2 * SPDX-License-Identifier: MIT
4 * Copyright © 2012-2014 Intel Corporation
6 * Based on amdgpu_mn, which bears the following notice:
8 * Copyright 2014 Advanced Micro Devices, Inc.
11 * Permission is hereby granted, free of charge, to any person obtaining a
12 * copy of this software and associated documentation files (the
13 * "Software"), to deal in the Software without restriction, including
14 * without limitation the rights to use, copy, modify, merge, publish,
15 * distribute, sub license, and/or sell copies of the Software, and to
16 * permit persons to whom the Software is furnished to do so, subject to
17 * the following conditions:
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
27 * The above copyright notice and this permission notice (including the
28 * next paragraph) shall be included in all copies or substantial portions
37 #include <linux/mmu_context.h>
38 #include <linux/mempolicy.h>
39 #include <linux/swap.h>
40 #include <linux/sched/mm.h>
43 #include "i915_gem_ioctls.h"
44 #include "i915_gem_object.h"
45 #include "i915_scatterlist.h"
47 #ifdef CONFIG_MMU_NOTIFIER
50 * i915_gem_userptr_invalidate - callback to notify about mm change
52 * @mni: the range (mm) is about to update
53 * @range: details on the invalidation
54 * @cur_seq: Value to pass to mmu_interval_set_seq()
56 * Block for operations on BOs to finish and mark pages as accessed and
59 static bool i915_gem_userptr_invalidate(struct mmu_interval_notifier *mni,
60 const struct mmu_notifier_range *range,
61 unsigned long cur_seq)
63 struct drm_i915_gem_object *obj = container_of(mni, struct drm_i915_gem_object, userptr.notifier);
64 struct drm_i915_private *i915 = to_i915(obj->base.dev);
67 if (!mmu_notifier_range_blockable(range))
70 spin_lock(&i915->mm.notifier_lock);
72 mmu_interval_set_seq(mni, cur_seq);
74 spin_unlock(&i915->mm.notifier_lock);
77 * We don't wait when the process is exiting. This is valid
78 * because the object will be cleaned up anyway.
80 * This is also temporarily required as a hack, because we
81 * cannot currently force non-consistent batch buffers to preempt
82 * and reschedule by waiting on it, hanging processes on exit.
84 if (current->flags & PF_EXITING)
87 /* we will unbind on next submission, still have userptr pins */
88 r = dma_resv_wait_timeout_rcu(obj->base.resv, true, false,
89 MAX_SCHEDULE_TIMEOUT);
91 drm_err(&i915->drm, "(%ld) failed to wait for idle\n", r);
96 static const struct mmu_interval_notifier_ops i915_gem_userptr_notifier_ops = {
97 .invalidate = i915_gem_userptr_invalidate,
101 i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj)
103 return mmu_interval_notifier_insert(&obj->userptr.notifier, current->mm,
104 obj->userptr.ptr, obj->base.size,
105 &i915_gem_userptr_notifier_ops);
108 static void i915_gem_object_userptr_drop_ref(struct drm_i915_gem_object *obj)
110 struct drm_i915_private *i915 = to_i915(obj->base.dev);
111 struct page **pvec = NULL;
113 spin_lock(&i915->mm.notifier_lock);
114 if (!--obj->userptr.page_ref) {
115 pvec = obj->userptr.pvec;
116 obj->userptr.pvec = NULL;
118 GEM_BUG_ON(obj->userptr.page_ref < 0);
119 spin_unlock(&i915->mm.notifier_lock);
122 const unsigned long num_pages = obj->base.size >> PAGE_SHIFT;
124 unpin_user_pages(pvec, num_pages);
129 static int i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
131 struct drm_i915_private *i915 = to_i915(obj->base.dev);
