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 };
  66
  67 static u8 get_nic_interface(struct mlx5_core_dev *dev)
  68 {
  69         return (ioread32be(&dev->iseg->cmdq_addr_l_sz) >> 8) & 3;
  70 }
  71
  72 static void trigger_cmd_completions(struct mlx5_core_dev *dev)
  73 {
  74         unsigned long flags;
  75         u64 vector;
  76
  77         /* wait for pending handlers to complete */
  78         synchronize_irq(dev->priv.msix_arr[MLX5_EQ_VEC_CMD].vector);
  79         spin_lock_irqsave(&dev->cmd.alloc_lock, flags);
  80         vector = ~dev->cmd.bitmask & ((1ul << (1 << dev->cmd.log_sz)) - 1);
  81         if (!vector)
  82                 goto no_trig;
  83
  84         vector |= MLX5_TRIGGERED_CMD_COMP;
  85         spin_unlock_irqrestore(&dev->cmd.alloc_lock, flags);
  86
  87         mlx5_core_dbg(dev, "vector 0x%llx\n", vector);
  88         mlx5_cmd_comp_handler(dev, vector);
  89         return;
  90
  91 no_trig:
  92         spin_unlock_irqrestore(&dev->cmd.alloc_lock, flags);
  93 }
  94
  95 static int in_fatal(struct mlx5_core_dev *dev)
  96 {
  97         struct mlx5_core_health *health = &dev->priv.health;
  98         struct health_buffer __iomem *h = health->health;
  99
 100         if (get_nic_interface(dev) == MLX5_NIC_IFC_DISABLED)
 101                 return 1;
 102
 103         if (ioread32be(&h->fw_ver) == 0xffffffff)
 104                 return 1;
 105
 106         return 0;
 107 }
 108
 109 void mlx5_enter_error_state(struct mlx5_core_dev *dev)
 110 {
 111         mutex_lock(&dev->intf_state_mutex);
 112         if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)
 113                 goto unlock;
 114
 115         mlx5_core_err(dev, "start\n");
 116         if (pci_channel_offline(dev->pdev) || in_fatal(dev)) {
 117                 dev->state = MLX5_DEVICE_STATE_INTERNAL_ERROR;
 118                 trigger_cmd_completions(dev);
 119         }
 120
 121         mlx5_core_event(dev, MLX5_DEV_EVENT_SYS_ERROR, 0);
 122         mlx5_core_err(dev, "end\n");
 123
 124 unlock:
 125         mutex_unlock(&dev->intf_state_mutex);
 126 }
 127
 128 static void mlx5_handle_bad_state(struct mlx5_core_dev *dev)
 129 {
 130         u8 nic_interface = get_nic_interface(dev);
 131
 132         switch (nic_interface) {
 133         case MLX5_NIC_IFC_FULL:
 134                 mlx5_core_warn(dev, "Expected to see disabled NIC but it is full driver\n");
 135                 break;
 136
 137         case MLX5_NIC_IFC_DISABLED:
 138                 mlx5_core_warn(dev, "starting teardown\n");
 139                 break;
 140
 141         case MLX5_NIC_IFC_NO_DRAM_NIC:
 142                 mlx5_core_warn(dev, "Expected to see disabled NIC but it is no dram nic\n");
 143                 break;
 144         default:
 145                 mlx5_core_warn(dev, "Expected to see disabled NIC but it is has invalid value %d\n",
 146                                nic_interface);
 147         }
 148
 149         mlx5_disable_device(dev);
 150 }
 151
 152 static void health_care(struct work_struct *work)
 153 {
 154         struct mlx5_core_health *health;
 155         struct mlx5_core_dev *dev;
 156         struct mlx5_priv *priv;
 157
 158         health = container_of(work, struct mlx5_core_health, work);
 159         priv = container_of(health, struct mlx5_priv, health);
 160         dev = container_of(priv, struct mlx5_core_dev, priv);
 161         mlx5_core_warn(dev, "handling bad device here\n");
 162         mlx5_handle_bad_state(dev);
 163 }
 164
 165 static const char *hsynd_str(u8 synd)
 166 {
 167         switch (synd) {
 168         case MLX5_HEALTH_SYNDR_FW_ERR:
 169                 return "firmware internal error";
 170         case MLX5_HEALTH_SYNDR_IRISC_ERR:
 171                 return "irisc not responding";
 172         case MLX5_HEALTH_SYNDR_HW_UNRECOVERABLE_ERR:
 173                 return "unrecoverable hardware error";
 174         case MLX5_HEALTH_SYNDR_CRC_ERR:
 175                 return "firmware CRC error";
 176         case MLX5_HEALTH_SYNDR_FETCH_PCI_ERR:
 177                 return "ICM fetch PCI error";
 178         case MLX5_HEALTH_SYNDR_HW_FTL_ERR:
 179                 return "HW fatal error\n";
 180         case MLX5_HEALTH_SYNDR_ASYNC_EQ_OVERRUN_ERR:
 181                 return "async EQ buffer overrun";
 182         case MLX5_HEALTH_SYNDR_EQ_ERR:
 183                 return "EQ error";
 184         case MLX5_HEALTH_SYNDR_EQ_INV:
 185                 return "Invalid EQ referenced";
 186         case MLX5_HEALTH_SYNDR_FFSER_ERR:
 187                 return "FFSER error";
 188         case MLX5_HEALTH_SYNDR_HIGH_TEMP:
 189                 return "High temperature";
 190         default:
 191                 return "unrecognized error";
 192         }
 193 }
 194
 195 static u16 get_maj(u32 fw)
 196 {
 197         return fw >> 28;
 198 }
 199
 200 static u16 get_min(u32 fw)
 201 {
 202         return fw >> 16 & 0xfff;
 203 }
 204
 205 static u16 get_sub(u32 fw)
 206 {
 207         return fw & 0xffff;
 208 }
 209
 210 static void print_health_info(struct mlx5_core_dev *dev)
 211 {
 212         struct mlx5_core_health *health = &dev->priv.health;
 213         struct health_buffer __iomem *h = health->health;
 214         char fw_str[18];
 215         u32 fw;
 216         int i;
 217
 218         /* If the syndrom is 0, the device is OK and no need to print buffer */
 219         if (!ioread8(&h->synd))
 220                 return;
 221
 222         for (i = 0; i < ARRAY_SIZE(h->assert_var); i++)
 223                 dev_err(&dev->pdev->dev, "assert_var[%d] 0x%08x\n", i, ioread32be(h->assert_var + i));
 224
 225         dev_err(&dev->pdev->dev, "assert_exit_ptr 0x%08x\n", ioread32be(&h->assert_exit_ptr));
 226         dev_err(&dev->pdev->dev, "assert_callra 0x%08x\n", ioread32be(&h->assert_callra));
 227         fw = ioread32be(&h->fw_ver);
 228         sprintf(fw_str, "%d.%d.%d", get_maj(fw), get_min(fw), get_sub(fw));
 229         dev_err(&dev->pdev->dev, "fw_ver %s\n", fw_str);
 230         dev_err(&dev->pdev->dev, "hw_id 0x%08x\n", ioread32be(&h->hw_id));
 231         dev_err(&dev->pdev->dev, "irisc_index %d\n", ioread8(&h->irisc_index));
 232         dev_err(&dev->pdev->dev, "synd 0x%x: %s\n", ioread8(&h->synd), hsynd_str(ioread8(&h->synd)));
 233         dev_err(&dev->pdev->dev, "ext_synd 0x%04x\n", ioread16be(&h->ext_synd));
 234 }
 235
 236 static unsigned long get_next_poll_jiffies(void)
 237 {
 238         unsigned long next;
 239
 240         get_random_bytes(&next, sizeof(next));
 241         next %= HZ;
 242         next += jiffies + MLX5_HEALTH_POLL_INTERVAL;
 243
 244         return next;
 245 }
 246
 247 static void poll_health(unsigned long data)
 248 {
 249         struct mlx5_core_dev *dev = (struct mlx5_core_dev *)data;
 250         struct mlx5_core_health *health = &dev->priv.health;
 251         u32 count;
 252
 253         if (dev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
 254                 mod_timer(&health->timer, get_next_poll_jiffies());
 255                 return;
 256         }
 257
 258         count = ioread32be(health->health_counter);
 259         if (count == health->prev)
 260                 ++health->miss_counter;
 261         else
 262                 health->miss_counter = 0;
 263
 264         health->prev = count;
 265         if (health->miss_counter == MAX_MISSES) {
 266                 dev_err(&dev->pdev->dev, "device's health compromised - reached miss count\n");
 267                 print_health_info(dev);
 268         } else {
 269                 mod_timer(&health->timer, get_next_poll_jiffies());
 270         }
 271
 272         if (in_fatal(dev) && !health->sick) {
 273                 health->sick = true;
 274                 print_health_info(dev);
 275                 queue_work(health->wq, &health->work);
 276         }
 277 }
 278
 279 void mlx5_start_health_poll(struct mlx5_core_dev *dev)
 280 {
 281         struct mlx5_core_health *health = &dev->priv.health;
 282
 283         init_timer(&health->timer);
 284         health->health = &dev->iseg->health;
 285         health->health_counter = &dev->iseg->health_counter;
 286
 287         health->timer.data = (unsigned long)dev;
 288         health->timer.function = poll_health;
 289         health->timer.expires = round_jiffies(jiffies + MLX5_HEALTH_POLL_INTERVAL);
 290         add_timer(&health->timer);
 291 }
 292
 293 void mlx5_stop_health_poll(struct mlx5_core_dev *dev)
 294 {
 295         struct mlx5_core_health *health = &dev->priv.health;
 296
 297         del_timer_sync(&health->timer);
 298 }
 299
 300 void mlx5_health_cleanup(struct mlx5_core_dev *dev)
 301 {
 302         struct mlx5_core_health *health = &dev->priv.health;
 303
 304         destroy_workqueue(health->wq);
 305 }
 306
 307 int mlx5_health_init(struct mlx5_core_dev *dev)
 308 {
 309         struct mlx5_core_health *health;
 310         char *name;
 311
 312         health = &dev->priv.health;
 313         name = kmalloc(64, GFP_KERNEL);
 314         if (!name)
 315                 return -ENOMEM;
 316
 317         strcpy(name, "mlx5_health");
 318         strcat(name, dev_name(&dev->pdev->dev));
 319         health->wq = create_singlethread_workqueue(name);
 320         kfree(name);
 321         if (!health->wq)
 322                 return -ENOMEM;
 323
 324         INIT_WORK(&health->work, health_care);
 325
 326         return 0;
 327 }