drm/amdkfd: reserve the BO before validating it
Subsystems (3)
PATCH v2 1/1 drm/amdkfd: reserve the BO before validating it
v2:
Avoid unmapping attachment twice when ERESTARTSYS.
[ 41.708711] WARNING: CPU: 0 PID: 1463 at drivers/gpu/drm/ttm/ttm_bo.c:846 ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.708989] Call Trace: [ 41.708992] <TASK> [ 41.708996] ? show_regs+0x6c/0x80 [ 41.709000] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709008] ? __warn+0x93/0x190 [ 41.709014] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709024] ? report_bug+0x1f9/0x210 [ 41.709035] ? handle_bug+0x46/0x80 [ 41.709041] ? exc_invalid_op+0x1d/0x80 [ 41.709048] ? asm_exc_invalid_op+0x1f/0x30 [ 41.709057] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80 [amdgpu] [ 41.709185] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709197] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80 [amdgpu] [ 41.709337] ? srso_alias_return_thunk+0x5/0x7f [ 41.709346] kfd_mem_dmaunmap_attachment+0x9e/0x1e0 [amdgpu] [ 41.709467] amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x56/0x80 [amdgpu] [ 41.709586] kfd_ioctl_unmap_memory_from_gpu+0x1b7/0x300 [amdgpu] [ 41.709710] kfd_ioctl+0x1ec/0x650 [amdgpu] [ 41.709822] ? __pfx_kfd_ioctl_unmap_memory_from_gpu+0x10/0x10 [amdgpu] [ 41.709945] ? srso_alias_return_thunk+0x5/0x7f [ 41.709949] ? tomoyo_file_ioctl+0x20/0x30 [ 41.709959] __x64_sys_ioctl+0x9c/0xd0 [ 41.709967] do_syscall_64+0x3f/0x90 [ 41.709973] entry_SYSCALL_64_after_hwframe+0x6e/0xd8
--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 28 +++++++++++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 4 ++- 3 files changed, 29 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -1
@@ -311,7 +311,7 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv);int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu (struct amdgpu_device *adev,struct kgd_mem *mem,void *drm_priv); -void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv); +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv);int amdgpu_amdkfd_gpuvm_sync_memory (struct amdgpu_device *adev,struct kgd_mem *mem,bool intr); int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +25 -3
@@ -2088,21 +2088,43 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(return ret; } -void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv) +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv) {struct kfd_mem_attachment *entry;struct amdgpu_vm *vm; +bool reserved = false; +int ret =0 ; vm =drm_priv_to_vm (drm_priv);mutex_lock (&mem->lock);list_for_each_entry (entry, &mem->attachments, list) { -if (entry->bo_va->base.vm == vm) -kfd_mem_dmaunmap_attachment (mem, entry); +if (entry->bo_va->base.vm != vm) +continue ; +if (entry->type ==KFD_MEM_ATT_SHARED || + entry->type ==KFD_MEM_ATT_DMABUF ) +continue ; +if (!entry->bo_va->base.bo->tbo.ttm->sg) +continue ; + +if (!reserved) { + ret =amdgpu_bo_reserve (mem->bo, true); +if (ret) + goto out; + reserved = true; + } + +kfd_mem_dmaunmap_attachment (mem, entry); } +if (reserved) +amdgpu_bo_unreserve (mem->bo); + +out:mutex_unlock (&mem->lock); + +return ret; } int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3 -1
@@ -1442,7 +1442,9 @@ static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,kfd_flush_tlb (peer_pdd,TLB_FLUSH_HEAVYWEIGHT );/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */ -amdgpu_amdkfd_gpuvm_dmaunmap_mem (mem, peer_pdd->drm_priv); + err =amdgpu_amdkfd_gpuvm_dmaunmap_mem (mem, peer_pdd->drm_priv); +if (err) + goto sync_memory_failed; }mutex_unlock (&p->mutex);
Signature · 1 line
2.25.1
Felix Kuehling
maintainerQuoted 71 lines · Lang Yu
Fixes 410f08516e0f ("drm/amdkfd: Move dma unmapping after TLB flush")v2:
Avoid unmapping attachment twice when ERESTARTSYS.[ 41.708711] WARNING: CPU: 0 PID: 1463 at drivers/gpu/drm/ttm/ttm_bo.c:846 ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.708989] Call Trace: [ 41.708992] <TASK> [ 41.708996] ? show_regs+0x6c/0x80 [ 41.709000] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709008] ? __warn+0x93/0x190 [ 41.709014] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709024] ? report_bug+0x1f9/0x210 [ 41.709035] ? handle_bug+0x46/0x80 [ 41.709041] ? exc_invalid_op+0x1d/0x80 [ 41.709048] ? asm_exc_invalid_op+0x1f/0x30 [ 41.709057] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80 [amdgpu] [ 41.709185] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709197] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80 [amdgpu] [ 41.709337] ? srso_alias_return_thunk+0x5/0x7f [ 41.709346] kfd_mem_dmaunmap_attachment+0x9e/0x1e0 [amdgpu] [ 41.709467] amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x56/0x80 [amdgpu] [ 41.709586] kfd_ioctl_unmap_memory_from_gpu+0x1b7/0x300 [amdgpu] [ 41.709710] kfd_ioctl+0x1ec/0x650 [amdgpu] [ 41.709822] ? __pfx_kfd_ioctl_unmap_memory_from_gpu+0x10/0x10 [amdgpu] [ 41.709945] ? srso_alias_return_thunk+0x5/0x7f [ 41.709949] ? tomoyo_file_ioctl+0x20/0x30 [ 41.709959] __x64_sys_ioctl+0x9c/0xd0 [ 41.709967] do_syscall_64+0x3f/0x90 [ 41.709973] entry_SYSCALL_64_after_hwframe+0x6e/0xd8Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 28 +++++++++++++++++-- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 4 ++- 3 files changed, 29 insertions(+), 5 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1