132 const unsigned long num_pages = obj->base.size >> PAGE_SHIFT;
133 unsigned int max_segment = i915_sg_segment_size();
135 unsigned int sg_page_sizes;
136 struct scatterlist *sg;
140 st = kmalloc(sizeof(*st), GFP_KERNEL);
144 spin_lock(&i915->mm.notifier_lock);
145 if (GEM_WARN_ON(!obj->userptr.page_ref)) {
146 spin_unlock(&i915->mm.notifier_lock);
151 obj->userptr.page_ref++;
152 pvec = obj->userptr.pvec;
153 spin_unlock(&i915->mm.notifier_lock);
156 sg = __sg_alloc_table_from_pages(st, pvec, num_pages, 0,
157 num_pages << PAGE_SHIFT, max_segment,
158 NULL, 0, GFP_KERNEL);
164 ret = i915_gem_gtt_prepare_pages(obj, st);
168 if (max_segment > PAGE_SIZE) {
169 max_segment = PAGE_SIZE;
176 sg_page_sizes = i915_sg_page_sizes(st->sgl);
178 __i915_gem_object_set_pages(obj, st, sg_page_sizes);
183 i915_gem_object_userptr_drop_ref(obj);
190 i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
191 struct sg_table *pages)
193 struct sgt_iter sgt_iter;
199 __i915_gem_object_release_shmem(obj, pages, true);
200 i915_gem_gtt_finish_pages(obj, pages);
203 * We always mark objects as dirty when they are used by the GPU,
204 * just in case. However, if we set the vma as being read-only we know
205 * that the object will never have been written to.
207 if (i915_gem_object_is_readonly(obj))
208 obj->mm.dirty = false;
210 for_each_sgt_page(page, sgt_iter, pages) {
211 if (obj->mm.dirty && trylock_page(page)) {
213 * As this may not be anonymous memory (e.g. shmem)
214 * but exist on a real mapping, we have to lock
215 * the page in order to dirty it -- holding
216 * the page reference is not sufficient to
217 * prevent the inode from being truncated.
218 * Play safe and take the lock.
222 * The mmu-notifier can be invalidated for a
223 * migrate_page, that is alreadying holding the lock
224 * on the page. Such a try_to_unmap() will result
225 * in us calling put_pages() and so recursively try
226 * to lock the page. We avoid that deadlock with
227 * a trylock_page() and in exchange we risk missing
228 * some page dirtying.
230 set_page_dirty(page);
234 mark_page_accessed(page);
236 obj->mm.dirty = false;
238 sg_free_table(pages);
241 i915_gem_object_userptr_drop_ref(obj);
244 static int i915_gem_object_userptr_unbind(struct drm_i915_gem_object *obj, bool get_pages)
246 struct sg_table *pages;
249 err = i915_gem_object_unbind(obj, I915_GEM_OBJECT_UNBIND_ACTIVE);
253 if (GEM_WARN_ON(i915_gem_object_has_pinned_pages(obj)))
256 assert_object_held(obj);
258 pages = __i915_gem_object_unset_pages(obj);
259 if (!IS_ERR_OR_NULL(pages))
260 i915_gem_userptr_put_pages(obj, pages);
263 err = ____i915_gem_object_get_pages(obj);
268 int i915_gem_object_userptr_submit_init(struct drm_i915_gem_object *obj)
270 struct drm_i915_private *i915 = to_i915(obj->base.dev);
271 const unsigned long num_pages = obj->base.size >> PAGE_SHIFT;
273 unsigned int gup_flags = 0;
274 unsigned long notifier_seq;
277 if (obj->userptr.notifier.mm != current->mm)
280 ret = i915_gem_object_lock_interruptible(obj, NULL);
284 /* optimistically try to preserve current pages while unlocked */
285 if (i915_gem_object_has_pages(obj) &&
286 !mmu_interval_check_retry(&obj->userptr.notifier,
287 obj->userptr.notifier_seq)) {
288 spin_lock(&i915->mm.notifier_lock);
289 if (obj->userptr.pvec &&
290 !mmu_interval_read_retry(&obj->userptr.notifier,
291 obj->userptr.notifier_seq)) {
292 obj->userptr.page_ref++;
