2 * Copyright (c) 2013-2017, Mellanox Technologies, Ltd. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
13 * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <linux/kernel.h>
29 #include <linux/module.h>
30 #include <linux/random.h>
31 #include <linux/vmalloc.h>
32 #include <linux/hardirq.h>
33 #include <linux/delay.h>
34 #include <dev/mlx5/driver.h>
35 #include <dev/mlx5/mlx5_ifc.h>
36 #include "mlx5_core.h"
38 #define MLX5_HEALTH_POLL_INTERVAL (2 * HZ)
42 MLX5_DROP_NEW_HEALTH_WORK,
43 MLX5_DROP_NEW_RECOVERY_WORK,
47 MLX5_SENSOR_NO_ERR = 0,
48 MLX5_SENSOR_PCI_COMM_ERR = 1,
49 MLX5_SENSOR_PCI_ERR = 2,
50 MLX5_SENSOR_NIC_DISABLED = 3,
51 MLX5_SENSOR_NIC_SW_RESET = 4,
52 MLX5_SENSOR_FW_SYND_RFR = 5,
55 static int mlx5_fw_reset_enable = 1;
56 SYSCTL_INT(_hw_mlx5, OID_AUTO, fw_reset_enable, CTLFLAG_RWTUN,
57 &mlx5_fw_reset_enable, 0,
58 "Enable firmware reset");
60 static unsigned int sw_reset_to = 1200;
61 SYSCTL_UINT(_hw_mlx5, OID_AUTO, sw_reset_timeout, CTLFLAG_RWTUN,
63 "Minimum timeout in seconds between two firmware resets");
66 static int lock_sem_sw_reset(struct mlx5_core_dev *dev)
71 ret = -mlx5_vsc_lock(dev);
73 mlx5_core_warn(dev, "Timed out locking gateway %d\n", ret);
77 ret = -mlx5_vsc_lock_addr_space(dev, MLX5_SEMAPHORE_SW_RESET);
80 mlx5_core_dbg(dev, "SW reset FW semaphore already locked, another function will handle the reset\n");
82 mlx5_core_warn(dev, "SW reset semaphore lock return %d\n", ret);
85 /* Unlock GW access */
91 static int unlock_sem_sw_reset(struct mlx5_core_dev *dev)
96 ret = -mlx5_vsc_lock(dev);
98 mlx5_core_warn(dev, "Timed out locking gateway %d\n", ret);
102 ret = -mlx5_vsc_unlock_addr_space(dev, MLX5_SEMAPHORE_SW_RESET);
104 /* Unlock GW access */
105 mlx5_vsc_unlock(dev);
110 u8 mlx5_get_nic_state(struct mlx5_core_dev *dev)
112 return (ioread32be(&dev->iseg->cmdq_addr_l_sz) >> 8) & 7;
115 void mlx5_set_nic_state(struct mlx5_core_dev *dev, u8 state)
117 u32 cur_cmdq_addr_l_sz;
119 cur_cmdq_addr_l_sz = ioread32be(&dev->iseg->cmdq_addr_l_sz);
120 iowrite32be((cur_cmdq_addr_l_sz & 0xFFFFF000) |
121 state << MLX5_NIC_IFC_OFFSET,
122 &dev->iseg->cmdq_addr_l_sz);
125 static bool sensor_fw_synd_rfr(struct mlx5_core_dev *dev)
127 struct mlx5_core_health *health = &dev->priv.health;
128 struct mlx5_health_buffer __iomem *h = health->health;
129 u32 rfr = ioread32be(&h->rfr) >> MLX5_RFR_OFFSET;
130 u8 synd = ioread8(&h->synd);
133 mlx5_core_dbg(dev, "FW requests reset, synd: %d\n", synd);
137 static void mlx5_trigger_cmd_completions(struct mlx5_core_dev *dev)
142 /* wait for pending handlers to complete */
143 synchronize_irq(dev->priv.msix_arr[MLX5_EQ_VEC_CMD].vector);
144 spin_lock_irqsave(&dev->cmd.alloc_lock, flags);
