[PATCH 2/3] drm/xe: Clear scratch page on vm_bind
Zeng, Oak
oak.zeng at intel.com
Wed Feb 19 20:19:10 UTC 2025
> -----Original Message-----
> From: Brost, Matthew <matthew.brost at intel.com>
> Sent: February 19, 2025 12:47 PM
> To: Zeng, Oak <oak.zeng at intel.com>
> Cc: intel-xe at lists.freedesktop.org;
> Thomas.Hellstrom at linux.intel.com; Cavitt, Jonathan
> <jonathan.cavitt at intel.com>
> Subject: Re: [PATCH 2/3] drm/xe: Clear scratch page on vm_bind
>
> On Wed, Feb 12, 2025 at 09:23:30PM -0500, Oak Zeng wrote:
> > When a vm runs under fault mode, if scratch page is enabled, we
> need
> > to clear the scratch page mapping on vm_bind for the vm_bind
> address
> > range. Under fault mode, we depend on recoverable page fault to
> > establish mapping in page table. If scratch page is not cleared, GPU
> > access of address won't cause page fault because it always hits the
> > existing scratch page mapping.
> >
> > When vm_bind with IMMEDIATE flag, there is no need of clearing as
> > immediate bind can overwrite the scratch page mapping.
> >
> > So far only is xe2 and xe3 products are allowed to enable scratch
> page
> > under fault mode. On other platform we don't allow scratch page
> under
> > fault mode, so no need of such clearing.
> >
> > v2: Rework vm_bind pipeline to clear scratch page mapping. This is
> similar
> > to a map operation, with the exception that PTEs are cleared
> instead of
> > pointing to valid physical pages. (Matt, Thomas)
> >
> > TLB invalidation is needed after clear scratch page mapping as larger
> > scratch page mapping could be backed by physical page and cached
> in
> > TLB. (Matt, Thomas)
> >
> > v3: Fix the case of clearing huge pte (Thomas)
> >
> > Improve commit message (Thomas)
> >
> > v4: TLB invalidation on all LR cases, not only the clear on bind
> > cases (Thomas)
> >
> > Signed-off-by: Oak Zeng <oak.zeng at intel.com>
> > ---
> > drivers/gpu/drm/xe/xe_pt.c | 53
> ++++++++++++++++++++++++--------
> > drivers/gpu/drm/xe/xe_pt_types.h | 2 ++
> > drivers/gpu/drm/xe/xe_vm.c | 29 ++++++++++++++---
> > drivers/gpu/drm/xe/xe_vm_types.h | 2 ++
> > 4 files changed, 70 insertions(+), 16 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/xe/xe_pt.c
> b/drivers/gpu/drm/xe/xe_pt.c
> > index 1ddcc7e79a93..319769056893 100644
> > --- a/drivers/gpu/drm/xe/xe_pt.c
> > +++ b/drivers/gpu/drm/xe/xe_pt.c
> > @@ -268,6 +268,8 @@ struct xe_pt_stage_bind_walk {
> > * granularity.
> > */
> > bool needs_64K;
> > + /* @clear_pt: clear page table entries during the bind walk */
> > + bool clear_pt;
> > /**
> > * @vma: VMA being mapped
> > */
> > @@ -415,6 +417,10 @@ static bool xe_pt_hugepte_possible(u64
> addr, u64 next, unsigned int level,
> > if (xe_vma_is_null(xe_walk->vma))
> > return true;
> >
> > + /* if we are clearing page table, no dma addresses*/
> > + if (xe_walk->clear_pt)
> > + return true;
> > +
> > /* Is the DMA address huge PTE size aligned? */
> > size = next - addr;
> > dma = addr - xe_walk->va_curs_start +
> xe_res_dma(xe_walk->curs);
> > @@ -497,16 +503,19 @@ xe_pt_stage_bind_entry(struct xe_ptw
> *parent, pgoff_t offset,
> >
> > XE_WARN_ON(xe_walk->va_curs_start != addr);
> >
> > - pte = vm->pt_ops->pte_encode_vma(is_null ? 0 :
> > - xe_res_dma(curs) +
> xe_walk->dma_offset,
> > - xe_walk->vma,
> pat_index, level);
> > + pte = vm->pt_ops->pte_encode_vma(is_null ||
> xe_walk->clear_pt ?
