blob: 0efee879df23d9aa06818c31e43decb81b7bb31a [file] [log] [blame]
Chris Wilson05235c52016-07-20 09:21:08 +01001/*
2 * Copyright © 2008-2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 */
24
25#ifndef I915_GEM_REQUEST_H
26#define I915_GEM_REQUEST_H
27
Chris Wilsonf54d1862016-10-25 13:00:45 +010028#include <linux/dma-fence.h>
Chris Wilson04769652016-07-20 09:21:11 +010029
30#include "i915_gem.h"
Chris Wilson5590af32016-09-09 14:11:54 +010031#include "i915_sw_fence.h"
Chris Wilson04769652016-07-20 09:21:11 +010032
Joonas Lahtinenb42fe9c2016-11-11 12:43:54 +020033struct drm_file;
34struct drm_i915_gem_object;
Chris Wilson56299fb2017-02-27 20:58:48 +000035struct drm_i915_gem_request;
Joonas Lahtinenb42fe9c2016-11-11 12:43:54 +020036
Chris Wilsondcff85c2016-08-05 10:14:11 +010037struct intel_wait {
38 struct rb_node node;
39 struct task_struct *tsk;
Chris Wilson56299fb2017-02-27 20:58:48 +000040 struct drm_i915_gem_request *request;
Chris Wilsondcff85c2016-08-05 10:14:11 +010041 u32 seqno;
42};
43
44struct intel_signal_node {
45 struct rb_node node;
46 struct intel_wait wait;
47};
48
Chris Wilson52e54202016-11-14 20:41:02 +000049struct i915_dependency {
50 struct i915_priotree *signaler;
51 struct list_head signal_link;
52 struct list_head wait_link;
Chris Wilson20311bd2016-11-14 20:41:03 +000053 struct list_head dfs_link;
Chris Wilson52e54202016-11-14 20:41:02 +000054 unsigned long flags;
55#define I915_DEPENDENCY_ALLOC BIT(0)
56};
57
58/* Requests exist in a complex web of interdependencies. Each request
59 * has to wait for some other request to complete before it is ready to be run
60 * (e.g. we have to wait until the pixels have been rendering into a texture
61 * before we can copy from it). We track the readiness of a request in terms
62 * of fences, but we also need to keep the dependency tree for the lifetime
63 * of the request (beyond the life of an individual fence). We use the tree
64 * at various points to reorder the requests whilst keeping the requests
65 * in order with respect to their various dependencies.
66 */
67struct i915_priotree {
68 struct list_head signalers_list; /* those before us, we depend upon */
69 struct list_head waiters_list; /* those after us, they depend upon us */
Chris Wilson20311bd2016-11-14 20:41:03 +000070 struct rb_node node;
71 int priority;
72#define I915_PRIORITY_MAX 1024
73#define I915_PRIORITY_MIN (-I915_PRIORITY_MAX)
Chris Wilson52e54202016-11-14 20:41:02 +000074};
75
Chris Wilson05235c52016-07-20 09:21:08 +010076/**
77 * Request queue structure.
78 *
79 * The request queue allows us to note sequence numbers that have been emitted
80 * and may be associated with active buffers to be retired.
81 *
82 * By keeping this list, we can avoid having to do questionable sequence
83 * number comparisons on buffer last_read|write_seqno. It also allows an
84 * emission time to be associated with the request for tracking how far ahead
85 * of the GPU the submission is.
86 *
Chris Wilson5a198b82016-08-09 09:23:34 +010087 * When modifying this structure be very aware that we perform a lockless
88 * RCU lookup of it that may race against reallocation of the struct
89 * from the slab freelist. We intentionally do not zero the structure on
90 * allocation so that the lookup can use the dangling pointers (and is
91 * cogniscent that those pointers may be wrong). Instead, everything that
92 * needs to be initialised must be done so explicitly.
93 *
Chris Wilson04769652016-07-20 09:21:11 +010094 * The requests are reference counted.
Chris Wilson05235c52016-07-20 09:21:08 +010095 */
96struct drm_i915_gem_request {
Chris Wilsonf54d1862016-10-25 13:00:45 +010097 struct dma_fence fence;
Chris Wilson04769652016-07-20 09:21:11 +010098 spinlock_t lock;
Chris Wilson05235c52016-07-20 09:21:08 +010099
100 /** On Which ring this request was generated */
101 struct drm_i915_private *i915;
102
103 /**
104 * Context and ring buffer related to this request
105 * Contexts are refcounted, so when this request is associated with a
106 * context, we must increment the context's refcount, to guarantee that
107 * it persists while any request is linked to it. Requests themselves
108 * are also refcounted, so the request will only be freed when the last
109 * reference to it is dismissed, and the code in
110 * i915_gem_request_free() will then decrement the refcount on the
111 * context.