145 vector = ~dev->cmd.bitmask & ((1ul << (1 << dev->cmd.log_sz)) - 1);
149 vector |= MLX5_TRIGGERED_CMD_COMP;
150 spin_unlock_irqrestore(&dev->cmd.alloc_lock, flags);
152 mlx5_core_dbg(dev, "vector 0x%jx\n", (uintmax_t)vector);
153 mlx5_cmd_comp_handler(dev, vector, MLX5_CMD_MODE_EVENTS);
157 spin_unlock_irqrestore(&dev->cmd.alloc_lock, flags);
160 static bool sensor_pci_no_comm(struct mlx5_core_dev *dev)
162 struct mlx5_core_health *health = &dev->priv.health;
163 struct mlx5_health_buffer __iomem *h = health->health;
164 bool err = ioread32be(&h->fw_ver) == 0xffffffff;
169 static bool sensor_nic_disabled(struct mlx5_core_dev *dev)
171 return mlx5_get_nic_state(dev) == MLX5_NIC_IFC_DISABLED;
174 static bool sensor_nic_sw_reset(struct mlx5_core_dev *dev)
176 return mlx5_get_nic_state(dev) == MLX5_NIC_IFC_SW_RESET;
179 static u32 check_fatal_sensors(struct mlx5_core_dev *dev)
181 if (sensor_pci_no_comm(dev))
182 return MLX5_SENSOR_PCI_COMM_ERR;
183 if (pci_channel_offline(dev->pdev))
184 return MLX5_SENSOR_PCI_ERR;
185 if (sensor_nic_disabled(dev))
186 return MLX5_SENSOR_NIC_DISABLED;
187 if (sensor_nic_sw_reset(dev))
188 return MLX5_SENSOR_NIC_SW_RESET;
189 if (sensor_fw_synd_rfr(dev))
190 return MLX5_SENSOR_FW_SYND_RFR;
192 return MLX5_SENSOR_NO_ERR;
195 static void reset_fw_if_needed(struct mlx5_core_dev *dev)
198 u32 cmdq_addr, fatal_error;
200 if (!mlx5_fw_reset_enable)
202 supported = (ioread32be(&dev->iseg->initializing) >>
203 MLX5_FW_RESET_SUPPORTED_OFFSET) & 1;
207 /* The reset only needs to be issued by one PF. The health buffer is
208 * shared between all functions, and will be cleared during a reset.
209 * Check again to avoid a redundant 2nd reset. If the fatal erros was
210 * PCI related a reset won't help.
212 fatal_error = check_fatal_sensors(dev);
213 if (fatal_error == MLX5_SENSOR_PCI_COMM_ERR ||
214 fatal_error == MLX5_SENSOR_NIC_DISABLED ||
215 fatal_error == MLX5_SENSOR_NIC_SW_RESET) {
216 mlx5_core_warn(dev, "Not issuing FW reset. Either it's already done or won't help.\n");
220 mlx5_core_warn(dev, "Issuing FW Reset\n");
221 /* Write the NIC interface field to initiate the reset, the command
222 * interface address also resides here, don't overwrite it.
224 cmdq_addr = ioread32be(&dev->iseg->cmdq_addr_l_sz);
225 iowrite32be((cmdq_addr & 0xFFFFF000) |
226 MLX5_NIC_IFC_SW_RESET << MLX5_NIC_IFC_OFFSET,
227 &dev->iseg->cmdq_addr_l_sz);
231 mlx5_health_allow_reset(struct mlx5_core_dev *dev)
233 struct mlx5_core_health *health = &dev->priv.health;
237 if (health->last_reset_req != 0) {
238 delta = ticks - health->last_reset_req;
240 ret = delta >= sw_reset_to;
246 * In principle, ticks may be 0. Setting it to off by one (-1)
247 * to prevent certain reset in next request.