> > + 0 : xe_res_dma(curs) + xe_walk-
> >dma_offset,
> > + xe_walk->vma, pat_index, level);
> > pte |= xe_walk->default_pte;
> >
> > + if (xe_walk->clear_pt)
> > + pte = 0;
>
> Is there a way to move this clear to later in the pipeline? e.g. At the
> step which actually programs the page tables via a GPU job or the CPU?
Isn't the design here generate all the page table entries in this stage_bind stage,
And actual programming of page table using cpu/gpu just focus on writing the
Page table entries to the actual page table? From this perspective, it is more
Natural to set the entries to 0 here for clear_pt case.
Of cause you can do it later, but it is at the cost of sacrifice the design.
> I
> ask as I think our userptr code is broken for invalidations which could
> hook into clearing to fix this.
Can you elaborate how the userptr invalidation is broken? Is it broken by this patch?
I don't follow here. As I understand it, the userptr invalidation doesn't go to this
Stage_bind function. It goes to xe_pt_zap_ptes. So I don't follow how userptr invalidate
Is broken...
Oak
Also I think prefetch for SVM could also
> make use of this too.
>
> I believe for this use case this patch is correct though.
>
> Matt
>
> > +
> > /*
> > * Set the XE_PTE_PS64 hint if possible, otherwise if
> > * this device *requires* 64K PTE size for VRAM, fail.
> > */
> > - if (level == 0 && !xe_parent->is_compact) {
> > + if (level == 0 && !xe_parent->is_compact
> && !xe_walk->clear_pt) {
> > if (xe_pt_is_pte_ps64K(addr, next, xe_walk))
> {
> > xe_walk->vma->gpuva.flags |=
> XE_VMA_PTE_64K;
> > pte |= XE_PTE_PS64;
> > @@ -519,7 +528,7 @@ xe_pt_stage_bind_entry(struct xe_ptw
> *parent, pgoff_t offset,
> > if (unlikely(ret))
> > return ret;
> >
> > - if (!is_null)
> > + if (!is_null && !xe_walk->clear_pt)
> > xe_res_next(curs, next - addr);
> > xe_walk->va_curs_start = next;
> > xe_walk->vma->gpuva.flags |= (XE_VMA_PTE_4K <<
> level);
> > @@ -589,6 +598,7 @@ static const struct xe_pt_walk_ops
> xe_pt_stage_bind_ops = {
> > * @vma: The vma indicating the address range.
> > * @entries: Storage for the update entries used for connecting the
> tree to
> > * the main tree at commit time.
> > + * @clear_pt: Clear the page table entries.
> > * @num_entries: On output contains the number of @entries
> used.
> > *
> > * This function builds a disconnected page-table tree for a given
> address
> > @@ -602,7 +612,8 @@ static const struct xe_pt_walk_ops
> xe_pt_stage_bind_ops = {
> > */
> > static int
> > xe_pt_stage_bind(struct xe_tile *tile, struct xe_vma *vma,
> > - struct xe_vm_pgtable_update *entries, u32
> *num_entries)
> > + struct xe_vm_pgtable_update *entries,
> > + bool clear_pt, u32 *num_entries)
> > {
> > struct xe_device *xe = tile_to_xe(tile);
> > struct xe_bo *bo = xe_vma_bo(vma);
> > @@ -622,10 +633,19 @@ xe_pt_stage_bind(struct xe_tile *tile,
> struct xe_vma *vma,
> > .vma = vma,
> > .wupd.entries = entries,
> > .needs_64K = (xe_vma_vm(vma)->flags &
> XE_VM_FLAG_64K) && is_devmem,
> > + .clear_pt = clear_pt,
> > };
> > struct xe_pt *pt = xe_vma_vm(vma)->pt_root[tile->id];
> > int ret;
> >
> > + if (clear_pt) {
> > + ret = xe_pt_walk_range(&pt->base, pt->level,
