[CI 2/7] drm/xe: Rework xe_vm_rebind()
Thomas Hellström
thomas.hellstrom at linux.intel.com
Tue Mar 12 15:34:54 UTC 2024
Rework xe_vm_rebind() to attach its resulting fence to the vm directly,
and have exec- and rebind worker only deal with the vm's rebind fence.
Also ensure that any new rebind fence completes *after* the previous
one so we don't lose dependencies. Use dma_fence_chain when
necessary.
Signed-off-by: Thomas Hellström <thomas.hellstrom at linux.intel.com>
---
drivers/gpu/drm/xe/xe_exec.c | 17 ++-------
drivers/gpu/drm/xe/xe_vm.c | 73 +++++++++++++++++++++++++++---------
drivers/gpu/drm/xe/xe_vm.h | 2 +-
3 files changed, 61 insertions(+), 31 deletions(-)
diff --git a/drivers/gpu/drm/xe/xe_exec.c b/drivers/gpu/drm/xe/xe_exec.c
index 952496c6260d..5124663a9694 100644
--- a/drivers/gpu/drm/xe/xe_exec.c
+++ b/drivers/gpu/drm/xe/xe_exec.c
@@ -152,7 +152,6 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
struct drm_exec *exec = &vm_exec.exec;
u32 i, num_syncs = 0, num_ufence = 0;
struct xe_sched_job *job;
- struct dma_fence *rebind_fence;
struct xe_vm *vm;
bool write_locked, skip_retry = false;
ktime_t end = 0;
@@ -289,20 +288,12 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
* Rebind any invalidated userptr or evicted BOs in the VM, non-compute
* VM mode only.
*/
- rebind_fence = xe_vm_rebind(vm, false);
- if (IS_ERR(rebind_fence)) {
- err = PTR_ERR(rebind_fence);
- goto err_put_job;
+ if (!xe_vm_in_lr_mode(vm)) {
+ err = xe_vm_rebind(vm, false);
+ if (err)
+ goto err_put_job;
}
- /*
- * We store the rebind_fence in the VM so subsequent execs don't get
- * scheduled before the rebinds of userptrs / evicted BOs is complete.
- */
- if (rebind_fence) {
- dma_fence_put(vm->rebind_fence);
- vm->rebind_fence = rebind_fence;
- }
if (vm->rebind_fence) {
if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
&vm->rebind_fence->flags)) {
diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
index 643b3701a738..18a8e1137016 100644
--- a/drivers/gpu/drm/xe/xe_vm.c
+++ b/drivers/gpu/drm/xe/xe_vm.c
@@ -15,6 +15,7 @@
#include <drm/xe_drm.h>
#include <linux/ascii85.h>
#include <linux/delay.h>
+#include <linux/dma-fence-chain.h>
#include <linux/kthread.h>
#include <linux/mm.h>
#include <linux/swap.h>
@@ -522,7 +523,6 @@ static void preempt_rebind_work_func(struct work_struct *w)
{
struct xe_vm *vm = container_of(w, struct xe_vm, preempt.rebind_work);
struct drm_exec exec;
- struct dma_fence *rebind_fence;
unsigned int fence_count = 0;
LIST_HEAD(preempt_fences);
ktime_t end = 0;
@@ -568,15 +568,14 @@ static void preempt_rebind_work_func(struct work_struct *w)
if (err)
goto out_unlock;
- rebind_fence = xe_vm_rebind(vm, true);
- if (IS_ERR(rebind_fence)) {
- err = PTR_ERR(rebind_fence);
+ err = xe_vm_rebind(vm, true);
+ if (err)
goto out_unlock;
- }
- if (rebind_fence) {
- dma_fence_wait(rebind_fence, false);
- dma_fence_put(rebind_fence);
+ if (vm->rebind_fence) {
+ dma_fence_wait(vm->rebind_fence, false);
+ dma_fence_put(vm->rebind_fence);
+ vm->rebind_fence = NULL;
}
/* Wait on munmap style VM unbinds */
@@ -756,32 +755,72 @@ xe_vm_bind_vma(struct xe_vma *vma, struct xe_exec_queue *q,
struct xe_sync_entry *syncs, u32 num_syncs,
bool first_op, bool last_op);
-struct dma_fence *xe_vm_rebind(struct xe_vm *vm, bool rebind_worker)
+int xe_vm_rebind(struct xe_vm *vm, bool rebind_worker)
{
struct dma_fence *fence = NULL;
struct xe_vma *vma, *next;
+ struct dma_fence_chain *chain = NULL;
lockdep_assert_held(&vm->lock);
- if (xe_vm_in_lr_mode(vm) && !rebind_worker)
- return NULL;
-
xe_vm_assert_held(vm);
+
list_for_each_entry_safe(vma, next, &vm->rebind_list,
combined_links.rebind) {
+ struct dma_fence *old = vm->rebind_fence;
+ bool old_signaled = !old || dma_fence_is_signaled(old);
+
xe_assert(vm->xe, vma->tile_present);
- list_del_init(&vma->combined_links.rebind);
- dma_fence_put(fence);
if (rebind_worker)
trace_xe_vma_rebind_worker(vma);
else
trace_xe_vma_rebind_exec(vma);
+
+ if (!old_signaled && !chain) {
+ chain = kzalloc(sizeof(*chain), GFP_NOWAIT | __GFP_ACCOUNT);
+ if (!chain) {
+ int err = dma_fence_wait(old, true);
+
+ if (err)
+ return err;
+ old_signaled = true;
+ }
+ }
+
+ if (old_signaled) {
+ dma_fence_put(old);
+ old = NULL;
+ vm->rebind_fence = NULL;
+ }
+
fence = xe_vm_bind_vma(vma, NULL, NULL, 0, false, false);
- if (IS_ERR(fence))
- return fence;
+ if (IS_ERR(fence)) {
+ kfree(chain);
+ return PTR_ERR(fence);
+ }
+
+ list_del_init(&vma->combined_links.rebind);
+ if (fence && !dma_fence_is_signaled(fence)) {
+ if (!old_signaled) {
+ if (fence->context == old->context) {
+ dma_fence_put(old);
+ vm->rebind_fence = fence;
+ } else {
+ dma_fence_chain_init(chain, old, fence, old->seqno + 1);
+ vm->rebind_fence = &chain->base;
+ chain = NULL;
+ }
+ fence = NULL;
+ } else {
+ vm->rebind_fence = fence;
+ fence = NULL;
+ }
+ }
+ dma_fence_put(fence);
}
+ kfree(chain);
- return fence;
+ return 0;
}
static void xe_vma_free(struct xe_vma *vma)
diff --git a/drivers/gpu/drm/xe/xe_vm.h b/drivers/gpu/drm/xe/xe_vm.h
index 6df1f1c7f85d..4853354336f2 100644
--- a/drivers/gpu/drm/xe/xe_vm.h
+++ b/drivers/gpu/drm/xe/xe_vm.h
@@ -207,7 +207,7 @@ int __xe_vm_userptr_needs_repin(struct xe_vm *vm);
int xe_vm_userptr_check_repin(struct xe_vm *vm);
-struct dma_fence *xe_vm_rebind(struct xe_vm *vm, bool rebind_worker);
+int xe_vm_rebind(struct xe_vm *vm, bool rebind_worker);
int xe_vm_invalidate_vma(struct xe_vma *vma);
--
2.44.0
More information about the Intel-xe
mailing list