On Thu, Apr 24, 2025 at 01:18:34PM +0100, Matthew Auld wrote:
> Goal here is cut over to gpusvm and remove xe_hmm, relying instead on
> common code. The core facilities we need are get_pages(), unmap_pages()
> and free_pages() for a given useptr range, plus a vm level notifier
> lock, which is now provided by gpusvm.
>
> v2:
> - Reuse the same SVM vm struct we use for full SVM, that way we can
> use the same lock (Matt B & Himal)
> v3:
> - Re-use svm_init/fini for userptr.
>
> Signed-off-by: Matthew Auld <matthew.a...@intel.com>
> Cc: Himal Prasad Ghimiray <himal.prasad.ghimi...@intel.com>
> Cc: Thomas Hellström <thomas.hellst...@linux.intel.com>
> Cc: Matthew Brost <matthew.br...@intel.com>
> ---
> drivers/gpu/drm/xe/Kconfig | 3 +-
> drivers/gpu/drm/xe/Makefile | 1 -
> drivers/gpu/drm/xe/xe_exec.c | 4 +-
> drivers/gpu/drm/xe/xe_hmm.c | 325 -------------------------------
> drivers/gpu/drm/xe/xe_hmm.h | 18 --
> drivers/gpu/drm/xe/xe_pt.c | 22 +--
> drivers/gpu/drm/xe/xe_svm.c | 32 +--
> drivers/gpu/drm/xe/xe_svm.h | 6 +-
> drivers/gpu/drm/xe/xe_vm.c | 86 ++++----
> drivers/gpu/drm/xe/xe_vm_types.h | 26 +--
> 10 files changed, 92 insertions(+), 431 deletions(-)
> delete mode 100644 drivers/gpu/drm/xe/xe_hmm.c
> delete mode 100644 drivers/gpu/drm/xe/xe_hmm.h
>
> diff --git a/drivers/gpu/drm/xe/Kconfig b/drivers/gpu/drm/xe/Kconfig
> index 9bce047901b2..1e63dde76c55 100644
> --- a/drivers/gpu/drm/xe/Kconfig
> +++ b/drivers/gpu/drm/xe/Kconfig
> @@ -43,7 +43,7 @@ config DRM_XE
> select MMU_NOTIFIER
> select WANT_DEV_COREDUMP
> select AUXILIARY_BUS
> - select HMM_MIRROR
> + select DRM_GPUSVM
select DRM_GPUSVM if !UML and DEVICE_PRIVATE
GPUSVM won't compile on UML. That might have other implictions
throughout the driver (e.g. compile out parts of userptr, disable
userptr if at IOCTLs if DRM_GPUSVM is not selected, etc...).
Rest of the patch LGTM.
Matt
> help
> Experimental driver for Intel Xe series GPUs
>
> @@ -79,7 +79,6 @@ config DRM_XE_GPUSVM
> depends on !UML
> depends on DEVICE_PRIVATE
> default y
> - select DRM_GPUSVM
> help
> Enable this option if you want support for CPU to GPU address
> mirroring.
> diff --git a/drivers/gpu/drm/xe/Makefile b/drivers/gpu/drm/xe/Makefile
> index 3ecac0a38b82..f1123be158f8 100644
> --- a/drivers/gpu/drm/xe/Makefile
> +++ b/drivers/gpu/drm/xe/Makefile
> @@ -124,7 +124,6 @@ xe-y += xe_bb.o \
> xe_wait_user_fence.o \
> xe_wopcm.o
>
> -xe-$(CONFIG_HMM_MIRROR) += xe_hmm.o
> xe-$(CONFIG_DRM_XE_GPUSVM) += xe_svm.o
>
> # graphics hardware monitoring (HWMON) support
> diff --git a/drivers/gpu/drm/xe/xe_exec.c b/drivers/gpu/drm/xe/xe_exec.c
> index b75adfc99fb7..c0ce681076d5 100644
> --- a/drivers/gpu/drm/xe/xe_exec.c
> +++ b/drivers/gpu/drm/xe/xe_exec.c
> @@ -294,7 +294,7 @@ int xe_exec_ioctl(struct drm_device *dev, void *data,
> struct drm_file *file)
> if (err)
> goto err_put_job;
>
> - err = down_read_interruptible(&vm->userptr.notifier_lock);
> + err = down_read_interruptible(&vm->svm.gpusvm.notifier_lock);
> if (err)
> goto err_put_job;
>
> @@ -336,7 +336,7 @@ int xe_exec_ioctl(struct drm_device *dev, void *data,
> struct drm_file *file)
>
> err_repin:
> if (!xe_vm_in_lr_mode(vm))
> - up_read(&vm->userptr.notifier_lock);
> + up_read(&vm->svm.gpusvm.notifier_lock);
> err_put_job:
> if (err)
> xe_sched_job_put(job);
> diff --git a/drivers/gpu/drm/xe/xe_hmm.c b/drivers/gpu/drm/xe/xe_hmm.c
> deleted file mode 100644
> index 57b71956ddf4..000000000000
> --- a/drivers/gpu/drm/xe/xe_hmm.c
> +++ /dev/null
> @@ -1,325 +0,0 @@
> -// SPDX-License-Identifier: MIT
> -/*
> - * Copyright © 2024 Intel Corporation
> - */
> -
> -#include <linux/scatterlist.h>
> -#include <linux/mmu_notifier.h>
> -#include <linux/dma-mapping.h>
> -#include <linux/memremap.h>
> -#include <linux/swap.h>
> -#include <linux/hmm.h>
> -#include <linux/mm.h>
> -#include "xe_hmm.h"
> -#include "xe_vm.h"
> -#include "xe_bo.h"
> -
> -static u64 xe_npages_in_range(unsigned long start, unsigned long end)
> -{
> - return (end - start) >> PAGE_SHIFT;
> -}
> -
> -static int xe_alloc_sg(struct xe_device *xe, struct sg_table *st,
> - struct hmm_range *range, struct rw_semaphore
> *notifier_sem)
> -{
> - unsigned long i, npages, hmm_pfn;
> - unsigned long num_chunks = 0;
> - int ret;
> -
> - /* HMM docs says this is needed. */
> - ret = down_read_interruptible(notifier_sem);
> - if (ret)
> - return ret;
> -
> - if (mmu_interval_read_retry(range->notifier, range->notifier_seq)) {
> - up_read(notifier_sem);
> - return -EAGAIN;
> - }
> -
> - npages = xe_npages_in_range(range->start, range->end);
> - for (i = 0; i < npages;) {
> - unsigned long len;
> -
> - hmm_pfn = range->hmm_pfns[i];
> - xe_assert(xe, hmm_pfn & HMM_PFN_VALID);
> -