294 /* We can keep using the current binding, this is the fastpath */
297 spin_unlock(&i915->mm.notifier_lock);
301 /* Make sure userptr is unbound for next attempt, so we don't use stale pages. */
302 ret = i915_gem_object_userptr_unbind(obj, false);
304 i915_gem_object_unlock(obj);
311 notifier_seq = mmu_interval_read_begin(&obj->userptr.notifier);
313 pvec = kvmalloc_array(num_pages, sizeof(struct page *), GFP_KERNEL);
317 if (!i915_gem_object_is_readonly(obj))
318 gup_flags |= FOLL_WRITE;
321 while (pinned < num_pages) {
322 ret = pin_user_pages_fast(obj->userptr.ptr + pinned * PAGE_SIZE,
323 num_pages - pinned, gup_flags,
332 spin_lock(&i915->mm.notifier_lock);
334 if (mmu_interval_read_retry(&obj->userptr.notifier,
335 !obj->userptr.page_ref ? notifier_seq :
336 obj->userptr.notifier_seq)) {
341 if (!obj->userptr.page_ref++) {
342 obj->userptr.pvec = pvec;
343 obj->userptr.notifier_seq = notifier_seq;
349 spin_unlock(&i915->mm.notifier_lock);
353 unpin_user_pages(pvec, pinned);
360 int i915_gem_object_userptr_submit_done(struct drm_i915_gem_object *obj)
362 if (mmu_interval_read_retry(&obj->userptr.notifier,
363 obj->userptr.notifier_seq)) {
364 /* We collided with the mmu notifier, need to retry */
372 void i915_gem_object_userptr_submit_fini(struct drm_i915_gem_object *obj)
374 i915_gem_object_userptr_drop_ref(obj);
377 int i915_gem_object_userptr_validate(struct drm_i915_gem_object *obj)
381 err = i915_gem_object_userptr_submit_init(obj);
385 err = i915_gem_object_lock_interruptible(obj, NULL);
388 * Since we only check validity, not use the pages,
389 * it doesn't matter if we collide with the mmu notifier,
390 * and -EAGAIN handling is not required.
392 err = i915_gem_object_pin_pages(obj);
394 i915_gem_object_unpin_pages(obj);
396 i915_gem_object_unlock(obj);
399 i915_gem_object_userptr_submit_fini(obj);
404 i915_gem_userptr_release(struct drm_i915_gem_object *obj)
406 GEM_WARN_ON(obj->userptr.page_ref);
408 mmu_interval_notifier_remove(&obj->userptr.notifier);
409 obj->userptr.notifier.mm = NULL;
413 i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
415 drm_dbg(obj->base.dev, "Exporting userptr no longer allowed\n");
421 i915_gem_userptr_pwrite(struct drm_i915_gem_object *obj,
422 const struct drm_i915_gem_pwrite *args)
424 drm_dbg(obj->base.dev, "pwrite to userptr no longer allowed\n");
430 i915_gem_userptr_pread(struct drm_i915_gem_object *obj,
431 const struct drm_i915_gem_pread *args)
433 drm_dbg(obj->base.dev, "pread from userptr no longer allowed\n");
438 static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
439 .name = "i915_gem_object_userptr",
440 .flags = I915_GEM_OBJECT_IS_SHRINKABLE |
441 I915_GEM_OBJECT_NO_MMAP |
442 I915_GEM_OBJECT_IS_PROXY,
443 .get_pages = i915_gem_userptr_get_pages,
444 .put_pages = i915_gem_userptr_put_pages,
445 .dmabuf_export = i915_gem_userptr_dmabuf_export,
446 .pwrite = i915_gem_userptr_pwrite,
447 .pread = i915_gem_userptr_pread,
448 .release = i915_gem_userptr_release,
454 * Creates a new mm object that wraps some normal memory from the process
455 * context - user memory.
457 * We impose several restrictions upon the memory being mapped
459 * 1. It must be page aligned (both start/end addresses, i.e ptr and size).
460 * 2. It must be normal system memory, not a pointer into another map of IO
461 * space (e.g. it must not be a GTT mmapping of another object).
