From d9483ecd327b7537c6a51cab515b5faad21b8200 Mon Sep 17 00:00:00 2001 From: Christian König Date: Wed, 9 Nov 2022 12:28:46 +0100 Subject: drm/amdgpu: rename the files for HMM handling MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Clean that up a bit, no functional change. Signed-off-by: Christian König Reviewed-by: Alex Deucher Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c | 245 ++++++++++++++++++++++++++++++++ 1 file changed, 245 insertions(+) create mode 100644 drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c new file mode 100644 index 000000000000..a68072f766c7 --- /dev/null +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c @@ -0,0 +1,245 @@ +/* + * Copyright 2014 Advanced Micro Devices, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL + * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, + * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR + * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE + * USE OR OTHER DEALINGS IN THE SOFTWARE. + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + */ +/* + * Authors: + * Christian König + */ + +/** + * DOC: MMU Notifier + * + * For coherent userptr handling registers an MMU notifier to inform the driver + * about updates on the page tables of a process. + * + * When somebody tries to invalidate the page tables we block the update until + * all operations on the pages in question are completed, then those pages are + * marked as accessed and also dirty if it wasn't a read only access. + * + * New command submissions using the userptrs in question are delayed until all + * page table invalidation are completed and we once more see a coherent process + * address space. + */ + +#include +#include +#include + +#include "amdgpu.h" +#include "amdgpu_amdkfd.h" +#include "amdgpu_hmm.h" + +/** + * amdgpu_hmm_invalidate_gfx - callback to notify about mm change + * + * @mni: the range (mm) is about to update + * @range: details on the invalidation + * @cur_seq: Value to pass to mmu_interval_set_seq() + * + * Block for operations on BOs to finish and mark pages as accessed and + * potentially dirty. + */ +static bool amdgpu_hmm_invalidate_gfx(struct mmu_interval_notifier *mni, + const struct mmu_notifier_range *range, + unsigned long cur_seq) +{ + struct amdgpu_bo *bo = container_of(mni, struct amdgpu_bo, notifier); + struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); + long r; + + if (!mmu_notifier_range_blockable(range)) + return false; + + mutex_lock(&adev->notifier_lock); + + mmu_interval_set_seq(mni, cur_seq); + + r = dma_resv_wait_timeout(bo->tbo.base.resv, DMA_RESV_USAGE_BOOKKEEP, + false, MAX_SCHEDULE_TIMEOUT); + mutex_unlock(&adev->notifier_lock); + if (r <= 0) + DRM_ERROR("(%ld) failed to wait for user bo\n", r); + return true; +} + +static const struct mmu_interval_notifier_ops amdgpu_hmm_gfx_ops = { + .invalidate = amdgpu_hmm_invalidate_gfx, +}; + +/** + * amdgpu_hmm_invalidate_hsa - callback to notify about mm change + * + * @mni: the range (mm) is about to update + * @range: details on the invalidation + * @cur_seq: Value to pass to mmu_interval_set_seq() + * + * We temporarily evict the BO attached to this range. This necessitates + * evicting all user-mode queues of the process. + */ +static bool amdgpu_hmm_invalidate_hsa(struct mmu_interval_notifier *mni, + const struct mmu_notifier_range *range, + unsigned long cur_seq) +{ + struct amdgpu_bo *bo = container_of(mni, struct amdgpu_bo, notifier); + struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); + + if (!mmu_notifier_range_blockable(range)) + return false; + + mutex_lock(&adev->notifier_lock); + + mmu_interval_set_seq(mni, cur_seq); + + amdgpu_amdkfd_evict_userptr(bo->kfd_bo, bo->notifier.mm); + mutex_unlock(&adev->notifier_lock); + + return true; +} + +static const struct mmu_interval_notifier_ops amdgpu_hmm_hsa_ops = { + .invalidate = amdgpu_hmm_invalidate_hsa, +}; + +/** + * amdgpu_hmm_register - register a BO for notifier updates + * + * @bo: amdgpu buffer object + * @addr: userptr addr we should monitor + * + * Registers a mmu_notifier for the given BO at the specified address. + * Returns 0 on success, -ERRNO if anything goes wrong. + */ +int amdgpu_hmm_register(struct amdgpu_bo *bo, unsigned long addr) +{ + if (bo->kfd_bo) + return