On Thu, 2025-12-18 at 10:33 -0800, Matthew Brost wrote:
> On Thu, Dec 18, 2025 at 05:20:40PM +0100, Thomas Hellström wrote:
> > In situations where no system memory is migrated to devmem, and in
> > upcoming patches where another GPU is performing the migration to
> > the newly allocated devmem buffer, there is nothing to ensure any
> > ongoing clear to the devmem allocation or async eviction from the
> > devmem allocation is complete.
> >
> > Address that by passing a struct dma_fence down to the copy
> > functions, and ensure it is waited for before migration is marked
> > complete.
> >
> > v3:
> > - New patch.
> > v4:
> > - Update the logic used for determining when to wait for the
> > pre_migrate_fence.
> > - Update the logic used for determining when to warn for the
> > pre_migrate_fence since the scheduler fences apparently
> > can signal out-of-order.
> > v5:
> > - Fix a UAF (CI)
> > - Remove references to source P2P migration (Himal)
> > - Put the pre_migrate_fence after migration.
> >
> > Fixes: c5b3eb5a906c ("drm/xe: Add GPUSVM device memory copy vfunc
> > functions")
> > Cc: Matthew Brost <[email protected]>
> > Cc: <[email protected]> # v6.15+
> > Signed-off-by: Thomas Hellström <[email protected]>
> > ---
> > drivers/gpu/drm/drm_pagemap.c | 17 ++++++---
> > drivers/gpu/drm/xe/xe_svm.c | 65 ++++++++++++++++++++++++++++++-
> > ----
> > include/drm/drm_pagemap.h | 17 +++++++--
> > 3 files changed, 83 insertions(+), 16 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/drm_pagemap.c
> > b/drivers/gpu/drm/drm_pagemap.c
> > index 4cf8f54e5a27..ac3832f85190 100644
> > --- a/drivers/gpu/drm/drm_pagemap.c
> > +++ b/drivers/gpu/drm/drm_pagemap.c
> > @@ -3,6 +3,7 @@
> > * Copyright © 2024-2025 Intel Corporation
> > */
> >
> > +#include <linux/dma-fence.h>
> > #include <linux/dma-mapping.h>
> > #include <linux/migrate.h>
> > #include <linux/pagemap.h>
> > @@ -408,10 +409,14 @@ int drm_pagemap_migrate_to_devmem(struct
> > drm_pagemap_devmem *devmem_allocation,
> > drm_pagemap_get_devmem_page(page, zdd);
> > }
> >
> > - err = ops->copy_to_devmem(pages, pagemap_addr, npages);
> > + err = ops->copy_to_devmem(pages, pagemap_addr, npages,
> > + devmem_allocation-
> > >pre_migrate_fence);
> > if (err)
> > goto err_finalize;
> >
> > + dma_fence_put(devmem_allocation->pre_migrate_fence);
> > + devmem_allocation->pre_migrate_fence = NULL;
> > +
> > /* Upon success bind devmem allocation to range and zdd */
> > devmem_allocation->timeslice_expiration = get_jiffies_64()
> > +
> > msecs_to_jiffies(timeslice_ms);
> > @@ -596,7 +601,7 @@ int drm_pagemap_evict_to_ram(struct
> > drm_pagemap_devmem *devmem_allocation)
> > for (i = 0; i < npages; ++i)
> > pages[i] = migrate_pfn_to_page(src[i]);
> >
> > - err = ops->copy_to_ram(pages, pagemap_addr, npages);
> > + err = ops->copy_to_ram(pages, pagemap_addr, npages, NULL);
> > if (err)
> > goto err_finalize;
> >
> > @@ -719,7 +724,7 @@ static int __drm_pagemap_migrate_to_ram(struct
> > vm_area_struct *vas,
> > for (i = 0; i < npages; ++i)
> > pages[i] = migrate_pfn_to_page(migrate.src[i]);
> >
> > - err = ops->copy_to_ram(pages, pagemap_addr, npages);
> > + err = ops->copy_to_ram(pages, pagemap_addr, npages, NULL);
> > if (err)
> > goto err_finalize;
> >
> > @@ -800,11 +805,14 @@
> > EXPORT_SYMBOL_GPL(drm_pagemap_pagemap_ops_get);
> > * @ops: Pointer to the operations structure for GPU SVM device
> > memory
> > * @dpagemap: The struct drm_pagemap we're allocating from.
