drivers/net/ethernet/mellanox/mlx5/core/health.c

   1 /*
   2  * Copyright (c) 2013-2015, Mellanox Technologies. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <linux/kernel.h>
  34 #include <linux/module.h>
  35 #include <linux/random.h>
  36 #include <linux/vmalloc.h>
  37 #include <linux/hardirq.h>
  38 #include <linux/mlx5/driver.h>
  39 #include <linux/mlx5/cmd.h>
  40 #include "mlx5_core.h"
  41
  42 enum {
  43         MLX5_HEALTH_POLL_INTERVAL       = 2 * HZ,
  44         MAX_MISSES                      = 3,
  45 };
  46
  47 enum {
  48         MLX5_HEALTH_SYNDR_FW_ERR                = 0x1,
  49         MLX5_HEALTH_SYNDR_IRISC_ERR             = 0x7,
  50         MLX5_HEALTH_SYNDR_HW_UNRECOVERABLE_ERR  = 0x8,
  51         MLX5_HEALTH_SYNDR_CRC_ERR               = 0x9,
  52         MLX5_HEALTH_SYNDR_FETCH_PCI_ERR         = 0xa,
  53         MLX5_HEALTH_SYNDR_HW_FTL_ERR            = 0xb,
  54         MLX5_HEALTH_SYNDR_ASYNC_EQ_OVERRUN_ERR  = 0xc,
  55         MLX5_HEALTH_SYNDR_EQ_ERR                = 0xd,
  56         MLX5_HEALTH_SYNDR_EQ_INV                = 0xe,
  57         MLX5_HEALTH_SYNDR_FFSER_ERR             = 0xf,
  58         MLX5_HEALTH_SYNDR_HIGH_TEMP             = 0x10
  59 };
  60
  61 enum {
  62         MLX5_NIC_IFC_FULL               = 0,
  63         MLX5_NIC_IFC_DISABLED           = 1,
  64         MLX5_NIC_IFC_NO_DRAM_NIC        = 2,
  65         MLX5_NIC_IFC_INVALID            = 3
  66 };
  67
  68 enum {
  69         MLX5_DROP_NEW_HEALTH_WORK,
  70 };
  71
  72 static u8 get_nic_state(struct mlx5_core_dev *dev)
  73 {
  74         return (ioread32be(&dev->iseg->cmdq_addr_l_sz) >> 8) & 3;
  75 }
  76
  77 static void trigger_cmd_completions(struct mlx5_core_dev *dev)
  78 {
  79         unsigned long flags;
  80         u64 vector;
  81
  82         /* wait for pending handlers to complete */
  83         synchronize_irq(dev->priv.msix_arr[MLX5_EQ_VEC_CMD].vector);
  84         spin_lock_irqsave(&dev->cmd.alloc_lock, flags);
  85         vector = ~dev->cmd.bitmask & ((1ul << (1 << dev->cmd.log_sz)) - 1);
  86         if (!vector)
  87                 goto no_trig;
  88
  89         vector |= MLX5_TRIGGERED_CMD_COMP;
  90         spin_unlock_irqrestore(&dev->cmd.alloc_lock, flags);
  91
  92         mlx5_core_dbg(dev, "vector 0x%llx\n", vector);
  93         mlx5_cmd_comp_handler(dev, vector);
  94         return;
  95
  96 no_trig:
  97         spin_unlock_irqrestore(&dev->cmd.alloc_lock, flags);
  98 }
  99
 100 static int in_fatal(struct mlx5_core_dev *dev)
 101 {
 102         struct mlx5_core_health *health = &dev->priv.health;
 103         struct health_buffer __iomem *h = health->health;
 104
 105         if (get_nic_state(dev) == MLX5_NIC_IFC_DISABLED)
 106                 return 1;
 107
 108         if (ioread32be(&h->fw_ver) == 0xffffffff)
 109                 return 1;
 110
 111         return 0;
 112 }
 113
 114 void mlx5_enter_error_state(struct mlx5_core_dev *dev)
 115 {
 116         mutex_lock(&dev->intf_state_mutex);
 117         if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)
 118                 goto unlock;
 119
 120         mlx5_core_err(dev, "start\n");
 121         if (pci_channel_offline(dev->pdev) || in_fatal(dev)) {
 122                 dev->state = MLX5_DEVICE_STATE_INTERNAL_ERROR;
 123                 trigger_cmd_completions(dev);
