| 1 | /* SPDX-License-Identifier: MIT */ |
| 2 | /* |
| 3 | * Copyright © 2019 Intel Corporation |
| 4 | */ |
| 5 | |
| 6 | #ifndef __I915_GEM_CONTEXT_TYPES_H__ |
| 7 | #define __I915_GEM_CONTEXT_TYPES_H__ |
| 8 | |
| 9 | #include <linux/atomic.h> |
| 10 | #include <linux/list.h> |
| 11 | #include <linux/llist.h> |
| 12 | #include <linux/kref.h> |
| 13 | #include <linux/mutex.h> |
| 14 | #include <linux/radix-tree.h> |
| 15 | #include <linux/rbtree.h> |
| 16 | #include <linux/rcupdate.h> |
| 17 | #include <linux/types.h> |
| 18 | |
| 19 | #include "gt/intel_context_types.h" |
| 20 | |
| 21 | #include "i915_scheduler.h" |
| 22 | #include "i915_sw_fence.h" |
| 23 | |
| 24 | struct pid; |
| 25 | |
| 26 | struct drm_i915_private; |
| 27 | struct drm_i915_file_private; |
| 28 | struct i915_address_space; |
| 29 | struct intel_timeline; |
| 30 | struct intel_ring; |
| 31 | |
| 32 | /** |
| 33 | * struct i915_gem_engines - A set of engines |
| 34 | */ |
| 35 | struct i915_gem_engines { |
| 36 | union { |
| 37 | /** @link: Link in i915_gem_context::stale::engines */ |
| 38 | struct list_head link; |
| 39 | |
| 40 | /** @rcu: RCU to use when freeing */ |
| 41 | struct rcu_head rcu; |
| 42 | }; |
| 43 | |
| 44 | /** @fence: Fence used for delayed destruction of engines */ |
| 45 | struct i915_sw_fence fence; |
| 46 | |
| 47 | /** @ctx: i915_gem_context backpointer */ |
| 48 | struct i915_gem_context *ctx; |
| 49 | |
| 50 | /** @num_engines: Number of engines in this set */ |
| 51 | unsigned int num_engines; |
| 52 | |
| 53 | /** @engines: Array of engines */ |
| 54 | struct intel_context *engines[]; |
| 55 | }; |
| 56 | |
| 57 | /** |
| 58 | * struct i915_gem_engines_iter - Iterator for an i915_gem_engines set |
| 59 | */ |
| 60 | struct i915_gem_engines_iter { |
| 61 | /** @idx: Index into i915_gem_engines::engines */ |
| 62 | unsigned int idx; |
| 63 | |
| 64 | /** @engines: Engine set being iterated */ |
| 65 | const struct i915_gem_engines *engines; |
| 66 | }; |
| 67 | |
| 68 | /** |
| 69 | * enum i915_gem_engine_type - Describes the type of an i915_gem_proto_engine |
| 70 | */ |
| 71 | enum i915_gem_engine_type { |
| 72 | /** @I915_GEM_ENGINE_TYPE_INVALID: An invalid engine */ |
| 73 | I915_GEM_ENGINE_TYPE_INVALID = 0, |
| 74 | |
| 75 | /** @I915_GEM_ENGINE_TYPE_PHYSICAL: A single physical engine */ |
| 76 | I915_GEM_ENGINE_TYPE_PHYSICAL, |
| 77 | |
| 78 | /** @I915_GEM_ENGINE_TYPE_BALANCED: A load-balanced engine set */ |
| 79 | I915_GEM_ENGINE_TYPE_BALANCED, |
| 80 | |
| 81 | /** @I915_GEM_ENGINE_TYPE_PARALLEL: A parallel engine set */ |
| 82 | I915_GEM_ENGINE_TYPE_PARALLEL, |
| 83 | }; |
| 84 | |
| 85 | /** |
| 86 | * struct i915_gem_proto_engine - prototype engine |
| 87 | * |
| 88 | * This struct describes an engine that a context may contain. Engines |
| 89 | * have four types: |
| 90 | * |
| 91 | * - I915_GEM_ENGINE_TYPE_INVALID: Invalid engines can be created but they |
| 92 | * show up as a NULL in i915_gem_engines::engines[i] and any attempt to |
| 93 | * use them by the user results in -EINVAL. They are also useful during |
| 94 | * proto-context construction because the client may create invalid |
| 95 | * engines and then set them up later as virtual engines. |
| 96 | * |
| 97 | * - I915_GEM_ENGINE_TYPE_PHYSICAL: A single physical engine, described by |
| 98 | * i915_gem_proto_engine::engine. |
| 99 | * |
| 100 | * - I915_GEM_ENGINE_TYPE_BALANCED: A load-balanced engine set, described |
| 101 | * i915_gem_proto_engine::num_siblings and i915_gem_proto_engine::siblings. |