112 */
113 struct i915_gem_context *ctx;
114 struct intel_engine_cs *engine;
Chris Wilson7e37f882016-08-02 22:50:21 +0100115 struct intel_ring *ring;
Chris Wilson73cb9702016-10-28 13:58:46 +0100116 struct intel_timeline *timeline;
Chris Wilson05235c52016-07-20 09:21:08 +0100117 struct intel_signal_node signaling;
118
Chris Wilson23902e42016-11-14 20:40:58 +0000119 /* Fences for the various phases in the request's lifetime.
120 *
121 * The submit fence is used to await upon all of the request's
122 * dependencies. When it is signaled, the request is ready to run.
123 * It is used by the driver to then queue the request for execution.
Chris Wilson23902e42016-11-14 20:40:58 +0000124 */
Chris Wilson5590af32016-09-09 14:11:54 +0100125 struct i915_sw_fence submit;
Chris Wilson0a046a02016-09-09 14:12:00 +0100126 wait_queue_t submitq;
Chris Wilsonfe497892017-02-23 07:44:13 +0000127 wait_queue_head_t execute;
Chris Wilson5590af32016-09-09 14:11:54 +0100128
Chris Wilson52e54202016-11-14 20:41:02 +0000129 /* A list of everyone we wait upon, and everyone who waits upon us.
130 * Even though we will not be submitted to the hardware before the
131 * submit fence is signaled (it waits for all external events as well
132 * as our own requests), the scheduler still needs to know the
133 * dependency tree for the lifetime of the request (from execbuf
134 * to retirement), i.e. bidirectional dependency information for the
135 * request not tied to individual fences.
136 */
137 struct i915_priotree priotree;
138 struct i915_dependency dep;
139
Chris Wilson754c9fd2017-02-23 07:44:14 +0000140 /** GEM sequence number associated with this request on the
141 * global execution timeline. It is zero when the request is not
142 * on the HW queue (i.e. not on the engine timeline list).
143 * Its value is guarded by the timeline spinlock.
144 */
Chris Wilson65e47602016-10-28 13:58:49 +0100145 u32 global_seqno;
146
Chris Wilsona52abd22016-09-09 14:11:43 +0100147 /** Position in the ring of the start of the request */
Chris Wilson05235c52016-07-20 09:21:08 +0100148 u32 head;
149
150 /**
Chris Wilsona52abd22016-09-09 14:11:43 +0100151 * Position in the ring of the start of the postfix.
152 * This is required to calculate the maximum available ring space
153 * without overwriting the postfix.
Chris Wilson05235c52016-07-20 09:21:08 +0100154 */
155 u32 postfix;
156
Chris Wilsona52abd22016-09-09 14:11:43 +0100157 /** Position in the ring of the end of the whole request */
Chris Wilson05235c52016-07-20 09:21:08 +0100158 u32 tail;
159
Chris Wilsona52abd22016-09-09 14:11:43 +0100160 /** Position in the ring of the end of any workarounds after the tail */
161 u32 wa_tail;
162
163 /** Preallocate space in the ring for the emitting the request */
Chris Wilson05235c52016-07-20 09:21:08 +0100164 u32 reserved_space;
165
Chris Wilson05235c52016-07-20 09:21:08 +0100166 /** Batch buffer related to this request if any (used for
167 * error state dump only).
168 */
Chris Wilson058d88c2016-08-15 10:49:06 +0100169 struct i915_vma *batch;
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100170 struct list_head active_list;
Chris Wilson05235c52016-07-20 09:21:08 +0100171
172 /** Time at which this request was emitted, in jiffies. */
173 unsigned long emitted_jiffies;
174
Chris Wilsonefdf7c02016-08-04 07:52:33 +0100175 /** engine->request_list entry for this request */
176 struct list_head link;
Chris Wilson05235c52016-07-20 09:21:08 +0100177
Chris Wilson675d9ad2016-08-04 07:52:36 +0100178 /** ring->request_list entry for this request */
179 struct list_head ring_link;
180
Chris Wilson05235c52016-07-20 09:21:08 +0100181 struct drm_i915_file_private *file_priv;
182 /** file_priv list entry for this request */
183 struct list_head client_list;
Chris Wilson05235c52016-07-20 09:21:08 +0100184};
185