> xe_vma_start(vma),
> > + xe_vma_end(vma),
> &xe_walk.base);
> > +
> > + *num_entries = xe_walk.wupd.num_used_entries;
> > + return ret;
> > + }
> > +
> > /**
> > * Default atomic expectations for different allocation
> scenarios are as follows:
> > *
> > @@ -981,12 +1001,14 @@ static void xe_pt_free_bind(struct
> xe_vm_pgtable_update *entries,
> >
> > static int
> > xe_pt_prepare_bind(struct xe_tile *tile, struct xe_vma *vma,
> > - struct xe_vm_pgtable_update *entries, u32
> *num_entries)
> > + struct xe_vm_pgtable_update *entries,
> > + bool invalidate_on_bind, u32 *num_entries)
> > {
> > int err;
> >
> > *num_entries = 0;
> > - err = xe_pt_stage_bind(tile, vma, entries, num_entries);
> > + err = xe_pt_stage_bind(tile, vma, entries, invalidate_on_bind,
> > + num_entries);
> > if (!err)
> > xe_tile_assert(tile, *num_entries);
> >
> > @@ -1661,6 +1683,7 @@ static int bind_op_prepare(struct xe_vm
> *vm, struct xe_tile *tile,
> > return err;
> >
> > err = xe_pt_prepare_bind(tile, vma, pt_op->entries,
> > + pt_update_ops->invalidate_on_bind,
> > &pt_op->num_entries);
> > if (!err) {
> > xe_tile_assert(tile, pt_op->num_entries <=
> > @@ -1681,11 +1704,11 @@ static int bind_op_prepare(struct
> xe_vm *vm, struct xe_tile *tile,
> > * If !rebind, and scratch enabled VMs, there is a
> chance the scratch
> > * PTE is already cached in the TLB so it needs to be
> invalidated.
> > * On !LR VMs this is done in the ring ops preceding a
> batch, but on
> > - * non-faulting LR, in particular on user-space batch
> buffer chaining,
> > - * it needs to be done here.
> > + * LR, in particular on user-space batch buffer chaining,
> it needs to
> > + * be done here.
> > */
> > if ((!pt_op->rebind && xe_vm_has_scratch(vm) &&
> > - xe_vm_in_preempt_fence_mode(vm)))
> > + xe_vm_in_lr_mode(vm)))
> > pt_update_ops->needs_invalidation = true;
> > else if (pt_op->rebind && !xe_vm_in_lr_mode(vm))
> > /* We bump also if batch_invalidate_tlb is
> true */
> > @@ -1759,9 +1782,13 @@ static int op_prepare(struct xe_vm *vm,
> >
> > switch (op->base.op) {
> > case DRM_GPUVA_OP_MAP:
> > - if (!op->map.immediate &&
> xe_vm_in_fault_mode(vm))
> > + if (!op->map.immediate &&
> xe_vm_in_fault_mode(vm) &&
> > + !op->map.invalidate_on_bind)
> > break;
> >
> > + if (op->map.invalidate_on_bind)
> > + pt_update_ops->invalidate_on_bind = true;
> > +
> > err = bind_op_prepare(vm, tile, pt_update_ops, op-
> >map.vma);
> > pt_update_ops->wait_vm_kernel = true;
> > break;
> > @@ -1871,6 +1898,8 @@ static void bind_op_commit(struct xe_vm
> *vm, struct xe_tile *tile,
> > }
> > vma->tile_present |= BIT(tile->id);
> > vma->tile_staged &= ~BIT(tile->id);
> > + if (pt_update_ops->invalidate_on_bind)
> > + vma->tile_invalidated |= BIT(tile->id);
> > if (xe_vma_is_userptr(vma)) {
> > lockdep_assert_held_read(&vm-
> >userptr.notifier_lock);
> > to_userptr_vma(vma)->userptr.initial_bind = true;
> > diff --git a/drivers/gpu/drm/xe/xe_pt_types.h
> b/drivers/gpu/drm/xe/xe_pt_types.h
> > index 384cc04de719..3d0aa2a5102e 100644
> > --- a/drivers/gpu/drm/xe/xe_pt_types.h
> > +++ b/drivers/gpu/drm/xe/xe_pt_types.h
> > @@ -108,6 +108,8 @@ struct xe_vm_pgtable_update_ops {
> > bool needs_userptr_lock;