| 102 | * |
| 103 | * - I915_GEM_ENGINE_TYPE_PARALLEL: A parallel submission engine set, described |
| 104 | * i915_gem_proto_engine::width, i915_gem_proto_engine::num_siblings, and |
| 105 | * i915_gem_proto_engine::siblings. |
| 106 | */ |
| 107 | struct i915_gem_proto_engine { |
| 108 | /** @type: Type of this engine */ |
| 109 | enum i915_gem_engine_type type; |
| 110 | |
| 111 | /** @engine: Engine, for physical */ |
| 112 | struct intel_engine_cs *engine; |
| 113 | |
| 114 | /** @num_siblings: Number of balanced or parallel siblings */ |
| 115 | unsigned int num_siblings; |
| 116 | |
| 117 | /** @width: Width of each sibling */ |
| 118 | unsigned int width; |
| 119 | |
| 120 | /** @siblings: Balanced siblings or num_siblings * width for parallel */ |
| 121 | struct intel_engine_cs **siblings; |
| 122 | |
| 123 | /** @sseu: Client-set SSEU parameters */ |
| 124 | struct intel_sseu sseu; |
| 125 | }; |
| 126 | |
| 127 | /** |
| 128 | * struct i915_gem_proto_context - prototype context |
| 129 | * |
| 130 | * The struct i915_gem_proto_context represents the creation parameters for |
| 131 | * a struct i915_gem_context. This is used to gather parameters provided |
| 132 | * either through creation flags or via SET_CONTEXT_PARAM so that, when we |
| 133 | * create the final i915_gem_context, those parameters can be immutable. |
| 134 | * |
| 135 | * The context uAPI allows for two methods of setting context parameters: |
| 136 | * SET_CONTEXT_PARAM and CONTEXT_CREATE_EXT_SETPARAM. The former is |
| 137 | * allowed to be called at any time while the later happens as part of |
| 138 | * GEM_CONTEXT_CREATE. When these were initially added, Currently, |
| 139 | * everything settable via one is settable via the other. While some |
| 140 | * params are fairly simple and setting them on a live context is harmless |
| 141 | * such the context priority, others are far trickier such as the VM or the |
| 142 | * set of engines. To avoid some truly nasty race conditions, we don't |
| 143 | * allow setting the VM or the set of engines on live contexts. |
| 144 | * |
| 145 | * The way we dealt with this without breaking older userspace that sets |
| 146 | * the VM or engine set via SET_CONTEXT_PARAM is to delay the creation of |
| 147 | * the actual context until after the client is done configuring it with |
| 148 | * SET_CONTEXT_PARAM. From the perspective of the client, it has the same |
| 149 | * u32 context ID the whole time. From the perspective of i915, however, |
| 150 | * it's an i915_gem_proto_context right up until the point where we attempt |
| 151 | * to do something which the proto-context can't handle at which point the |
| 152 | * real context gets created. |
| 153 | * |
| 154 | * This is accomplished via a little xarray dance. When GEM_CONTEXT_CREATE |
| 155 | * is called, we create a proto-context, reserve a slot in context_xa but |
| 156 | * leave it NULL, the proto-context in the corresponding slot in |
| 157 | * proto_context_xa. Then, whenever we go to look up a context, we first |
| 158 | * check context_xa. If it's there, we return the i915_gem_context and |
| 159 | * we're done. If it's not, we look in proto_context_xa and, if we find it |
| 160 | * there, we create the actual context and kill the proto-context. |
| 161 | * |
| 162 | * At the time we made this change (April, 2021), we did a fairly complete |
| 163 | * audit of existing userspace to ensure this wouldn't break anything: |
| 164 | * |
| 165 | * - Mesa/i965 didn't use the engines or VM APIs at all |