Chris Wilsonf54d1862016-10-25 13:00:45 +0100186extern const struct dma_fence_ops i915_fence_ops;
Chris Wilson04769652016-07-20 09:21:11 +0100187
Chris Wilsonb52992c2016-10-28 13:58:24 +0100188static inline bool dma_fence_is_i915(const struct dma_fence *fence)
Chris Wilson04769652016-07-20 09:21:11 +0100189{
190 return fence->ops == &i915_fence_ops;
191}
192
Chris Wilson05235c52016-07-20 09:21:08 +0100193struct drm_i915_gem_request * __must_check
194i915_gem_request_alloc(struct intel_engine_cs *engine,
195 struct i915_gem_context *ctx);
Chris Wilson05235c52016-07-20 09:21:08 +0100196int i915_gem_request_add_to_client(struct drm_i915_gem_request *req,
197 struct drm_file *file);
198void i915_gem_request_retire_upto(struct drm_i915_gem_request *req);
199
Chris Wilson05235c52016-07-20 09:21:08 +0100200static inline struct drm_i915_gem_request *
Chris Wilsonf54d1862016-10-25 13:00:45 +0100201to_request(struct dma_fence *fence)
Chris Wilson04769652016-07-20 09:21:11 +0100202{
203 /* We assume that NULL fence/request are interoperable */
204 BUILD_BUG_ON(offsetof(struct drm_i915_gem_request, fence) != 0);
Chris Wilsonb52992c2016-10-28 13:58:24 +0100205 GEM_BUG_ON(fence && !dma_fence_is_i915(fence));
Chris Wilson04769652016-07-20 09:21:11 +0100206 return container_of(fence, struct drm_i915_gem_request, fence);
207}
208
209static inline struct drm_i915_gem_request *
Chris Wilsone8a261e2016-07-20 13:31:49 +0100210i915_gem_request_get(struct drm_i915_gem_request *req)
Chris Wilson05235c52016-07-20 09:21:08 +0100211{
Chris Wilsonf54d1862016-10-25 13:00:45 +0100212 return to_request(dma_fence_get(&req->fence));
Chris Wilson05235c52016-07-20 09:21:08 +0100213}
214
Chris Wilson0eafec62016-08-04 16:32:41 +0100215static inline struct drm_i915_gem_request *
216i915_gem_request_get_rcu(struct drm_i915_gem_request *req)
217{
Chris Wilsonf54d1862016-10-25 13:00:45 +0100218 return to_request(dma_fence_get_rcu(&req->fence));
Chris Wilson0eafec62016-08-04 16:32:41 +0100219}
220
Chris Wilson05235c52016-07-20 09:21:08 +0100221static inline void
Chris Wilsone8a261e2016-07-20 13:31:49 +0100222i915_gem_request_put(struct drm_i915_gem_request *req)
Chris Wilson05235c52016-07-20 09:21:08 +0100223{
Chris Wilsonf54d1862016-10-25 13:00:45 +0100224 dma_fence_put(&req->fence);
Chris Wilson05235c52016-07-20 09:21:08 +0100225}
226
227static inline void i915_gem_request_assign(struct drm_i915_gem_request **pdst,
228 struct drm_i915_gem_request *src)
229{
230 if (src)
Chris Wilsone8a261e2016-07-20 13:31:49 +0100231 i915_gem_request_get(src);
Chris Wilson05235c52016-07-20 09:21:08 +0100232
233 if (*pdst)
Chris Wilsone8a261e2016-07-20 13:31:49 +0100234 i915_gem_request_put(*pdst);
Chris Wilson05235c52016-07-20 09:21:08 +0100235
236 *pdst = src;
237}
238
Chris Wilson754c9fd2017-02-23 07:44:14 +0000239/**
240 * i915_gem_request_global_seqno - report the current global seqno
241 * @request - the request
242 *
243 * A request is assigned a global seqno only when it is on the hardware
244 * execution queue. The global seqno can be used to maintain a list of
245 * requests on the same engine in retirement order, for example for
246 * constructing a priority queue for waiting. Prior to its execution, or
247 * if it is subsequently removed in the event of preemption, its global
248 * seqno is zero. As both insertion and removal from the execution queue
249 * may operate in IRQ context, it is not guarded by the usual struct_mutex
250 * BKL. Instead those relying on the global seqno must be prepared for its
251 * value to change between reads. Only when the request is complete can
252 * the global seqno be stable (due to the memory barriers on submitting
253 * the commands to the hardware to write the breadcrumb, if the HWS shows
254 * that it has passed the global seqno and the global seqno is unchanged
255 * after the read, it is indeed complete).