mmu_interval_notifier_insert(&bo->notifier, current->mm, + addr, amdgpu_bo_size(bo), + &amdgpu_hmm_hsa_ops); + return mmu_interval_notifier_insert(&bo->notifier, current->mm, addr, + amdgpu_bo_size(bo), + &amdgpu_hmm_gfx_ops); +} + +/** + * amdgpu_hmm_unregister - unregister a BO for notifier updates + * + * @bo: amdgpu buffer object + * + * Remove any registration of mmu notifier updates from the buffer object. + */ +void amdgpu_hmm_unregister(struct amdgpu_bo *bo) +{ + if (!bo->notifier.mm) + return; + mmu_interval_notifier_remove(&bo->notifier); + bo->notifier.mm = NULL; +} + +int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier, + struct mm_struct *mm, struct page **pages, + uint64_t start, uint64_t npages, + struct hmm_range **phmm_range, bool readonly, + bool mmap_locked, void *owner) +{ + struct hmm_range *hmm_range; + unsigned long timeout; + unsigned long i; + unsigned long *pfns; + int r = 0; + + hmm_range = kzalloc(sizeof(*hmm_range), GFP_KERNEL); + if (unlikely(!hmm_range)) + return -ENOMEM; + + pfns = kvmalloc_array(npages, sizeof(*pfns), GFP_KERNEL); + if (unlikely(!pfns)) { + r = -ENOMEM; + goto out_free_range; + } + + hmm_range->notifier = notifier; + hmm_range->default_flags = HMM_PFN_REQ_FAULT; + if (!readonly) + hmm_range->default_flags |= HMM_PFN_REQ_WRITE; + hmm_range->hmm_pfns = pfns; + hmm_range->start = start; + hmm_range->end = start + npages * PAGE_SIZE; + hmm_range->dev_private_owner = owner; + + /* Assuming 512MB takes maxmium 1 second to fault page address */ + timeout = max(npages >> 17, 1ULL) * HMM_RANGE_DEFAULT_TIMEOUT; + timeout = jiffies + msecs_to_jiffies(timeout); + +retry: + hmm_range->notifier_seq = mmu_interval_read_begin(notifier); + + if (likely(!mmap_locked)) + mmap_read_lock(mm); + + r = hmm_range_fault(hmm_range); + + if (likely(!mmap_locked)) + mmap_read_unlock(mm); + if (unlikely(r)) { + /* + * FIXME: This timeout should encompass the retry from + * mmu_interval_read_retry() as well. + */ + if (r == -EBUSY && !time_after(jiffies, timeout)) + goto retry; + goto out_free_pfns; + } + + /* + * Due to default_flags, all pages are HMM_PFN_VALID or + * hmm_range_fault() fails. FIXME: The pages cannot be touched outside + * the notifier_lock, and mmu_interval_read_retry() must be done first. + */ + for (i = 0; pages && i < npages; i++) + pages[i] = hmm_pfn_to_page(pfns[i]); + + *phmm_range = hmm_range; + + return 0; + +out_free_pfns: + kvfree(pfns); +out_free_range: + kfree(hmm_range); + + return r; +} + +int amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range) +{ + int r; + + r = mmu_interval_read_retry(hmm_range->notifier, + hmm_range->notifier_seq); + kvfree(hmm_range->hmm_pfns); + kfree(hmm_range); + + return r; +} -- cgit From d4cbff464d2932a71d7d3b7d17ffef7700b58edd Mon Sep 17 00:00:00 2001 From: Christian König Date: Wed, 9 Nov 2022 12:41:08 +0100 Subject: drm/amdgpu: cleanup amdgpu_hmm_range_get_pages MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Remove unused parameters and cleanup dead code. Signed-off-by: Christian König Reviewed-by: Alex Deucher Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c | 14 +++----------- drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h | 7 +++---- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 5 ++--- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 6 +++--- 4 files changed, 11 insertions(+), 21 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c index a68072f766c7..a48ea62b12b0 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c @@ -158,10 +158,9 @@ void amdgpu_hmm_unregister(struct amdgpu_bo *bo) } int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier, - struct mm_struct *mm, struct page **pages, - uint64_t start, uint64_t npages, - struct hmm_range **phmm_range, bool readonly, - bool mmap_locked, void *owner) + uint64_t start, uint64_t npages, bool readonly, + void *owner, struct page **pages, + struct hmm_range **phmm_range) { struct hmm_range *hmm_range; unsigned long timeout; @@ -194,14 +193,7 @@ int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier, retry: hmm_range->notifier_seq = mmu_interval_read_begin(notifier); - - if (likely(!mmap_locked)) - mmap_read_lock(mm); - r = hmm_range_fault(hmm_range); - - if (likely(!mmap_locked)) - mmap_read_unlock(mm); if (unlikely(r)) { /* * FIXME: This timeout should