| 166 | * |
| 167 | * - Mesa/ANV used the engines API but via CONTEXT_CREATE_EXT_SETPARAM and |
| 168 | * didn't use the VM API. |
| 169 | * |
| 170 | * - Mesa/iris didn't use the engines or VM APIs at all |
| 171 | * |
| 172 | * - The open-source compute-runtime didn't yet use the engines API but |
| 173 | * did use the VM API via SET_CONTEXT_PARAM. However, CONTEXT_SETPARAM |
| 174 | * was always the second ioctl on that context, immediately following |
| 175 | * GEM_CONTEXT_CREATE. |
| 176 | * |
| 177 | * - The media driver sets engines and bonding/balancing via |
| 178 | * SET_CONTEXT_PARAM. However, CONTEXT_SETPARAM to set the VM was |
| 179 | * always the second ioctl on that context, immediately following |
| 180 | * GEM_CONTEXT_CREATE and setting engines immediately followed that. |
| 181 | * |
| 182 | * In order for this dance to work properly, any modification to an |
| 183 | * i915_gem_proto_context that is exposed to the client via |
| 184 | * drm_i915_file_private::proto_context_xa must be guarded by |
| 185 | * drm_i915_file_private::proto_context_lock. The exception is when a |
| 186 | * proto-context has not yet been exposed such as when handling |
| 187 | * CONTEXT_CREATE_SET_PARAM during GEM_CONTEXT_CREATE. |
| 188 | */ |
| 189 | struct i915_gem_proto_context { |
| 190 | /** @fpriv: Client which creates the context */ |
| 191 | struct drm_i915_file_private *fpriv; |
| 192 | |
| 193 | /** @vm: See &i915_gem_context.vm */ |
| 194 | struct i915_address_space *vm; |
| 195 | |
| 196 | /** @user_flags: See &i915_gem_context.user_flags */ |
| 197 | unsigned long user_flags; |
| 198 | |
| 199 | /** @sched: See &i915_gem_context.sched */ |
| 200 | struct i915_sched_attr sched; |
| 201 | |
| 202 | /** @num_user_engines: Number of user-specified engines or -1 */ |
| 203 | int num_user_engines; |
| 204 | |
| 205 | /** @user_engines: User-specified engines */ |
| 206 | struct i915_gem_proto_engine *user_engines; |
| 207 | |
| 208 | /** @legacy_rcs_sseu: Client-set SSEU parameters for the legacy RCS */ |
| 209 | struct intel_sseu legacy_rcs_sseu; |
| 210 | |
| 211 | /** @single_timeline: See See &i915_gem_context.syncobj */ |
| 212 | bool single_timeline; |
| 213 | |
| 214 | /** @uses_protected_content: See &i915_gem_context.uses_protected_content */ |
| 215 | bool uses_protected_content; |
| 216 | |
| 217 | /** @pxp_wakeref: See &i915_gem_context.pxp_wakeref */ |
| 218 | intel_wakeref_t pxp_wakeref; |
| 219 | }; |
| 220 | |
| 221 | /** |
| 222 | * struct i915_gem_context - client state |
| 223 | * |
| 224 | * The struct i915_gem_context represents the combined view of the driver and |
| 225 | * logical hardware state for a particular client. |
| 226 | */ |
| 227 | struct i915_gem_context { |
| 228 | /** @i915: i915 device backpointer */ |
| 229 | struct drm_i915_private *i915; |
| 230 | |
| 231 | /** @file_priv: owning file descriptor */ |
| 232 | struct drm_i915_file_private *file_priv; |
| 233 | |
| 234 | /** |
| 235 | * @engines: User defined engines for this context |
| 236 | * |
| 237 | * Various uAPI offer the ability to lookup up an |
| 238 | * index from this array to select an engine operate on. |
| 239 | * |
| 240 | * Multiple logically distinct instances of the same engine |
| 241 | * may be defined in the array, as well as composite virtual |
| 242 | * engines. |
| 243 | * |
| 244 | * Execbuf uses the I915_EXEC_RING_MASK as an index into this |
| 245 | * array to select which HW context + engine to execute on. For |
| 246 | * the default array, the user_ring_map[] is used to translate |