256 */
257static u32
258i915_gem_request_global_seqno(const struct drm_i915_gem_request *request)
259{
260 return READ_ONCE(request->global_seqno);
261}
262
Chris Wilsona2bc4692016-09-09 14:11:56 +0100263int
264i915_gem_request_await_object(struct drm_i915_gem_request *to,
265 struct drm_i915_gem_object *obj,
266 bool write);
Chris Wilsonb52992c2016-10-28 13:58:24 +0100267int i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req,
268 struct dma_fence *fence);
Chris Wilsona2bc4692016-09-09 14:11:56 +0100269
Chris Wilson17f298cf2016-08-10 13:41:46 +0100270void __i915_add_request(struct drm_i915_gem_request *req, bool flush_caches);
Chris Wilson05235c52016-07-20 09:21:08 +0100271#define i915_add_request(req) \
Chris Wilson17f298cf2016-08-10 13:41:46 +0100272 __i915_add_request(req, true)
Chris Wilson05235c52016-07-20 09:21:08 +0100273#define i915_add_request_no_flush(req) \
Chris Wilson17f298cf2016-08-10 13:41:46 +0100274 __i915_add_request(req, false)
Chris Wilson05235c52016-07-20 09:21:08 +0100275
Chris Wilsond55ac5b2016-11-14 20:40:59 +0000276void __i915_gem_request_submit(struct drm_i915_gem_request *request);
277void i915_gem_request_submit(struct drm_i915_gem_request *request);
278
Chris Wilsond6a22892017-02-23 07:44:17 +0000279void __i915_gem_request_unsubmit(struct drm_i915_gem_request *request);
280void i915_gem_request_unsubmit(struct drm_i915_gem_request *request);
281
Chris Wilson05235c52016-07-20 09:21:08 +0100282struct intel_rps_client;
Chris Wilson42df2712016-07-20 09:21:12 +0100283#define NO_WAITBOOST ERR_PTR(-1)
284#define IS_RPS_CLIENT(p) (!IS_ERR(p))
285#define IS_RPS_USER(p) (!IS_ERR_OR_NULL(p))
Chris Wilson05235c52016-07-20 09:21:08 +0100286
Chris Wilsone95433c2016-10-28 13:58:27 +0100287long i915_wait_request(struct drm_i915_gem_request *req,
288 unsigned int flags,
289 long timeout)
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100290 __attribute__((nonnull(1)));
Chris Wilson22dd3bb2016-09-09 14:11:50 +0100291#define I915_WAIT_INTERRUPTIBLE BIT(0)
292#define I915_WAIT_LOCKED BIT(1) /* struct_mutex held, handle GPU reset */
Chris Wilsone95433c2016-10-28 13:58:27 +0100293#define I915_WAIT_ALL BIT(2) /* used by i915_gem_object_wait() */
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100294
Chris Wilson05235c52016-07-20 09:21:08 +0100295static inline u32 intel_engine_get_seqno(struct intel_engine_cs *engine);
296
297/**
298 * Returns true if seq1 is later than seq2.
299 */
300static inline bool i915_seqno_passed(u32 seq1, u32 seq2)
301{
302 return (s32)(seq1 - seq2) >= 0;
303}
304
305static inline bool
Chris Wilson754c9fd2017-02-23 07:44:14 +0000306__i915_gem_request_started(const struct drm_i915_gem_request *req, u32 seqno)
Chris Wilson05235c52016-07-20 09:21:08 +0100307{
Chris Wilson754c9fd2017-02-23 07:44:14 +0000308 GEM_BUG_ON(!seqno);
Chris Wilson05235c52016-07-20 09:21:08 +0100309 return i915_seqno_passed(intel_engine_get_seqno(req->engine),
Chris Wilson754c9fd2017-02-23 07:44:14 +0000310 seqno - 1);
Chris Wilson05235c52016-07-20 09:21:08 +0100311}
312
313static inline bool
Chris Wilson65e47602016-10-28 13:58:49 +0100314i915_gem_request_started(const struct drm_i915_gem_request *req)
315{
Chris Wilson754c9fd2017-02-23 07:44:14 +0000316 u32 seqno;
317
318 seqno = i915_gem_request_global_seqno(req);
319 if (!seqno)
Chris Wilson65e47602016-10-28 13:58:49 +0100320 return false;
321
Chris Wilson754c9fd2017-02-23 07:44:14 +0000322 return __i915_gem_request_started(req, seqno);
Chris Wilson65e47602016-10-28 13:58:49 +0100323}
324
325static inline bool
Chris Wilson754c9fd2017-02-23 07:44:14 +0000326__i915_gem_request_completed(const struct drm_i915_gem_request *req, u32 seqno)
Chris Wilson65e47602016-10-28 13:58:49 +0100327{
Chris Wilson754c9fd2017-02-23 07:44:14 +0000328 GEM_BUG_ON(!seqno);
329 return i915_seqno_passed(intel_engine_get_seqno(req->engine), seqno) &&
330 seqno == i915_gem_request_global_seqno(req);
Chris Wilson65e47602016-10-28 13:58:49 +0100331}
332
333static inline bool
Chris Wilson05235c52016-07-20 09:21:08 +0100334i915_gem_request_completed(const struct drm_i915_gem_request *req)
335{
Chris Wilson754c9fd2017-02-23 07:44:14 +0000336 u32 seqno;
337
338 seqno = i915_gem_request_global_seqno(req);
339 if (!seqno)
Chris Wilson65e47602016-10-28 13:58:49 +0100340 return false;
341
Chris Wilson754c9fd2017-02-23 07:44:14 +0000342 return __i915_gem_request_completed(req, seqno);
Chris Wilson05235c52016-07-20 09:21:08 +0100343}
344
345bool __i915_spin_request(const struct drm_i915_gem_request *request,
Chris Wilson754c9fd2017-02-23 07:44:14 +0000346 u32 seqno, int state, unsigned long timeout_us);
Chris Wilson05235c52016-07-20 09:21:08 +0100347static inline bool i915_spin_request(const struct drm_i915_gem_request *request,
348 int state, unsigned long timeout_us)
349{
Chris Wilson754c9fd2017-02-23 07:44:14 +0000350 u32 seqno;
351
352 seqno = i915_gem_request_global_seqno(request);
353 return (__i915_gem_request_started(request, seqno) &&
354 __i915_spin_request(request, seqno, state, timeout_us));
Chris Wilson05235c52016-07-20 09:21:08 +0100355}
356
Chris Wilson381f3712016-08-04 07:52:29 +0100357/* We treat requests as fences. This is not be to confused with our
358 * "fence registers" but pipeline synchronisation objects ala GL_ARB_sync.