-1@@ -311,7 +311,7 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct amdgpu_device *adev, struct kgd_mem *mem, void *drm_priv);int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu (struct amdgpu_device *adev,struct kgd_mem *mem,void *drm_priv); -void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv); +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv);int amdgpu_amdkfd_gpuvm_sync_memory (struct amdgpu_device *adev,struct kgd_mem *mem,bool intr); int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +3
-1@@ -2088,21 +2088,43 @@ int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(return ret; } -void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv) +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv) {struct kfd_mem_attachment *entry;struct amdgpu_vm *vm; +bool reserved = false; +int ret =0 ; vm =drm_priv_to_vm (drm_priv);mutex_lock (&mem->lock);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( · line 2109
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +7
-2list_for_each_entry (entry, &mem->attachments, list) { -if (entry->bo_va->base.vm == vm) -kfd_mem_dmaunmap_attachment (mem, entry); +if (entry->bo_va->base.vm != vm) +continue ; +if (entry->type ==KFD_MEM_ATT_SHARED || + entry->type ==KFD_MEM_ATT_DMABUF ) +continue ; +if (!entry->bo_va->base.bo->tbo.ttm->sg) +continue ;
You're going to great lengths to avoid the reservation when it's not needed by kfd_mem_dmaunmap_attachment. However, this feels a bit fragile. Any changes in the kfd_mem_dmaunmap_* functions could break this.
Quoted 26 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +15
-0+ +if (!reserved) { + ret =amdgpu_bo_reserve (mem->bo, true); +if (ret) + goto out; + reserved = true; + } + +kfd_mem_dmaunmap_attachment (mem, entry); } +if (reserved) +amdgpu_bo_unreserve (mem->bo); + +out:mutex_unlock (&mem->lock); + +return ret; } int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, · line 1447
Quoted 7 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +3
-1@@ -1442,7 +1442,9 @@ static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,kfd_flush_tlb (peer_pdd,TLB_FLUSH_HEAVYWEIGHT );/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */ -amdgpu_amdkfd_gpuvm_dmaunmap_mem (mem, peer_pdd->drm_priv); + err =amdgpu_amdkfd_gpuvm_dmaunmap_mem (mem, peer_pdd->drm_priv); +if (err) + goto sync_memory_failed;
This handles the case that the system call got interrupted. But you're not handling the restart correctly. When the ioctl is restarted, you don't know how many dmaunmaps are already done. So you'd need to make sure that repeating the dmaunmap is safe in all cases. Or what David tried earlier, find a way to track the unmapping so you only try to dmaunmap on GPUs where it's actually dmamapped.
Regards,
Felix
Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0}mutex_unlock (&p->mutex);
[AMD Official Use Only - General]
Quoted 146 lines
-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, January 25, 2024 5:41 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Francis, David <[email protected]>
Subject: Re: [PATCH v2] drm/amdkfd: reserve the BO before validating itQuoted 6 lines · Lang Yu
Fixes 410f08516e0f ("drm/amdkfd: Move dma unmapping after TLB flush")v2:
Avoid unmapping attachment twice when ERESTARTSYS.[ 41.708711] WARNING: CPU: 0 PID: 1463 at
drivers/gpu/drm/ttm/ttm_bo.c:846ttm_bo_validate+0x146/0x1b0[ttm]Quoted 11 lines
[ 41.708989] Call Trace: [ 41.708992] <TASK> [ 41.708996] ? show_regs+0x6c/0x80 [ 41.709000] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709008] ? __warn+0x93/0x190 [ 41.709014] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709024] ? report_bug+0x1f9/0x210 [ 41.709035] ? handle_bug+0x46/0x80 [ 41.709041] ? exc_invalid_op+0x1d/0x80 [ 41.709048] ? asm_exc_invalid_op+0x1f/0x30 [ 41.709057] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80[amdgpu]
Quoted 2 lines
[ 41.709185] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709197] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80[amdgpu]
Quoted 3 lines
[ 41.709337] ? srso_alias_return_thunk+0x5/0x7f [ 41.709346] kfd_mem_dmaunmap_attachment+0x9e/0x1e0 [amdgpu] [ 41.709467] amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x56/0x80[amdgpu]
Quoted 3 lines
[ 41.709586] kfd_ioctl_unmap_memory_from_gpu+0x1b7/0x300 [amdgpu] [ 41.709710] kfd_ioctl+0x1ec/0x650 [amdgpu] [ 41.709822] ? __pfx_kfd_ioctl_unmap_memory_from_gpu+0x10/0x10[amdgpu]
Quoted 10 lines
[ 41.709945] ? srso_alias_return_thunk+0x5/0x7f [ 41.709949] ? tomoyo_file_ioctl+0x20/0x30 [ 41.709959] __x64_sys_ioctl+0x9c/0xd0 [ 41.709967] do_syscall_64+0x3f/0x90 [ 41.709973] entry_SYSCALL_64_after_hwframe+0x6e/0xd8Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 28+++++++++++++++++--
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int
Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 4 ++- 3 files changed, 29 insertions(+), 5 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0@@ -311,7 +311,7 @@ intamdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct
amdgpu_device*adev,drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 311
Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0struct kgd_mem *mem, void
*drm_priv);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 313
Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu ( struct amdgpu_device *adev, struct kgd_mem *mem, void
*drm_priv);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int
Quoted 13 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +2
-1-void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv); +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void +*drm_priv);int amdgpu_amdkfd_gpuvm_sync_memory (struct amdgpu_device *adev,struct kgd_mem *mem,bool intr); int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0
-0@@ -2088,21 +2088,43 @@ intamdgpu_amdkfd_gpuvm_map_memory_to_gpu(
Quoted 17 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +4
-1return ret; } -void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv) +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void +*drm_priv) {struct kfd_mem_attachment *entry;struct amdgpu_vm *vm; +bool reserved = false; +int ret =0 ; vm =drm_priv_to_vm (drm_priv);mutex_lock (&mem->lock);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( · line 2109
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +7
-2list_for_each_entry (entry, &mem->attachments, list) { -if (entry->bo_va->base.vm == vm) -kfd_mem_dmaunmap_attachment (mem, entry); +if (entry->bo_va->base.vm != vm) +continue ; +if (entry->type ==KFD_MEM_ATT_SHARED || + entry->type ==KFD_MEM_ATT_DMABUF ) +continue ; +if (!entry->bo_va->base.bo->tbo.ttm->sg) +continue ;You're going to great lengths to avoid the reservation when it's not needed by
kfd_mem_dmaunmap_attachment. However, this feels a bit fragile. Any changes in the kfd_mem_dmaunmap_* functions could break this.drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct
Quoted 27 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +15
-0+ +if (!reserved) { + ret =amdgpu_bo_reserve (mem->bo, true); +if (ret) + goto out; + reserved = true; + } + +kfd_mem_dmaunmap_attachment (mem, entry); } +if (reserved) +amdgpu_bo_unreserve (mem->bo); + +out:mutex_unlock (&mem->lock); + +return ret; } int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0@@ -1442,7 +1442,9 @@ static int kfd_ioctl_unmap_memory_from_gpu(structfile *filep,
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0kfd_flush_tlb (peer_pdd,TLB_FLUSH_HEAVYWEIGHT );/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct · line 1446
Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1
-1-amdgpu_amdkfd_gpuvm_dmaunmap_mem (mem, peer_pdd- drm_priv); + err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem,
peer_pdd->drm_priv);
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, · line 1447
Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2
-0+if (err) + goto sync_memory_failed;This handles the case that the system call got interrupted. But you're not handling the restart correctly. When the ioctl is restarted, you don't know how many dmaunmaps are already done. So you'd need to make sure that repeating the dmaunmap is safe in all cases. Or what David tried earlier, find a way to track the unmapping so you only try to dmaunmap on GPUs where it's actually dmamapped.
From previous discussion, no one likes add another variable to track the unmappings. So I'm avoiding adding another variable.
Actually, all memory attachments use sg_table, ttm->sg is NULL? can be used as an indicator to see whether an attachment is already unmapped. That already unmapped will not be repeated.
Regards,
Lang
Quoted 7 lines
Regards,
FelixQuoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0}mutex_unlock(&p->mutex);
Felix Kuehling
maintainerQuoted 154 lines · Lang
[AMD Official Use Only - General]
Quoted 144 lines
-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, January 25, 2024 5:41 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Francis, David <[email protected]>
Subject: Re: [PATCH v2] drm/amdkfd: reserve the BO before validating itQuoted 6 lines · Lang Yu
Fixes 410f08516e0f ("drm/amdkfd: Move dma unmapping after TLB flush")v2:
Avoid unmapping attachment twice when ERESTARTSYS.[ 41.708711] WARNING: CPU: 0 PID: 1463 at
drivers/gpu/drm/ttm/ttm_bo.c:846ttm_bo_validate+0x146/0x1b0[ttm]Quoted 11 lines
[ 41.708989] Call Trace: [ 41.708992] <TASK> [ 41.708996] ? show_regs+0x6c/0x80 [ 41.709000] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709008] ? __warn+0x93/0x190 [ 41.709014] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709024] ? report_bug+0x1f9/0x210 [ 41.709035] ? handle_bug+0x46/0x80 [ 41.709041] ? exc_invalid_op+0x1d/0x80 [ 41.709048] ? asm_exc_invalid_op+0x1f/0x30 [ 41.709057] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80[amdgpu]
Quoted 2 lines
[ 41.709185] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709197] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80[amdgpu]
