drm/amdgpu: Update PDEs flush TLB if PDEs is moved
Subsystems (2)
PATCH 1/1 drm/amdgpu: Update PDEs flush TLB if PDEs is moved
Update PDEs, PTEs don't need flush TLB after updating mapping, this will remove the unnecessary TLB flush to reduce map to GPUs time.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4 -3
@@ -755,6 +755,10 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev, goto error;list_for_each_entry (entry, &vm->relocated, vm_status) { +/* vm_flush_needed after updating moved PDEs */ +if (entry->moved) +atomic64_inc (&vm->tlb_seq); + r =amdgpu_vm_pde_update (¶ms, entry);if (r) goto error;
@@ -764,9 +768,6 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev,if (r) goto error; -/* vm_flush_needed after updating PDEs */ -atomic64_inc (&vm->tlb_seq); -while (!list_empty (&vm->relocated)) { entry = list_first_entry(&vm->relocated, struct amdgpu_vm_bo_base,
Signature · 1 line
2.35.1
Am 2022-06-01 um 19:12 schrieb Philip Yang:
Quoted 2 lines
Update PDEs, PTEs don't need flush TLB after updating mapping, this will remove the unnecessary TLB flush to reduce map to GPUs time.
This description is unclear. I think what this change does is, flush TLBs when existing PDEs are updated (because a PTB or PDB moved). But it avoids unnecessary TLB flushes when new PDBs or PTBs are added to the page table, which commonly happens when memory is mapped for the first time.
Regards,
Felix
Quoted 32 lines
Suggested-by Christian König <[email protected]>Signed-off-by Philip Yang <[email protected]>--- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4
-3@@ -755,6 +755,10 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev, goto error;list_for_each_entry (entry, &vm->relocated, vm_status) { +/* vm_flush_needed after updating moved PDEs */ +if (entry->moved) +atomic64_inc (&vm->tlb_seq); + r =amdgpu_vm_pde_update (¶ms, entry);if (r) goto error;@@ -764,9 +768,6 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev,if (r) goto error; -/* vm_flush_needed after updating PDEs */ -atomic64_inc (&vm->tlb_seq); -while (!list_empty (&vm->relocated)) { entry = list_first_entry(&vm->relocated, struct amdgpu_vm_bo_base,
Christian König
maintainerAm 02.06.22 um 01:19 schrieb Felix Kuehling:
Quoted 5 lines
Am 2022-06-01 um 19:12 schrieb Philip Yang:
Quoted 2 lines
Update PDEs, PTEs don't need flush TLB after updating mapping, this will remove the unnecessary TLB flush to reduce map to GPUs time.
This description is unclear. I think what this change does is, flush TLBs when existing PDEs are updated (because a PTB or PDB moved). But it avoids unnecessary TLB flushes when new PDBs or PTBs are added to the page table, which commonly happens when memory is mapped for the first time.
Yes, agree. Additional to that (I know reviewing my own suggestion) setting a local variable and then incrementing the atomic only once might be better because atomics are barriers on x86.
Regards,
Christian.
Quoted 37 lines
Regards,
FelixQuoted 32 lines
Suggested-by Christian König <[email protected]>Signed-off-by Philip Yang <[email protected]>---
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c| 7 ++++---1 file changed, 4 insertions(+), 3 deletions(-)drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4
-0@@ -755,6 +755,10 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev, goto error; list_for_each_entry(entry, &vm->relocated, vm_status) { + /* vm_flush_needed after updating moved PDEs */ + if (entry->moved) + atomic64_inc(&vm->tlb_seq); + r = amdgpu_vm_pde_update(¶ms, entry); if (r) goto error;@@ -764,9 +768,6 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev, if (r) goto error;- /*
vm_flush_neededafter updating PDEs */ -atomic64_inc(&vm->tlb_seq);
-
while (!list_empty(&vm->relocated)) { entry = list_first_entry(&vm->relocated, structamdgpu_vm_bo_base,