[PATCH v3 10/19] drm/xe: Implement madvise ioctl for xe
Matthew Brost
matthew.brost at intel.com
Fri May 30 06:36:13 UTC 2025
On Thu, May 29, 2025 at 03:43:40PM -0700, Matthew Brost wrote:
> On Tue, May 27, 2025 at 10:09:54PM +0530, Himal Prasad Ghimiray wrote:
> > This driver-specific ioctl enables UMDs to control the memory attributes
> > for GPU VMAs within a specified input range. If the start or end
> > addresses fall within an existing VMA, the VMA is split accordingly. The
> > attributes of the VMA are modified as provided by the users. The old
> > mappings of the VMAs are invalidated, and TLB invalidation is performed
> > if necessary.
> >
> > v2(Matthew brost)
> > - xe_vm_in_fault_mode can't be enabled by Mesa, hence allow ioctl in non
> > fault mode too
> > - fix tlb invalidation skip for same ranges in multiple op
> > - use helper for tlb invalidation
> > - use xe_svm_notifier_lock/unlock helper
> > - s/lockdep_assert_held/lockdep_assert_held_write
> > - Add kernel-doc
> >
> > Signed-off-by: Himal Prasad Ghimiray <himal.prasad.ghimiray at intel.com>
> > ---
> > drivers/gpu/drm/xe/Makefile | 1 +
> > drivers/gpu/drm/xe/xe_device.c | 2 +
> > drivers/gpu/drm/xe/xe_vm_madvise.c | 264 +++++++++++++++++++++++++++++
> > drivers/gpu/drm/xe/xe_vm_madvise.h | 15 ++
> > 4 files changed, 282 insertions(+)
> > create mode 100644 drivers/gpu/drm/xe/xe_vm_madvise.c
> > create mode 100644 drivers/gpu/drm/xe/xe_vm_madvise.h
> >
> > diff --git a/drivers/gpu/drm/xe/Makefile b/drivers/gpu/drm/xe/Makefile
> > index c5d6681645ed..dc64bdcddfdc 100644
> > --- a/drivers/gpu/drm/xe/Makefile
> > +++ b/drivers/gpu/drm/xe/Makefile
> > @@ -117,6 +117,7 @@ xe-y += xe_bb.o \
> > xe_uc.o \
> > xe_uc_fw.o \
> > xe_vm.o \
> > + xe_vm_madvise.o \
> > xe_vram.o \
> > xe_vram_freq.o \
> > xe_vsec.o \
> > diff --git a/drivers/gpu/drm/xe/xe_device.c b/drivers/gpu/drm/xe/xe_device.c
> > index d4b6e623aa48..b9791c614749 100644
> > --- a/drivers/gpu/drm/xe/xe_device.c
> > +++ b/drivers/gpu/drm/xe/xe_device.c
> > @@ -61,6 +61,7 @@
> > #include "xe_ttm_stolen_mgr.h"
> > #include "xe_ttm_sys_mgr.h"
> > #include "xe_vm.h"
> > +#include "xe_vm_madvise.h"
> > #include "xe_vram.h"
> > #include "xe_vsec.h"
> > #include "xe_wait_user_fence.h"
> > @@ -197,6 +198,7 @@ static const struct drm_ioctl_desc xe_ioctls[] = {
> > DRM_IOCTL_DEF_DRV(XE_WAIT_USER_FENCE, xe_wait_user_fence_ioctl,
> > DRM_RENDER_ALLOW),
> > DRM_IOCTL_DEF_DRV(XE_OBSERVATION, xe_observation_ioctl, DRM_RENDER_ALLOW),
> > + DRM_IOCTL_DEF_DRV(XE_MADVISE, xe_vm_madvise_ioctl, DRM_RENDER_ALLOW),
> > };
> >
> > static long xe_drm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
> > diff --git a/drivers/gpu/drm/xe/xe_vm_madvise.c b/drivers/gpu/drm/xe/xe_vm_madvise.c
> > new file mode 100644
> > index 000000000000..f7edefe5f6cf
> > --- /dev/null
> > +++ b/drivers/gpu/drm/xe/xe_vm_madvise.c
> > @@ -0,0 +1,264 @@
> > +// SPDX-License-Identifier: MIT
> > +/*
> > + * Copyright © 2024 Intel Corporation
> > + */
> > +
> > +#include "xe_vm_madvise.h"
> > +
> > +#include <linux/nospec.h>
> > +#include <drm/ttm/ttm_tt.h>
> > +#include <drm/xe_drm.h>
> > +
> > +#include "xe_bo.h"
> > +#include "xe_gt_tlb_invalidation.h"
> > +#include "xe_pt.h"
> > +#include "xe_svm.h"
> > +
> > +static struct xe_vma **get_vmas(struct xe_vm *vm, int *num_vmas,
> > + u64 addr, u64 range)
> > +{
> > + struct xe_vma **vmas, **__vmas;
> > + struct drm_gpuva *gpuva;
> > + int max_vmas = 8;
> > +
> > + lockdep_assert_held(&vm->lock);
> > +
> > + *num_vmas = 0;
> > + vmas = kmalloc_array(max_vmas, sizeof(*vmas), GFP_KERNEL);