> - len = 1UL << hmm_pfn_to_map_order(hmm_pfn);
> -
> - /* If order > 0 the page may extend beyond range->start */
> - len -= (hmm_pfn & ~HMM_PFN_FLAGS) & (len - 1);
> - i += len;
> - num_chunks++;
> - }
> - up_read(notifier_sem);
> -
> - return sg_alloc_table(st, num_chunks, GFP_KERNEL);
> -}
> -
> -/**
> - * xe_build_sg() - build a scatter gather table for all the physical
> pages/pfn
> - * in a hmm_range. dma-map pages if necessary. dma-address is save in sg
> table
> - * and will be used to program GPU page table later.
> - * @xe: the xe device who will access the dma-address in sg table
> - * @range: the hmm range that we build the sg table from. range->hmm_pfns[]
> - * has the pfn numbers of pages that back up this hmm address range.
> - * @st: pointer to the sg table.
> - * @notifier_sem: The xe notifier lock.
> - * @write: whether we write to this range. This decides dma map direction
> - * for system pages. If write we map it bi-diretional; otherwise
> - * DMA_TO_DEVICE
> - *
> - * All the contiguous pfns will be collapsed into one entry in
> - * the scatter gather table. This is for the purpose of efficiently
> - * programming GPU page table.
> - *
> - * The dma_address in the sg table will later be used by GPU to
> - * access memory. So if the memory is system memory, we need to
> - * do a dma-mapping so it can be accessed by GPU/DMA.
> - *
> - * FIXME: This function currently only support pages in system
> - * memory. If the memory is GPU local memory (of the GPU who
> - * is going to access memory), we need gpu dpa (device physical
> - * address), and there is no need of dma-mapping. This is TBD.
> - *
> - * FIXME: dma-mapping for peer gpu device to access remote gpu's
> - * memory. Add this when you support p2p
> - *
> - * This function allocates the storage of the sg table. It is
> - * caller's responsibility to free it calling sg_free_table.
> - *
> - * Returns 0 if successful; -ENOMEM if fails to allocate memory
> - */
> -static int xe_build_sg(struct xe_device *xe, struct hmm_range *range,
> - struct sg_table *st,
> - struct rw_semaphore *notifier_sem,
> - bool write)
> -{
> - unsigned long npages = xe_npages_in_range(range->start, range->end);
> - struct device *dev = xe->drm.dev;
> - struct scatterlist *sgl;
> - struct page *page;
> - unsigned long i, j;
> -
> - lockdep_assert_held(notifier_sem);
> -
> - i = 0;
> - for_each_sg(st->sgl, sgl, st->nents, j) {
> - unsigned long hmm_pfn, size;
> -
> - hmm_pfn = range->hmm_pfns[i];
> - page = hmm_pfn_to_page(hmm_pfn);
> - xe_assert(xe, !is_device_private_page(page));
> -
> - size = 1UL << hmm_pfn_to_map_order(hmm_pfn);
> - size -= page_to_pfn(page) & (size - 1);
> - i += size;
> -
> - if (unlikely(j == st->nents - 1)) {
> - xe_assert(xe, i >= npages);
> - if (i > npages)
> - size -= (i - npages);
> -
> - sg_mark_end(sgl);
> - } else {
> - xe_assert(xe, i < npages);
> - }
> -
> - sg_set_page(sgl, page, size << PAGE_SHIFT, 0);
> - }
> -
> - return dma_map_sgtable(dev, st, write ? DMA_BIDIRECTIONAL :
> DMA_TO_DEVICE,
> - DMA_ATTR_SKIP_CPU_SYNC |
> DMA_ATTR_NO_KERNEL_MAPPING);
> -}
> -
> -static void xe_hmm_userptr_set_mapped(struct xe_userptr_vma *uvma)
> -{
> - struct xe_userptr *userptr = &uvma->userptr;
> - struct xe_vm *vm = xe_vma_vm(&uvma->vma);
> -
> - lockdep_assert_held_write(&vm->lock);
> - lockdep_assert_held(&vm->userptr.notifier_lock);
> -
> - mutex_lock(&userptr->unmap_mutex);
> - xe_assert(vm->xe, !userptr->mapped);
> - userptr->mapped = true;
> - mutex_unlock(&userptr->unmap_mutex);
> -}
> -
> -void xe_hmm_userptr_unmap(struct xe_userptr_vma *uvma)
> -{
> - struct xe_userptr *userptr = &uvma->userptr;
> - struct xe_vma *vma = &uvma->vma;
> - bool write = !xe_vma_read_only(vma);
> - struct xe_vm *vm = xe_vma_vm(vma);
> - struct xe_device *xe = vm->xe;
> -
> - if (!lockdep_is_held_type(&vm->userptr.notifier_lock, 0) &&
> - !lockdep_is_held_type(&vm->lock, 0) &&
> - !(vma->gpuva.flags & XE_VMA_DESTROYED)) {
> - /* Don't unmap in exec critical section. */
> - xe_vm_assert_held(vm);
> - /* Don't unmap while mapping the sg. */
> - lockdep_assert_held(&vm->lock);
> - }
> -
> - mutex_lock(&userptr->unmap_mutex);
> - if (userptr->sg && userptr->mapped)
> - dma_unmap_sgtable(xe->drm.dev, userptr->sg,
> - write ? DMA_BIDIRECTIONAL : DMA_TO_DEVICE, 0);
> - userptr->mapped = false;
> - mutex_unlock(&userptr->unmap_mutex);
> -}
> -
> -/**
> - * xe_hmm_userptr_free_sg() - Free the scatter gather table of userptr
> - * @uvma: the userptr vma which hold the scatter gather table
> - *
> - * With function xe_userptr_populate_range, we allocate storage of
> - * the userptr sg table. This is a helper function to free this
> - * sg table, and dma unmap the address in the table.