Quoted 3 lines
[ 41.709337] ? srso_alias_return_thunk+0x5/0x7f [ 41.709346] kfd_mem_dmaunmap_attachment+0x9e/0x1e0 [amdgpu] [ 41.709467] amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x56/0x80[amdgpu]
Quoted 3 lines
[ 41.709586] kfd_ioctl_unmap_memory_from_gpu+0x1b7/0x300 [amdgpu] [ 41.709710] kfd_ioctl+0x1ec/0x650 [amdgpu] [ 41.709822] ? __pfx_kfd_ioctl_unmap_memory_from_gpu+0x10/0x10[amdgpu]
Quoted 10 lines
[ 41.709945] ? srso_alias_return_thunk+0x5/0x7f [ 41.709949] ? tomoyo_file_ioctl+0x20/0x30 [ 41.709959] __x64_sys_ioctl+0x9c/0xd0 [ 41.709967] do_syscall_64+0x3f/0x90 [ 41.709973] entry_SYSCALL_64_after_hwframe+0x6e/0xd8Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 28+++++++++++++++++--
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int
Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 4 ++- 3 files changed, 29 insertions(+), 5 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0@@ -311,7 +311,7 @@ intamdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct
amdgpu_device*adev,drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 311
Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0struct kgd_mem *mem, void
*drm_priv);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 313
Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu ( struct amdgpu_device *adev, struct kgd_mem *mem, void
*drm_priv);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int
Quoted 13 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +2
-1-void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv); +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void +*drm_priv);int amdgpu_amdkfd_gpuvm_sync_memory (struct amdgpu_device *adev,struct kgd_mem *mem,bool intr); int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0
-0@@ -2088,21 +2088,43 @@ intamdgpu_amdkfd_gpuvm_map_memory_to_gpu(
Quoted 17 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +4
-1return ret; } -void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv) +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void +*drm_priv) {struct kfd_mem_attachment *entry;struct amdgpu_vm *vm; +bool reserved = false; +int ret =0 ; vm =drm_priv_to_vm (drm_priv);mutex_lock (&mem->lock);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( · line 2109
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +7
-2list_for_each_entry (entry, &mem->attachments, list) { -if (entry->bo_va->base.vm == vm) -kfd_mem_dmaunmap_attachment (mem, entry); +if (entry->bo_va->base.vm != vm) +continue ; +if (entry->type ==KFD_MEM_ATT_SHARED || + entry->type ==KFD_MEM_ATT_DMABUF ) +continue ; +if (!entry->bo_va->base.bo->tbo.ttm->sg) +continue ;You're going to great lengths to avoid the reservation when it's not needed by
kfd_mem_dmaunmap_attachment. However, this feels a bit fragile. Any changes in the kfd_mem_dmaunmap_* functions could break this.drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct
Quoted 27 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +15
-0+ +if (!reserved) { + ret =amdgpu_bo_reserve (mem->bo, true); +if (ret) + goto out; + reserved = true; + } + +kfd_mem_dmaunmap_attachment (mem, entry); } +if (reserved) +amdgpu_bo_unreserve (mem->bo); + +out:mutex_unlock (&mem->lock); + +return ret; } int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0@@ -1442,7 +1442,9 @@ static int kfd_ioctl_unmap_memory_from_gpu(structfile *filep,
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep,
Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0kfd_flush_tlb (peer_pdd,TLB_FLUSH_HEAVYWEIGHT );/* Remove dma mapping after tlb flush to avoid IO_PAGE_FAULT */
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct · line 1446
Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +1
-1-amdgpu_amdkfd_gpuvm_dmaunmap_mem (mem, peer_pdd- drm_priv); + err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem,Quoted 8 lines
peer_pdd->drm_priv);drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, · line 1447
Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2
-0+if (err) + goto sync_memory_failed;This handles the case that the system call got interrupted. But you're not handling the restart correctly. When the ioctl is restarted, you don't know how many dmaunmaps are already done. So you'd need to make sure that repeating the dmaunmap is safe in all cases. Or what David tried earlier, find a way to track the unmapping so you only try to dmaunmap on GPUs where it's actually dmamapped.
Quoted 4 lines
From previous discussion, no one likes add another variable to track the unmappings. So I'm avoiding adding another variable.
Actually, all memory attachments use
sg_table,ttm->sgis NULL? can be used as an indicator to see whether an attachment is already unmapped. That already unmapped will not be repeated.
I think that should work. I'd add the checks in kfd_mem_dmaunmap_userptr and kfd_mem_dmaunmap_sg_bo, where we also set ttm->sg to NULL. In fact, both those functions already have that check. So looks like it should handle the the restart correctly with your patch.