249 health->last_reset_req = ticks ? : -1;
251 mlx5_core_warn(dev, "Firmware reset elided due to "
252 "auto-reset frequency threshold.\n");
256 #define MLX5_CRDUMP_WAIT_MS 60000
257 #define MLX5_FW_RESET_WAIT_MS 1000
258 #define MLX5_NIC_STATE_POLL_MS 5
259 void mlx5_enter_error_state(struct mlx5_core_dev *dev, bool force)
261 int end, delay_ms = MLX5_CRDUMP_WAIT_MS;
265 fatal_error = check_fatal_sensors(dev);
267 if (fatal_error || force) {
268 if (xchg(&dev->state, MLX5_DEVICE_STATE_INTERNAL_ERROR) ==
269 MLX5_DEVICE_STATE_INTERNAL_ERROR)
272 mlx5_core_err(dev, "internal state error detected\n");
273 mlx5_trigger_cmd_completions(dev);
276 mutex_lock(&dev->intf_state_mutex);
281 if (fatal_error == MLX5_SENSOR_FW_SYND_RFR &&
282 mlx5_health_allow_reset(dev)) {
283 /* Get cr-dump and reset FW semaphore */
284 if (mlx5_core_is_pf(dev))
285 lock = lock_sem_sw_reset(dev);
287 /* Execute cr-dump and SW reset */
288 if (lock != -EBUSY) {
290 reset_fw_if_needed(dev);
291 delay_ms = MLX5_FW_RESET_WAIT_MS;
295 /* Recover from SW reset */
296 end = jiffies + msecs_to_jiffies(delay_ms);
298 if (sensor_nic_disabled(dev))
301 msleep(MLX5_NIC_STATE_POLL_MS);
302 } while (!time_after(jiffies, end));
304 if (!sensor_nic_disabled(dev)) {
305 dev_err(&dev->pdev->dev, "NIC IFC still %d after %ums.\n",
306 mlx5_get_nic_state(dev), delay_ms);
309 /* Release FW semaphore if you are the lock owner */
311 unlock_sem_sw_reset(dev);
313 mlx5_core_err(dev, "system error event triggered\n");
316 mlx5_core_event(dev, MLX5_DEV_EVENT_SYS_ERROR, 1);
317 mutex_unlock(&dev->intf_state_mutex);
320 static void mlx5_handle_bad_state(struct mlx5_core_dev *dev)
322 u8 nic_mode = mlx5_get_nic_state(dev);
324 if (nic_mode == MLX5_NIC_IFC_SW_RESET) {
325 /* The IFC mode field is 3 bits, so it will read 0x7 in two cases:
326 * 1. PCI has been disabled (ie. PCI-AER, PF driver unloaded
327 * and this is a VF), this is not recoverable by SW reset.
328 * Logging of this is handled elsewhere.
329 * 2. FW reset has been issued by another function, driver can
330 * be reloaded to recover after the mode switches to
331 * MLX5_NIC_IFC_DISABLED.
333 if (dev->priv.health.fatal_error != MLX5_SENSOR_PCI_COMM_ERR)
334 mlx5_core_warn(dev, "NIC SW reset is already progress\n");
336 mlx5_core_warn(dev, "Communication with FW over the PCI link is down\n");
338 mlx5_core_warn(dev, "NIC mode %d\n", nic_mode);
341 mlx5_disable_device(dev);
344 #define MLX5_FW_RESET_WAIT_MS 1000
345 #define MLX5_NIC_STATE_POLL_MS 5
346 static void health_recover(struct work_struct *work)
348 unsigned long end = jiffies + msecs_to_jiffies(MLX5_FW_RESET_WAIT_MS);
349 struct mlx5_core_health *health;
350 struct delayed_work *dwork;
351 struct mlx5_core_dev *dev;
352 struct mlx5_priv *priv;
356 dwork = container_of(work, struct delayed_work, work);
357 health = container_of(dwork, struct mlx5_core_health, recover_work);
358 priv = container_of(health, struct mlx5_priv, health);
359 dev = container_of(priv, struct mlx5_core_dev, priv);
361 mtx_lock(&Giant); /* XXX newbus needs this */
363 if (sensor_pci_no_comm(dev)) {
364 dev_err(&dev->pdev->dev, "health recovery flow aborted, PCI reads still not working\n");
368 nic_mode = mlx5_get_nic_state(dev);
369 while (nic_mode != MLX5_NIC_IFC_DISABLED &&
370 !time_after(jiffies, end)) {
371 msleep(MLX5_NIC_STATE_POLL_MS);
372 nic_mode = mlx5_get_nic_state(dev);
375 if (nic_mode != MLX5_NIC_IFC_DISABLED) {
376 dev_err(&dev->pdev->dev, "health recovery flow aborted, unexpected NIC IFC mode %d.\n",
382 dev_err(&dev->pdev->dev, "starting health recovery flow\n");
383 mlx5_recover_device(dev);
389 /* How much time to wait until health resetting the driver (in msecs) */
390 #define MLX5_RECOVERY_DELAY_MSECS 60000
391 #define MLX5_RECOVERY_NO_DELAY 0
392 static unsigned long get_recovery_delay(struct mlx5_core_dev *dev)
394 return dev->priv.health.fatal_error == MLX5_SENSOR_PCI_ERR ||
395 dev->priv.health.fatal_error == MLX5_SENSOR_PCI_COMM_ERR ?