> - */
> -void xe_hmm_userptr_free_sg(struct xe_userptr_vma *uvma)
> -{
> - struct xe_userptr *userptr = &uvma->userptr;
> -
> - xe_assert(xe_vma_vm(&uvma->vma)->xe, userptr->sg);
> - xe_hmm_userptr_unmap(uvma);
> - sg_free_table(userptr->sg);
> - userptr->sg = NULL;
> -}
> -
> -/**
> - * xe_hmm_userptr_populate_range() - Populate physical pages of a virtual
> - * address range
> - *
> - * @uvma: userptr vma which has information of the range to populate.
> - * @is_mm_mmap_locked: True if mmap_read_lock is already acquired by caller.
> - *
> - * This function populate the physical pages of a virtual
> - * address range. The populated physical pages is saved in
> - * userptr's sg table. It is similar to get_user_pages but call
> - * hmm_range_fault.
> - *
> - * This function also read mmu notifier sequence # (
> - * mmu_interval_read_begin), for the purpose of later
> - * comparison (through mmu_interval_read_retry).
> - *
> - * This must be called with mmap read or write lock held.
> - *
> - * This function allocates the storage of the userptr sg table.
> - * It is caller's responsibility to free it calling sg_free_table.
> - *
> - * returns: 0 for success; negative error no on failure
> - */
> -int xe_hmm_userptr_populate_range(struct xe_userptr_vma *uvma,
> - bool is_mm_mmap_locked)
> -{
> - unsigned long timeout =
> - jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
> - unsigned long *pfns;
> - struct xe_userptr *userptr;
> - struct xe_vma *vma = &uvma->vma;
> - u64 userptr_start = xe_vma_userptr(vma);
> - u64 userptr_end = userptr_start + xe_vma_size(vma);
> - struct xe_vm *vm = xe_vma_vm(vma);
> - struct hmm_range hmm_range = {
> - .pfn_flags_mask = 0, /* ignore pfns */
> - .default_flags = HMM_PFN_REQ_FAULT,
> - .start = userptr_start,
> - .end = userptr_end,
> - .notifier = &uvma->userptr.notifier,
> - .dev_private_owner = vm->xe,
> - };
> - bool write = !xe_vma_read_only(vma);
> - unsigned long notifier_seq;
> - u64 npages;
> - int ret;
> -
> - userptr = &uvma->userptr;
> -
> - if (is_mm_mmap_locked)
> - mmap_assert_locked(userptr->notifier.mm);
> -
> - if (vma->gpuva.flags & XE_VMA_DESTROYED)
> - return 0;
> -
> - notifier_seq = mmu_interval_read_begin(&userptr->notifier);
> - if (notifier_seq == userptr->notifier_seq)
> - return 0;
> -
> - if (userptr->sg)
> - xe_hmm_userptr_free_sg(uvma);
> -
> - npages = xe_npages_in_range(userptr_start, userptr_end);
> - pfns = kvmalloc_array(npages, sizeof(*pfns), GFP_KERNEL);
> - if (unlikely(!pfns))
> - return -ENOMEM;
> -
> - if (write)
> - hmm_range.default_flags |= HMM_PFN_REQ_WRITE;
> -
> - if (!mmget_not_zero(userptr->notifier.mm)) {
> - ret = -EFAULT;
> - goto free_pfns;
> - }
> -
> - hmm_range.hmm_pfns = pfns;
> -
> - while (true) {
> - hmm_range.notifier_seq =
> mmu_interval_read_begin(&userptr->notifier);
> -
> - if (!is_mm_mmap_locked)
> - mmap_read_lock(userptr->notifier.mm);
> -
> - ret = hmm_range_fault(&hmm_range);
> -
> - if (!is_mm_mmap_locked)
> - mmap_read_unlock(userptr->notifier.mm);
> -
> - if (ret == -EBUSY) {
> - if (time_after(jiffies, timeout))
> - break;
> -
> - continue;
> - }
> - break;
> - }
> -
> - mmput(userptr->notifier.mm);
> -
> - if (ret)
> - goto free_pfns;
> -
> - ret = xe_alloc_sg(vm->xe, &userptr->sgt, &hmm_range,
> &vm->userptr.notifier_lock);
> - if (ret)
> - goto free_pfns;
> -
> - ret = down_read_interruptible(&vm->userptr.notifier_lock);
> - if (ret)
> - goto free_st;
> -
> - if (mmu_interval_read_retry(hmm_range.notifier,
> hmm_range.notifier_seq)) {
> - ret = -EAGAIN;
> - goto out_unlock;
> - }
> -
> - ret = xe_build_sg(vm->xe, &hmm_range, &userptr->sgt,
> - &vm->userptr.notifier_lock, write);
> - if (ret)
> - goto out_unlock;
> -
> - userptr->sg = &userptr->sgt;