Regards,
Felix
Quoted 11 lines
Regards,
LangQuoted 7 lines
Regards,
FelixQuoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0}mutex_unlock(&p->mutex);
[AMD Official Use Only - General]
Quoted 172 lines
-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Saturday, January 27, 2024 3:22 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Francis, David <[email protected]>
Subject: Re: [PATCH v2] drm/amdkfd: reserve the BO before validating itQuoted 159 lines · Lang
[AMD Official Use Only - General]
Quoted 157 lines
-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, January 25, 2024 5:41 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Francis, David <[email protected]>
Subject: Re: [PATCH v2] drm/amdkfd: reserve the BO before validating itQuoted 7 lines · Lang Yu
Fixes 410f08516e0f ("drm/amdkfd: Move dma unmapping after TLBflush")
v2:
Avoid unmapping attachment twice when ERESTARTSYS.[ 41.708711] WARNING: CPU: 0 PID: 1463 at
drivers/gpu/drm/ttm/ttm_bo.c:846ttm_bo_validate+0x146/0x1b0[ttm]Quoted 11 lines
[ 41.708989] Call Trace: [ 41.708992] <TASK> [ 41.708996] ? show_regs+0x6c/0x80 [ 41.709000] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709008] ? __warn+0x93/0x190 [ 41.709014] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709024] ? report_bug+0x1f9/0x210 [ 41.709035] ? handle_bug+0x46/0x80 [ 41.709041] ? exc_invalid_op+0x1d/0x80 [ 41.709048] ? asm_exc_invalid_op+0x1f/0x30 [ 41.709057] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80[amdgpu]
Quoted 2 lines
[ 41.709185] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709197] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80[amdgpu]
Quoted 3 lines
[ 41.709337] ? srso_alias_return_thunk+0x5/0x7f [ 41.709346] kfd_mem_dmaunmap_attachment+0x9e/0x1e0 [amdgpu] [ 41.709467] amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x56/0x80[amdgpu]
Quoted 3 lines
[ 41.709586] kfd_ioctl_unmap_memory_from_gpu+0x1b7/0x300 [amdgpu] [ 41.709710] kfd_ioctl+0x1ec/0x650 [amdgpu] [ 41.709822] ? __pfx_kfd_ioctl_unmap_memory_from_gpu+0x10/0x10[amdgpu]
Quoted 10 lines
[ 41.709945] ? srso_alias_return_thunk+0x5/0x7f [ 41.709949] ? tomoyo_file_ioctl+0x20/0x30 [ 41.709959] __x64_sys_ioctl+0x9c/0xd0 [ 41.709967] do_syscall_64+0x3f/0x90 [ 41.709973] entry_SYSCALL_64_after_hwframe+0x6e/0xd8Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 28+++++++++++++++++--
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int
Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 4 ++- 3 files changed, 29 insertions(+), 5 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0@@ -311,7 +311,7 @@ intamdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct
amdgpu_device*adev,drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 311
Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0struct kgd_mem *mem, void
*drm_priv);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 313
Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu ( struct amdgpu_device *adev, struct kgd_mem *mem, void
*drm_priv);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int
Quoted 13 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +2
-1-void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv); +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void +*drm_priv);int amdgpu_amdkfd_gpuvm_sync_memory (struct amdgpu_device *adev,struct kgd_mem *mem,bool intr); int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0
-0@@ -2088,21 +2088,43 @@ intamdgpu_amdkfd_gpuvm_map_memory_to_gpu(
Quoted 17 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +4
-1return ret; } -void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv) +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void +*drm_priv) {struct kfd_mem_attachment *entry;struct amdgpu_vm *vm; +bool reserved = false; +int ret =0 ; vm =drm_priv_to_vm (drm_priv);mutex_lock (&mem->lock);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( · line 2109
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +7
-2list_for_each_entry (entry, &mem->attachments, list) { -if (entry->bo_va->base.vm == vm) -kfd_mem_dmaunmap_attachment (mem, entry); +if (entry->bo_va->base.vm != vm) +continue ; +if (entry->type ==KFD_MEM_ATT_SHARED || + entry->type ==KFD_MEM_ATT_DMABUF ) +continue ; +if (!entry->bo_va->base.bo->tbo.ttm->sg) +continue ;You're going to great lengths to avoid the reservation when it's not needed by
kfd_mem_dmaunmap_attachment. However, this feels a bit fragile. Any changes in the kfd_mem_dmaunmap_* functions could break this.Quoted 27 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +15
-0+ +if (!reserved) { + ret =amdgpu_bo_reserve (mem->bo, true); +if (ret) + goto out; + reserved = true; + } + +kfd_mem_dmaunmap_attachment (mem, entry); } +if (reserved) +amdgpu_bo_unreserve (mem->bo); + +out:mutex_unlock (&mem->lock); + +return ret; } int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0@@ -1442,7 +1442,9 @@ static intQuoted 1 line
kfd_ioctl_unmap_memory_from_gpu(struct
file *filep,
Quoted 5 lines
kfd_flush_tlb(peer_pdd,
TLB_FLUSH_HEAVYWEIGHT);/* Remove dma mapping after tlb flush to avoid
IO_PAGE_FAULT*/
Quoted 3 lines
- amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd-
drm_priv);
+ err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem,
peer_pdd->drm_priv);drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, · line 1447
Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2
-0+if (err) + goto sync_memory_failed;This handles the case that the system call got interrupted. But you're not handling the restart correctly. When the ioctl is restarted, you don't know how many dmaunmaps are already done. So you'd need to make sure that repeating the dmaunmap is safe in all cases. Or what David tried earlier, find a way to track the unmapping so you
only try to dmaunmap on GPUs where it's actually dmamapped.
Quoted 1 line
From previous discussion, no one likes add another variable to track the
unmappings. So I'm avoiding adding another variable.
Quoted 1 line
Quoted 8 lines
Quoted 1 line
Actually, all memory attachments use
sg_table,ttm->sgis NULL? can be used asan indicator to see whether an attachment is already unmapped.
Quoted 1 line
That already unmapped will not be repeated.
I think that should work. I'd add the checks in
kfd_mem_dmaunmap_userptrandkfd_mem_dmaunmap_sg_bo, where we also setttm->sgto NULL. In fact, both those functions already have that check. So looks like it should handle the the restart correctly with your patch.
Yes, both kfd_mem_dmaunmap_userptr() and kfd_mem_dmaunmap_sg_bo() have NULL check for ttm->sg.
And dmabuf also have this check in amdgpu_ttm_backend_unbind(). So dmaunmap won't be repeated actually.
Then the benefits of handling ERESTARTSYS is avoiding amdgpu_bo_reserve().
What do you think? It's worth avoiding reservation in this case?