462 * 3. We only allow a bo as large as we could in theory map into the GTT,
463 * that is we limit the size to the total size of the GTT.
464 * 4. The bo is marked as being snoopable. The backing pages are left
465 * accessible directly by the CPU, but reads and writes by the GPU may
466 * incur the cost of a snoop (unless you have an LLC architecture).
468 * Synchronisation between multiple users and the GPU is left to userspace
469 * through the normal set-domain-ioctl. The kernel will enforce that the
470 * GPU relinquishes the VMA before it is returned back to the system
471 * i.e. upon free(), munmap() or process termination. However, the userspace
472 * malloc() library may not immediately relinquish the VMA after free() and
473 * instead reuse it whilst the GPU is still reading and writing to the VMA.
476 * Also note, that the object created here is not currently a "first class"
477 * object, in that several ioctls are banned. These are the CPU access
478 * ioctls: mmap(), pwrite and pread. In practice, you are expected to use
479 * direct access via your pointer rather than use those ioctls. Another
480 * restriction is that we do not allow userptr surfaces to be pinned to the
481 * hardware and so we reject any attempt to create a framebuffer out of a
484 * If you think this is a good interface to use to pass GPU memory between
485 * drivers, please use dma-buf instead. In fact, wherever possible use
489 i915_gem_userptr_ioctl(struct drm_device *dev,
491 struct drm_file *file)
493 static struct lock_class_key __maybe_unused lock_class;
494 struct drm_i915_private *dev_priv = to_i915(dev);
495 struct drm_i915_gem_userptr *args = data;
496 struct drm_i915_gem_object __maybe_unused *obj;
497 int __maybe_unused ret;
498 u32 __maybe_unused handle;
500 if (!HAS_LLC(dev_priv) && !HAS_SNOOP(dev_priv)) {
501 /* We cannot support coherent userptr objects on hw without
502 * LLC and broken snooping.
507 if (args->flags & ~(I915_USERPTR_READ_ONLY |
508 I915_USERPTR_UNSYNCHRONIZED))
511 if (i915_gem_object_size_2big(args->user_size))
514 if (!args->user_size)
517 if (offset_in_page(args->user_ptr | args->user_size))
520 if (!access_ok((char __user *)(unsigned long)args->user_ptr, args->user_size))
523 if (args->flags & I915_USERPTR_UNSYNCHRONIZED)
526 if (args->flags & I915_USERPTR_READ_ONLY) {
528 * On almost all of the older hw, we cannot tell the GPU that
529 * a page is readonly.
531 if (!dev_priv->gt.vm->has_read_only)
535 #ifdef CONFIG_MMU_NOTIFIER
536 obj = i915_gem_object_alloc();
540 drm_gem_private_object_init(dev, &obj->base, args->user_size);
541 i915_gem_object_init(obj, &i915_gem_userptr_ops, &lock_class,
542 I915_BO_ALLOC_STRUCT_PAGE);
543 obj->read_domains = I915_GEM_DOMAIN_CPU;
544 obj->write_domain = I915_GEM_DOMAIN_CPU;
545 i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
547 obj->userptr.ptr = args->user_ptr;
548 obj->userptr.notifier_seq = ULONG_MAX;
549 if (args->flags & I915_USERPTR_READ_ONLY)
550 i915_gem_object_set_readonly(obj);
552 /* And keep a pointer to the current->mm for resolving the user pages
553 * at binding. This means that we need to hook into the mmu_notifier
554 * in order to detect if the mmu is destroyed.
556 ret = i915_gem_userptr_init__mmu_notifier(obj);
558 ret = drm_gem_handle_create(file, &obj->base, &handle);
560 /* drop reference from allocate - handle holds it now */
561 i915_gem_object_put(obj);
565 args->handle = handle;
572 int i915_gem_init_userptr(struct drm_i915_private *dev_priv)
574 #ifdef CONFIG_MMU_NOTIFIER
575 spin_lock_init(&dev_priv->mm.notifier_lock);
581 void i915_gem_cleanup_userptr(struct drm_i915_private *dev_priv)