[igt-dev] [PATCH 3/5] xe_vm: Add mmap / munmap sections that split large pages
Rodrigo Vivi
rodrigo.vivi at intel.com
Thu Jul 13 15:30:19 UTC 2023
On Mon, Jul 10, 2023 at 07:58:54AM -0700, Matthew Brost wrote:
> Splitting large pages involves using dma-resv slots for ordering, make
> sure this works.
>
> Signed-off-by: Matthew Brost <matthew.brost at intel.com>
> ---
> tests/xe/xe_vm.c | 81 +++++++++++++++++++++++++++++++++++-------------
> 1 file changed, 60 insertions(+), 21 deletions(-)
>
> diff --git a/tests/xe/xe_vm.c b/tests/xe/xe_vm.c
> index 434bdb332..5fd511f0b 100644
> --- a/tests/xe/xe_vm.c
> +++ b/tests/xe/xe_vm.c
> @@ -1252,7 +1252,8 @@ static void *hammer_thread(void *tdata)
> #define MAP_FLAG_USERPTR (0x1 << 0)
> #define MAP_FLAG_INVALIDATE (0x1 << 1)
> #define MAP_FLAG_HAMMER_FIRST_PAGE (0x1 << 2)
> -
> +#define MAP_FLAG_LARGE_PAGE (0x1 << 3)
> +#define MAP_FLAG_LARGE_PAGE_NO_SPLIT (0x1 << 4)
>
> /**
> * SUBTEST: munmap-style-unbind-%s
> @@ -1305,12 +1306,16 @@ static void *hammer_thread(void *tdata)
> * userptr inval many either side full
> * @userptr-inval-many-end: userptr inval many end
> * @userptr-inval-many-front: userptr inval many front
> + * @either-side-partial-large-page-hammer:
> + * either side partial large page hammer
> + * @either-side-partial-split-page-hammer:
> + * either side partial split page hammer
> */
>
> static void
> test_munmap_style_unbind(int fd, struct drm_xe_engine_class_instance *eci,
> int bo_n_pages, int n_binds,
> - int unbind_n_page_offfset, int unbind_n_pages,
> + int unbind_n_page_offset, int unbind_n_pages,
> unsigned int flags)
> {
> struct drm_xe_sync sync[2] = {
> @@ -1322,7 +1327,7 @@ test_munmap_style_unbind(int fd, struct drm_xe_engine_class_instance *eci,
> .num_syncs = 2,
> .syncs = to_user_pointer(sync),
> };
> - uint64_t addr = 0x1a0000, base_addr = 0x1a0000;
> + uint64_t addr = 0x1a00000, base_addr = 0x1a00000;
> uint32_t vm;
> uint32_t engine;
> size_t bo_size;
> @@ -1340,6 +1345,14 @@ test_munmap_style_unbind(int fd, struct drm_xe_engine_class_instance *eci,
> struct thread_data t;
> pthread_barrier_t barrier;
> int exit = 0;
> + int n_page_per_2mb = 0x200000 / xe_get_default_alignment(fd);
> +
> + if (flags & MAP_FLAG_LARGE_PAGE) {
> + bo_n_pages *= n_page_per_2mb;
> + unbind_n_pages *= n_page_per_2mb;
> + if (flags & MAP_FLAG_LARGE_PAGE_NO_SPLIT)
> + unbind_n_page_offset *= n_page_per_2mb;
> + }
>
> vm = xe_vm_create(fd, DRM_XE_VM_CREATE_ASYNC_BIND_OPS, 0);
> bo_size = page_size * bo_n_pages;
> @@ -1426,7 +1439,7 @@ test_munmap_style_unbind(int fd, struct drm_xe_engine_class_instance *eci,
> sync[0].flags |= DRM_XE_SYNC_SIGNAL;
> sync[1].flags &= ~DRM_XE_SYNC_SIGNAL;
> xe_vm_unbind_async(fd, vm, 0, 0,
> - addr + unbind_n_page_offfset * page_size,
> + addr + unbind_n_page_offset * page_size,
> unbind_n_pages * page_size, sync, 2);
>
> igt_assert(syncobj_wait(fd, &sync[0].handle, 1, INT64_MAX, 0, NULL));
> @@ -1455,8 +1468,8 @@ try_again_after_invalidate:
> data = map + i * page_size;
> addr += page_size;
>
> - if (i < unbind_n_page_offfset ||
> - i + 1 > unbind_n_page_offfset + unbind_n_pages) {
> + if (i < unbind_n_page_offset ||
> + i + 1 > unbind_n_page_offset + unbind_n_pages) {
> b = 0;
> data->batch[b++] = MI_STORE_DWORD_IMM_GEN4;
> data->batch[b++] = sdi_addr;
> @@ -1481,8 +1494,8 @@ try_again_after_invalidate:
>
> /* Verify all pages still bound written */
> for (i = 0; i < n_binds; ++i) {
> - if (i < unbind_n_page_offfset ||
> - i + 1 > unbind_n_page_offfset + unbind_n_pages) {
> + if (i < unbind_n_page_offset ||
> + i + 1 > unbind_n_page_offset + unbind_n_pages) {
> data = map + i * page_size;
> igt_assert_eq(data->data, 0xc0ffee);
> }
> @@ -1511,13 +1524,13 @@ try_again_after_invalidate:
> sync[0].flags |= DRM_XE_SYNC_SIGNAL;
> if (flags & MAP_FLAG_USERPTR)
> xe_vm_bind_userptr_async(fd, vm, 0,
> - addr + unbind_n_page_offfset * page_size,
> - addr + unbind_n_page_offfset * page_size,
> + addr + unbind_n_page_offset * page_size,
> + addr + unbind_n_page_offset * page_size,
> unbind_n_pages * page_size, sync, 1);
> else
> xe_vm_bind_async(fd, vm, 0, bo,
> - unbind_n_page_offfset * page_size,
> - addr + unbind_n_page_offfset * page_size,
> + unbind_n_page_offset * page_size,
> + addr + unbind_n_page_offset * page_size,
> unbind_n_pages * page_size, sync, 1);
>
> /* Verify we can use every page */
> @@ -1594,11 +1607,15 @@ try_again_after_invalidate:
> * @userptr-one-partial: userptr one partial
> * @userptr-either-side-partial: userptr either side partial
> * @userptr-either-side-full: userptr either side full
> + * @either-side-partial-large-page-hammer:
> + * either side partial large page hammer
> + * @either-side-partial-split-page-hammer:
> + * either side partial split page hammer
> */
>
> static void
> test_mmap_style_bind(int fd, struct drm_xe_engine_class_instance *eci,
> - int bo_n_pages, int n_binds, int unbind_n_page_offfset,
> + int bo_n_pages, int n_binds, int unbind_n_page_offset,
> int unbind_n_pages, unsigned int flags)
> {
> struct drm_xe_sync sync[2] = {
> @@ -1610,7 +1627,7 @@ test_mmap_style_bind(int fd, struct drm_xe_engine_class_instance *eci,
> .num_syncs = 2,
> .syncs = to_user_pointer(sync),
> };
> - uint64_t addr = 0x1a0000, base_addr = 0x1a0000;
> + uint64_t addr = 0x1a00000, base_addr = 0x1a00000;
why do we need to change the offset here?
> uint32_t vm;
> uint32_t engine;
> size_t bo_size;
> @@ -1627,6 +1644,14 @@ test_mmap_style_bind(int fd, struct drm_xe_engine_class_instance *eci,
> struct thread_data t;
> pthread_barrier_t barrier;
> int exit = 0;
> + int n_page_per_2mb = 0x200000 / xe_get_default_alignment(fd);
> +
> + if (flags & MAP_FLAG_LARGE_PAGE) {
> + bo_n_pages *= n_page_per_2mb;
> + unbind_n_pages *= n_page_per_2mb;
> + if (flags & MAP_FLAG_LARGE_PAGE_NO_SPLIT)
> + unbind_n_page_offset *= n_page_per_2mb;
> + }
>
> vm = xe_vm_create(fd, DRM_XE_VM_CREATE_ASYNC_BIND_OPS, 0);
> bo_size = page_size * bo_n_pages;
> @@ -1721,13 +1746,13 @@ test_mmap_style_bind(int fd, struct drm_xe_engine_class_instance *eci,
> sync[1].flags &= ~DRM_XE_SYNC_SIGNAL;
> if (flags & MAP_FLAG_USERPTR)
> xe_vm_bind_userptr_async(fd, vm, 0, addr + bo_size +
> - unbind_n_page_offfset * page_size,
> - addr + unbind_n_page_offfset * page_size,
> + unbind_n_page_offset * page_size,
> + addr + unbind_n_page_offset * page_size,
> unbind_n_pages * page_size, sync, 2);
> else
> xe_vm_bind_async(fd, vm, 0, bo1,
> - unbind_n_page_offfset * page_size,
> - addr + unbind_n_page_offfset * page_size,
> + unbind_n_page_offset * page_size,
> + addr + unbind_n_page_offset * page_size,
probably the typo fix could be a separated patch to avoid noise here...