Regards,
Lang
Quoted 15 lines
Regards,
FelixQuoted 11 lines
Regards,
LangQuoted 7 lines
Regards,
FelixQuoted 3 lines
} mutex_unlock(&p->mutex);
Felix Kuehling
maintainerQuoted 180 lines · Lang
[AMD Official Use Only - General]
Quoted 170 lines
-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Saturday, January 27, 2024 3:22 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Francis, David <[email protected]>
Subject: Re: [PATCH v2] drm/amdkfd: reserve the BO before validating itQuoted 159 lines · Lang
[AMD Official Use Only - General]
Quoted 157 lines
-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, January 25, 2024 5:41 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Francis, David <[email protected]>
Subject: Re: [PATCH v2] drm/amdkfd: reserve the BO before validating itQuoted 7 lines · Lang Yu
Fixes 410f08516e0f ("drm/amdkfd: Move dma unmapping after TLBflush")
v2:
Avoid unmapping attachment twice when ERESTARTSYS.[ 41.708711] WARNING: CPU: 0 PID: 1463 at
drivers/gpu/drm/ttm/ttm_bo.c:846ttm_bo_validate+0x146/0x1b0[ttm]Quoted 11 lines
[ 41.708989] Call Trace: [ 41.708992] <TASK> [ 41.708996] ? show_regs+0x6c/0x80 [ 41.709000] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709008] ? __warn+0x93/0x190 [ 41.709014] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709024] ? report_bug+0x1f9/0x210 [ 41.709035] ? handle_bug+0x46/0x80 [ 41.709041] ? exc_invalid_op+0x1d/0x80 [ 41.709048] ? asm_exc_invalid_op+0x1f/0x30 [ 41.709057] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80[amdgpu]
Quoted 2 lines
[ 41.709185] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709197] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80[amdgpu]
Quoted 3 lines
[ 41.709337] ? srso_alias_return_thunk+0x5/0x7f [ 41.709346] kfd_mem_dmaunmap_attachment+0x9e/0x1e0 [amdgpu] [ 41.709467] amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x56/0x80[amdgpu]
Quoted 3 lines
[ 41.709586] kfd_ioctl_unmap_memory_from_gpu+0x1b7/0x300 [amdgpu] [ 41.709710] kfd_ioctl+0x1ec/0x650 [amdgpu] [ 41.709822] ? __pfx_kfd_ioctl_unmap_memory_from_gpu+0x10/0x10[amdgpu]
Quoted 10 lines
[ 41.709945] ? srso_alias_return_thunk+0x5/0x7f [ 41.709949] ? tomoyo_file_ioctl+0x20/0x30 [ 41.709959] __x64_sys_ioctl+0x9c/0xd0 [ 41.709967] do_syscall_64+0x3f/0x90 [ 41.709973] entry_SYSCALL_64_after_hwframe+0x6e/0xd8Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 28+++++++++++++++++--
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int
Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 4 ++- 3 files changed, 29 insertions(+), 5 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0@@ -311,7 +311,7 @@ intamdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct
amdgpu_device*adev,drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 311
Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0struct kgd_mem *mem, void
*drm_priv);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 313
Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu ( struct amdgpu_device *adev, struct kgd_mem *mem, void
*drm_priv);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int
Quoted 13 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +2
-1-void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv); +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void +*drm_priv);int amdgpu_amdkfd_gpuvm_sync_memory (struct amdgpu_device *adev,struct kgd_mem *mem,bool intr); int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem *mem,drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0
-0@@ -2088,21 +2088,43 @@ intamdgpu_amdkfd_gpuvm_map_memory_to_gpu(
Quoted 17 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +4
-1return ret; } -void amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void *drm_priv) +int amdgpu_amdkfd_gpuvm_dmaunmap_mem (struct kgd_mem *mem,void +*drm_priv) {struct kfd_mem_attachment *entry;struct amdgpu_vm *vm; +bool reserved = false; +int ret =0 ; vm =drm_priv_to_vm (drm_priv);mutex_lock (&mem->lock);drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_map_memory_to_gpu( · line 2109
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +7
-2list_for_each_entry (entry, &mem->attachments, list) { -if (entry->bo_va->base.vm == vm) -kfd_mem_dmaunmap_attachment (mem, entry); +if (entry->bo_va->base.vm != vm) +continue ; +if (entry->type ==KFD_MEM_ATT_SHARED || + entry->type ==KFD_MEM_ATT_DMABUF ) +continue ; +if (!entry->bo_va->base.bo->tbo.ttm->sg) +continue ;You're going to great lengths to avoid the reservation when it's not needed by
kfd_mem_dmaunmap_attachment. However, this feels a bit fragile. Any changes in the kfd_mem_dmaunmap_* functions could break this.Quoted 27 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +15
-0+ +if (!reserved) { + ret =amdgpu_bo_reserve (mem->bo, true); +if (ret) + goto out; + reserved = true; + } + +kfd_mem_dmaunmap_attachment (mem, entry); } +if (reserved) +amdgpu_bo_unreserve (mem->bo); + +out:mutex_unlock (&mem->lock); + +return ret; } int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0@@ -1442,7 +1442,9 @@ static intQuoted 1 line
kfd_ioctl_unmap_memory_from_gpu(struct
file *filep,
Quoted 5 lines
kfd_flush_tlb(peer_pdd,
TLB_FLUSH_HEAVYWEIGHT);/* Remove dma mapping after tlb flush to avoid
IO_PAGE_FAULT*/
Quoted 3 lines
- amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd-
drm_priv);
+ err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem,
peer_pdd->drm_priv);drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, · line 1447
Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2
-0+if (err) + goto sync_memory_failed;This handles the case that the system call got interrupted. But you're not handling the restart correctly. When the ioctl is restarted, you don't know how many dmaunmaps are already done. So you'd need to make sure that repeating the dmaunmap is safe in all cases. Or what David tried earlier, find a way to track the unmapping so you
only try to dmaunmap on GPUs where it's actually dmamapped.
