2 * Copyright (C) 2013 Red Hat
3 * Author: Rob Clark <robdclark@gmail.com>
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published by
7 * the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * You should have received a copy of the GNU General Public License along with
15 * this program. If not, see <http://www.gnu.org/licenses/>.
18 #include <linux/sync_file.h>
23 #include "msm_gpu_trace.h"
26 * Cmdstream submission:
29 /* make sure these don't conflict w/ MSM_SUBMIT_BO_x */
30 #define BO_VALID 0x8000 /* is current addr in cmdstream correct/valid? */
31 #define BO_LOCKED 0x4000
32 #define BO_PINNED 0x2000
34 static struct msm_gem_submit *submit_create(struct drm_device *dev,
35 struct msm_gpu *gpu, struct msm_gpu_submitqueue *queue,
36 uint32_t nr_bos, uint32_t nr_cmds)
38 struct msm_gem_submit *submit;
39 uint64_t sz = sizeof(*submit) + ((u64)nr_bos * sizeof(submit->bos[0])) +
40 ((u64)nr_cmds * sizeof(submit->cmd[0]));
45 submit = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
52 submit->cmd = (void *)&submit->bos[nr_bos];
53 submit->queue = queue;
54 submit->ring = gpu->rb[queue->prio];
56 /* initially, until copy_from_user() and bo lookup succeeds: */
60 INIT_LIST_HEAD(&submit->node);
61 INIT_LIST_HEAD(&submit->bo_list);
62 ww_acquire_init(&submit->ticket, &reservation_ww_class);
67 void msm_gem_submit_free(struct msm_gem_submit *submit)
69 dma_fence_put(submit->fence);
70 list_del(&submit->node);
72 msm_submitqueue_put(submit->queue);
77 static int submit_lookup_objects(struct msm_gem_submit *submit,
78 struct drm_msm_gem_submit *args, struct drm_file *file)
83 for (i = 0; i < args->nr_bos; i++) {
84 struct drm_msm_gem_submit_bo submit_bo;
85 void __user *userptr =
86 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo)));
88 /* make sure we don't have garbage flags, in case we hit
89 * error path before flags is initialized:
91 submit->bos[i].flags = 0;
93 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) {
99 /* at least one of READ and/or WRITE flags should be set: */
100 #define MANDATORY_FLAGS (MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE)
102 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) ||
103 !(submit_bo.flags & MANDATORY_FLAGS)) {
104 DRM_ERROR("invalid flags: %x\n", submit_bo.flags);
110 submit->bos[i].handle = submit_bo.handle;
111 submit->bos[i].flags = submit_bo.flags;
112 /* in validate_objects() we figure out if this is true: */
113 submit->bos[i].iova = submit_bo.presumed;
116 spin_lock(&file->table_lock);
118 for (i = 0; i < args->nr_bos; i++) {
119 struct drm_gem_object *obj;
120 struct msm_gem_object *msm_obj;
122 /* normally use drm_gem_object_lookup(), but for bulk lookup
123 * all under single table_lock just hit object_idr directly:
125 obj = idr_find(&file->object_idr, submit->bos[i].handle);
127 DRM_ERROR("invalid handle %u at index %u\n", submit->bos[i].handle, i);
132 msm_obj = to_msm_bo(obj);
134 if (!list_empty(&msm_obj->submit_entry)) {
135 DRM_ERROR("handle %u at index %u already on submit list\n",
136 submit->bos[i].handle, i);
141 drm_gem_object_get(obj);
143 submit->bos[i].obj = msm_obj;
145 list_add_tail(&msm_obj->submit_entry, &submit->bo_list);
149 spin_unlock(&file->table_lock);
157 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit,
160 struct msm_gem_object *msm_obj = submit->bos[i].obj;
162 if (submit->bos[i].flags & BO_PINNED)
163 msm_gem_unpin_iova(&msm_obj->base, submit->gpu->aspace);
165 if (submit->bos[i].flags & BO_LOCKED)
166 ww_mutex_unlock(&msm_obj->base.resv->lock);
168 if (backoff && !(submit->bos[i].flags & BO_VALID))
169 submit->bos[i].iova = 0;
171 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED);
174 /* This is where we make sure all the bo's are reserved and pin'd: */
175 static int submit_lock_objects(struct msm_gem_submit *submit)
177 int contended, slow_locked = -1, i, ret = 0;
180 for (i = 0; i < submit->nr_bos; i++) {