encompass the retry from diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h index 4e596a16d288..13ed94d3b01b 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h @@ -31,10 +31,9 @@ #include int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier, - struct mm_struct *mm, struct page **pages, - uint64_t start, uint64_t npages, - struct hmm_range **phmm_range, bool readonly, - bool mmap_locked, void *owner); + uint64_t start, uint64_t npages, bool readonly, + void *owner, struct page **pages, + struct hmm_range **phmm_range); int amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range); #if defined(CONFIG_HMM_MIRROR) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index ddb13b18ed7b..c8169cea3032 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -692,9 +692,8 @@ int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages, } readonly = amdgpu_ttm_tt_is_readonly(ttm); - r = amdgpu_hmm_range_get_pages(&bo->notifier, mm, pages, start, - ttm->num_pages, range, readonly, - true, NULL); + r = amdgpu_hmm_range_get_pages(&bo->notifier, start, ttm->num_pages, + readonly, NULL, pages, range); out_unlock: mmap_read_unlock(mm); if (r) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 1cf7dcb22e3f..814f99888ab1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -1596,9 +1596,9 @@ static int svm_range_validate_and_map(struct mm_struct *mm, next = min(vma->vm_end, end); npages = (next - addr) >> PAGE_SHIFT; WRITE_ONCE(p->svms.faulting_task, current); - r = amdgpu_hmm_range_get_pages(&prange->notifier, mm, NULL, - addr, npages, &hmm_range, - readonly, true, owner); + r = amdgpu_hmm_range_get_pages(&prange->notifier, addr, npages, + readonly, owner, NULL, + &hmm_range); WRITE_ONCE(p->svms.faulting_task, NULL); if (r) { pr_debug("failed %d to get svm range pages\n", r); -- cgit From c1420a5dd4dd2a1ffc86965ad1c6f439afb130bc Mon Sep 17 00:00:00 2001 From: James Zhu Date: Thu, 17 Nov 2022 15:57:52 -0500 Subject: drm/amdgpu: fix stall on CPU when allocate large system memory -v2: 1. rename variable to redue confuse 2. optimize the code -v3: move new define out of the middle of the code -v4: squash in minmax error fix (Luben) When applications try to allocate large system (more than > 128GB), "stall cpu" is reported. for such large system memory, walk_page_range takes more than 20s usually. The warning message can be removed when splitting hmm range into smaller ones which is not more 64GB for each walk_page_range. [ 164.437617] amdgpu:amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu:1753: amdgpu: create BO VA 0x7f63c7a00000 size 0x2f16000000 domain CPU [ 164.488847] amdgpu:amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu:1785: amdgpu: creating userptr BO for user_addr = 7f63c7a00000 [ 185.439116] rcu: INFO: rcu_sched self-detected stall on CPU [ 185.439125] rcu: 8-....: (20999 ticks this GP) idle=e22/1/0x4000000000000000 softirq=2242/2242 fqs=5249 [ 185.439137] (t=21000 jiffies g=6325 q=1215) [ 185.439141] NMI backtrace for cpu 8 [ 185.439143] CPU: 8 PID: 3470 Comm: kfdtest Kdump: loaded Tainted: G O 5.12.0-0_fbk5_zion_rc1_5697_g2c723fb88626 #1 [ 185.439147] Hardware name: HPE ProLiant XL675d Gen10 Plus/ProLiant XL675d Gen10 Plus, BIOS A47 11/06/2020 [ 185.439150] Call Trace: [ 185.439153] [ 185.439157] dump_stack+0x64/0x7c [ 185.439163] nmi_cpu_backtrace.cold.7+0x30/0x65 [ 185.439165] ? lapic_can_unplug_cpu+0x70/0x70 [ 185.439170] nmi_trigger_cpumask_backtrace+0xf9/0x100 [ 185.439174] rcu_dump_cpu_stacks+0xc5/0xf5 [ 185.439178] rcu_sched_clock_irq.cold.97+0x112/0x38c [ 185.439182] ? tick_sched_handle.isra.21+0x50/0x50 [ 185.439185] update_process_times+0x8c/0xc0 [ 185.439189] tick_sched_timer+0x63/0x70 [ 185.439192] __hrtimer_run_queues+0xff/0x250 [ 185.439195] hrtimer_interrupt+0xf4/0x200 [ 185.439199] __sysvec_apic_timer_interrupt+0x51/0xd0 [ 185.439201] sysvec_apic_timer_interrupt+0x69/0x90 [ 185.439206] [ 185.439207] asm_sysvec_apic_timer_interrupt+0x12/0x20 [ 185.439211] RIP: 0010:clear_page_rep+0x7/0x10 [ 185.439214] Code: e8 fe 7c 51 00 44 89 e2 48 89 ee 48 89 df e8 60 ff ff ff c6 03 00 5b 5d 41 5c c3 cc cc cc cc cc cc cc cc b9 00 02 00 00 31 c0 48 ab c3 0f 1f 44 00 00 31 c0 b9 40 00 00 00 66 0f 1f 84 00 00 [ 185.439218] RSP: 0018:ffffc9000f58f818 EFLAGS: 00000246 [ 185.439220] RAX: 0000000000000000 RBX: 