the indicated PE for
1527 * testing purpose.
1528 */
eeh_pe_inject_err(struct eeh_pe * pe,int type,int func,unsigned long addr,unsigned long mask)1529 int eeh_pe_inject_err(struct eeh_pe *pe, int type, int func,
1530 unsigned long addr, unsigned long mask)
1531 {
1532 /* Invalid PE ? */
1533 if (!pe)
1534 return -ENODEV;
1535
1536 /* Unsupported operation ? */
1537 if (!eeh_ops || !eeh_ops->err_inject)
1538 return -ENOENT;
1539
1540 /* Check on PCI error function */
1541 if (func < EEH_ERR_FUNC_MIN || func > EEH_ERR_FUNC_MAX)
1542 return -EINVAL;
1543
1544 return eeh_ops->err_inject(pe, type, func, addr, mask);
1545 }
1546 EXPORT_SYMBOL_GPL(eeh_pe_inject_err);
1547
1548 #ifdef CONFIG_PROC_FS
proc_eeh_show(struct seq_file * m,void * v)1549 static int proc_eeh_show(struct seq_file *m, void *v)
1550 {
1551 if (!eeh_enabled()) {
1552 seq_printf(m, "EEH Subsystem is globally disabled\n");
1553 seq_printf(m, "eeh_total_mmio_ffs=%llu\n", eeh_stats.total_mmio_ffs);
1554 } else {
1555 seq_printf(m, "EEH Subsystem is enabled\n");
1556 seq_printf(m,
1557 "no device=%llu\n"
1558 "no device node=%llu\n"
1559 "no config address=%llu\n"
1560 "check not wanted=%llu\n"
1561 "eeh_total_mmio_ffs=%llu\n"
1562 "eeh_false_positives=%llu\n"
1563 "eeh_slot_resets=%llu\n",
1564 eeh_stats.no_device,
1565 eeh_stats.no_dn,
1566 eeh_stats.no_cfg_addr,
1567 eeh_stats.ignored_check,
1568 eeh_stats.total_mmio_ffs,
1569 eeh_stats.false_positives,
1570 eeh_stats.slot_resets);
1571 }
1572
1573 return 0;
1574 }
1575 #endif /* CONFIG_PROC_FS */
1576
eeh_break_device(struct pci_dev * pdev)1577 static int eeh_break_device(struct pci_dev *pdev)
1578 {
1579 struct resource *bar = NULL;
1580 void __iomem *mapped;
1581 u16 old, bit;
1582 int i, pos;
1583
1584 /* Do we have an MMIO BAR to disable? */
1585 for (i = 0; i <= PCI_STD_RESOURCE_END; i++) {
1586 struct resource *r = &pdev->resource[i];
1587
1588 if (!r->flags || !r->start)
1589 continue;
1590 if (r->flags & IORESOURCE_IO)
1591 continue;
1592 if (r->flags & IORESOURCE_UNSET)
1593 continue;
1594
1595 bar = r;
1596 break;
1597 }
1598
1599 if (!bar) {
1600 pci_err(pdev, "Unable to find Memory BAR to cause EEH with\n");
1601 return -ENXIO;
1602 }
1603
1604 pci_err(pdev, "Going to break: %pR\n", bar);
1605
1606 if (pdev->is_virtfn) {
1607 #ifndef CONFIG_PCI_IOV
1608 return -ENXIO;
1609 #else
1610 /*
1611 * VFs don't have a per-function COMMAND register, so the best
1612 * we can do is clear the Memory Space Enable bit in the PF's
1613 * SRIOV control reg.
1614 *
1615 * Unfortunately, this requires that we have a PF (i.e doesn't
1616 * work for a passed-through VF) and it has the potential side
1617 * effect of also causing an EEH on every other VF under the
1618 * PF. Oh well.