> > + if (!vmas)
> > + return NULL;
> > +
> > + vm_dbg(&vm->xe->drm, "VMA's in range: start=0x%016llx, end=0x%016llx", addr, addr + range);
> > +
> > + drm_gpuvm_for_each_va_range(gpuva, &vm->gpuvm, addr, addr + range) {
> > + struct xe_vma *vma = gpuva_to_vma(gpuva);
> > +
> > + if (*num_vmas == max_vmas) {
> > + max_vmas <<= 1;
> > + __vmas = krealloc(vmas, max_vmas * sizeof(*vmas), GFP_KERNEL);
> > + if (!__vmas) {
> > + kfree(vmas);
> > + return NULL;
> > + }
> > + vmas = __vmas;
> > + }
> > +
> > + vmas[*num_vmas] = vma;
> > + (*num_vmas)++;
> > + }
> > +
> > + vm_dbg(&vm->xe->drm, "*num_vmas = %d\n", *num_vmas);
> > +
> > + if (!*num_vmas) {
> > + kfree(vmas);
> > + return NULL;
> > + }
> > +
> > + return vmas;
> > +}
> > +
> > +static int madvise_preferred_mem_loc(struct xe_device *xe, struct xe_vm *vm,
> > + struct xe_vma **vmas, int num_vmas,
> > + struct drm_xe_madvise_ops ops)
> > +{
> > + /* Implementation pending */
> > + return 0;
> > +}
> > +
> > +static int madvise_atomic(struct xe_device *xe, struct xe_vm *vm,
> > + struct xe_vma **vmas, int num_vmas,
> > + struct drm_xe_madvise_ops ops)
> > +{
> > + /* Implementation pending */
> > + return 0;
> > +}
> > +
> > +static int madvise_pat_index(struct xe_device *xe, struct xe_vm *vm,
> > + struct xe_vma **vmas, int num_vmas,
> > + struct drm_xe_madvise_ops ops)
> > +{
> > + /* Implementation pending */
> > + return 0;
> > +}
> > +
> > +static int madvise_purgeable_state(struct xe_device *xe, struct xe_vm *vm,
> > + struct xe_vma **vmas, int num_vmas,
> > + struct drm_xe_madvise_ops ops)
> > +{
> > + /* Implementation pending */
> > + return 0;
> > +}
> > +
> > +typedef int (*madvise_func)(struct xe_device *xe, struct xe_vm *vm,
> > + struct xe_vma **vmas, int num_vmas, struct drm_xe_madvise_ops ops);
> > +
>
> See my latest replies in patch #19, if possible, making these functions
> so they can't fail would be best.
>
> > +static const madvise_func madvise_funcs[] = {
> > + [DRM_XE_VMA_ATTR_PREFERRED_LOC] = madvise_preferred_mem_loc,
> > + [DRM_XE_VMA_ATTR_ATOMIC] = madvise_atomic,
> > + [DRM_XE_VMA_ATTR_PAT] = madvise_pat_index,
> > + [DRM_XE_VMA_ATTR_PURGEABLE_STATE] = madvise_purgeable_state,
> > +};
> > +
> > +static void xe_zap_ptes_in_madvise_range(struct xe_vm *vm, u64 start, u64 end, u8 *tile_mask)
> > +{
> > + struct drm_gpuva *gpuva;
> > + struct xe_tile *tile;
> > + u8 id;
> > +
> > + lockdep_assert_held_write(&vm->lock);
> > +
> > + if (dma_resv_wait_timeout(xe_vm_resv(vm), DMA_RESV_USAGE_BOOKKEEP,
> > + false, MAX_SCHEDULE_TIMEOUT) <= 0)
> > + XE_WARN_ON(1);
> > +
> > + *tile_mask = xe_svm_ranges_zap_ptes_in_range(vm, start, end);
> > +
> > + drm_gpuvm_for_each_va_range(gpuva, &vm->gpuvm, start, end) {
> > + struct xe_vma *vma = gpuva_to_vma(gpuva);
> > +
> > + if (xe_vma_is_cpu_addr_mirror(vma))
> > + continue;
> > +
> > + if (xe_vma_is_userptr(vma)) {
> > + WARN_ON_ONCE(!mmu_interval_check_retry
> > + (&to_userptr_vma(vma)->userptr.notifier,
> > + to_userptr_vma(vma)->userptr.notifier_seq));
> > +
> > + WARN_ON_ONCE(!dma_resv_test_signaled(xe_vm_resv(xe_vma_vm(vma)),
> > + DMA_RESV_USAGE_BOOKKEEP));
> > + }
> > +
>
> I think the similar code in xe_vm_invalidate_vma is a bit stale and
> contains remnants from when we issued the userptr TLB invalidation
> outside of the notifier lock in the MMU notifier. There, it should be:
>
> if (xe_vma_is_userptr())
> lockdep_assert(userptr notifier lock held);
> extra sanity checks
>
> So with that, you need the userptr notifier lock in read mode here.