0000000000000881 RCX: 000000000000005c [ 185.439223] RDX: 0000000000100dca RSI: 0000000000000000 RDI: ffff88a59e0e5d20 [ 185.439225] RBP: ffffea0096783940 R08: ffff888118c35280 R09: ffffea0096783940 [ 185.439227] R10: ffff888000000000 R11: 0000160000000000 R12: ffffea0096783980 [ 185.439228] R13: ffffea0096783940 R14: ffff88b07fdfdd00 R15: 0000000000000000 [ 185.439232] prep_new_page+0x81/0xc0 [ 185.439236] get_page_from_freelist+0x13be/0x16f0 [ 185.439240] ? release_pages+0x16a/0x4a0 [ 185.439244] __alloc_pages_nodemask+0x1ae/0x340 [ 185.439247] alloc_pages_vma+0x74/0x1e0 [ 185.439251] __handle_mm_fault+0xafe/0x1360 [ 185.439255] handle_mm_fault+0xc3/0x280 [ 185.439257] hmm_vma_fault.isra.22+0x49/0x90 [ 185.439261] __walk_page_range+0x692/0x9b0 [ 185.439265] walk_page_range+0x9b/0x120 [ 185.439269] hmm_range_fault+0x4f/0x90 [ 185.439274] amdgpu_hmm_range_get_pages+0x24f/0x260 [amdgpu] [ 185.439463] amdgpu_ttm_tt_get_user_pages+0xc2/0x190 [amdgpu] [ 185.439603] amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu+0x49f/0x7a0 [amdgpu] [ 185.439774] kfd_ioctl_alloc_memory_of_gpu+0xfb/0x410 [amdgpu] Signed-off-by: James Zhu Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c | 50 +++++++++++++++++++++++---------- 1 file changed, 35 insertions(+), 15 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c index a48ea62b12b0..65715cb395d8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c @@ -51,6 +51,8 @@ #include "amdgpu_amdkfd.h" #include "amdgpu_hmm.h" +#define MAX_WALK_BYTE (2UL << 30) + /** * amdgpu_hmm_invalidate_gfx - callback to notify about mm change * @@ -163,6 +165,7 @@ int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier, struct hmm_range **phmm_range) { struct hmm_range *hmm_range; + unsigned long end; unsigned long timeout; unsigned long i; unsigned long *pfns; @@ -184,25 +187,42 @@ int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier, hmm_range->default_flags |= HMM_PFN_REQ_WRITE; hmm_range->hmm_pfns = pfns; hmm_range->start = start; - hmm_range->end = start + npages * PAGE_SIZE; + end = start + npages * PAGE_SIZE; hmm_range->dev_private_owner = owner; - /* Assuming 512MB takes maxmium 1 second to fault page address */ - timeout = max(npages >> 17, 1ULL) * HMM_RANGE_DEFAULT_TIMEOUT; - timeout = jiffies + msecs_to_jiffies(timeout); + do { + hmm_range->end = min(hmm_range->start + MAX_WALK_BYTE, end); + + pr_debug("hmm range: start = 0x%lx, end = 0x%lx", + hmm_range->start, hmm_range->end); + + /* Assuming 512MB takes maxmium 1 second to fault page address */ + timeout = max((hmm_range->end - hmm_range->start) >> 29, 1UL); + timeout *= HMM_RANGE_DEFAULT_TIMEOUT; + timeout = jiffies + msecs_to_jiffies(timeout); retry: - hmm_range->notifier_seq = mmu_interval_read_begin(notifier); - r = hmm_range_fault(hmm_range); - if (unlikely(r)) { - /* - * FIXME: This timeout should encompass the retry from - * mmu_interval_read_retry() as well. - */ - if (r == -EBUSY && !time_after(jiffies, timeout)) - goto retry; - goto out_free_pfns; - } + hmm_range->notifier_seq = mmu_interval_read_begin(notifier); + r = hmm_range_fault(hmm_range); + if (unlikely(r)) { + /* + * FIXME: This timeout should encompass the retry from + * mmu_interval_read_retry() as well. + */ + if (r == -EBUSY && !time_after(jiffies, timeout)) + goto retry; + goto out_free_pfns; + } + + if (hmm_range->end == end) + break; + hmm_range->hmm_pfns += MAX_WALK_BYTE >> PAGE_SHIFT; + hmm_range->start = hmm_range->end; + schedule(); + } while (hmm_range->end < end); + + hmm_range->start = start; + hmm_range->hmm_pfns = pfns; /* * Due to default_flags, all pages are HMM_PFN_VALID or -- cgit From f95f51a4c3357eabf74fe14ab7daa5b5c0422b27 Mon Sep 17 00:00:00 2001 From: Felix Kuehling Date: Wed, 21 Apr 2021 21:09:54 -0400 Subject: drm/amdgpu: Add notifier lock for KFD userptrs Add a per-process MMU notifier lock for processing notifiers from userptrs. Use that lock to properly synchronize page table updates with MMU notifiers. Signed-off-by: Felix Kuehling Reviewed-by: Xiaogang Chen Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 13 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 212 +++++++++++++++-------- drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c | 12 +- drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h | 3 