359 * We use the fences to synchronize access from the CPU with activity on the
360 * GPU, for example, we should not rewrite an object's PTE whilst the GPU
361 * is reading them. We also track fences at a higher level to provide
362 * implicit synchronisation around GEM objects, e.g. set-domain will wait
363 * for outstanding GPU rendering before marking the object ready for CPU
364 * access, or a pageflip will wait until the GPU is complete before showing
365 * the frame on the scanout.
366 *
367 * In order to use a fence, the object must track the fence it needs to
368 * serialise with. For example, GEM objects want to track both read and
369 * write access so that we can perform concurrent read operations between
370 * the CPU and GPU engines, as well as waiting for all rendering to
371 * complete, or waiting for the last GPU user of a "fence register". The
372 * object then embeds a #i915_gem_active to track the most recent (in
373 * retirement order) request relevant for the desired mode of access.
374 * The #i915_gem_active is updated with i915_gem_active_set() to track the
375 * most recent fence request, typically this is done as part of
376 * i915_vma_move_to_active().
377 *
378 * When the #i915_gem_active completes (is retired), it will
379 * signal its completion to the owner through a callback as well as mark
380 * itself as idle (i915_gem_active.request == NULL). The owner
381 * can then perform any action, such as delayed freeing of an active
382 * resource including itself.
383 */
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100384struct i915_gem_active;
385
386typedef void (*i915_gem_retire_fn)(struct i915_gem_active *,
387 struct drm_i915_gem_request *);
388
Chris Wilson381f3712016-08-04 07:52:29 +0100389struct i915_gem_active {
Chris Wilson0eafec62016-08-04 16:32:41 +0100390 struct drm_i915_gem_request __rcu *request;
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100391 struct list_head link;
392 i915_gem_retire_fn retire;
Chris Wilson381f3712016-08-04 07:52:29 +0100393};
394
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100395void i915_gem_retire_noop(struct i915_gem_active *,
396 struct drm_i915_gem_request *request);
397
398/**
399 * init_request_active - prepares the activity tracker for use
400 * @active - the active tracker
401 * @func - a callback when then the tracker is retired (becomes idle),
402 * can be NULL
403 *
404 * init_request_active() prepares the embedded @active struct for use as
405 * an activity tracker, that is for tracking the last known active request
406 * associated with it. When the last request becomes idle, when it is retired
407 * after completion, the optional callback @func is invoked.
408 */
409static inline void
410init_request_active(struct i915_gem_active *active,
411 i915_gem_retire_fn retire)
412{
413 INIT_LIST_HEAD(&active->link);
414 active->retire = retire ?: i915_gem_retire_noop;
415}
416
Chris Wilson27c01aa2016-08-04 07:52:30 +0100417/**
418 * i915_gem_active_set - updates the tracker to watch the current request
419 * @active - the active tracker
420 * @request - the request to watch
421 *
422 * i915_gem_active_set() watches the given @request for completion. Whilst
423 * that @request is busy, the @active reports busy. When that @request is
424 * retired, the @active tracker is updated to report idle.
425 */
Chris Wilson381f3712016-08-04 07:52:29 +0100426static inline void
427i915_gem_active_set(struct i915_gem_active *active,
428 struct drm_i915_gem_request *request)
429{
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100430 list_move(&active->link, &request->active_list);
Chris Wilson0eafec62016-08-04 16:32:41 +0100431 rcu_assign_pointer(active->request, request);
Chris Wilson381f3712016-08-04 07:52:29 +0100432}
433
Ville Syrjäläecd9caa02016-12-07 17:56:47 +0000434/**
435 * i915_gem_active_set_retire_fn - updates the retirement callback
436 * @active - the active tracker
437 * @fn - the routine called when the request is retired
438 * @mutex - struct_mutex used to guard retirements
439 *
440 * i915_gem_active_set_retire_fn() updates the function pointer that
441 * is called when the final request associated with the @active tracker
442 * is retired.