1619 */
1620 pdev = pdev->physfn;
1621 if (!pdev)
1622 return -ENXIO; /* passed through VFs have no PF */
1623
1624 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_SRIOV);
1625 pos += PCI_SRIOV_CTRL;
1626 bit = PCI_SRIOV_CTRL_MSE;
1627 #endif /* !CONFIG_PCI_IOV */
1628 } else {
1629 bit = PCI_COMMAND_MEMORY;
1630 pos = PCI_COMMAND;
1631 }
1632
1633 /*
1634 * Process here is:
1635 *
1636 * 1. Disable Memory space.
1637 *
1638 * 2. Perform an MMIO to the device. This should result in an error
1639 * (CA / UR) being raised by the device which results in an EEH
1640 * PE freeze. Using the in_8() accessor skips the eeh detection hook
1641 * so the freeze hook so the EEH Detection machinery won't be
1642 * triggered here. This is to match the usual behaviour of EEH
1643 * where the HW will asynchronously freeze a PE and it's up to
1644 * the kernel to notice and deal with it.
1645 *
1646 * 3. Turn Memory space back on. This is more important for VFs
1647 * since recovery will probably fail if we don't. For normal
1648 * the COMMAND register is reset as a part of re-initialising
1649 * the device.
1650 *
1651 * Breaking stuff is the point so who cares if it's racy ;)
1652 */
1653 pci_read_config_word(pdev, pos, &old);
1654
1655 mapped = ioremap(bar->start, PAGE_SIZE);
1656 if (!mapped) {
1657 pci_err(pdev, "Unable to map MMIO BAR %pR\n", bar);
1658 return -ENXIO;
1659 }
1660
1661 pci_write_config_word(pdev, pos, old & ~bit);
1662 in_8(mapped);
1663 pci_write_config_word(pdev, pos, old);
1664
1665 iounmap(mapped);
1666
1667 return 0;
1668 }
1669
eeh_pe_inject_mmio_error(struct pci_dev * pdev)1670 int eeh_pe_inject_mmio_error(struct pci_dev *pdev)
1671 {
1672 return eeh_break_device(pdev);
1673 }
1674
1675 #ifdef CONFIG_DEBUG_FS
1676
1677
eeh_debug_lookup_pdev(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1678 static struct pci_dev *eeh_debug_lookup_pdev(struct file *filp,
1679 const char __user *user_buf,
1680 size_t count, loff_t *ppos)
1681 {
1682 uint32_t domain, bus, dev, fn;
1683 struct pci_dev *pdev;
1684 char buf[20];
1685 int ret;
1686
1687 memset(buf, 0, sizeof(buf));
1688 ret = simple_write_to_buffer(buf, sizeof(buf)-1, ppos, user_buf, count);
1689 if (!ret)
1690 return ERR_PTR(-EFAULT);
1691
1692 ret = sscanf(buf, "%x:%x:%x.%x", &domain, &bus, &dev, &fn);
1693 if (ret != 4) {
1694 pr_err("%s: expected 4 args, got %d\n", __func__, ret);
1695 return ERR_PTR(-EINVAL);
1696 }
1697
1698 pdev = pci_get_domain_bus_and_slot(domain, bus, (dev << 3) | fn);
1699 if (!pdev)
1700 return ERR_PTR(-ENODEV);
1701
1702 return pdev;
1703 }
1704
eeh_enable_dbgfs_set(void * data,u64 val)1705 static int eeh_enable_dbgfs_set(void *data, u64 val)
1706 {
1707 if (val)
1708 eeh_clear_flag(EEH_FORCE_DISABLED);
1709 else
1710 eeh_add_flag(EEH_FORCE_DISABLED);
1711
1712 return 0;
1713 }
1714
eeh_enable_dbgfs_get(void * data,u64 * val)1715 static int eeh_enable_dbgfs_get(void *data, u64 *val)
1716 {
1717 if (eeh_enabled())
1718 *val = 0x1ul;
1719 else
1720 *val = 0x0ul;
1721 return 0;
1722 }
1723
1724 DEFINE_DEBUGFS_ATTRIBUTE(eeh_enable_dbgfs_ops, eeh_enable_dbgfs_get,
1725 eeh_enable_dbgfs_set, "0x%llx\n");
1726
eeh_force_recover_write(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1727 static ssize_t eeh_force_recover_write(struct file *filp,
1728 const char __user *user_buf,
1729 size_t count, loff_t *ppos)
1730 {
1731 struct pci_controller *hose;
1732 uint32_t phbid, pe_no;
1733 struct eeh_pe *pe;
1734 char buf[20];
1735 int ret;
1736
1737 ret = simple_write_to_buffer(buf, sizeof(buf), ppos, user_buf, count);
1738 if (!ret)
1739 return -EFAULT;
1740
1741 /*
1742 * When PE is NULL the event is a "special" event. Rather than
1743 * recovering a specific PE it forces the EEH core to scan for failed
1744 * PHBs and recovers each. This needs to be done before any device
1745 * recoveries can occur.