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 17 +- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h | 6 + 6 files changed, 172 insertions(+), 91 deletions(-) (limited to 'drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h index f50e3ba4d7a5..589939631ed4 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h @@ -29,6 +29,7 @@ #include #include #include +#include #include #include #include "amdgpu_sync.h" @@ -65,6 +66,7 @@ struct kgd_mem { struct mutex lock; struct amdgpu_bo *bo; struct dma_buf *dmabuf; + struct hmm_range *range; struct list_head attachments; /* protected by amdkfd_process_info.lock */ struct ttm_validate_buffer validate_list; @@ -75,7 +77,7 @@ struct kgd_mem { uint32_t alloc_flags; - atomic_t invalid; + uint32_t invalid; struct amdkfd_process_info *process_info; struct amdgpu_sync sync; @@ -131,7 +133,8 @@ struct amdkfd_process_info { struct amdgpu_amdkfd_fence *eviction_fence; /* MMU-notifier related fields */ - atomic_t evicted_bos; + struct mutex notifier_lock; + uint32_t evicted_bos; struct delayed_work restore_userptr_work; struct pid *pid; bool block_mmu_notifications; @@ -180,7 +183,8 @@ int kfd_debugfs_kfd_mem_limits(struct seq_file *m, void *data); bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm); struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f); int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo); -int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm); +int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni, + unsigned long cur_seq, struct kgd_mem *mem); #else static inline bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm) @@ -201,7 +205,8 @@ int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo) } static inline -int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm) +int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni, + unsigned long cur_seq, struct kgd_mem *mem) { return 0; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index 8782916e64a0..0a854bb8b47e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -964,7 +964,9 @@ static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr, * later stage when it is scheduled by another ioctl called by * CRIU master process for the target pid for restore. */ - atomic_inc(&mem->invalid); + mutex_lock(&process_info->notifier_lock); + mem->invalid++; + mutex_unlock(&process_info->notifier_lock); mutex_unlock(&process_info->lock); return 0; } @@ -1301,6 +1303,7 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, return -ENOMEM; mutex_init(&info->lock); + mutex_init(&info->notifier_lock); INIT_LIST_HEAD(&info->vm_list_head); INIT_LIST_HEAD(&info->kfd_bo_list); INIT_LIST_HEAD(&info->userptr_valid_list); @@ -1317,7 +1320,6 @@ static int init_kfd_vm(struct amdgpu_vm *vm, void **process_info, } info->pid = get_task_pid(current->group_leader, PIDTYPE_PID); - atomic_set(&info->evicted_bos, 0); INIT_DELAYED_WORK(&info->restore_userptr_work, amdgpu_amdkfd_restore_userptr_worker); @@ -1372,6 +1374,7 @@ reserve_pd_fail: put_pid(info->pid); create_evict_fence_fail: mutex_destroy(&info->lock); + mutex_destroy(&info->notifier_lock); kfree(info); } return ret; @@ -1496,6 +1499,7 @@ void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev, cancel_delayed_work_sync(&process_info->restore_userptr_work); put_pid(process_info->pid); mutex_destroy(&process_info->lock); + mutex_destroy(&process_info->notifier_lock); kfree(process_info); } } @@ -1548,7 +1552,9 @@ int amdgpu_amdkfd_criu_resume(void *p) mutex_lock(&pinfo->lock); pr_debug("scheduling work\n"); - atomic_inc(&pinfo->evicted_bos); + mutex_lock(&pinfo->notifier_lock); + pinfo->evicted_bos++; + mutex_unlock(&pinfo->notifier_lock); if (!READ_ONCE(pinfo->block_mmu_notifications)) { ret = -EINVAL; goto out_unlock; @@ -1773,8 +1779,13 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu( list_del(&bo_list_entry->head); mutex_unlock(&process_info->lock); - /* No more MMU notifiers */ - amdgpu_hmm_unregister(mem->bo); + /* Cleanup user pages and MMU notifiers */ + if (amdgpu_ttm_tt_get_usermm(mem->bo->tbo.ttm)) { + amdgpu_hmm_unregister(mem->bo); + mutex_lock(&process_info->notifier_lock); + amdgpu_ttm_tt_discard_user_pages(mem->bo->tbo.ttm, mem->range); + mutex_unlock(&process_info->notifier_lock); + } ret = reserve_bo_and_cond_vms(mem, NULL, BO_VM_ALL, &ctx); if (unlikely(ret)) @@ -1864,6 +1875,16 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( */ mutex_lock(&mem->process_info->lock); + /* Lock notifier lock. If we find an invalid userptr BO, we can be + * sure that the MMU notifier is no longer running + * concurrently and the queues are actually stopped + */ + if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) { + mutex_lock(&mem->process_info->notifier_lock); + is_invalid_userptr = !!mem->invalid; + mutex_unlock(&mem->process_info->notifier_lock); + } + mutex_lock(&mem->lock); domain = mem->domain; @@ -2241,34 +2262,38 @@ int amdgpu_amdkfd_gpuvm_import_dmabuf(struct amdgpu_device *adev, * * Runs in MMU notifier, may be in RECLAIM_FS context. This means it * cannot do any memory allocations, and cannot take any locks that - * are held elsewhere while allocating memory. Therefore this is as - * simple as possible, using atomic counters. + * are held elsewhere while allocating memory. * * It doesn't do anything to the BO itself. The real work happens in * restore, where we get updated page addresses. This function only * ensures that GPU access to the BO is stopped. */ -int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, - struct mm_struct *mm) +int amdgpu_amdkfd_evict_userptr(struct mmu_interval_notifier *mni, + unsigned long cur_seq, struct kgd_mem *mem) { struct amdkfd_process_info *process_info = mem->process_info; - int evicted_bos; int r = 0; - /* Do not process MMU notifications until stage-4 IOCTL is received */ + /* Do not process MMU notifications during CRIU restore until + * KFD_CRIU_OP_RESUME IOCTL is received + */ if (READ_ONCE(process_info->block_mmu_notifications)) return 0; - atomic_inc(&mem->invalid); - evicted_bos = atomic_inc_return(&process_info->evicted_bos); - if (evicted_bos == 1) { + mutex_lock(&process_info->notifier_lock); + mmu_interval_set_seq(mni, cur_seq); + + mem->invalid++; + if (++process_info->evicted_bos == 1) { /* First eviction, stop the queues */ - r = kgd2kfd_quiesce_mm(mm, KFD_QUEUE_EVICTION_TRIGGER_USERPTR); + r = kgd2kfd_quiesce_mm(mni->mm, + KFD_QUEUE_EVICTION_TRIGGER_USERPTR); if (r) pr_err("Failed to quiesce KFD\n"); schedule_delayed_work(&process_info->restore_userptr_work, msecs_to_jiffies(AMDGPU_USERPTR_RESTORE_DELAY_MS)); } + mutex_unlock(&process_info->notifier_lock); return r; } @@ -2285,54 +2310,58 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info, struct kgd_mem *mem, *tmp_mem; struct amdgpu_bo *bo; struct ttm_operation_ctx ctx = { false, false }; - int invalid, ret; + uint32_t invalid; + int ret = 0; - /* Move all invalidated BOs to the userptr_inval_list and - * release their user pages by migration to the CPU domain - */ + mutex_lock(&process_info->notifier_lock); + + /* Move all invalidated BOs to the userptr_inval_list */ list_for_each_entry_safe(mem, tmp_mem, &process_info->userptr_valid_list, - validate_list.head) { - if (!atomic_read(&mem->invalid)) - continue; /* BO is still valid */ - - bo = mem->bo; - - if (amdgpu_bo_reserve(bo, true)) - return -EAGAIN; - amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); - ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); - amdgpu_bo_unreserve(bo); - if (ret) { - pr_err("%s: Failed to invalidate userptr BO\n", - __func__); - return -EAGAIN; - } - - list_move_tail(&mem->validate_list.head, - &process_info->userptr_inval_list); - } - - if (list_empty(&process_info->userptr_inval_list)) - return 0; /* All evicted userptr BOs were freed */ + validate_list.head) + if (mem->invalid) + list_move_tail(&mem->validate_list.head, + &process_info->userptr_inval_list); /* Go through userptr_inval_list and update any invalid user_pages */ list_for_each_entry(mem, &process_info->userptr_inval_list, validate_list.head) { - struct hmm_range *range; - - invalid = atomic_read(&mem->invalid); + invalid = mem->invalid; if (!invalid) /* BO hasn't been invalidated since the last - * revalidation attempt. Keep its BO list. + * revalidation attempt. Keep its page list. */ continue; bo = mem->bo; + amdgpu_ttm_tt_discard_user_pages(bo->tbo.ttm, mem->range); + mem->range = NULL; + + /* BO reservations and getting user pages (hmm_range_fault) + * must happen outside the notifier lock + */ + mutex_unlock(&process_info->notifier_lock); + + /* Move the BO to system (CPU) domain if