| Message ID | 20241019192030.1505020-5-matthew.brost@intel.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | Fix non-contiguous VRAM BO access in Xe | expand |
On 19/10/2024 20:20, Matthew Brost wrote: > Non-contiguous mapping of BO in VRAM doesn't work, use ttm_bo_access > instead. > > v2: > - Fix error handling > > Fixes: 0eb2a18a8fad ("drm/xe: Implement VM snapshot support for BO's and userptr") > Suggested-by: Matthew Auld <matthew.auld@intel.com> > Signed-off-by: Matthew Brost <matthew.brost@intel.com> The other user looks to be the clear color stuff for display. See intel_bo_read_from_page(). I think that is also potentially busted? > --- > drivers/gpu/drm/xe/xe_vm.c | 17 ++++++----------- > 1 file changed, 6 insertions(+), 11 deletions(-) > > diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c > index c99380271de6..c8782da3a5c3 100644 > --- a/drivers/gpu/drm/xe/xe_vm.c > +++ b/drivers/gpu/drm/xe/xe_vm.c > @@ -3303,7 +3303,6 @@ void xe_vm_snapshot_capture_delayed(struct xe_vm_snapshot *snap) > > for (int i = 0; i < snap->num_snaps; i++) { > struct xe_bo *bo = snap->snap[i].bo; > - struct iosys_map src; > int err; > > if (IS_ERR(snap->snap[i].data)) > @@ -3316,16 +3315,12 @@ void xe_vm_snapshot_capture_delayed(struct xe_vm_snapshot *snap) > } > > if (bo) { > - xe_bo_lock(bo, false); I think we still need the lock, or is that grabbed somewhere else? Also I guess CI doesn't currently hit this path? > - err = ttm_bo_vmap(&bo->ttm, &src); > - if (!err) { > - xe_map_memcpy_from(xe_bo_device(bo), > - snap->snap[i].data, > - &src, snap->snap[i].bo_ofs, > - snap->snap[i].len); > - ttm_bo_vunmap(&bo->ttm, &src); > - } > - xe_bo_unlock(bo); > + err = ttm_bo_access(&bo->ttm, snap->snap[i].bo_ofs, > + snap->snap[i].data, snap->snap[i].len, 0); > + if (!(err < 0) && err != snap->snap[i].len) > + err = -EIO; > + else if (!(err < 0)) > + err = 0; > } else { > void __user *userptr = (void __user *)(size_t)snap->snap[i].bo_ofs; >
On Mon, Oct 21, 2024 at 10:07:31AM +0100, Matthew Auld wrote: > On 19/10/2024 20:20, Matthew Brost wrote: > > Non-contiguous mapping of BO in VRAM doesn't work, use ttm_bo_access > > instead. > > > > v2: > > - Fix error handling > > > > Fixes: 0eb2a18a8fad ("drm/xe: Implement VM snapshot support for BO's and userptr") > > Suggested-by: Matthew Auld <matthew.auld@intel.com> > > Signed-off-by: Matthew Brost <matthew.brost@intel.com> > > The other user looks to be the clear color stuff for display. See > intel_bo_read_from_page(). I think that is also potentially busted? > I looked at display but missed this. Will fix in next rev. > > --- > > drivers/gpu/drm/xe/xe_vm.c | 17 ++++++----------- > > 1 file changed, 6 insertions(+), 11 deletions(-) > > > > diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c > > index c99380271de6..c8782da3a5c3 100644 > > --- a/drivers/gpu/drm/xe/xe_vm.c > > +++ b/drivers/gpu/drm/xe/xe_vm.c > > @@ -3303,7 +3303,6 @@ void xe_vm_snapshot_capture_delayed(struct xe_vm_snapshot *snap) > > for (int i = 0; i < snap->num_snaps; i++) { > > struct xe_bo *bo = snap->snap[i].bo; > > - struct iosys_map src; > > int err; > > if (IS_ERR(snap->snap[i].data)) > > @@ -3316,16 +3315,12 @@ void xe_vm_snapshot_capture_delayed(struct xe_vm_snapshot *snap) > > } > > if (bo) { > > - xe_bo_lock(bo, false); > > I think we still need the lock, or is that grabbed somewhere else? Also I > guess CI doesn't currently hit this path? > Yep. ttm_bo_access does a ttm_bo_reserve / ttm_bo_unreserve which is a lock / unlock. Matt > > - err = ttm_bo_vmap(&bo->ttm, &src); > > - if (!err) { > > - xe_map_memcpy_from(xe_bo_device(bo), > > - snap->snap[i].data, > > - &src, snap->snap[i].bo_ofs, > > - snap->snap[i].len); > > - ttm_bo_vunmap(&bo->ttm, &src); > > - } > > - xe_bo_unlock(bo); > > + err = ttm_bo_access(&bo->ttm, snap->snap[i].bo_ofs, > > + snap->snap[i].data, snap->snap[i].len, 0); > > + if (!(err < 0) && err != snap->snap[i].len) > > + err = -EIO; > > + else if (!(err < 0)) > > + err = 0; > > } else { > > void __user *userptr = (void __user *)(size_t)snap->snap[i].bo_ofs;
diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c index c99380271de6..c8782da3a5c3 100644 --- a/drivers/gpu/drm/xe/xe_vm.c +++ b/drivers/gpu/drm/xe/xe_vm.c @@ -3303,7 +3303,6 @@ void xe_vm_snapshot_capture_delayed(struct xe_vm_snapshot *snap) for (int i = 0; i < snap->num_snaps; i++) { struct xe_bo *bo = snap->snap[i].bo; - struct iosys_map src; int err; if (IS_ERR(snap->snap[i].data)) @@ -3316,16 +3315,12 @@ void xe_vm_snapshot_capture_delayed(struct xe_vm_snapshot *snap) } if (bo) { - xe_bo_lock(bo, false); - err = ttm_bo_vmap(&bo->ttm, &src); - if (!err) { - xe_map_memcpy_from(xe_bo_device(bo), - snap->snap[i].data, - &src, snap->snap[i].bo_ofs, - snap->snap[i].len); - ttm_bo_vunmap(&bo->ttm, &src); - } - xe_bo_unlock(bo); + err = ttm_bo_access(&bo->ttm, snap->snap[i].bo_ofs, + snap->snap[i].data, snap->snap[i].len, 0); + if (!(err < 0) && err != snap->snap[i].len) + err = -EIO; + else if (!(err < 0)) + err = 0; } else { void __user *userptr = (void __user *)(size_t)snap->snap[i].bo_ofs;
Non-contiguous mapping of BO in VRAM doesn't work, use ttm_bo_access instead. v2: - Fix error handling Fixes: 0eb2a18a8fad ("drm/xe: Implement VM snapshot support for BO's and userptr") Suggested-by: Matthew Auld <matthew.auld@intel.com> Signed-off-by: Matthew Brost <matthew.brost@intel.com> --- drivers/gpu/drm/xe/xe_vm.c | 17 ++++++----------- 1 file changed, 6 insertions(+), 11 deletions(-)