 124         }
 125
 126         mlx5_core_event(dev, MLX5_DEV_EVENT_SYS_ERROR, 0);
 127         mlx5_core_err(dev, "end\n");
 128
 129 unlock:
 130         mutex_unlock(&dev->intf_state_mutex);
 131 }
 132
 133 static void mlx5_handle_bad_state(struct mlx5_core_dev *dev)
 134 {
 135         u8 nic_interface = get_nic_state(dev);
 136
 137         switch (nic_interface) {
 138         case MLX5_NIC_IFC_FULL:
 139                 mlx5_core_warn(dev, "Expected to see disabled NIC but it is full driver\n");
 140                 break;
 141
 142         case MLX5_NIC_IFC_DISABLED:
 143                 mlx5_core_warn(dev, "starting teardown\n");
 144                 break;
 145
 146         case MLX5_NIC_IFC_NO_DRAM_NIC:
 147                 mlx5_core_warn(dev, "Expected to see disabled NIC but it is no dram nic\n");
 148                 break;
 149         default:
 150                 mlx5_core_warn(dev, "Expected to see disabled NIC but it is has invalid value %d\n",
 151                                nic_interface);
 152         }
 153
 154         mlx5_disable_device(dev);
 155 }
 156
 157 static void health_recover(struct work_struct *work)
 158 {
 159         struct mlx5_core_health *health;
 160         struct delayed_work *dwork;
 161         struct mlx5_core_dev *dev;
 162         struct mlx5_priv *priv;
 163         u8 nic_state;
 164
 165         dwork = container_of(work, struct delayed_work, work);
 166         health = container_of(dwork, struct mlx5_core_health, recover_work);
 167         priv = container_of(health, struct mlx5_priv, health);
 168         dev = container_of(priv, struct mlx5_core_dev, priv);
 169
 170         nic_state = get_nic_state(dev);
 171         if (nic_state == MLX5_NIC_IFC_INVALID) {
 172                 dev_err(&dev->pdev->dev, "health recovery flow aborted since the nic state is invalid\n");
 173                 return;
 174         }
 175
 176         dev_err(&dev->pdev->dev, "starting health recovery flow\n");
 177         mlx5_recover_device(dev);
 178 }
 179
 180 /* How much time to wait until health resetting the driver (in msecs) */
 181 #define MLX5_RECOVERY_DELAY_MSECS 60000
 182 static void health_care(struct work_struct *work)
 183 {
 184         unsigned long recover_delay = msecs_to_jiffies(MLX5_RECOVERY_DELAY_MSECS);
 185         struct mlx5_core_health *health;
 186         struct mlx5_core_dev *dev;
 187         struct mlx5_priv *priv;
 188
 189         health = container_of(work, struct mlx5_core_health, work);
 190         priv = container_of(health, struct mlx5_priv, health);
 191         dev = container_of(priv, struct mlx5_core_dev, priv);
 192         mlx5_core_warn(dev, "handling bad device here\n");
 193         mlx5_handle_bad_state(dev);
 194
 195         spin_lock(&health->wq_lock);
 196         if (!test_bit(MLX5_DROP_NEW_HEALTH_WORK, &health->flags))
 197                 schedule_delayed_work(&health->recover_work, recover_delay);
 198         else
 199                 dev_err(&dev->pdev->dev,
 200                         "new health works are not permitted at this stage\n");
 201         spin_unlock(&health->wq_lock);
 202 }
 203
 204 static const char *hsynd_str(u8 synd)
 205 {
 206         switch (synd) {
 207         case MLX5_HEALTH_SYNDR_FW_ERR:
 208                 return "firmware internal error";