396 MLX5_RECOVERY_DELAY_MSECS : MLX5_RECOVERY_NO_DELAY;
399 static void health_care(struct work_struct *work)
401 struct mlx5_core_health *health;
402 unsigned long recover_delay;
403 struct mlx5_core_dev *dev;
404 struct mlx5_priv *priv;
407 health = container_of(work, struct mlx5_core_health, work);
408 priv = container_of(health, struct mlx5_priv, health);
409 dev = container_of(priv, struct mlx5_core_dev, priv);
411 mlx5_core_warn(dev, "handling bad device here\n");
412 mlx5_handle_bad_state(dev);
413 recover_delay = msecs_to_jiffies(get_recovery_delay(dev));
415 spin_lock_irqsave(&health->wq_lock, flags);
416 if (!test_bit(MLX5_DROP_NEW_RECOVERY_WORK, &health->flags)) {
417 mlx5_core_warn(dev, "Scheduling recovery work with %lums delay\n",
419 schedule_delayed_work(&health->recover_work, recover_delay);
421 dev_err(&dev->pdev->dev,
422 "new health works are not permitted at this stage\n");
424 spin_unlock_irqrestore(&health->wq_lock, flags);
427 static int get_next_poll_jiffies(void)
431 get_random_bytes(&next, sizeof(next));
433 next += jiffies + MLX5_HEALTH_POLL_INTERVAL;
438 void mlx5_trigger_health_work(struct mlx5_core_dev *dev)
440 struct mlx5_core_health *health = &dev->priv.health;
443 spin_lock_irqsave(&health->wq_lock, flags);
444 if (!test_bit(MLX5_DROP_NEW_HEALTH_WORK, &health->flags))
445 queue_work(health->wq, &health->work);
447 dev_err(&dev->pdev->dev,
448 "new health works are not permitted at this stage\n");
449 spin_unlock_irqrestore(&health->wq_lock, flags);
452 static const char *hsynd_str(u8 synd)
455 case MLX5_HEALTH_SYNDR_FW_ERR:
456 return "firmware internal error";
457 case MLX5_HEALTH_SYNDR_IRISC_ERR:
458 return "irisc not responding";
459 case MLX5_HEALTH_SYNDR_HW_UNRECOVERABLE_ERR:
460 return "unrecoverable hardware error";
461 case MLX5_HEALTH_SYNDR_CRC_ERR:
462 return "firmware CRC error";
463 case MLX5_HEALTH_SYNDR_FETCH_PCI_ERR:
464 return "ICM fetch PCI error";
465 case MLX5_HEALTH_SYNDR_HW_FTL_ERR:
466 return "HW fatal error\n";
467 case MLX5_HEALTH_SYNDR_ASYNC_EQ_OVERRUN_ERR:
468 return "async EQ buffer overrun";
469 case MLX5_HEALTH_SYNDR_EQ_ERR:
471 case MLX5_HEALTH_SYNDR_EQ_INV:
472 return "Invalid EQ referenced";
473 case MLX5_HEALTH_SYNDR_FFSER_ERR:
474 return "FFSER error";
475 case MLX5_HEALTH_SYNDR_HIGH_TEMP:
476 return "High temprature";
478 return "unrecognized error";
482 static void print_health_info(struct mlx5_core_dev *dev)
484 struct mlx5_core_health *health = &dev->priv.health;
485 struct mlx5_health_buffer __iomem *h = health->health;
490 /* If the syndrom is 0, the device is OK and no need to print buffer */
491 if (!ioread8(&h->synd))
494 for (i = 0; i < ARRAY_SIZE(h->assert_var); i++)
495 printf("mlx5_core: INFO: ""assert_var[%d] 0x%08x\n", i, ioread32be(h->assert_var + i));
497 printf("mlx5_core: INFO: ""assert_exit_ptr 0x%08x\n", ioread32be(&h->assert_exit_ptr));
498 printf("mlx5_core: INFO: ""assert_callra 0x%08x\n", ioread32be(&h->assert_callra));
499 snprintf(fw_str, sizeof(fw_str), "%d.%d.%d", fw_rev_maj(dev), fw_rev_min(dev), fw_rev_sub(dev));
500 printf("mlx5_core: INFO: ""fw_ver %s\n", fw_str);
501 printf("mlx5_core: INFO: ""hw_id 0x%08x\n", ioread32be(&h->hw_id));
502 printf("mlx5_core: INFO: ""irisc_index %d\n", ioread8(&h->irisc_index));
503 printf("mlx5_core: INFO: ""synd 0x%x: %s\n", ioread8(&h->synd), hsynd_str(ioread8(&h->synd)));
504 printf("mlx5_core: INFO: ""ext_synd 0x%04x\n", ioread16be(&h->ext_synd));
