drivers/gpu/drm/i915/i915_gpu_error.h

   1 /*
   2  * SPDX-License-Identifier: MIT
   3  *
   4  * Copyright � 2008-2018 Intel Corporation
   5  */
   6
   7 #ifndef _I915_GPU_ERROR_H_
   8 #define _I915_GPU_ERROR_H_
   9
  10 #include <linux/kref.h>
  11 #include <linux/ktime.h>
  12 #include <linux/sched.h>
  13
  14 #include <drm/drm_mm.h>
  15
  16 #include "intel_device_info.h"
  17 #include "intel_ringbuffer.h"
  18 #include "intel_uc_fw.h"
  19
  20 #include "i915_gem.h"
  21 #include "i915_gem_gtt.h"
  22 #include "i915_params.h"
  23 #include "i915_scheduler.h"
  24
  25 struct drm_i915_private;
  26 struct intel_overlay_error_state;
  27 struct intel_display_error_state;
  28
  29 struct i915_gpu_state {
  30         struct kref ref;
  31         ktime_t time;
  32         ktime_t boottime;
  33         ktime_t uptime;
  34         unsigned long capture;
  35         unsigned long epoch;
  36
  37         struct drm_i915_private *i915;
  38
  39         char error_msg[128];
  40         bool simulated;
  41         bool awake;
  42         bool wakelock;
  43         bool suspended;
  44         int iommu;
  45         u32 reset_count;
  46         u32 suspend_count;
  47         struct intel_device_info device_info;
  48         struct intel_driver_caps driver_caps;
  49         struct i915_params params;
  50
  51         struct i915_error_uc {
  52                 struct intel_uc_fw guc_fw;
  53                 struct intel_uc_fw huc_fw;
  54                 struct drm_i915_error_object *guc_log;
  55         } uc;
  56
  57         /* Generic register state */
  58         u32 eir;
  59         u32 pgtbl_er;
  60         u32 ier;
  61         u32 gtier[6], ngtier;
  62         u32 ccid;
  63         u32 derrmr;
  64         u32 forcewake;
  65         u32 error; /* gen6+ */
  66         u32 err_int; /* gen7 */
  67         u32 fault_data0; /* gen8, gen9 */
  68         u32 fault_data1; /* gen8, gen9 */
  69         u32 done_reg;
  70         u32 gac_eco;
  71         u32 gam_ecochk;
  72         u32 gab_ctl;
  73         u32 gfx_mode;
  74
  75         u32 nfence;
  76         u64 fence[I915_MAX_NUM_FENCES];
  77         struct intel_overlay_error_state *overlay;
  78         struct intel_display_error_state *display;
  79
  80         struct drm_i915_error_engine {
  81                 int engine_id;
  82                 /* Software tracked state */
  83                 bool idle;
  84                 bool waiting;
  85                 int num_waiters;
  86                 unsigned long hangcheck_timestamp;
  87                 bool hangcheck_stalled;
  88                 enum intel_engine_hangcheck_action hangcheck_action;
  89                 struct i915_address_space *vm;
  90                 int num_requests;
  91                 u32 reset_count;
  92
  93                 /* position of active request inside the ring */
  94                 u32 rq_head, rq_post, rq_tail;
  95
  96                 /* our own tracking of ring head and tail */
  97                 u32 cpu_ring_head;
  98                 u32 cpu_ring_tail;
  99
 100                 u32 last_seqno;
 101
 102                 /* Register state */
 103                 u32 start;
 104                 u32 tail;
 105                 u32 head;
 106                 u32 ctl;
 107                 u32 mode;
 108                 u32 hws;
 109                 u32 ipeir;
 110                 u32 ipehr;
 111                 u32 bbstate;
 112                 u32 instpm;
 113                 u32 instps;
 114                 u32 seqno;
 115                 u64 bbaddr;
 116                 u64 acthd;
 117                 u32 fault_reg;
 118                 u64 faddr;
 119                 u32 rc_psmi; /* sleep state */
 120                 u32 semaphore_mboxes[I915_NUM_ENGINES - 1];
 121                 struct intel_instdone instdone;
 122
 123                 struct drm_i915_error_context {
 124                         char comm[TASK_COMM_LEN];
 125                         pid_t pid;
 126                         u32 handle;
 127                         u32 hw_id;
 128                         int ban_score;
 129                         int active;
 130                         int guilty;
 131                         bool bannable;
 132                         struct i915_sched_attr sched_attr;
 133                 } context;
 134
 135                 struct drm_i915_error_object {
 136                         u64 gtt_offset;
 137                         u64 gtt_size;
 138                         int num_pages;
 139                         int page_count;
 140                         int unused;