1746 */
1747 if (!strncmp(buf, "hwcheck", 7)) {
1748 __eeh_send_failure_event(NULL);
1749 return count;
1750 }
1751
1752 ret = sscanf(buf, "%x:%x", &phbid, &pe_no);
1753 if (ret != 2)
1754 return -EINVAL;
1755
1756 hose = pci_find_controller_for_domain(phbid);
1757 if (!hose)
1758 return -ENODEV;
1759
1760 /* Retrieve PE */
1761 pe = eeh_pe_get(hose, pe_no);
1762 if (!pe)
1763 return -ENODEV;
1764
1765 /*
1766 * We don't do any state checking here since the detection
1767 * process is async to the recovery process. The recovery
1768 * thread *should* not break even if we schedule a recovery
1769 * from an odd state (e.g. PE removed, or recovery of a
1770 * non-isolated PE)
1771 */
1772 __eeh_send_failure_event(pe);
1773
1774 return ret < 0 ? ret : count;
1775 }
1776
1777 static const struct file_operations eeh_force_recover_fops = {
1778 .open = simple_open,
1779 .write = eeh_force_recover_write,
1780 };
1781
eeh_debugfs_dev_usage(struct file * filp,char __user * user_buf,size_t count,loff_t * ppos)1782 static ssize_t eeh_debugfs_dev_usage(struct file *filp,
1783 char __user *user_buf,
1784 size_t count, loff_t *ppos)
1785 {
1786 static const char usage[] = "input format: <domain>:<bus>:<dev>.<fn>\n";
1787
1788 return simple_read_from_buffer(user_buf, count, ppos,
1789 usage, sizeof(usage) - 1);
1790 }
1791
eeh_dev_check_write(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1792 static ssize_t eeh_dev_check_write(struct file *filp,
1793 const char __user *user_buf,
1794 size_t count, loff_t *ppos)
1795 {
1796 struct pci_dev *pdev;
1797 struct eeh_dev *edev;
1798 int ret;
1799
1800 pdev = eeh_debug_lookup_pdev(filp, user_buf, count, ppos);
1801 if (IS_ERR(pdev))
1802 return PTR_ERR(pdev);
1803
1804 edev = pci_dev_to_eeh_dev(pdev);
1805 if (!edev) {
1806 pci_err(pdev, "No eeh_dev for this device!\n");
1807 pci_dev_put(pdev);
1808 return -ENODEV;
1809 }
1810
1811 ret = eeh_dev_check_failure(edev);
1812 pci_info(pdev, "eeh_dev_check_failure(%s) = %d\n",
1813 pci_name(pdev), ret);
1814
1815 pci_dev_put(pdev);
1816
1817 return count;
1818 }
1819
1820 static const struct file_operations eeh_dev_check_fops = {
1821 .open = simple_open,
1822 .write = eeh_dev_check_write,
1823 .read = eeh_debugfs_dev_usage,
1824 };
1825
eeh_dev_break_write(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1826 static ssize_t eeh_dev_break_write(struct file *filp,
1827 const char __user *user_buf,
1828 size_t count, loff_t *ppos)
1829 {
1830 struct pci_dev *pdev;
1831 int ret;
1832
1833 pdev = eeh_debug_lookup_pdev(filp, user_buf, count, ppos);
1834 if (IS_ERR(pdev))
1835 return PTR_ERR(pdev);
1836
1837 ret = eeh_break_device(pdev);
1838 pci_dev_put(pdev);
1839
1840 if (ret < 0)
1841 return ret;
1842
1843 return count;