necessary to unmap + * and free the SG table + */ + if (bo->tbo.resource->mem_type != TTM_PL_SYSTEM) { + if (amdgpu_bo_reserve(bo, true)) + return -EAGAIN; + amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_CPU); + ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx); + amdgpu_bo_unreserve(bo); + if (ret) { + pr_err("%s: Failed to invalidate userptr BO\n", + __func__); + return -EAGAIN; + } + } + /* Get updated user pages */ ret = amdgpu_ttm_tt_get_user_pages(bo, bo->tbo.ttm->pages, - &range); + &mem->range); if (ret) { pr_debug("Failed %d to get user pages\n", ret); @@ -2345,30 +2374,32 @@ static int update_invalid_user_pages(struct amdkfd_process_info *process_info, */ if (ret != -EFAULT) return ret; - } else { - /* - * FIXME: Cannot ignore the return code, must hold - * notifier_lock - */ - amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm, range); + ret = 0; } + mutex_lock(&process_info->notifier_lock); + /* Mark the BO as valid unless it was invalidated * again concurrently. */ - if (atomic_cmpxchg(&mem->invalid, invalid, 0) != invalid) - return -EAGAIN; + if (mem->invalid != invalid) { + ret = -EAGAIN; + goto unlock_out; + } + mem->invalid = 0; } - return 0; +unlock_out: + mutex_unlock(&process_info->notifier_lock); + + return ret; } /* Validate invalid userptr BOs * - * Validates BOs on the userptr_inval_list, and moves them back to the - * userptr_valid_list. Also updates GPUVM page tables with new page - * addresses and waits for the page table updates to complete. + * Validates BOs on the userptr_inval_list. Also updates GPUVM page tables + * with new page addresses and waits for the page table updates to complete. */ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info) { @@ -2439,9 +2470,6 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info) } } - list_move_tail(&mem->validate_list.head, - &process_info->userptr_valid_list); - /* Update mapping. If the BO was not validated * (because we couldn't get user pages), this will * clear the page table entries, which will result in @@ -2457,7 +2485,9 @@ static int validate_invalid_user_pages(struct amdkfd_process_info *process_info) if (ret) { pr_err("%s: update PTE failed\n", __func__); /* make sure this gets validated again */ - atomic_inc(&mem->invalid); + mutex_lock(&process_info->notifier_lock); + mem->invalid++; + mutex_unlock(&process_info->notifier_lock); goto unreserve_out; } } @@ -2477,6 +2507,36 @@ out_no_mem: return ret; } +/* Confirm that all user pages are valid while holding the notifier lock + * + * Moves valid BOs from the userptr_inval_list back to userptr_val_list. + */ +static int confirm_valid_user_pages_locked(struct amdkfd_process_info *process_info) +{ + struct kgd_mem *mem, *tmp_mem; + int ret = 0; + + list_for_each_entry_safe(mem, tmp_mem, + &process_info->userptr_inval_list, + validate_list.head) { + bool valid = amdgpu_ttm_tt_get_user_pages_done( + mem->bo->tbo.ttm, mem->range); + + mem->range = NULL; + if (!valid) { + WARN(!mem->invalid, "Invalid BO not marked invalid"); + ret = -EAGAIN; + continue; + } + WARN(mem->invalid, "Valid BO is marked invalid"); + + list_move_tail(&mem->validate_list.head, + &process_info->userptr_valid_list); + } + + return ret; +} + /* Worker callback to restore evicted userptr BOs * * Tries to update and validate all userptr BOs. If successful and no @@ -2491,9 +2551,11 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work) restore_userptr_work); struct task_struct *usertask; struct mm_struct *mm; - int evicted_bos; + uint32_t evicted_bos; - evicted_bos = atomic_read(&process_info->evicted_bos); + mutex_lock(&process_info->notifier_lock); + evicted_bos = process_info->evicted_bos; + mutex_unlock(&process_info->notifier_lock); if (!evicted_bos) return; @@ -2516,9 +2578,6 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work) * and we can just restart the queues. */ if (!list_empty(&process_info->userptr_inval_list)) { - if (atomic_read(&process_info->evicted_bos) != evicted_bos) - goto unlock_out; /* Concurrent eviction, try again */ - if (validate_invalid_user_pages(process_info)) goto unlock_out; } @@ -2527,10 +2586,17 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work) * be a first eviction that calls quiesce_mm. The eviction * reference counting inside KFD will handle this