>
> Also, I think the first WARN_ON is likely to always trigger here, as the
> CPU pages are likely valid, so that one should be removed.
>
> In xe_vm_invalidate_vma, the extra tests are protected by the
> PROVE_LOCKING Kconfig option. So if you keep the dma-resv check, I’d
> recommend hiding it behind PROVE_LOCKING.
>
> Note: speaking of the userptr notifier lock, Matt Auld has a patch
> series [1] to unify this with the GPU SVM notifier lock. If that lands
> before your changes, then you should use the SVM notifier lock in read
> mode for userptr here.
>
> [1] https://patchwork.freedesktop.org/series/146553/
>
> > + if (xe_vma_bo(vma))
> > + xe_bo_lock(xe_vma_bo(vma), false);
> > +
>
> Ah yes, you do need the BO lock here to prevent races from BO moves
> issuing a zap. Maybe we can update the xe_pt_zap_ptes lockdep assertions
> for userptr/BOs, along with some comments? We could do that
> independently of this series. I suggest this because I had forgotten how
> this worked in the previous revision. Maybe I can post this
> independently.
>
I posted a patch for this [2] and also realized a few more things. These
locks need to be held for an extended period, as detailed here [3].
Additionally, you don’t set vma->tile_invalidated here — you’ll need to
do that. Make sure to use WRITE_ONCE when doing so, and include a pairs
with comment like in [2] and in the existing SVM code for the
tile_invalidated / tile_present ranges.
Matt
[2] https://patchwork.freedesktop.org/series/149676/
[3] https://patchwork.freedesktop.org/patch/655898/?series=149550&rev=1#comment_1201162
> > + for_each_tile(tile, vm->xe, id) {
> > + if (xe_pt_zap_ptes(tile, vma))
> > + *tile_mask |= BIT(id);
> > + }
> > +
> > + if (xe_vma_bo(vma))
> > + xe_bo_unlock(xe_vma_bo(vma));
> > + }
> > +}
> > +
> > +static int xe_vm_invalidate_madvise_range(struct xe_vm *vm, u64 start, u64 end)
> > +{
> > + u8 tile_mask = 0;
> > +
> > + xe_zap_ptes_in_madvise_range(vm, start, end, &tile_mask);
> > + if (!tile_mask)
> > + return 0;
> > +
> > + xe_device_wmb(vm->xe);
> > +
> > + return xe_vm_range_tilemask_tlb_invalidation(vm, start, end, tile_mask);
> > +}
> > +
> > +static int input_ranges_same(struct drm_xe_madvise_ops *old,
> > + struct drm_xe_madvise_ops *new)
> > +{
> > + return (new->start == old->start && new->range == old->range);
> > +}
> > +
> > +/**
> > + * xe_vm_madvise_ioctl - Handle MADVise ioctl for a VM
> > + * @dev: DRM device pointer
> > + * @data: Pointer to ioctl data (drm_xe_madvise*)
> > + * @file: DRM file pointer
> > + *
> > + * Handles the MADVISE ioctl to provide memory advice for vma's within
> > + * input range.
> > + *
> > + * Return: 0 on success or a negative error code on failure.