| 247 | * the legacy uABI onto the appropriate index (e.g. both |
| 248 | * I915_EXEC_DEFAULT and I915_EXEC_RENDER select the same |
| 249 | * context, and I915_EXEC_BSD is weird). For a user defined |
| 250 | * array, execbuf uses I915_EXEC_RING_MASK as a plain index. |
| 251 | * |
| 252 | * User defined by I915_CONTEXT_PARAM_ENGINE (when the |
| 253 | * CONTEXT_USER_ENGINES flag is set). |
| 254 | */ |
| 255 | struct i915_gem_engines __rcu *engines; |
| 256 | |
| 257 | /** @engines_mutex: guards writes to engines */ |
| 258 | struct mutex engines_mutex; |
| 259 | |
| 260 | /** |
| 261 | * @syncobj: Shared timeline syncobj |
| 262 | * |
| 263 | * When the SHARED_TIMELINE flag is set on context creation, we |
| 264 | * emulate a single timeline across all engines using this syncobj. |
| 265 | * For every execbuffer2 call, this syncobj is used as both an in- |
| 266 | * and out-fence. Unlike the real intel_timeline, this doesn't |
| 267 | * provide perfect atomic in-order guarantees if the client races |
| 268 | * with itself by calling execbuffer2 twice concurrently. However, |
| 269 | * if userspace races with itself, that's not likely to yield well- |
| 270 | * defined results anyway so we choose to not care. |
| 271 | */ |
| 272 | struct drm_syncobj *syncobj; |
| 273 | |
| 274 | /** |
| 275 | * @vm: unique address space (GTT) |
| 276 | * |
| 277 | * In full-ppgtt mode, each context has its own address space ensuring |
| 278 | * complete separation of one client from all others. |
| 279 | * |
| 280 | * In other modes, this is a NULL pointer with the expectation that |
| 281 | * the caller uses the shared global GTT. |
| 282 | */ |
| 283 | struct i915_address_space *vm; |
| 284 | |
| 285 | /** |
| 286 | * @pid: process id of creator |
| 287 | * |
| 288 | * Note that who created the context may not be the principle user, |
| 289 | * as the context may be shared across a local socket. However, |
| 290 | * that should only affect the default context, all contexts created |
| 291 | * explicitly by the client are expected to be isolated. |
| 292 | */ |
| 293 | struct pid *pid; |
| 294 | |
| 295 | /** @link: place with &drm_i915_private.context_list */ |
| 296 | struct list_head link; |
| 297 | |
| 298 | /** @client: struct i915_drm_client */ |
| 299 | struct i915_drm_client *client; |
| 300 | |
| 301 | /** @client_link: for linking onto &i915_drm_client.ctx_list */ |
| 302 | struct list_head client_link; |
| 303 | |
| 304 | /** |
| 305 | * @ref: reference count |
| 306 | * |
| 307 | * A reference to a context is held by both the client who created it |
| 308 | * and on each request submitted to the hardware using the request |
| 309 | * (to ensure the hardware has access to the state until it has |
| 310 | * finished all pending writes). See i915_gem_context_get() and |
| 311 | * i915_gem_context_put() for access. |
| 312 | */ |
| 313 | struct kref ref; |
| 314 | |
| 315 | /** |
| 316 | * @release_work: |
| 317 | * |
| 318 | * Work item for deferred cleanup, since i915_gem_context_put() tends to |
| 319 | * be called from hardirq context. |
| 320 | * |
| 321 | * FIXME: The only real reason for this is &i915_gem_engines.fence, all |
| 322 | * other callers are from process context and need at most some mild |
| 323 | * shuffling to pull the i915_gem_context_put() call out of a spinlock. |
| 324 | */ |
| 325 | struct work_struct release_work; |
| 326 | |
| 327 | /** |
| 328 | * @rcu: rcu_head for deferred freeing. |
| 329 | */ |
| 330 | struct rcu_head rcu; |
| 331 | |