> - xe_hmm_userptr_set_mapped(uvma);
> - userptr->notifier_seq = hmm_range.notifier_seq;
> - up_read(&vm->userptr.notifier_lock);
> - kvfree(pfns);
> - return 0;
> -
> -out_unlock:
> - up_read(&vm->userptr.notifier_lock);
> -free_st:
> - sg_free_table(&userptr->sgt);
> -free_pfns:
> - kvfree(pfns);
> - return ret;
> -}
> diff --git a/drivers/gpu/drm/xe/xe_hmm.h b/drivers/gpu/drm/xe/xe_hmm.h
> deleted file mode 100644
> index 0ea98d8e7bbc..000000000000
> --- a/drivers/gpu/drm/xe/xe_hmm.h
> +++ /dev/null
> @@ -1,18 +0,0 @@
> -/* SPDX-License-Identifier: MIT
> - *
> - * Copyright © 2024 Intel Corporation
> - */
> -
> -#ifndef _XE_HMM_H_
> -#define _XE_HMM_H_
> -
> -#include <linux/types.h>
> -
> -struct xe_userptr_vma;
> -
> -int xe_hmm_userptr_populate_range(struct xe_userptr_vma *uvma, bool
> is_mm_mmap_locked);
> -
> -void xe_hmm_userptr_free_sg(struct xe_userptr_vma *uvma);
> -
> -void xe_hmm_userptr_unmap(struct xe_userptr_vma *uvma);
> -#endif
> diff --git a/drivers/gpu/drm/xe/xe_pt.c b/drivers/gpu/drm/xe/xe_pt.c
> index 5cccfd9cc3e9..ff898e518afd 100644
> --- a/drivers/gpu/drm/xe/xe_pt.c
> +++ b/drivers/gpu/drm/xe/xe_pt.c
> @@ -756,8 +756,8 @@ xe_pt_stage_bind(struct xe_tile *tile, struct xe_vma *vma,
>
> if (!xe_vma_is_null(vma) && !range) {
> if (xe_vma_is_userptr(vma))
> - xe_res_first_sg(to_userptr_vma(vma)->userptr.sg, 0,
> - xe_vma_size(vma), &curs);
> +
> xe_res_first_dma(to_userptr_vma(vma)->userptr.pages.dma_addr, 0,
> + xe_vma_size(vma), &curs);
> else if (xe_bo_is_vram(bo) || xe_bo_is_stolen(bo))
> xe_res_first(bo->ttm.resource, xe_vma_bo_offset(vma),
> xe_vma_size(vma), &curs);
> @@ -1028,7 +1028,7 @@ static void xe_pt_commit_locks_assert(struct xe_vma
> *vma)
> xe_pt_commit_prepare_locks_assert(vma);
>
> if (xe_vma_is_userptr(vma))
> - lockdep_assert_held_read(&vm->userptr.notifier_lock);
> + lockdep_assert_held_read(&vm->svm.gpusvm.notifier_lock);
> }
>
> static void xe_pt_commit(struct xe_vma *vma,
> @@ -1368,7 +1368,7 @@ static int vma_check_userptr(struct xe_vm *vm, struct
> xe_vma *vma,
> struct xe_userptr_vma *uvma;
> unsigned long notifier_seq;
>
> - lockdep_assert_held_read(&vm->userptr.notifier_lock);
> + lockdep_assert_held_read(&vm->svm.gpusvm.notifier_lock);
>
> if (!xe_vma_is_userptr(vma))
> return 0;
> @@ -1377,7 +1377,7 @@ static int vma_check_userptr(struct xe_vm *vm, struct
> xe_vma *vma,
> if (xe_pt_userptr_inject_eagain(uvma))
> xe_vma_userptr_force_invalidate(uvma);
>
> - notifier_seq = uvma->userptr.notifier_seq;
> + notifier_seq = uvma->userptr.pages.notifier_seq;
>
> if (!mmu_interval_read_retry(&uvma->userptr.notifier,
> notifier_seq))
> @@ -1398,7 +1398,7 @@ static int op_check_userptr(struct xe_vm *vm, struct
> xe_vma_op *op,
> {
> int err = 0;
>
> - lockdep_assert_held_read(&vm->userptr.notifier_lock);
> + lockdep_assert_held_read(&vm->svm.gpusvm.notifier_lock);
>
> switch (op->base.op) {
> case DRM_GPUVA_OP_MAP:
> @@ -1439,12 +1439,12 @@ static int xe_pt_userptr_pre_commit(struct
> xe_migrate_pt_update *pt_update)
> if (err)
> return err;
>
> - down_read(&vm->userptr.notifier_lock);
> + down_read(&vm->svm.gpusvm.notifier_lock);
>
> list_for_each_entry(op, &vops->list, link) {
> err = op_check_userptr(vm, op, pt_update_ops);
> if (err) {
> - up_read(&vm->userptr.notifier_lock);
> + up_read(&vm->svm.gpusvm.notifier_lock);
> break;
> }
> }
> @@ -2171,7 +2171,7 @@ static void bind_op_commit(struct xe_vm *vm, struct
> xe_tile *tile,
> if (invalidate_on_bind)
> vma->tile_invalidated |= BIT(tile->id);
> if (xe_vma_is_userptr(vma)) {
> - lockdep_assert_held_read(&vm->userptr.notifier_lock);
> + lockdep_assert_held_read(&vm->svm.gpusvm.notifier_lock);
> to_userptr_vma(vma)->userptr.initial_bind = true;
> }
>