 141                         u32 *pages[0];
 142                 } *ringbuffer, *batchbuffer, *wa_batchbuffer, *ctx, *hws_page;
 143
 144                 struct drm_i915_error_object **user_bo;
 145                 long user_bo_count;
 146
 147                 struct drm_i915_error_object *wa_ctx;
 148                 struct drm_i915_error_object *default_state;
 149
 150                 struct drm_i915_error_request {
 151                         long jiffies;
 152                         pid_t pid;
 153                         u32 context;
 154                         int ban_score;
 155                         u32 seqno;
 156                         u32 start;
 157                         u32 head;
 158                         u32 tail;
 159                         struct i915_sched_attr sched_attr;
 160                 } *requests, execlist[EXECLIST_MAX_PORTS];
 161                 unsigned int num_ports;
 162
 163                 struct drm_i915_error_waiter {
 164                         char comm[TASK_COMM_LEN];
 165                         pid_t pid;
 166                         u32 seqno;
 167                 } *waiters;
 168
 169                 struct {
 170                         u32 gfx_mode;
 171                         union {
 172                                 u64 pdp[4];
 173                                 u32 pp_dir_base;
 174                         };
 175                 } vm_info;
 176         } engine[I915_NUM_ENGINES];
 177
 178         struct drm_i915_error_buffer {
 179                 u32 size;
 180                 u32 name;
 181                 u32 wseqno;
 182                 u64 gtt_offset;
 183                 u32 read_domains;
 184                 u32 write_domain;
 185                 s32 fence_reg:I915_MAX_NUM_FENCE_BITS;
 186                 u32 tiling:2;
 187                 u32 dirty:1;
 188                 u32 purgeable:1;
 189                 u32 userptr:1;
 190                 s32 engine:4;
 191                 u32 cache_level:3;
 192         } *active_bo[I915_NUM_ENGINES], *pinned_bo;
 193         u32 active_bo_count[I915_NUM_ENGINES], pinned_bo_count;
 194         struct i915_address_space *active_vm[I915_NUM_ENGINES];
 195 };
 196
 197 struct i915_gpu_error {
 198         /* For hangcheck timer */
 199 #define DRM_I915_HANGCHECK_PERIOD 1500 /* in ms */
 200 #define DRM_I915_HANGCHECK_JIFFIES msecs_to_jiffies(DRM_I915_HANGCHECK_PERIOD)
 201
 202         struct delayed_work hangcheck_work;
 203
 204         /* For reset and error_state handling. */
 205         spinlock_t lock;
 206         /* Protected by the above dev->gpu_error.lock. */
 207         struct i915_gpu_state *first_error;
 208
 209         atomic_t pending_fb_pin;
 210
 211         unsigned long missed_irq_rings;
 212
 213         /**
 214          * State variable controlling the reset flow and count
 215          *
 216          * This is a counter which gets incremented when reset is triggered,
 217          *
 218          * Before the reset commences, the I915_RESET_BACKOFF bit is set
 219          * meaning that any waiters holding onto the struct_mutex should
 220          * relinquish the lock immediately in order for the reset to start.
 221          *
 222          * If reset is not completed successfully, the I915_WEDGE bit is
 223          * set meaning that hardware is terminally sour and there is no
 224          * recovery. All waiters on the reset_queue will be woken when
 225          * that happens.
 226          *
 227          * This counter is used by the wait_seqno code to notice that reset
 228          * event happened and it needs to restart the entire ioctl (since most
 229          * likely the seqno it waited for won't ever signal anytime soon).
 230          *
 231          * This is important for lock-free wait paths, where no contended lock
 232          * naturally enforces the correct ordering between the bail-out of the
 233          * waiter and the gpu reset work code.
 234          */
 235         unsigned long reset_count;
 236
 237         /**
 238          * flags: Control various stages of the GPU reset
 239          *
 240          * #I915_RESET_BACKOFF - When we start a reset, we want to stop any
 241          * other users acquiring the struct_mutex. To do this we set the
 242          * #I915_RESET_BACKOFF bit in the error flags when we detect a reset
 243          * and then check for that bit before acquiring the struct_mutex (in
 244          * i915_mutex_lock_interruptible()?). I915_RESET_BACKOFF serves a
 245          * secondary role in preventing two concurrent global reset attempts.