 209         case MLX5_HEALTH_SYNDR_IRISC_ERR:
 210                 return "irisc not responding";
 211         case MLX5_HEALTH_SYNDR_HW_UNRECOVERABLE_ERR:
 212                 return "unrecoverable hardware error";
 213         case MLX5_HEALTH_SYNDR_CRC_ERR:
 214                 return "firmware CRC error";
 215         case MLX5_HEALTH_SYNDR_FETCH_PCI_ERR:
 216                 return "ICM fetch PCI error";
 217         case MLX5_HEALTH_SYNDR_HW_FTL_ERR:
 218                 return "HW fatal error\n";
 219         case MLX5_HEALTH_SYNDR_ASYNC_EQ_OVERRUN_ERR:
 220                 return "async EQ buffer overrun";
 221         case MLX5_HEALTH_SYNDR_EQ_ERR:
 222                 return "EQ error";
 223         case MLX5_HEALTH_SYNDR_EQ_INV:
 224                 return "Invalid EQ referenced";
 225         case MLX5_HEALTH_SYNDR_FFSER_ERR:
 226                 return "FFSER error";
 227         case MLX5_HEALTH_SYNDR_HIGH_TEMP:
 228                 return "High temperature";
 229         default:
 230                 return "unrecognized error";
 231         }
 232 }
 233
 234 static u16 get_maj(u32 fw)
 235 {
 236         return fw >> 28;
 237 }
 238
 239 static u16 get_min(u32 fw)
 240 {
 241         return fw >> 16 & 0xfff;
 242 }
 243
 244 static u16 get_sub(u32 fw)
 245 {
 246         return fw & 0xffff;
 247 }
 248
 249 static void print_health_info(struct mlx5_core_dev *dev)
 250 {
 251         struct mlx5_core_health *health = &dev->priv.health;
 252         struct health_buffer __iomem *h = health->health;
 253         char fw_str[18];
 254         u32 fw;
 255         int i;
 256
 257         /* If the syndrom is 0, the device is OK and no need to print buffer */
 258         if (!ioread8(&h->synd))
 259                 return;
 260
 261         for (i = 0; i < ARRAY_SIZE(h->assert_var); i++)
 262                 dev_err(&dev->pdev->dev, "assert_var[%d] 0x%08x\n", i, ioread32be(h->assert_var + i));
 263
 264         dev_err(&dev->pdev->dev, "assert_exit_ptr 0x%08x\n", ioread32be(&h->assert_exit_ptr));
 265         dev_err(&dev->pdev->dev, "assert_callra 0x%08x\n", ioread32be(&h->assert_callra));
 266         fw = ioread32be(&h->fw_ver);
 267         sprintf(fw_str, "%d.%d.%d", get_maj(fw), get_min(fw), get_sub(fw));
 268         dev_err(&dev->pdev->dev, "fw_ver %s\n", fw_str);
 269         dev_err(&dev->pdev->dev, "hw_id 0x%08x\n", ioread32be(&h->hw_id));
 270         dev_err(&dev->pdev->dev, "irisc_index %d\n", ioread8(&h->irisc_index));
 271         dev_err(&dev->pdev->dev, "synd 0x%x: %s\n", ioread8(&h->synd), hsynd_str(ioread8(&h->synd)));
 272         dev_err(&dev->pdev->dev, "ext_synd 0x%04x\n", ioread16be(&h->ext_synd));
 273 }
 274
 275 static unsigned long get_next_poll_jiffies(void)
 276 {
 277         unsigned long next;
 278
 279         get_random_bytes(&next, sizeof(next));
 280         next %= HZ;
 281         next += jiffies + MLX5_HEALTH_POLL_INTERVAL;
 282
 283         return next;
 284 }
 285
 286 static void poll_health(unsigned long data)
 287 {
 288         struct mlx5_core_dev *dev = (struct mlx5_core_dev *)data;
 289         struct mlx5_core_health *health = &dev->priv.health;
 290         u32 count;
 291
 292         if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
 293                 mod_timer(&health->timer, get_next_poll_jiffies());
 294                 return;