> @@ -2207,7 +2207,7 @@ static void unbind_op_commit(struct xe_vm *vm, struct
> xe_tile *tile,
> if (!vma->tile_present) {
> list_del_init(&vma->combined_links.rebind);
> if (xe_vma_is_userptr(vma)) {
> - lockdep_assert_held_read(&vm->userptr.notifier_lock);
> + lockdep_assert_held_read(&vm->svm.gpusvm.notifier_lock);
>
> spin_lock(&vm->userptr.invalidated_lock);
>
> list_del_init(&to_userptr_vma(vma)->userptr.invalidate_link);
> @@ -2456,7 +2456,7 @@ xe_pt_update_ops_run(struct xe_tile *tile, struct
> xe_vma_ops *vops)
> if (pt_update_ops->needs_svm_lock)
> xe_svm_notifier_unlock(vm);
> if (pt_update_ops->needs_userptr_lock)
> - up_read(&vm->userptr.notifier_lock);
> + up_read(&vm->svm.gpusvm.notifier_lock);
>
> return fence;
>
> diff --git a/drivers/gpu/drm/xe/xe_svm.c b/drivers/gpu/drm/xe/xe_svm.c
> index f60c7f587cd6..10ec237ec88c 100644
> --- a/drivers/gpu/drm/xe/xe_svm.c
> +++ b/drivers/gpu/drm/xe/xe_svm.c
> @@ -606,22 +606,26 @@ int xe_svm_init(struct xe_vm *vm)
> {
> int err;
>
> - spin_lock_init(&vm->svm.garbage_collector.lock);
> - INIT_LIST_HEAD(&vm->svm.garbage_collector.range_list);
> - INIT_WORK(&vm->svm.garbage_collector.work,
> - xe_svm_garbage_collector_work_func);
> + if (vm->flags & XE_VM_FLAG_FAULT_MODE) {
> + spin_lock_init(&vm->svm.garbage_collector.lock);
> + INIT_LIST_HEAD(&vm->svm.garbage_collector.range_list);
> + INIT_WORK(&vm->svm.garbage_collector.work,
> + xe_svm_garbage_collector_work_func);
>
> - err = drm_gpusvm_init(&vm->svm.gpusvm, "Xe SVM", &vm->xe->drm,
> - current->mm, xe_svm_devm_owner(vm->xe), 0,
> - vm->size, xe_modparam.svm_notifier_size * SZ_1M,
> - &gpusvm_ops, fault_chunk_sizes,
> - ARRAY_SIZE(fault_chunk_sizes));
> - if (err)
> - return err;
> + err = drm_gpusvm_init(&vm->svm.gpusvm, "Xe SVM", &vm->xe->drm,
> + current->mm, xe_svm_devm_owner(vm->xe), 0,
> + vm->size,
> + xe_modparam.svm_notifier_size * SZ_1M,
> + &gpusvm_ops, fault_chunk_sizes,
> + ARRAY_SIZE(fault_chunk_sizes));
> + drm_gpusvm_driver_set_lock(&vm->svm.gpusvm, &vm->lock);
> + } else {
> + err = drm_gpusvm_init(&vm->svm.gpusvm, "Xe SVM (simple)",
> + &vm->xe->drm, NULL, NULL, 0, 0, 0, NULL,
> + NULL, 0);
> + }
>
> - drm_gpusvm_driver_set_lock(&vm->svm.gpusvm, &vm->lock);
> -
> - return 0;
> + return err;
> }
>
> /**
> diff --git a/drivers/gpu/drm/xe/xe_svm.h b/drivers/gpu/drm/xe/xe_svm.h
> index 8bc83de87250..923cb074d0cb 100644
> --- a/drivers/gpu/drm/xe/xe_svm.h
> +++ b/drivers/gpu/drm/xe/xe_svm.h
> @@ -134,12 +134,16 @@ int xe_devm_add(struct xe_tile *tile, struct
> xe_vram_region *vr)
> static inline
> int xe_svm_init(struct xe_vm *vm)
> {
> - return 0;
> + return drm_gpusvm_init(&vm->svm.gpusvm, "Xe SVM (simple)", &vm->xe->drm,
> + NULL, NULL, 0, 0, 0, NULL, NULL, 0);
> }
>
> static inline
> void xe_svm_fini(struct xe_vm *vm)
> {
> + xe_assert(vm->xe, xe_vm_is_closed(vm));
> +
> + drm_gpusvm_fini(&vm->svm.gpusvm);
> }
>
> static inline
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index 0c69ef6b5ec5..b900afe61931 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -40,7 +40,6 @@
> #include "xe_sync.h"
> #include "xe_trace_bo.h"
> #include "xe_wa.h"
> -#include "xe_hmm.h"
>
> static struct drm_gem_object *xe_vm_obj(struct xe_vm *vm)
> {
> @@ -53,7 +52,7 @@ static struct drm_gem_object *xe_vm_obj(struct xe_vm *vm)
> *
> * Check if the userptr vma has been invalidated since last successful
> * repin. The check is advisory only and can the function can be called
> - * without the vm->userptr.notifier_lock held. There is no guarantee that the
> + * without the vm->svm.gpusvm.notifier_lock held. There is no guarantee that
> the
> * vma userptr will remain valid after a lockless check, so typically
> * the call needs to be followed by a proper check under the notifier_lock.