443 */
444static inline void
445i915_gem_active_set_retire_fn(struct i915_gem_active *active,
446 i915_gem_retire_fn fn,
447 struct mutex *mutex)
448{
449 lockdep_assert_held(mutex);
450 active->retire = fn ?: i915_gem_retire_noop;
451}
452
Chris Wilsond72d9082016-08-04 07:52:31 +0100453static inline struct drm_i915_gem_request *
454__i915_gem_active_peek(const struct i915_gem_active *active)
455{
Chris Wilson0eafec62016-08-04 16:32:41 +0100456 /* Inside the error capture (running with the driver in an unknown
457 * state), we want to bend the rules slightly (a lot).
458 *
459 * Work is in progress to make it safer, in the meantime this keeps
460 * the known issue from spamming the logs.
461 */
462 return rcu_dereference_protected(active->request, 1);
Chris Wilsond72d9082016-08-04 07:52:31 +0100463}
464
Chris Wilson27c01aa2016-08-04 07:52:30 +0100465/**
Chris Wilson385384a2016-08-09 08:37:01 +0100466 * i915_gem_active_raw - return the active request
467 * @active - the active tracker
468 *
469 * i915_gem_active_raw() returns the current request being tracked, or NULL.
470 * It does not obtain a reference on the request for the caller, so the caller
471 * must hold struct_mutex.
472 */
473static inline struct drm_i915_gem_request *
474i915_gem_active_raw(const struct i915_gem_active *active, struct mutex *mutex)
475{
476 return rcu_dereference_protected(active->request,
477 lockdep_is_held(mutex));
478}
479
480/**
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100481 * i915_gem_active_peek - report the active request being monitored
Chris Wilson27c01aa2016-08-04 07:52:30 +0100482 * @active - the active tracker
483 *
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100484 * i915_gem_active_peek() returns the current request being tracked if
485 * still active, or NULL. It does not obtain a reference on the request
486 * for the caller, so the caller must hold struct_mutex.
Chris Wilson27c01aa2016-08-04 07:52:30 +0100487 */
488static inline struct drm_i915_gem_request *
Chris Wilsond72d9082016-08-04 07:52:31 +0100489i915_gem_active_peek(const struct i915_gem_active *active, struct mutex *mutex)
Chris Wilson27c01aa2016-08-04 07:52:30 +0100490{
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100491 struct drm_i915_gem_request *request;
492
Chris Wilson385384a2016-08-09 08:37:01 +0100493 request = i915_gem_active_raw(active, mutex);
Chris Wilson0eafec62016-08-04 16:32:41 +0100494 if (!request || i915_gem_request_completed(request))
495 return NULL;
496
497 return request;
498}
499
500/**
Chris Wilson27c01aa2016-08-04 07:52:30 +0100501 * i915_gem_active_get - return a reference to the active request
502 * @active - the active tracker
503 *
504 * i915_gem_active_get() returns a reference to the active request, or NULL
505 * if the active tracker is idle. The caller must hold struct_mutex.
506 */
507static inline struct drm_i915_gem_request *
Chris Wilsond72d9082016-08-04 07:52:31 +0100508i915_gem_active_get(const struct i915_gem_active *active, struct mutex *mutex)
Chris Wilson27c01aa2016-08-04 07:52:30 +0100509{
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100510 return i915_gem_request_get(i915_gem_active_peek(active, mutex));
Chris Wilson27c01aa2016-08-04 07:52:30 +0100511}
512
513/**
Chris Wilson0eafec62016-08-04 16:32:41 +0100514 * __i915_gem_active_get_rcu - return a reference to the active request
515 * @active - the active tracker
516 *
517 * __i915_gem_active_get() returns a reference to the active request, or NULL
518 * if the active tracker is idle. The caller must hold the RCU read lock, but
519 * the returned pointer is safe to use outside of RCU.
520 */
521static inline struct drm_i915_gem_request *
522__i915_gem_active_get_rcu(const struct i915_gem_active *active)
523{
524 /* Performing a lockless retrieval of the active request is super
525 * tricky. SLAB_DESTROY_BY_RCU merely guarantees that the backing
526 * slab of request objects will not be freed whilst we hold the
527 * RCU read lock. It does not guarantee that the request itself
528 * will not be freed and then *reused*. Viz,
529 *
530 * Thread A Thread B
531 *
532 * req = active.request
533 * retire(req) -> free(req);
534 * (req is now first on the slab freelist)
535 * active.request = NULL
536 *
537 * req = new submission on a new object
538 * ref(req)
539 *
540 * To prevent the request from being reused whilst the caller
541 * uses it, we take a reference like normal. Whilst acquiring
542 * the reference we check that it is not in a destroyed state
543 * (refcnt == 0). That prevents the request being reallocated
544 * whilst the caller holds on to it. To check that the request
545 * was not reallocated as we acquired the reference we have to
546 * check that our request remains the active request across
547 * the lookup, in the same manner as a seqlock. The visibility
548 * of the pointer versus the reference counting is controlled
549 * by using RCU barriers (rcu_dereference and rcu_assign_pointer).