1844 }
1845
1846 static const struct file_operations eeh_dev_break_fops = {
1847 .open = simple_open,
1848 .write = eeh_dev_break_write,
1849 .read = eeh_debugfs_dev_usage,
1850 };
1851
eeh_dev_can_recover(struct file * filp,const char __user * user_buf,size_t count,loff_t * ppos)1852 static ssize_t eeh_dev_can_recover(struct file *filp,
1853 const char __user *user_buf,
1854 size_t count, loff_t *ppos)
1855 {
1856 struct pci_driver *drv;
1857 struct pci_dev *pdev;
1858 size_t ret;
1859
1860 pdev = eeh_debug_lookup_pdev(filp, user_buf, count, ppos);
1861 if (IS_ERR(pdev))
1862 return PTR_ERR(pdev);
1863
1864 /*
1865 * In order for error recovery to work the driver needs to implement
1866 * .error_detected(), so it can quiesce IO to the device, and
1867 * .slot_reset() so it can re-initialise the device after a reset.
1868 *
1869 * Ideally they'd implement .resume() too, but some drivers which
1870 * we need to support (notably IPR) don't so I guess we can tolerate
1871 * that.
1872 *
1873 * .mmio_enabled() is mostly there as a work-around for devices which
1874 * take forever to re-init after a hot reset. Implementing that is
1875 * strictly optional.
1876 */
1877 drv = pci_dev_driver(pdev);
1878 if (drv &&
1879 drv->err_handler &&
1880 drv->err_handler->error_detected &&
1881 drv->err_handler->slot_reset) {
1882 ret = count;
1883 } else {
1884 ret = -EOPNOTSUPP;
1885 }
1886
1887 pci_dev_put(pdev);
1888
1889 return ret;
1890 }
1891
1892 static const struct file_operations eeh_dev_can_recover_fops = {
1893 .open = simple_open,
1894 .write = eeh_dev_can_recover,
1895 .read = eeh_debugfs_dev_usage,
1896 };
1897
1898 #endif
1899
eeh_init_proc(void)1900 static int __init eeh_init_proc(void)
1901 {
1902 if (machine_is(pseries) || machine_is(powernv)) {
1903 proc_create_single("powerpc/eeh", 0, NULL, proc_eeh_show);
1904 #ifdef CONFIG_DEBUG_FS
1905 debugfs_create_file_unsafe("eeh_enable", 0600,
1906 arch_debugfs_dir, NULL,
1907 &eeh_enable_dbgfs_ops);
1908 debugfs_create_u32("eeh_max_freezes", 0600,
1909 arch_debugfs_dir, &eeh_max_freezes);
1910 debugfs_create_bool("eeh_disable_recovery", 0600,
1911 arch_debugfs_dir,
1912 &eeh_debugfs_no_recover);
1913 debugfs_create_file_unsafe("eeh_dev_check", 0600,
1914 arch_debugfs_dir, NULL,
1915 &eeh_dev_check_fops);
1916 debugfs_create_file_unsafe("eeh_dev_break", 0600,
1917 arch_debugfs_dir, NULL,
1918 &eeh_dev_break_fops);
1919 debugfs_create_file_unsafe("eeh_force_recover", 0600,
1920 arch_debugfs_dir, NULL,
1921 &eeh_force_recover_fops);
1922 debugfs_create_file_unsafe("eeh_dev_can_recover", 0600,
1923 arch_debugfs_dir, NULL,
1924 &eeh_dev_can_recover_fops);
1925 eeh_cache_debugfs_init();
1926 #endif
1927 }
1928
1929 return 0;
1930 }
1931 __initcall(eeh_init_proc);
1932