181 struct msm_gem_object *msm_obj = submit->bos[i].obj;
183 if (slow_locked == i)
188 if (!(submit->bos[i].flags & BO_LOCKED)) {
189 ret = ww_mutex_lock_interruptible(&msm_obj->base.resv->lock,
193 submit->bos[i].flags |= BO_LOCKED;
197 ww_acquire_done(&submit->ticket);
203 submit_unlock_unpin_bo(submit, i, true);
206 submit_unlock_unpin_bo(submit, slow_locked, true);
208 if (ret == -EDEADLK) {
209 struct msm_gem_object *msm_obj = submit->bos[contended].obj;
210 /* we lost out in a seqno race, lock and retry.. */
211 ret = ww_mutex_lock_slow_interruptible(&msm_obj->base.resv->lock,
214 submit->bos[contended].flags |= BO_LOCKED;
215 slow_locked = contended;
223 static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
227 for (i = 0; i < submit->nr_bos; i++) {
228 struct msm_gem_object *msm_obj = submit->bos[i].obj;
229 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE;
232 /* NOTE: _reserve_shared() must happen before
233 * _add_shared_fence(), which makes this a slightly
234 * strange place to call it. OTOH this is a
235 * convenient can-fail point to hook it in.
237 ret = reservation_object_reserve_shared(msm_obj->base.resv,
246 ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,
255 static int submit_pin_objects(struct msm_gem_submit *submit)
259 submit->valid = true;
261 for (i = 0; i < submit->nr_bos; i++) {
262 struct msm_gem_object *msm_obj = submit->bos[i].obj;
265 /* if locking succeeded, pin bo: */
266 ret = msm_gem_get_and_pin_iova(&msm_obj->base,
267 submit->gpu->aspace, &iova);
272 submit->bos[i].flags |= BO_PINNED;
274 if (iova == submit->bos[i].iova) {
275 submit->bos[i].flags |= BO_VALID;
277 submit->bos[i].iova = iova;
278 /* iova changed, so address in cmdstream is not valid: */
279 submit->bos[i].flags &= ~BO_VALID;
280 submit->valid = false;
287 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx,
288 struct msm_gem_object **obj, uint64_t *iova, bool *valid)
290 if (idx >= submit->nr_bos) {
291 DRM_ERROR("invalid buffer index: %u (out of %u)\n",
292 idx, submit->nr_bos);
297 *obj = submit->bos[idx].obj;
299 *iova = submit->bos[idx].iova;
301 *valid = !!(submit->bos[idx].flags & BO_VALID);
306 /* process the reloc's and patch up the cmdstream as needed: */
307 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj,
308 uint32_t offset, uint32_t nr_relocs, uint64_t relocs)
310 uint32_t i, last_offset = 0;
318 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset);
322 /* For now, just map the entire thing. Eventually we probably
323 * to do it page-by-page, w/ kmap() if not vmap()d..
325 ptr = msm_gem_get_vaddr(&obj->base);
329 DBG("failed to map: %d", ret);
333 for (i = 0; i < nr_relocs; i++) {
334 struct drm_msm_gem_submit_reloc submit_reloc;
335 void __user *userptr =
336 u64_to_user_ptr(relocs + (i * sizeof(submit_reloc)));
341 if (copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc))) {
346 if (submit_reloc.submit_offset % 4) {
347 DRM_ERROR("non-aligned reloc offset: %u\n",
348 submit_reloc.submit_offset);
353 /* offset in dwords: */
354 off = submit_reloc.submit_offset / 4;
356 if ((off >= (obj->base.size / 4)) ||
357 (off < last_offset)) {
358 DRM_ERROR("invalid offset %u at reloc %u\n", off, i);
363 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid);
370 iova += submit_reloc.reloc_offset;
372 if (submit_reloc.shift < 0)
373 iova >>= -submit_reloc.shift;
375 iova <<= submit_reloc.shift;
377 ptr[off] = iova | submit_reloc.or;
383 msm_gem_put_vaddr(&obj->base);
388 static void submit_cleanup(struct msm_gem_submit *submit)
392 for (i = 0; i < submit->nr_bos; i++) {
393 struct msm_gem_object *msm_obj = submit->bos[i].obj;
394 submit_unlock_unpin_bo(submit, i, false);
395 list_del_init(&msm_obj->submit_entry);
396 drm_gem_object_put(&msm_obj->base);
399 ww_acquire_fini(&submit->ticket);
402 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
403 struct drm_file *file)