 295         }
 296
 297         count = ioread32be(health->health_counter);
 298         if (count == health->prev)
 299                 ++health->miss_counter;
 300         else
 301                 health->miss_counter = 0;
 302
 303         health->prev = count;
 304         if (health->miss_counter == MAX_MISSES) {
 305                 dev_err(&dev->pdev->dev, "device's health compromised - reached miss count\n");
 306                 print_health_info(dev);
 307         } else {
 308                 mod_timer(&health->timer, get_next_poll_jiffies());
 309         }
 310
 311         if (in_fatal(dev) && !health->sick) {
 312                 health->sick = true;
 313                 print_health_info(dev);
 314                 spin_lock(&health->wq_lock);
 315                 if (!test_bit(MLX5_DROP_NEW_HEALTH_WORK, &health->flags))
 316                         queue_work(health->wq, &health->work);
 317                 else
 318                         dev_err(&dev->pdev->dev,
 319                                 "new health works are not permitted at this stage\n");
 320                 spin_unlock(&health->wq_lock);
 321         }
 322 }
 323
 324 void mlx5_start_health_poll(struct mlx5_core_dev *dev)
 325 {
 326         struct mlx5_core_health *health = &dev->priv.health;
 327
 328         init_timer(&health->timer);
 329         health->sick = 0;
 330         clear_bit(MLX5_DROP_NEW_HEALTH_WORK, &health->flags);
 331         health->health = &dev->iseg->health;
 332         health->health_counter = &dev->iseg->health_counter;
 333
 334         health->timer.data = (unsigned long)dev;
 335         health->timer.function = poll_health;
 336         health->timer.expires = round_jiffies(jiffies + MLX5_HEALTH_POLL_INTERVAL);
 337         add_timer(&health->timer);
 338 }
 339
 340 void mlx5_stop_health_poll(struct mlx5_core_dev *dev)
 341 {
 342         struct mlx5_core_health *health = &dev->priv.health;
 343
 344         del_timer_sync(&health->timer);
 345 }
 346
 347 void mlx5_drain_health_wq(struct mlx5_core_dev *dev)
 348 {
 349         struct mlx5_core_health *health = &dev->priv.health;
 350
 351         spin_lock(&health->wq_lock);
 352         set_bit(MLX5_DROP_NEW_HEALTH_WORK, &health->flags);
 353         spin_unlock(&health->wq_lock);
 354         cancel_delayed_work_sync(&health->recover_work);
 355         cancel_work_sync(&health->work);
 356 }
 357
 358 void mlx5_health_cleanup(struct mlx5_core_dev *dev)
 359 {
 360         struct mlx5_core_health *health = &dev->priv.health;
 361
 362         destroy_workqueue(health->wq);
 363 }
 364
 365 int mlx5_health_init(struct mlx5_core_dev *dev)
 366 {
 367         struct mlx5_core_health *health;
 368         char *name;
 369
 370         health = &dev->priv.health;
 371         name = kmalloc(64, GFP_KERNEL);
 372         if (!name)
 373                 return -ENOMEM;
 374
 375         strcpy(name, "mlx5_health");
 376         strcat(name, dev_name(&dev->pdev->dev));
 377         health->wq = create_singlethread_workqueue(name);
 378         kfree(name);
 379         if (!health->wq)
 380                 return -ENOMEM;
 381         spin_lock_init(&health->wq_lock);
 382         INIT_WORK(&health->work, health_care);
 383         INIT_DELAYED_WORK(&health->recover_work, health_recover);
 384
 385         return 0;
 386 }