> *
> @@ -62,7 +61,7 @@ static struct drm_gem_object *xe_vm_obj(struct xe_vm *vm)
> int xe_vma_userptr_check_repin(struct xe_userptr_vma *uvma)
> {
> return mmu_interval_check_retry(&uvma->userptr.notifier,
> - uvma->userptr.notifier_seq) ?
> + uvma->userptr.pages.notifier_seq) ?
> -EAGAIN : 0;
> }
>
> @@ -71,11 +70,22 @@ int xe_vma_userptr_pin_pages(struct xe_userptr_vma *uvma)
> struct xe_vma *vma = &uvma->vma;
> struct xe_vm *vm = xe_vma_vm(vma);
> struct xe_device *xe = vm->xe;
> + struct drm_gpusvm_ctx ctx = {
> + .read_only = xe_vma_read_only(vma),
> + };
>
> lockdep_assert_held(&vm->lock);
> xe_assert(xe, xe_vma_is_userptr(vma));
>
> - return xe_hmm_userptr_populate_range(uvma, false);
> + if (vma->gpuva.flags & XE_VMA_DESTROYED)
> + return 0;
> +
> + return drm_gpusvm_get_pages(&vm->svm.gpusvm, &uvma->userptr.pages,
> + uvma->userptr.notifier.mm,
> + &uvma->userptr.notifier,
> + xe_vma_userptr(vma),
> + xe_vma_userptr(vma) + xe_vma_size(vma),
> + &ctx);
> }
>
> static bool preempt_fences_waiting(struct xe_vm *vm)
> @@ -249,7 +259,7 @@ int xe_vm_add_compute_exec_queue(struct xe_vm *vm, struct
> xe_exec_queue *q)
> ++vm->preempt.num_exec_queues;
> q->lr.pfence = pfence;
>
> - down_read(&vm->userptr.notifier_lock);
> + down_read(&vm->svm.gpusvm.notifier_lock);
>
> drm_gpuvm_resv_add_fence(&vm->gpuvm, exec, pfence,
> DMA_RESV_USAGE_BOOKKEEP,
> DMA_RESV_USAGE_BOOKKEEP);
> @@ -263,7 +273,7 @@ int xe_vm_add_compute_exec_queue(struct xe_vm *vm, struct
> xe_exec_queue *q)
> if (wait)
> dma_fence_enable_sw_signaling(pfence);
>
> - up_read(&vm->userptr.notifier_lock);
> + up_read(&vm->svm.gpusvm.notifier_lock);
>
> out_fini:
> drm_exec_fini(exec);
> @@ -305,14 +315,14 @@ void xe_vm_remove_compute_exec_queue(struct xe_vm *vm,
> struct xe_exec_queue *q)
> * @vm: The VM.
> *
> * This function checks for whether the VM has userptrs that need repinning,
> - * and provides a release-type barrier on the userptr.notifier_lock after
> + * and provides a release-type barrier on the svm.gpusvm.notifier_lock after
> * checking.
> *
> * Return: 0 if there are no userptrs needing repinning, -EAGAIN if there
> are.
> */
> int __xe_vm_userptr_needs_repin(struct xe_vm *vm)
> {
> - lockdep_assert_held_read(&vm->userptr.notifier_lock);
> + lockdep_assert_held_read(&vm->svm.gpusvm.notifier_lock);
>
> return (list_empty(&vm->userptr.repin_list) &&
> list_empty(&vm->userptr.invalidated)) ? 0 : -EAGAIN;
> @@ -546,9 +556,9 @@ static void preempt_rebind_work_func(struct work_struct
> *w)
> (!(__tries)++ || __xe_vm_userptr_needs_repin(__vm)) : \
> __xe_vm_userptr_needs_repin(__vm))
>
> - down_read(&vm->userptr.notifier_lock);
> + down_read(&vm->svm.gpusvm.notifier_lock);
> if (retry_required(tries, vm)) {
> - up_read(&vm->userptr.notifier_lock);
> + up_read(&vm->svm.gpusvm.notifier_lock);
> err = -EAGAIN;
> goto out_unlock;
> }
> @@ -562,7 +572,7 @@ static void preempt_rebind_work_func(struct work_struct
> *w)
> /* Point of no return. */
> arm_preempt_fences(vm, &preempt_fences);
> resume_and_reinstall_preempt_fences(vm, &exec);
> - up_read(&vm->userptr.notifier_lock);
> + up_read(&vm->svm.gpusvm.notifier_lock);
>
> out_unlock:
> drm_exec_fini(&exec);
> @@ -589,6 +599,9 @@ static void __vma_userptr_invalidate(struct xe_vm *vm,
> struct xe_userptr_vma *uv
> struct xe_vma *vma = &uvma->vma;
> struct dma_resv_iter cursor;
> struct dma_fence *fence;
> + struct drm_gpusvm_ctx ctx = {
> + .in_notifier = true,
> + };
> long err;
>
> /*
> @@ -625,7 +638,8 @@ static void __vma_userptr_invalidate(struct xe_vm *vm,
> struct xe_userptr_vma *uv
> XE_WARN_ON(err);
> }
>
> - xe_hmm_userptr_unmap(uvma);
> + drm_gpusvm_unmap_pages(&vm->svm.gpusvm, &uvma->userptr.pages,
> + xe_vma_size(vma) >> PAGE_SHIFT, &ctx);
> }
>
> static bool vma_userptr_invalidate(struct mmu_interval_notifier *mni,
> @@ -646,11 +660,11 @@ static bool vma_userptr_invalidate(struct