Quoted 1 line
From previous discussion, no one likes add another variable to track the
Quoted 8 lines
unmappings. So I'm avoiding adding another variable.
Quoted 1 line
Actually, all memory attachments use
sg_table,ttm->sgis NULL? can be used asan indicator to see whether an attachment is already unmapped.
Quoted 1 line
That already unmapped will not be repeated.
I think that should work. I'd add the checks in
kfd_mem_dmaunmap_userptrandkfd_mem_dmaunmap_sg_bo, where we also setttm->sgto NULL. In fact, both those functions already have that check. So looks like it should handle the the restart correctly with your patch.
Quoted 7 lines
Yes, both
kfd_mem_dmaunmap_userptr()andkfd_mem_dmaunmap_sg_bo()have NULL check forttm->sg.And dmabuf also have this check in
amdgpu_ttm_backend_unbind(). So dmaunmap won't be repeated actually.Then the benefits of handling ERESTARTSYS is avoiding
amdgpu_bo_reserve().What do you think? It's worth avoiding reservation in this case?
I don't think it's worth the trouble. In fact, to avoid race conditions, you probably should take the reservation anyway before looking at ttm->sg.
Regards,
Felix
Quoted 18 lines
Regards,
LangQuoted 14 lines
Regards,
FelixQuoted 10 lines
Regards,
LangQuoted 7 lines
Regards,
FelixQuoted 3 lines
} mutex_unlock(&p->mutex);
[Public]
Quoted 206 lines
-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Monday, January 29, 2024 10:58 PM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Francis, David <[email protected]>
Subject: Re: [PATCH v2] drm/amdkfd: reserve the BO before validating itQuoted 51 lines · Lang
[AMD Official Use Only - General]
Quoted 49 lines
-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Saturday, January 27, 2024 3:22 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Francis, David <[email protected]>
Subject: Re: [PATCH v2] drm/amdkfd: reserve the BO before validating itQuoted 39 lines · Lang
[AMD Official Use Only - General]
Quoted 37 lines
-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Thursday, January 25, 2024 5:41 AM
To: Yu, Lang <[email protected]>; [email protected]
Cc: Francis, David <[email protected]>
Subject: Re: [PATCH v2] drm/amdkfd: reserve the BO before validating itQuoted 7 lines · Lang Yu
Fixes 410f08516e0f ("drm/amdkfd: Move dma unmapping after TLBflush")
v2:
Avoid unmapping attachment twice when ERESTARTSYS.[ 41.708711] WARNING: CPU: 0 PID: 1463 at
drivers/gpu/drm/ttm/ttm_bo.c:846ttm_bo_validate+0x146/0x1b0[ttm]Quoted 11 lines
[ 41.708989] Call Trace: [ 41.708992] <TASK> [ 41.708996] ? show_regs+0x6c/0x80 [ 41.709000] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709008] ? __warn+0x93/0x190 [ 41.709014] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709024] ? report_bug+0x1f9/0x210 [ 41.709035] ? handle_bug+0x46/0x80 [ 41.709041] ? exc_invalid_op+0x1d/0x80 [ 41.709048] ? asm_exc_invalid_op+0x1f/0x30 [ 41.709057] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80[amdgpu]
Quoted 2 lines
[ 41.709185] ? ttm_bo_validate+0x146/0x1b0 [ttm] [ 41.709197] ? amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x2c/0x80[amdgpu]
Quoted 3 lines
[ 41.709337] ? srso_alias_return_thunk+0x5/0x7f [ 41.709346] kfd_mem_dmaunmap_attachment+0x9e/0x1e0 [amdgpu] [ 41.709467] amdgpu_amdkfd_gpuvm_dmaunmap_mem+0x56/0x80[amdgpu]
Quoted 1 line
[ 41.709586] kfd_ioctl_unmap_memory_from_gpu+0x1b7/0x300[amdgpu]
Quoted 17 lines
Quoted 17 lines
Quoted 17 lines
Quoted 17 lines
Quoted 2 lines
[ 41.709710] kfd_ioctl+0x1ec/0x650 [amdgpu] [ 41.709822] ? __pfx_kfd_ioctl_unmap_memory_from_gpu+0x10/0x10[amdgpu]
Quoted 10 lines
[ 41.709945] ? srso_alias_return_thunk+0x5/0x7f [ 41.709949] ? tomoyo_file_ioctl+0x20/0x30 [ 41.709959] __x64_sys_ioctl+0x9c/0xd0 [ 41.709967] do_syscall_64+0x3f/0x90 [ 41.709973] entry_SYSCALL_64_after_hwframe+0x6e/0xd8Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 28+++++++++++++++++--
Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 4 ++- 3 files changed, 29 insertions(+), 5 deletions(-)Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int
Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0@@ -311,7 +311,7 @@ intamdgpu_amdkfd_gpuvm_map_memory_to_gpu(struct
amdgpu_device*adev,
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h int · line 311
Quoted 1 line
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0struct kgd_mem *mem, void
*drm_priv);Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +0
-0int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu( struct amdgpu_device *adev, struct kgd_mem *mem,void
*drm_priv);Quoted 1 line
-void amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct
kgd_mem*mem,void
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
*drm_priv);
+int amdgpu_amdkfd_gpuvm_dmaunmap_mem(structkgd_mem*mem,void