case. */ - if (atomic_cmpxchg(&process_info->evicted_bos, evicted_bos, 0) != - evicted_bos) - goto unlock_out; - evicted_bos = 0; + mutex_lock(&process_info->notifier_lock); + if (process_info->evicted_bos != evicted_bos) + goto unlock_notifier_out; + + if (confirm_valid_user_pages_locked(process_info)) { + WARN(1, "User pages unexpectedly invalid"); + goto unlock_notifier_out; + } + + process_info->evicted_bos = evicted_bos = 0; + if (kgd2kfd_resume_mm(mm)) { pr_err("%s: Failed to resume KFD\n", __func__); /* No recovery from this failure. Probably the CP is @@ -2538,6 +2604,8 @@ static void amdgpu_amdkfd_restore_userptr_worker(struct work_struct *work) */ } +unlock_notifier_out: + mutex_unlock(&process_info->notifier_lock); unlock_out: mutex_unlock(&process_info->lock); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c index 65715cb395d8..2dadcfe43d03 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c @@ -105,17 +105,11 @@ static bool amdgpu_hmm_invalidate_hsa(struct mmu_interval_notifier *mni, unsigned long cur_seq) { struct amdgpu_bo *bo = container_of(mni, struct amdgpu_bo, notifier); - struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); if (!mmu_notifier_range_blockable(range)) return false; - mutex_lock(&adev->notifier_lock); - - mmu_interval_set_seq(mni, cur_seq); - - amdgpu_amdkfd_evict_userptr(bo->kfd_bo, bo->notifier.mm); - mutex_unlock(&adev->notifier_lock); + amdgpu_amdkfd_evict_userptr(mni, cur_seq, bo->kfd_bo); return true; } @@ -244,9 +238,9 @@ out_free_range: return r; } -int amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range) +bool amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range) { - int r; + bool r; r = mmu_interval_read_retry(hmm_range->notifier, hmm_range->notifier_seq); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h index 13ed94d3b01b..e2edcd010ccc 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.h @@ -29,12 +29,13 @@ #include #include #include +#include int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier, uint64_t start, uint64_t npages, bool readonly, void *owner, struct page **pages, struct hmm_range **phmm_range); -int amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range); +bool amdgpu_hmm_range_get_pages_done(struct hmm_range *hmm_range); #if defined(CONFIG_HMM_MIRROR) int amdgpu_hmm_register(struct amdgpu_bo *bo, unsigned long addr); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index b4236572eae1..f0e4c7309438 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -695,8 +695,19 @@ out_unlock: return r; } +/* amdgpu_ttm_tt_discard_user_pages - Discard range and pfn array allocations + */ +void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm, + struct hmm_range *range) +{ + struct amdgpu_ttm_tt *gtt = (void *)ttm; + + if (gtt && gtt->userptr && range) + amdgpu_hmm_range_get_pages_done(range); +} + /* - * amdgpu_ttm_tt_userptr_range_done - stop HMM track the CPU page table change + * amdgpu_ttm_tt_get_user_pages_done - stop HMM track the CPU page table change * Check if the pages backing this ttm range have been invalidated * * Returns: true if pages are still valid @@ -714,10 +725,6 @@ bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm, WARN_ONCE(!range->hmm_pfns, "No user pages to check\n"); - /* - * FIXME: Must always hold notifier_lock for this, and must - * not ignore the return code. - */ return !amdgpu_hmm_range_get_pages_done(range); } #endif diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h index b4d8ba2789f3..e2cd5894afc9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h @@ -159,6 +159,8 @@ uint64_t amdgpu_ttm_domain_start(struct amdgpu_device *adev, uint32_t type); #if IS_ENABLED(CONFIG_DRM_AMDGPU_USERPTR) int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, struct page **pages, struct hmm_range **range); +void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm, + struct hmm_range *range); bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm, struct hmm_range *range); #else @@ -168,6 +170,10 @@ static inline int amdgpu_ttm_tt_get_user_pages(struct amdgpu_bo *bo, { return -EPERM; } +static inline void amdgpu_ttm_tt_discard_user_pages(struct ttm_tt *ttm, + struct hmm_range *range) +{ +} static inline bool amdgpu_ttm_tt_get_user_pages_done(struct ttm_tt *ttm, struct hmm_range *range) { -- cgit