550 *
551 * In the middle of all that, we inspect whether the request is
552 * complete. Retiring is lazy so the request may be completed long
553 * before the active tracker is updated. Querying whether the
554 * request is complete is far cheaper (as it involves no locked
555 * instructions setting cachelines to exclusive) than acquiring
556 * the reference, so we do it first. The RCU read lock ensures the
557 * pointer dereference is valid, but does not ensure that the
558 * seqno nor HWS is the right one! However, if the request was
559 * reallocated, that means the active tracker's request was complete.
560 * If the new request is also complete, then both are and we can
561 * just report the active tracker is idle. If the new request is
562 * incomplete, then we acquire a reference on it and check that
563 * it remained the active request.
Chris Wilson5a198b82016-08-09 09:23:34 +0100564 *
565 * It is then imperative that we do not zero the request on
566 * reallocation, so that we can chase the dangling pointers!
567 * See i915_gem_request_alloc().
Chris Wilson0eafec62016-08-04 16:32:41 +0100568 */
569 do {
570 struct drm_i915_gem_request *request;
571
572 request = rcu_dereference(active->request);
573 if (!request || i915_gem_request_completed(request))
574 return NULL;
575
Daniel Vetterc75870d2016-08-22 10:55:22 +0200576 /* An especially silly compiler could decide to recompute the
577 * result of i915_gem_request_completed, more specifically
578 * re-emit the load for request->fence.seqno. A race would catch
579 * a later seqno value, which could flip the result from true to
580 * false. Which means part of the instructions below might not
581 * be executed, while later on instructions are executed. Due to
582 * barriers within the refcounting the inconsistency can't reach
583 * past the call to i915_gem_request_get_rcu, but not executing
584 * that while still executing i915_gem_request_put() creates
585 * havoc enough. Prevent this with a compiler barrier.
586 */
587 barrier();
588
Chris Wilson0eafec62016-08-04 16:32:41 +0100589 request = i915_gem_request_get_rcu(request);
590
591 /* What stops the following rcu_access_pointer() from occurring
592 * before the above i915_gem_request_get_rcu()? If we were
593 * to read the value before pausing to get the reference to
594 * the request, we may not notice a change in the active
595 * tracker.
596 *
597 * The rcu_access_pointer() is a mere compiler barrier, which
598 * means both the CPU and compiler are free to perform the
599 * memory read without constraint. The compiler only has to
600 * ensure that any operations after the rcu_access_pointer()
601 * occur afterwards in program order. This means the read may
602 * be performed earlier by an out-of-order CPU, or adventurous
603 * compiler.
604 *
605 * The atomic operation at the heart of
Chris Wilsonf54d1862016-10-25 13:00:45 +0100606 * i915_gem_request_get_rcu(), see dma_fence_get_rcu(), is
Chris Wilson0eafec62016-08-04 16:32:41 +0100607 * atomic_inc_not_zero() which is only a full memory barrier
608 * when successful. That is, if i915_gem_request_get_rcu()
609 * returns the request (and so with the reference counted
610 * incremented) then the following read for rcu_access_pointer()
611 * must occur after the atomic operation and so confirm
612 * that this request is the one currently being tracked.
Chris Wilsonedf6b762016-08-09 09:23:33 +0100613 *
614 * The corresponding write barrier is part of
615 * rcu_assign_pointer().
Chris Wilson0eafec62016-08-04 16:32:41 +0100616 */
617 if (!request || request == rcu_access_pointer(active->request))
618 return rcu_pointer_handoff(request);
619
620 i915_gem_request_put(request);
621 } while (1);
622}
623
624/**
625 * i915_gem_active_get_unlocked - return a reference to the active request
626 * @active - the active tracker
627 *
628 * i915_gem_active_get_unlocked() returns a reference to the active request,
629 * or NULL if the active tracker is idle. The reference is obtained under RCU,
630 * so no locking is required by the caller.
631 *
632 * The reference should be freed with i915_gem_request_put().
633 */
634static inline struct drm_i915_gem_request *
635i915_gem_active_get_unlocked(const struct i915_gem_active *active)
636{
637 struct drm_i915_gem_request *request;
638
639 rcu_read_lock();
640 request = __i915_gem_active_get_rcu(active);
641 rcu_read_unlock();
642
643 return request;
644}
645
646/**
Chris Wilson27c01aa2016-08-04 07:52:30 +0100647 * i915_gem_active_isset - report whether the active tracker is assigned
648 * @active - the active tracker
649 *
650 * i915_gem_active_isset() returns true if the active tracker is currently
651 * assigned to a request. Due to the lazy retiring, that request may be idle
652 * and this may report stale information.