> > * @size: Size of device memory allocation
> > + * @pre_migrate_fence: Fence to wait for or pipeline behind before
> > migration starts.
> > + * (May be NULL).
> > */
> > void drm_pagemap_devmem_init(struct drm_pagemap_devmem
> > *devmem_allocation,
> > struct device *dev, struct mm_struct
> > *mm,
> > const struct drm_pagemap_devmem_ops
> > *ops,
> > - struct drm_pagemap *dpagemap, size_t
> > size)
> > + struct drm_pagemap *dpagemap, size_t
> > size,
> > + struct dma_fence *pre_migrate_fence)
> > {
> > init_completion(&devmem_allocation->detached);
> > devmem_allocation->dev = dev;
> > @@ -812,6 +820,7 @@ void drm_pagemap_devmem_init(struct
> > drm_pagemap_devmem *devmem_allocation,
> > devmem_allocation->ops = ops;
> > devmem_allocation->dpagemap = dpagemap;
> > devmem_allocation->size = size;
> > + devmem_allocation->pre_migrate_fence = pre_migrate_fence;
> > }
> > EXPORT_SYMBOL_GPL(drm_pagemap_devmem_init);
> >
> > diff --git a/drivers/gpu/drm/xe/xe_svm.c
> > b/drivers/gpu/drm/xe/xe_svm.c
> > index bab8e6cbe53d..b806a1fce188 100644
> > --- a/drivers/gpu/drm/xe/xe_svm.c
> > +++ b/drivers/gpu/drm/xe/xe_svm.c
> > @@ -472,11 +472,12 @@ static void xe_svm_copy_us_stats_incr(struct
> > xe_gt *gt,
> >
> > static int xe_svm_copy(struct page **pages,
> > struct drm_pagemap_addr *pagemap_addr,
> > - unsigned long npages, const enum
> > xe_svm_copy_dir dir)
> > + unsigned long npages, const enum
> > xe_svm_copy_dir dir,
> > + struct dma_fence *pre_migrate_fence)
> > {
> > struct xe_vram_region *vr = NULL;
> > struct xe_gt *gt = NULL;
> > - struct xe_device *xe;
> > + struct xe_device *xe = NULL;
> > struct dma_fence *fence = NULL;
> > unsigned long i;
> > #define XE_VRAM_ADDR_INVALID ~0x0ull
> > @@ -485,6 +486,16 @@ static int xe_svm_copy(struct page **pages,
> > bool sram = dir == XE_SVM_COPY_TO_SRAM;
> > ktime_t start = xe_gt_stats_ktime_get();
> >
> > + if (pre_migrate_fence &&
> > dma_fence_is_container(pre_migrate_fence)) {
> > + /*
> > + * This would typically be a composite fence
> > operation on the destination memory.
> > + * Ensure that the other GPU operation on the
> > destination is complete.
> > + */
> > + err = dma_fence_wait(pre_migrate_fence, true);
> > + if (err)
> > + return err;
> > + }
> > +
>
> I'm not fully convienced this code works. Consider the case where we
> allocate memory a device A and we copying from device B. In this case
> device A issues the clear but device B issues the copy. The
> pre_migrate_fence is not going be a container and I believe our
> ordering
> breaks.
So the logic to handle that case was moved to the patch that enables
source migration, as per Himal's review comment. So consider this patch
only for destination migration where the devmem allocation is allocated
on the same gpu that migrates.
>
> Would it be simplier to pass in the pre_migrate_fence fence a
> dependency
> to the first copy job issued, then set it to NULL. The drm scheduler
> is
> smart enough to squash the input fence into a NOP if a copy / clear
> are
> from the same devices queues.
I intended to do that as a follow-up if needed but since this is a fix
that fixes an existing problem I wanted it to be lightweight. But let
me take a quick look at that and probably resend only that patch.
/Thomas