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
Quoted 12 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int
Quoted 11 lines
+*drm_priv); int amdgpu_amdkfd_gpuvm_sync_memory( struct amdgpu_device *adev, struct kgd_mem *mem, bool intr); int amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(struct kgd_mem*mem, diff --git
a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.cb/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.cindex6f3a4cb2a9ef..7a050d46fa4d100644drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0
-0@@ -2088,21 +2088,43 @@ intamdgpu_amdkfd_gpuvm_map_memory_to_gpu(
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_map_memory_to_gpu(
Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0
-1return ret; } -void amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem,void
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int · line 2092
Quoted 2 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1
-0*drm_priv) +int amdgpu_amdkfd_gpuvm_dmaunmap_mem(struct kgd_mem *mem,void
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
Quoted 21 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +10
-2+*drm_priv) {struct kfd_mem_attachment *entry;struct amdgpu_vm *vm; +bool reserved = false; +int ret =0 ; vm =drm_priv_to_vm (drm_priv);mutex_lock (&mem->lock);list_for_each_entry (entry, &mem->attachments, list) { -if (entry->bo_va->base.vm == vm) -kfd_mem_dmaunmap_attachment (mem, entry); +if (entry->bo_va->base.vm != vm) +continue ; +if (entry->type ==KFD_MEM_ATT_SHARED || + entry->type ==KFD_MEM_ATT_DMABUF ) +continue ; +if (!entry->bo_va->base.bo->tbo.ttm->sg) +continue ;Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
You're going to great lengths to avoid the reservation when it's not needed by
kfd_mem_dmaunmap_attachment. However, this feels a bit fragile. Any changes in the kfd_mem_dmaunmap_* functions could breakthis.
Quoted 51 lines
Quoted 51 lines
Quoted 40 lines
Quoted 40 lines
Quoted 28 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +15
-0+ +if (!reserved) { + ret =amdgpu_bo_reserve (mem->bo, true); +if (ret) + goto out; + reserved = true; + } + +kfd_mem_dmaunmap_attachment (mem, entry); } +if (reserved) +amdgpu_bo_unreserve (mem->bo); + +out:mutex_unlock (&mem->lock); + +return ret; } int amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-0@@ -1442,7 +1442,9 @@ static intQuoted 1 line
kfd_ioctl_unmap_memory_from_gpu(struct
file *filep,
Quoted 5 lines
kfd_flush_tlb(peer_pdd,
TLB_FLUSH_HEAVYWEIGHT);/* Remove dma mapping after tlb flush to avoid
IO_PAGE_FAULT*/
Quoted 3 lines
- amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem, peer_pdd-
drm_priv);
+ err = amdgpu_amdkfd_gpuvm_dmaunmap_mem(mem,
peer_pdd->drm_priv);Quoted 8 lines
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_unmap_memory_from_gpu(struct file *filep, · line 1447
Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2
-0+if (err) + goto sync_memory_failed;This handles the case that the system call got interrupted. But you're not handling the restart correctly. When the ioctl is restarted, you don't know how many dmaunmaps are already done. So you'd need to make sure that repeating the dmaunmap is safe in all cases. Or what David tried earlier, find a way to track the unmapping so you
only try to dmaunmap on GPUs where it's actually dmamapped.
Quoted 2 lines
From previous discussion, no one likes add another variable to
track theQuoted 7 lines
Quoted 7 lines
unmappings. So I'm avoiding adding another variable.
Quoted 2 lines
Actually, all memory attachments use
sg_table,ttm->sgis NULL? can be used asan indicator to see whether an attachment is already unmapped.
Quoted 1 line
That already unmapped will not be repeated.
I think that should work. I'd add the checks in
kfd_mem_dmaunmap_userptrandkfd_mem_dmaunmap_sg_bo, where wealso
Quoted 3 lines
Quoted 2 lines
set
ttm->sgto NULL. In fact, both those functions already have that check. So looks like it should handle the the restart correctly with your patch.Yes, both
kfd_mem_dmaunmap_userptr()andkfd_mem_dmaunmap_sg_bo()have NULL check for
ttm->sg.Quoted 2 lines
And dmabuf also have this check in
amdgpu_ttm_backend_unbind(). So
Quoted 8 lines
dmaunmap won't be repeated actually.
Quoted 4 lines
Then the benefits of handling ERESTARTSYS is avoiding
amdgpu_bo_reserve().What do you think? It's worth avoiding reservation in this case?
I don't think it's worth the trouble. In fact, to avoid race conditions, you probably should take the reservation anyway before looking at
ttm->sg.
Got it. Thanks.
Regards,
Lang
Quoted 22 lines
Regards,
FelixQuoted 18 lines
Regards,
LangQuoted 14 lines
Regards,
FelixQuoted 10 lines
Regards,
LangQuoted 7 lines
Regards,
FelixQuoted 3 lines
} mutex_unlock(&p->mutex);