> mmu_interval_notifier *mni,
> "NOTIFIER: addr=0x%016llx, range=0x%016llx",
> xe_vma_start(vma), xe_vma_size(vma));
>
> - down_write(&vm->userptr.notifier_lock);
> + down_write(&vm->svm.gpusvm.notifier_lock);
> mmu_interval_set_seq(mni, cur_seq);
>
> __vma_userptr_invalidate(vm, uvma);
> - up_write(&vm->userptr.notifier_lock);
> + up_write(&vm->svm.gpusvm.notifier_lock);
> trace_xe_vma_userptr_invalidate_complete(vma);
>
> return true;
> @@ -674,7 +688,7 @@ void xe_vma_userptr_force_invalidate(struct
> xe_userptr_vma *uvma)
> /* Protect against concurrent userptr pinning */
> lockdep_assert_held(&vm->lock);
> /* Protect against concurrent notifiers */
> - lockdep_assert_held(&vm->userptr.notifier_lock);
> + lockdep_assert_held(&vm->svm.gpusvm.notifier_lock);
> /*
> * Protect against concurrent instances of this function and
> * the critical exec sections
> @@ -747,7 +761,7 @@ int xe_vm_userptr_pin(struct xe_vm *vm)
> }
>
> if (err) {
> - down_write(&vm->userptr.notifier_lock);
> + down_write(&vm->svm.gpusvm.notifier_lock);
> spin_lock(&vm->userptr.invalidated_lock);
> list_for_each_entry_safe(uvma, next, &vm->userptr.repin_list,
> userptr.repin_link) {
> @@ -756,7 +770,7 @@ int xe_vm_userptr_pin(struct xe_vm *vm)
> &vm->userptr.invalidated);
> }
> spin_unlock(&vm->userptr.invalidated_lock);
> - up_write(&vm->userptr.notifier_lock);
> + up_write(&vm->svm.gpusvm.notifier_lock);
> }
> return err;
> }
> @@ -1222,7 +1236,6 @@ static struct xe_vma *xe_vma_create(struct xe_vm *vm,
> INIT_LIST_HEAD(&userptr->invalidate_link);
> INIT_LIST_HEAD(&userptr->repin_link);
> vma->gpuva.gem.offset = bo_offset_or_userptr;
> - mutex_init(&userptr->unmap_mutex);
>
> err = mmu_interval_notifier_insert(&userptr->notifier,
> current->mm,
> @@ -1233,7 +1246,7 @@ static struct xe_vma *xe_vma_create(struct xe_vm *vm,
> return ERR_PTR(err);
> }
>
> - userptr->notifier_seq = LONG_MAX;
> + userptr->pages.notifier_seq = LONG_MAX;
> }
>
> xe_vm_get(vm);
> @@ -1255,8 +1268,8 @@ static void xe_vma_destroy_late(struct xe_vma *vma)
> struct xe_userptr_vma *uvma = to_userptr_vma(vma);
> struct xe_userptr *userptr = &uvma->userptr;
>
> - if (userptr->sg)
> - xe_hmm_userptr_free_sg(uvma);
> + drm_gpusvm_free_pages(&vm->svm.gpusvm, &uvma->userptr.pages,
> + xe_vma_size(vma) >> PAGE_SHIFT);
>
> /*
> * Since userptr pages are not pinned, we can't remove
> @@ -1264,7 +1277,6 @@ static void xe_vma_destroy_late(struct xe_vma *vma)
> * them anymore
> */
> mmu_interval_notifier_remove(&userptr->notifier);
> - mutex_destroy(&userptr->unmap_mutex);
> xe_vm_put(vm);
> } else if (xe_vma_is_null(vma) || xe_vma_is_cpu_addr_mirror(vma)) {
> xe_vm_put(vm);
> @@ -1657,7 +1669,6 @@ struct xe_vm *xe_vm_create(struct xe_device *xe, u32
> flags)
>
> INIT_LIST_HEAD(&vm->userptr.repin_list);
> INIT_LIST_HEAD(&vm->userptr.invalidated);
> - init_rwsem(&vm->userptr.notifier_lock);
> spin_lock_init(&vm->userptr.invalidated_lock);
>
> ttm_lru_bulk_move_init(&vm->lru_bulk_move);
> @@ -1757,11 +1768,9 @@ struct xe_vm *xe_vm_create(struct xe_device *xe, u32
> flags)
> }
> }
>
> - if (flags & XE_VM_FLAG_FAULT_MODE) {
> - err = xe_svm_init(vm);
> - if (err)
> - goto err_close;
> - }
> + err = xe_svm_init(vm);
> + if (err)
> + goto err_close;
>
> if (number_tiles > 1)
> vm->composite_fence_ctx = dma_fence_context_alloc(1);
> @@ -1867,9 +1876,9 @@ void xe_vm_close_and_put(struct xe_vm *vm)
> vma = gpuva_to_vma(gpuva);
>
> if (xe_vma_has_no_bo(vma)) {
> - down_read(&vm->userptr.notifier_lock);
> + down_read(&vm->svm.gpusvm.notifier_lock);
> vma->gpuva.flags |= XE_VMA_DESTROYED;
> - up_read(&vm->userptr.notifier_lock);
> + up_read(&vm->svm.gpusvm.notifier_lock);
> }
>
> xe_vm_remove_vma(vm, vma);
> @@ -1913,8 +1922,7 @@ void xe_vm_close_and_put(struct xe_vm *vm)
> xe_vma_destroy_unlocked(vma);
> }
>
> - if (xe_vm_in_fault_mode(vm))
> - xe_svm_fini(vm);