505 fw = ioread32be(&h->fw_ver);
506 printf("mlx5_core: INFO: ""raw fw_ver 0x%08x\n", fw);
509 static void poll_health(unsigned long data)
511 struct mlx5_core_dev *dev = (struct mlx5_core_dev *)data;
512 struct mlx5_core_health *health = &dev->priv.health;
516 if (dev->state != MLX5_DEVICE_STATE_UP)
519 if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)
522 count = ioread32be(health->health_counter);
523 if (count == health->prev)
524 ++health->miss_counter;
526 health->miss_counter = 0;
528 health->prev = count;
529 if (health->miss_counter == MAX_MISSES) {
530 mlx5_core_err(dev, "device's health compromised - reached miss count\n");
531 print_health_info(dev);
534 fatal_error = check_fatal_sensors(dev);
536 if (fatal_error && !health->fatal_error) {
537 mlx5_core_err(dev, "Fatal error %u detected\n", fatal_error);
538 dev->priv.health.fatal_error = fatal_error;
539 print_health_info(dev);
540 mlx5_trigger_health_work(dev);
544 mod_timer(&health->timer, get_next_poll_jiffies());
547 void mlx5_start_health_poll(struct mlx5_core_dev *dev)
549 struct mlx5_core_health *health = &dev->priv.health;
551 init_timer(&health->timer);
552 health->fatal_error = MLX5_SENSOR_NO_ERR;
553 clear_bit(MLX5_DROP_NEW_HEALTH_WORK, &health->flags);
554 clear_bit(MLX5_DROP_NEW_RECOVERY_WORK, &health->flags);
555 health->health = &dev->iseg->health;
556 health->health_counter = &dev->iseg->health_counter;
558 setup_timer(&health->timer, poll_health, (unsigned long)dev);
559 mod_timer(&health->timer,
560 round_jiffies(jiffies + MLX5_HEALTH_POLL_INTERVAL));
563 void mlx5_stop_health_poll(struct mlx5_core_dev *dev, bool disable_health)
565 struct mlx5_core_health *health = &dev->priv.health;
568 if (disable_health) {
569 spin_lock_irqsave(&health->wq_lock, flags);
570 set_bit(MLX5_DROP_NEW_HEALTH_WORK, &health->flags);
571 set_bit(MLX5_DROP_NEW_RECOVERY_WORK, &health->flags);
572 spin_unlock_irqrestore(&health->wq_lock, flags);
575 del_timer_sync(&health->timer);
578 void mlx5_drain_health_wq(struct mlx5_core_dev *dev)
580 struct mlx5_core_health *health = &dev->priv.health;
583 spin_lock_irqsave(&health->wq_lock, flags);
584 set_bit(MLX5_DROP_NEW_HEALTH_WORK, &health->flags);
585 set_bit(MLX5_DROP_NEW_RECOVERY_WORK, &health->flags);
586 spin_unlock_irqrestore(&health->wq_lock, flags);
587 cancel_delayed_work_sync(&health->recover_work);
588 cancel_work_sync(&health->work);
591 void mlx5_drain_health_recovery(struct mlx5_core_dev *dev)
593 struct mlx5_core_health *health = &dev->priv.health;
596 spin_lock_irqsave(&health->wq_lock, flags);
597 set_bit(MLX5_DROP_NEW_RECOVERY_WORK, &health->flags);
598 spin_unlock_irqrestore(&health->wq_lock, flags);
599 cancel_delayed_work_sync(&dev->priv.health.recover_work);
602 void mlx5_health_cleanup(struct mlx5_core_dev *dev)
604 struct mlx5_core_health *health = &dev->priv.health;
606 destroy_workqueue(health->wq);
609 #define HEALTH_NAME "mlx5_health"
610 int mlx5_health_init(struct mlx5_core_dev *dev)
612 struct mlx5_core_health *health;
616 health = &dev->priv.health;
617 len = strlen(HEALTH_NAME) + strlen(dev_name(&dev->pdev->dev));
618 name = kmalloc(len + 1, GFP_KERNEL);
622 snprintf(name, len, "%s:%s", HEALTH_NAME, dev_name(&dev->pdev->dev));
623 health->wq = create_singlethread_workqueue(name);
628 spin_lock_init(&health->wq_lock);
629 INIT_WORK(&health->work, health_care);
630 INIT_DELAYED_WORK(&health->recover_work, health_recover);