> > /** @needs_invalidation: Needs invalidation */
> > bool needs_invalidation;
> > + /** @invalidate_on_bind: Invalidate the range before bind */
> > + bool invalidate_on_bind;
> > /**
> > * @wait_vm_bookkeep: PT operations need to wait until VM
> is idle
> > * (bookkeep dma-resv slots are idle) and stage all future VM
> activity
> > diff --git a/drivers/gpu/drm/xe/xe_vm.c
> b/drivers/gpu/drm/xe/xe_vm.c
> > index d664f2e418b2..813d893d9b63 100644
> > --- a/drivers/gpu/drm/xe/xe_vm.c
> > +++ b/drivers/gpu/drm/xe/xe_vm.c
> > @@ -1921,6 +1921,23 @@ static void print_op(struct xe_device *xe,
> struct drm_gpuva_op *op)
> > }
> > #endif
> >
> > +static bool __xe_vm_needs_clear_scratch_pages(struct xe_vm
> *vm, u32 bind_flags)
> > +{
> > + if (!xe_vm_in_fault_mode(vm))
> > + return false;
> > +
> > + if (!NEEDS_SCRATCH(vm->xe))
> > + return false;
> > +
> > + if (!xe_vm_has_scratch(vm))
> > + return false;
> > +
> > + if (bind_flags & DRM_XE_VM_BIND_FLAG_IMMEDIATE)
> > + return false;
> > +
> > + return true;
> > +}
> > +
> > /*
> > * Create operations list from IOCTL arguments, setup operations
> fields so parse
> > * and commit steps are decoupled from IOCTL arguments. This
> step can fail.
> > @@ -1991,6 +2008,8 @@ vm_bind_ioctl_ops_create(struct xe_vm
> *vm, struct xe_bo *bo,
> > op->map.is_null = flags &
> DRM_XE_VM_BIND_FLAG_NULL;
> > op->map.dumpable = flags &
> DRM_XE_VM_BIND_FLAG_DUMPABLE;
> > op->map.pat_index = pat_index;
> > + op->map.invalidate_on_bind =
> > +
> __xe_vm_needs_clear_scratch_pages(vm, flags);
> > } else if (__op->op == DRM_GPUVA_OP_PREFETCH) {
> > op->prefetch.region = prefetch_region;
> > }
> > @@ -2188,7 +2207,8 @@ static int vm_bind_ioctl_ops_parse(struct
> xe_vm *vm, struct drm_gpuva_ops *ops,
> > return PTR_ERR(vma);
> >
> > op->map.vma = vma;
> > - if (op->map.immediate
> || !xe_vm_in_fault_mode(vm))
> > + if (op->map.immediate
> || !xe_vm_in_fault_mode(vm) ||
> > + op->map.invalidate_on_bind)
> >
> xe_vma_ops_incr_pt_update_ops(vops,
> > op-
> >tile_mask);
> > break;
> > @@ -2416,9 +2436,10 @@ static int op_lock_and_prep(struct
> drm_exec *exec, struct xe_vm *vm,
> >
> > switch (op->base.op) {
> > case DRM_GPUVA_OP_MAP:
> > - err = vma_lock_and_validate(exec, op->map.vma,
> > - !xe_vm_in_fault_mode(vm)
> ||
> > - op->map.immediate);
> > + if (!op->map.invalidate_on_bind)
> > + err = vma_lock_and_validate(exec, op-
> >map.vma,
> > +
> !xe_vm_in_fault_mode(vm) ||
> > + op-
> >map.immediate);
> > break;
> > case DRM_GPUVA_OP_REMAP:
> > err = check_ufence(gpuva_to_vma(op-
> >base.remap.unmap->va));
> > diff --git a/drivers/gpu/drm/xe/xe_vm_types.h
> b/drivers/gpu/drm/xe/xe_vm_types.h
> > index 52467b9b5348..dace04f4ea5e 100644
> > --- a/drivers/gpu/drm/xe/xe_vm_types.h
> > +++ b/drivers/gpu/drm/xe/xe_vm_types.h
> > @@ -297,6 +297,8 @@ struct xe_vma_op_map {
> > bool is_null;
> > /** @dumpable: whether BO is dumped on GPU hang */
> > bool dumpable;
> > + /** @invalidate: invalidate the VMA before bind */
> > + bool invalidate_on_bind;
> > /** @pat_index: The pat index to use for this operation. */
> > u16 pat_index;
> > };
> > --
> > 2.26.3
> >
More information about the Intel-xe
mailing list