653 */
654static inline bool
655i915_gem_active_isset(const struct i915_gem_active *active)
656{
Chris Wilson0eafec62016-08-04 16:32:41 +0100657 return rcu_access_pointer(active->request);
Chris Wilson27c01aa2016-08-04 07:52:30 +0100658}
659
660/**
Chris Wilsond07f0e52016-10-28 13:58:44 +0100661 * i915_gem_active_wait - waits until the request is completed
Chris Wilson24676582016-08-05 10:14:06 +0100662 * @active - the active request on which to wait
Chris Wilsonea746f32016-09-09 14:11:49 +0100663 * @flags - how to wait
Chris Wilson24676582016-08-05 10:14:06 +0100664 * @timeout - how long to wait at most
665 * @rps - userspace client to charge for a waitboost
666 *
Chris Wilson2e369912016-10-28 13:58:28 +0100667 * i915_gem_active_wait() waits until the request is completed before
Chris Wilson24676582016-08-05 10:14:06 +0100668 * returning, without requiring any locks to be held. Note that it does not
669 * retire any requests before returning.
670 *
671 * This function relies on RCU in order to acquire the reference to the active
672 * request without holding any locks. See __i915_gem_active_get_rcu() for the
673 * glory details on how that is managed. Once the reference is acquired, we
674 * can then wait upon the request, and afterwards release our reference,
675 * free of any locking.
676 *
677 * This function wraps i915_wait_request(), see it for the full details on
678 * the arguments.
679 *
680 * Returns 0 if successful, or a negative error code.
681 */
682static inline int
Chris Wilson2e369912016-10-28 13:58:28 +0100683i915_gem_active_wait(const struct i915_gem_active *active, unsigned int flags)
Chris Wilson24676582016-08-05 10:14:06 +0100684{
685 struct drm_i915_gem_request *request;
Chris Wilsone95433c2016-10-28 13:58:27 +0100686 long ret = 0;
Chris Wilson24676582016-08-05 10:14:06 +0100687
688 request = i915_gem_active_get_unlocked(active);
689 if (request) {
Chris Wilsone95433c2016-10-28 13:58:27 +0100690 ret = i915_wait_request(request, flags, MAX_SCHEDULE_TIMEOUT);
Chris Wilson24676582016-08-05 10:14:06 +0100691 i915_gem_request_put(request);
692 }
693
Chris Wilsone95433c2016-10-28 13:58:27 +0100694 return ret < 0 ? ret : 0;
Chris Wilson24676582016-08-05 10:14:06 +0100695}
696
697/**
Chris Wilson27c01aa2016-08-04 07:52:30 +0100698 * i915_gem_active_retire - waits until the request is retired
699 * @active - the active request on which to wait
700 *
701 * i915_gem_active_retire() waits until the request is completed,
702 * and then ensures that at least the retirement handler for this
703 * @active tracker is called before returning. If the @active
704 * tracker is idle, the function returns immediately.
705 */
706static inline int __must_check
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100707i915_gem_active_retire(struct i915_gem_active *active,
Chris Wilsond72d9082016-08-04 07:52:31 +0100708 struct mutex *mutex)
Chris Wilson27c01aa2016-08-04 07:52:30 +0100709{
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100710 struct drm_i915_gem_request *request;
Chris Wilsone95433c2016-10-28 13:58:27 +0100711 long ret;
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100712
Chris Wilson385384a2016-08-09 08:37:01 +0100713 request = i915_gem_active_raw(active, mutex);
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100714 if (!request)
715 return 0;
716
Chris Wilson22dd3bb2016-09-09 14:11:50 +0100717 ret = i915_wait_request(request,
718 I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
Chris Wilsone95433c2016-10-28 13:58:27 +0100719 MAX_SCHEDULE_TIMEOUT);
720 if (ret < 0)
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100721 return ret;
722
723 list_del_init(&active->link);
Chris Wilson0eafec62016-08-04 16:32:41 +0100724 RCU_INIT_POINTER(active->request, NULL);
725
Chris Wilsonfa545cb2016-08-04 07:52:35 +0100726 active->retire(active, request);
727
728 return 0;
Chris Wilson27c01aa2016-08-04 07:52:30 +0100729}
730
Chris Wilson381f3712016-08-04 07:52:29 +0100731#define for_each_active(mask, idx) \
732 for (; mask ? idx = ffs(mask) - 1, 1 : 0; mask &= ~BIT(idx))
733
Chris Wilson05235c52016-07-20 09:21:08 +0100734#endif /* I915_GEM_REQUEST_H */