because now I'm wondering if the addr change is a change that is needed for
the large page or just something extra this patch covered like the typo here.
But anyway, it is already here, no need to change...
Everything looks okay on the test itself... and I will eventually get the
addr choices...
Reviewed-by: Rodrigo Vivi <rodrigo.vivi at intel.com>
> unbind_n_pages * page_size, sync, 2);
> igt_assert(syncobj_wait(fd, &sync[0].handle, 1, INT64_MAX, 0, NULL));
> igt_assert(syncobj_wait(fd, &sync[1].handle, 1, INT64_MAX, 0, NULL));
> @@ -1826,7 +1851,7 @@ igt_main
> const char *name;
> int bo_n_pages;
> int n_binds;
> - int unbind_n_page_offfset;
> + int unbind_n_page_offset;
> int unbind_n_pages;
> unsigned int flags;
> } munmap_sections[] = {
> @@ -1835,6 +1860,13 @@ igt_main
> { "either-side-partial", 4, 2, 1, 2, 0 },
> { "either-side-partial-hammer", 4, 2, 1, 2,
> MAP_FLAG_HAMMER_FIRST_PAGE },
> + { "either-side-partial-split-page-hammer", 4, 2, 1, 2,
> + MAP_FLAG_HAMMER_FIRST_PAGE |
> + MAP_FLAG_LARGE_PAGE },
> + { "either-side-partial-large-page-hammer", 4, 2, 1, 2,
> + MAP_FLAG_HAMMER_FIRST_PAGE |
> + MAP_FLAG_LARGE_PAGE |
> + MAP_FLAG_LARGE_PAGE_NO_SPLIT },
> { "either-side-full", 4, 4, 1, 2, 0 },
> { "end", 4, 2, 0, 3, 0 },
> { "front", 4, 2, 1, 3, 0 },
> @@ -1887,6 +1919,13 @@ igt_main
> { "either-side-full", 4, 4, 1, 2, 0 },
> { "either-side-partial-hammer", 4, 2, 1, 2,
> MAP_FLAG_HAMMER_FIRST_PAGE },
> + { "either-side-partial-split-page-hammer", 4, 2, 1, 2,
> + MAP_FLAG_HAMMER_FIRST_PAGE |
> + MAP_FLAG_LARGE_PAGE },
> + { "either-side-partial-large-page-hammer", 4, 2, 1, 2,
> + MAP_FLAG_HAMMER_FIRST_PAGE |
> + MAP_FLAG_LARGE_PAGE |
> + MAP_FLAG_LARGE_PAGE_NO_SPLIT },
> { "end", 4, 2, 0, 3, 0 },
> { "front", 4, 2, 1, 3, 0 },
> { "many-all", 4 * 8, 2 * 8, 0 * 8, 4 * 8, 0 },
> @@ -2111,7 +2150,7 @@ igt_main
> test_munmap_style_unbind(fd, hwe_non_copy,
> s->bo_n_pages,
> s->n_binds,
> - s->unbind_n_page_offfset,
> + s->unbind_n_page_offset,
> s->unbind_n_pages,
> s->flags);
> }
> @@ -2125,7 +2164,7 @@ igt_main
> test_mmap_style_bind(fd, hwe_non_copy,
> s->bo_n_pages,
> s->n_binds,
> - s->unbind_n_page_offfset,
> + s->unbind_n_page_offset,
> s->unbind_n_pages,
> s->flags);
> }
> --
> 2.34.1
>
More information about the igt-dev
mailing list