 246          *
 247          * #I915_RESET_HANDOFF - To perform the actual GPU reset, we need the
 248          * struct_mutex. We try to acquire the struct_mutex in the reset worker,
 249          * but it may be held by some long running waiter (that we cannot
 250          * interrupt without causing trouble). Once we are ready to do the GPU
 251          * reset, we set the I915_RESET_HANDOFF bit and wakeup any waiters. If
 252          * they already hold the struct_mutex and want to participate they can
 253          * inspect the bit and do the reset directly, otherwise the worker
 254          * waits for the struct_mutex.
 255          *
 256          * #I915_RESET_ENGINE[num_engines] - Since the driver doesn't need to
 257          * acquire the struct_mutex to reset an engine, we need an explicit
 258          * flag to prevent two concurrent reset attempts in the same engine.
 259          * As the number of engines continues to grow, allocate the flags from
 260          * the most significant bits.
 261          *
 262          * #I915_WEDGED - If reset fails and we can no longer use the GPU,
 263          * we set the #I915_WEDGED bit. Prior to command submission, e.g.
 264          * i915_request_alloc(), this bit is checked and the sequence
 265          * aborted (with -EIO reported to userspace) if set.
 266          */
 267         unsigned long flags;
 268 #define I915_RESET_BACKOFF      0
 269 #define I915_RESET_HANDOFF      1
 270 #define I915_RESET_MODESET      2
 271 #define I915_WEDGED             (BITS_PER_LONG - 1)
 272 #define I915_RESET_ENGINE       (I915_WEDGED - I915_NUM_ENGINES)
 273
 274         /** Number of times an engine has been reset */
 275         u32 reset_engine_count[I915_NUM_ENGINES];
 276
 277         /** Set of stalled engines with guilty requests, in the current reset */
 278         u32 stalled_mask;
 279
 280         /** Reason for the current *global* reset */
 281         const char *reason;
 282
 283         /**
 284          * Waitqueue to signal when a hang is detected. Used to for waiters
 285          * to release the struct_mutex for the reset to procede.
 286          */
 287         wait_queue_head_t wait_queue;
 288
 289         /**
 290          * Waitqueue to signal when the reset has completed. Used by clients
 291          * that wait for dev_priv->mm.wedged to settle.
 292          */
 293         wait_queue_head_t reset_queue;
 294
 295         /* For missed irq/seqno simulation. */
 296         unsigned long test_irq_rings;
 297 };
 298
 299 struct drm_i915_error_state_buf {
 300         struct drm_i915_private *i915;
 301         unsigned int bytes;
 302         unsigned int size;
 303         int err;
 304         u8 *buf;
 305         loff_t start;
 306         loff_t pos;
 307 };
 308
 309 #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR)
 310
 311 __printf(2, 3)
 312 void i915_error_printf(struct drm_i915_error_state_buf *e, const char *f, ...);
 313 int i915_error_state_to_str(struct drm_i915_error_state_buf *estr,
 314                             const struct i915_gpu_state *gpu);
 315 int i915_error_state_buf_init(struct drm_i915_error_state_buf *eb,
 316                               struct drm_i915_private *i915,
 317                               size_t count, loff_t pos);
 318
 319 static inline void
 320 i915_error_state_buf_release(struct drm_i915_error_state_buf *eb)
 321 {
 322         kfree(eb->buf);
 323 }
 324
 325 struct i915_gpu_state *i915_capture_gpu_state(struct drm_i915_private *i915);
 326 void i915_capture_error_state(struct drm_i915_private *dev_priv,
 327                               u32 engine_mask,
 328                               const char *error_msg);
 329
 330 static inline struct i915_gpu_state *
 331 i915_gpu_state_get(struct i915_gpu_state *gpu)
 332 {
 333         kref_get(&gpu->ref);
 334         return gpu;
 335 }
 336
 337 void __i915_gpu_state_free(struct kref *kref);
 338 static inline void i915_gpu_state_put(struct i915_gpu_state *gpu)
 339 {
 340         if (gpu)
 341                 kref_put(&gpu->ref, __i915_gpu_state_free);
 342 }
 343
 344 struct i915_gpu_state *i915_first_error_state(struct drm_i915_private *i915);
 345 void i915_reset_error_state(struct drm_i915_private *i915);
 346
 347 #else
 348
 349 static inline void i915_capture_error_state(struct drm_i915_private *dev_priv,
 350                                             u32 engine_mask,
 351                                             const char *error_msg)
 352 {
 353 }
 354
 355 static inline struct i915_gpu_state *
 356 i915_first_error_state(struct drm_i915_private *i915)
 357 {
 358         return NULL;
 359 }
 360
 361 static inline void i915_reset_error_state(struct drm_i915_private *i915)
 362 {
 363 }
 364
 365 #endif /* IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR) */
 366
 367 #endif /* _I915_GPU_ERROR_H_ */