| 332 | /** |
| 333 | * @user_flags: small set of booleans controlled by the user |
| 334 | */ |
| 335 | unsigned long user_flags; |
| 336 | #define UCONTEXT_NO_ERROR_CAPTURE 1 |
| 337 | #define UCONTEXT_BANNABLE 2 |
| 338 | #define UCONTEXT_RECOVERABLE 3 |
| 339 | #define UCONTEXT_PERSISTENCE 4 |
| 340 | #define UCONTEXT_LOW_LATENCY 5 |
| 341 | |
| 342 | /** |
| 343 | * @flags: small set of booleans |
| 344 | */ |
| 345 | unsigned long flags; |
| 346 | #define CONTEXT_CLOSED 0 |
| 347 | #define CONTEXT_USER_ENGINES 1 |
| 348 | |
| 349 | /** |
| 350 | * @uses_protected_content: context uses PXP-encrypted objects. |
| 351 | * |
| 352 | * This flag can only be set at ctx creation time and it's immutable for |
| 353 | * the lifetime of the context. See I915_CONTEXT_PARAM_PROTECTED_CONTENT |
| 354 | * in uapi/drm/i915_drm.h for more info on setting restrictions and |
| 355 | * expected behaviour of marked contexts. |
| 356 | */ |
| 357 | bool uses_protected_content; |
| 358 | |
| 359 | /** |
| 360 | * @pxp_wakeref: wakeref to keep the device awake when PXP is in use |
| 361 | * |
| 362 | * PXP sessions are invalidated when the device is suspended, which in |
| 363 | * turns invalidates all contexts and objects using it. To keep the |
| 364 | * flow simple, we keep the device awake when contexts using PXP objects |
| 365 | * are in use. It is expected that the userspace application only uses |
| 366 | * PXP when the display is on, so taking a wakeref here shouldn't worsen |
| 367 | * our power metrics. |
| 368 | */ |
| 369 | intel_wakeref_t pxp_wakeref; |
| 370 | |
| 371 | /** @mutex: guards everything that isn't engines or handles_vma */ |
| 372 | struct mutex mutex; |
| 373 | |
| 374 | /** @sched: scheduler parameters */ |
| 375 | struct i915_sched_attr sched; |
| 376 | |
| 377 | /** @guilty_count: How many times this context has caused a GPU hang. */ |
| 378 | atomic_t guilty_count; |
| 379 | /** |
| 380 | * @active_count: How many times this context was active during a GPU |
| 381 | * hang, but did not cause it. |
| 382 | */ |
| 383 | atomic_t active_count; |
| 384 | |
| 385 | /** |
| 386 | * @hang_timestamp: The last time(s) this context caused a GPU hang |
| 387 | */ |
| 388 | unsigned long hang_timestamp[2]; |
| 389 | #define CONTEXT_FAST_HANG_JIFFIES (120 * HZ) /* 3 hangs within 120s? Banned! */ |
| 390 | |
| 391 | /** @remap_slice: Bitmask of cache lines that need remapping */ |
| 392 | u8 remap_slice; |
| 393 | |
| 394 | /** |
| 395 | * @handles_vma: rbtree to look up our context specific obj/vma for |
| 396 | * the user handle. (user handles are per fd, but the binding is |
| 397 | * per vm, which may be one per context or shared with the global GTT) |
| 398 | */ |
| 399 | struct radix_tree_root handles_vma; |
| 400 | |
| 401 | /** @lut_mutex: Locks handles_vma */ |
| 402 | struct mutex lut_mutex; |
| 403 | |
| 404 | /** |
| 405 | * @name: arbitrary name, used for user debug |
| 406 | * |
| 407 | * A name is constructed for the context from the creator's process |
| 408 | * name, pid and user handle in order to uniquely identify the |
| 409 | * context in messages. |
| 410 | */ |
| 411 | char name[TASK_COMM_LEN + 8]; |
| 412 | |
| 413 | /** @stale: tracks stale engines to be destroyed */ |
| 414 | struct { |
| 415 | /** @stale.lock: guards engines */ |
| 416 | spinlock_t lock; |
| 417 | /** @stale.engines: list of stale engines */ |
| 418 | struct list_head engines; |
| 419 | } stale; |
| 420 | }; |
| 421 | |
| 422 | #endif /* __I915_GEM_CONTEXT_TYPES_H__ */ |
| 423 | |