405 static atomic_t ident = ATOMIC_INIT(0);
406 struct msm_drm_private *priv = dev->dev_private;
407 struct drm_msm_gem_submit *args = data;
408 struct msm_file_private *ctx = file->driver_priv;
409 struct msm_gem_submit *submit;
410 struct msm_gpu *gpu = priv->gpu;
411 struct sync_file *sync_file = NULL;
412 struct msm_gpu_submitqueue *queue;
413 struct msm_ringbuffer *ring;
414 int out_fence_fd = -1;
415 struct pid *pid = get_pid(task_pid(current));
421 /* for now, we just have 3d pipe.. eventually this would need to
422 * be more clever to dispatch to appropriate gpu module:
424 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0)
427 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS)
430 if (args->flags & MSM_SUBMIT_SUDO) {
431 if (!IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) ||
432 !capable(CAP_SYS_RAWIO))
436 queue = msm_submitqueue_get(ctx, args->queueid);
440 /* Get a unique identifier for the submission for logging purposes */
441 submitid = atomic_inc_return(&ident) - 1;
443 ring = gpu->rb[queue->prio];
444 trace_msm_gpu_submit(pid_nr(pid), ring->id, submitid,
445 args->nr_bos, args->nr_cmds);
447 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) {
448 struct dma_fence *in_fence;
450 in_fence = sync_file_get_fence(args->fence_fd);
456 * Wait if the fence is from a foreign context, or if the fence
457 * array contains any fence from a foreign context.
460 if (!dma_fence_match_context(in_fence, ring->fctx->context))
461 ret = dma_fence_wait(in_fence, true);
463 dma_fence_put(in_fence);
468 ret = mutex_lock_interruptible(&dev->struct_mutex);
472 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
473 out_fence_fd = get_unused_fd_flags(O_CLOEXEC);
474 if (out_fence_fd < 0) {
480 submit = submit_create(dev, gpu, queue, args->nr_bos, args->nr_cmds);
487 submit->ident = submitid;
489 if (args->flags & MSM_SUBMIT_SUDO)
490 submit->in_rb = true;
492 ret = submit_lookup_objects(submit, args, file);
496 ret = submit_lock_objects(submit);
500 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
504 ret = submit_pin_objects(submit);
508 for (i = 0; i < args->nr_cmds; i++) {
509 struct drm_msm_gem_submit_cmd submit_cmd;
510 void __user *userptr =
511 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
512 struct msm_gem_object *msm_obj;
515 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
521 /* validate input from userspace: */
522 switch (submit_cmd.type) {
523 case MSM_SUBMIT_CMD_BUF:
524 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
525 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
528 DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
533 ret = submit_bo(submit, submit_cmd.submit_idx,
534 &msm_obj, &iova, NULL);
538 if (submit_cmd.size % 4) {
539 DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
545 if (!submit_cmd.size ||
546 ((submit_cmd.size + submit_cmd.submit_offset) >
547 msm_obj->base.size)) {
548 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size);
553 submit->cmd[i].type = submit_cmd.type;
554 submit->cmd[i].size = submit_cmd.size / 4;
555 submit->cmd[i].iova = iova + submit_cmd.submit_offset;
556 submit->cmd[i].idx = submit_cmd.submit_idx;
561 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset,
562 submit_cmd.nr_relocs, submit_cmd.relocs);
569 submit->fence = msm_fence_alloc(ring->fctx);
570 if (IS_ERR(submit->fence)) {
571 ret = PTR_ERR(submit->fence);
572 submit->fence = NULL;
576 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
577 sync_file = sync_file_create(submit->fence);
584 msm_gpu_submit(gpu, submit, ctx);
586 args->fence = submit->fence->seqno;
588 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) {
589 fd_install(out_fence_fd, sync_file->file);
590 args->fence_fd = out_fence_fd;
594 submit_cleanup(submit);
596 msm_gem_submit_free(submit);
598 if (ret && (out_fence_fd >= 0))
599 put_unused_fd(out_fence_fd);
600 mutex_unlock(&dev->struct_mutex);