> > + */
> > +int xe_vm_madvise_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
> > +{
> > + struct xe_device *xe = to_xe_device(dev);
> > + struct xe_file *xef = to_xe_file(file);
> > + struct drm_xe_madvise_ops *advs_ops;
> > + struct drm_xe_madvise *args = data;
> > + struct xe_vm *vm;
> > + struct xe_vma **vmas = NULL;
> > + int num_vmas, err = 0;
> > + int i, j, attr_type;
> > + bool needs_invalidation;
> > +
> > + if (XE_IOCTL_DBG(xe, args->num_ops < 1))
> > + return -EINVAL;
> > +
> > + vm = xe_vm_lookup(xef, args->vm_id);
> > + if (XE_IOCTL_DBG(xe, !vm))
> > + return -EINVAL;
> > +
> > + down_write(&vm->lock);
> > +
> > + if (XE_IOCTL_DBG(xe, xe_vm_is_closed_or_banned(vm))) {
> > + err = -ENOENT;
> > + goto unlock_vm;
> > + }
> > +
> > + if (args->num_ops > 1) {
> > + u64 __user *madvise_user = u64_to_user_ptr(args->vector_of_ops);
> > +
> > + advs_ops = kvmalloc_array(args->num_ops, sizeof(struct drm_xe_madvise_ops),
> > + GFP_KERNEL | __GFP_ACCOUNT |
> > + __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
> > + if (!advs_ops) {
> > + err = args->num_ops > 1 ? -ENOBUFS : -ENOMEM;
> > + goto unlock_vm;
> > + }
> > +
> > + err = __copy_from_user(advs_ops, madvise_user,
> > + sizeof(struct drm_xe_madvise_ops) *
> > + args->num_ops);
> > + if (XE_IOCTL_DBG(xe, err)) {
> > + err = -EFAULT;
> > + goto free_advs_ops;
> > + }
> > + } else {
> > + advs_ops = &args->ops;
> > + }
> > +
>
> See my reply in patch #19, I think we should validate user input ahead
> of the below loop rather than failing mid-loop on bad user input.
>
> > + for (i = 0; i < args->num_ops; i++) {
> > + xe_vm_alloc_madvise_vma(vm, advs_ops[i].start, advs_ops[i].range);
> > +
> > + vmas = get_vmas(vm, &num_vmas, advs_ops[i].start, advs_ops[i].range);
> > + if (!vmas) {
> > + err = -ENOMEM;
> > + goto free_advs_ops;
> > + }
> > +
> > + attr_type = array_index_nospec(advs_ops[i].type, ARRAY_SIZE(madvise_funcs));
> > + err = madvise_funcs[attr_type](xe, vm, vmas, num_vmas, advs_ops[i]);
> > +
> > + kfree(vmas);
> > + vmas = NULL;
> > +
> > + if (err)
> > + goto free_advs_ops;
> > + }
> > +
> > + for (i = 0; i < args->num_ops; i++) {
> > + needs_invalidation = true;
> > + for (j = i + 1; j < args->num_ops; ++j) {
> > + if (input_ranges_same(&advs_ops[j], &advs_ops[i])) {
> > + needs_invalidation = false;
> > + break;
> > + }
> > + }
>
> I'd drop this extra check. The invalidation code already short ciruits
> on the tile_present / tile_invalidated bits in the range or VMA so I
> don't think an extra short circuit here buys us a ton.
>
> Matt
>
> > + if (needs_invalidation) {
> > + err = xe_vm_invalidate_madvise_range(vm, advs_ops[i].start,
> > + advs_ops[i].start + advs_ops[i].range);
> > + if (err)
> > + goto free_advs_ops;
> > + }
> > + }
> > +
> > +free_advs_ops:
> > + if (args->num_ops > 1)
> > + kvfree(advs_ops);
> > +unlock_vm:
> > + up_write(&vm->lock);
> > + xe_vm_put(vm);
> > + return err;
> > +}
> > diff --git a/drivers/gpu/drm/xe/xe_vm_madvise.h b/drivers/gpu/drm/xe/xe_vm_madvise.h
> > new file mode 100644
> > index 000000000000..c5cdd058c322
> > --- /dev/null
> > +++ b/drivers/gpu/drm/xe/xe_vm_madvise.h
> > @@ -0,0 +1,15 @@
> > +/* SPDX-License-Identifier: MIT */
> > +/*
> > + * Copyright © 2024 Intel Corporation
> > + */
> > +
> > +#ifndef _XE_VM_MADVISE_H_
> > +#define _XE_VM_MADVISE_H_
> > +
> > +struct drm_device;
> > +struct drm_file;
> > +
> > +int xe_vm_madvise_ioctl(struct drm_device *dev, void *data,
> > + struct drm_file *file);
> > +
> > +#endif
> > --
> > 2.34.1
> >
More information about the Intel-xe
mailing list