@@ -321,9 +321,10 @@ validate_init(struct nouveau_channel *chan, struct drm_file *file_priv,
uint32_t sequence;
int trycnt = 0;
int ret, i;
+ struct nouveau_bo *res_bo = NULL;
-retry:
sequence = atomic_add_return(1, &drm->ttm.validate_sequence);
+retry:
if (++trycnt > 100000) {
NV_ERROR(drm, "%s failed and gave up.\n", __func__);
return -EINVAL;
@@ -341,6 +342,11 @@ retry:
return -ENOENT;
}
nvbo = gem->driver_private;
+ if (nvbo == res_bo) {
+ res_bo = NULL;
+ drm_gem_object_unreference_unlocked(gem);
+ continue;
+ }
if (nvbo->reserved_by && nvbo->reserved_by == file_priv) {
NV_ERROR(drm, "multiple instances of buffer %d on "
@@ -353,15 +359,19 @@ retry:
ret = ttm_bo_reserve(&nvbo->bo, true, false, true, sequence);
if (ret) {
validate_fini(op, NULL);
- if (unlikely(ret == -EAGAIN))
- ret = ttm_bo_wait_unreserved(&nvbo->bo, true);
- drm_gem_object_unreference_unlocked(gem);
+ if (unlikely(ret == -EAGAIN)) {
+ sequence = atomic_add_return(1, &drm->ttm.validate_sequence);
+ ret = ttm_bo_reserve_slowpath(&nvbo->bo, true,
+ sequence);
+ if (!ret)
+ res_bo = nvbo;
+ }
if (unlikely(ret)) {
+ drm_gem_object_unreference_unlocked(gem);
if (ret != -ERESTARTSYS)
NV_ERROR(drm, "fail reserve\n");
return ret;
}
- goto retry;
}
b->user_priv = (uint64_t)(unsigned long)nvbo;
@@ -383,6 +393,8 @@ retry:
validate_fini(op, NULL);
return -EINVAL;
}
+ if (nvbo == res_bo)
+ goto retry;
}
return 0;
Similar rationale to the identical commit in drm/ttm. Instead of only waiting for unreservation, we make sure we actually own the reservation, then retry to get the rest. Changes since v1: - Increase the seqno before calling ttm_bo_reserve_slowpath Signed-off-by: Maarten Lankhorst <maarten.lankhorst@canonical.com> --- drivers/gpu/drm/nouveau/nouveau_gem.c | 22 +++++++++++++++++----- 1 file changed, 17 insertions(+), 5 deletions(-)