> + xe_svm_fini(vm);
>
> up_write(&vm->lock);
>
> @@ -2145,9 +2153,9 @@ static const u32 region_to_mem_type[] = {
> static void prep_vma_destroy(struct xe_vm *vm, struct xe_vma *vma,
> bool post_commit)
> {
> - down_read(&vm->userptr.notifier_lock);
> + down_read(&vm->svm.gpusvm.notifier_lock);
> vma->gpuva.flags |= XE_VMA_DESTROYED;
> - up_read(&vm->userptr.notifier_lock);
> + up_read(&vm->svm.gpusvm.notifier_lock);
> if (post_commit)
> xe_vm_remove_vma(vm, vma);
> }
> @@ -2643,9 +2651,9 @@ static void xe_vma_op_unwind(struct xe_vm *vm, struct
> xe_vma_op *op,
> struct xe_vma *vma = gpuva_to_vma(op->base.unmap.va);
>
> if (vma) {
> - down_read(&vm->userptr.notifier_lock);
> + down_read(&vm->svm.gpusvm.notifier_lock);
> vma->gpuva.flags &= ~XE_VMA_DESTROYED;
> - up_read(&vm->userptr.notifier_lock);
> + up_read(&vm->svm.gpusvm.notifier_lock);
> if (post_commit)
> xe_vm_insert_vma(vm, vma);
> }
> @@ -2664,9 +2672,9 @@ static void xe_vma_op_unwind(struct xe_vm *vm, struct
> xe_vma_op *op,
> xe_vma_destroy_unlocked(op->remap.next);
> }
> if (vma) {
> - down_read(&vm->userptr.notifier_lock);
> + down_read(&vm->svm.gpusvm.notifier_lock);
> vma->gpuva.flags &= ~XE_VMA_DESTROYED;
> - up_read(&vm->userptr.notifier_lock);
> + up_read(&vm->svm.gpusvm.notifier_lock);
> if (post_commit)
> xe_vm_insert_vma(vm, vma);
> }
> @@ -3648,7 +3656,7 @@ int xe_vm_invalidate_vma(struct xe_vma *vma)
> 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));
> +
> to_userptr_vma(vma)->userptr.pages.notifier_seq));
>
> WARN_ON_ONCE(!dma_resv_test_signaled(xe_vm_resv(xe_vma_vm(vma)),
>
> DMA_RESV_USAGE_BOOKKEEP));
>
> diff --git a/drivers/gpu/drm/xe/xe_vm_types.h
> b/drivers/gpu/drm/xe/xe_vm_types.h
> index 1662604c4486..b57c4299875a 100644
> --- a/drivers/gpu/drm/xe/xe_vm_types.h
> +++ b/drivers/gpu/drm/xe/xe_vm_types.h
> @@ -52,26 +52,21 @@ struct xe_userptr {
> struct list_head invalidate_link;
> /** @userptr: link into VM repin list if userptr. */
> struct list_head repin_link;
> + /**
> + * @pages: gpusvm pages for this user pointer.
> + */
> + struct drm_gpusvm_pages pages;
> /**
> * @notifier: MMU notifier for user pointer (invalidation call back)
> */
> struct mmu_interval_notifier notifier;
> - /** @sgt: storage for a scatter gather table */
> - struct sg_table sgt;
> - /** @sg: allocated scatter gather table */
> - struct sg_table *sg;
> - /** @notifier_seq: notifier sequence number */
> - unsigned long notifier_seq;
> - /** @unmap_mutex: Mutex protecting dma-unmapping */
> - struct mutex unmap_mutex;
> +
> /**
> * @initial_bind: user pointer has been bound at least once.
> - * write: vm->userptr.notifier_lock in read mode and vm->resv held.
> - * read: vm->userptr.notifier_lock in write mode or vm->resv held.
> + * write: vm->svm.gpusvm.notifier_lock in read mode and vm->resv held.
> + * read: vm->svm.gpusvm.notifier_lock in write mode or vm->resv held.
> */
> bool initial_bind;
> - /** @mapped: Whether the @sgt sg-table is dma-mapped. Protected by
> @unmap_mutex. */
> - bool mapped;
> #if IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT)
> u32 divisor;
> #endif
> @@ -109,7 +104,7 @@ struct xe_vma {
> /**
> * @tile_present: GT mask of binding are present for this VMA.
> * protected by vm->lock, vm->resv and for userptrs,
> - * vm->userptr.notifier_lock for writing. Needs either for reading,
> + * vm->svm.gpusvm.notifier_lock for writing. Needs either for reading,
> * but if reading is done under the vm->lock only, it needs to be held
> * in write mode.
> */
> @@ -243,11 +238,6 @@ struct xe_vm {
> * and needs repinning. Protected by @lock.
> */
> struct list_head repin_list;
> - /**
> - * @notifier_lock: protects notifier in write mode and
> - * submission in read mode.
> - */
> - struct rw_semaphore notifier_lock;
> /**
> * @userptr.invalidated_lock: Protects the
> * @userptr.invalidated list.
> --
> 2.49.0
>
