[igt-dev] [PATCH i-g-t 1/2] lib/xe/xe_spin: Integrate igt_spin_new with Xe.
Zbigniew Kempczyński
zbigniew.kempczynski at intel.com
Tue May 30 19:39:03 UTC 2023
On Tue, May 30, 2023 at 03:38:04PM +0530, sai.gowtham.ch at intel.com wrote:
> From: Sai Gowtham Ch <sai.gowtham.ch at intel.com>
>
> Extending the spin_create implementation and allocator handle support in xe,
> where it submits dummy work loads to engine. This Implementation is wrapped
> around vm_bind and unbind as we are supposed to do it manually for xe.
>
> Cc: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
> Signed-off-by: Sai Gowtham Ch <sai.gowtham.ch at intel.com>
> ---
> lib/igt_dummyload.c | 24 +++++++++---
> lib/igt_dummyload.h | 10 +++++
> lib/xe/xe_spin.c | 91 +++++++++++++++++++++++++++++++++++++++++++++
> lib/xe/xe_spin.h | 7 ++++
> 4 files changed, 126 insertions(+), 6 deletions(-)
>
> diff --git a/lib/igt_dummyload.c b/lib/igt_dummyload.c
> index 740a58f3..6e89b72d 100644
> --- a/lib/igt_dummyload.c
> +++ b/lib/igt_dummyload.c
> @@ -46,6 +46,7 @@
> #include "intel_reg.h"
> #include "ioctl_wrappers.h"
> #include "sw_sync.h"
> +#include "xe/xe_spin.h"
>
> /**
> * SECTION:igt_dummyload
> @@ -447,7 +448,10 @@ spin_create(int fd, const struct igt_spin_factory *opts)
> igt_spin_t *
> __igt_spin_factory(int fd, const struct igt_spin_factory *opts)
> {
> - return spin_create(fd, opts);
> + if (is_xe_device(fd))
> + return xe_spin_create(fd, opts);
> + else
> + return spin_create(fd, opts);
> }
>
> /**
> @@ -467,6 +471,11 @@ igt_spin_factory(int fd, const struct igt_spin_factory *opts)
> {
> igt_spin_t *spin;
>
> + if (is_xe_device(fd)) {
> + spin = xe_spin_create(fd, opts);
> + return spin;
> + }
> +
> if ((opts->flags & IGT_SPIN_POLL_RUN) && opts->engine != ALL_ENGINES) {
> unsigned int class;
>
> @@ -647,11 +656,14 @@ void igt_spin_free(int fd, igt_spin_t *spin)
> if (!spin)
> return;
>
> - pthread_mutex_lock(&list_lock);
> - igt_list_del(&spin->link);
> - pthread_mutex_unlock(&list_lock);
> -
> - __igt_spin_free(fd, spin);
> + if (is_xe_device(fd)) {
> + xe_spin_free(fd, spin);
> + } else {
> + pthread_mutex_lock(&list_lock);
> + igt_list_del(&spin->link);
> + pthread_mutex_unlock(&list_lock);
> + __igt_spin_free(fd, spin);
> + }
At the moment I'm not sure but I think for failed subtests we may
have spinners running and not cleared from igt_terminate_spinners().
I think you may add xe spinner to the list and handle this case.
> }
>
> void igt_terminate_spins(void)
> diff --git a/lib/igt_dummyload.h b/lib/igt_dummyload.h
> index b247ab02..7bcc7b56 100644
> --- a/lib/igt_dummyload.h
> +++ b/lib/igt_dummyload.h
> @@ -54,6 +54,8 @@ typedef struct igt_spin_factory {
> unsigned int flags;
> int fence;
> uint64_t ahnd;
> + struct drm_xe_engine_class_instance *hwe;
> + uint32_t vm;
> } igt_spin_factory_t;
>
> typedef struct igt_spin {
> @@ -83,6 +85,14 @@ typedef struct igt_spin {
> #define SPIN_CLFLUSH (1 << 0)
>
> struct igt_spin_factory opts;
> +
> + struct xe_spin *xe_spin;
> + size_t bo_size;
> + uint64_t address;
> + unsigned int engine;
> + uint32_t vm;
> + uint32_t syncobj;
> +
> } igt_spin_t;
>
>
> diff --git a/lib/xe/xe_spin.c b/lib/xe/xe_spin.c
> index 856d0ba2..bc0fbcc6 100644
> --- a/lib/xe/xe_spin.c
> +++ b/lib/xe/xe_spin.c
> @@ -15,6 +15,7 @@
> #include "intel_reg.h"
> #include "xe_ioctl.h"
> #include "xe_spin.h"
> +#include "lib/igt_dummyload.h"
>
> /**
> * xe_spin_init:
> @@ -82,6 +83,96 @@ void xe_spin_end(struct xe_spin *spin)
> spin->end = 0;
> }
>
> +igt_spin_t *
> +xe_spin_create(int fd, const struct igt_spin_factory *opt)
> +{
> + size_t bo_size = xe_get_default_alignment(fd);
> + uint32_t bo;
> + uint64_t ahnd = opt->ahnd, addr;
> + struct igt_spin *spin;
> + struct xe_spin *xe_spin;
> + struct drm_xe_sync sync = {
> + .flags = DRM_XE_SYNC_SYNCOBJ | DRM_XE_SYNC_SIGNAL,
> + };
> + struct drm_xe_exec exec = {
> + .num_batch_buffer = 1,
> + .num_syncs = 1,
> + .syncs = to_user_pointer(&sync),
> + };
> +
> + igt_assert(ahnd);
> + spin = calloc(1, sizeof(struct igt_spin));
> + igt_assert(spin);
> +
> + spin->syncobj = syncobj_create(fd, 0);
> + if (opt->engine) {
> + spin->opts.engine = opt->engine;
> + spin->opts.vm = opt->vm;
You may copy opt -> spin->opts (see how emit_recursive_batch() is doing this).
If user passed engine ensure vm is also not zeroed.
> +
> + spin->handle = xe_bo_create(fd, 0, spin->opts.vm, bo_size);
> + xe_spin = xe_bo_map(fd, spin->handle, bo_size);
> + addr = intel_allocator_alloc_with_strategy(ahnd, spin->handle, bo_size, 0, ALLOC_STRATEGY_LOW_TO_HIGH);
> + xe_vm_bind_sync(fd, spin->opts.vm, spin->handle, 0, addr, bo_size);
> +
> + xe_spin_init(xe_spin, addr, true);
> + exec.engine_id = spin->opts.engine;
> + exec.address = addr;
> + } else {
> + spin->vm = xe_vm_create(fd, 0, 0);
> + if (!opt->hwe)
> + spin->engine = xe_engine_create_class(fd, spin->vm, DRM_XE_ENGINE_CLASS_RENDER);
On PVC there's no RENDER engine.
> + else
> + spin->engine = xe_engine_create(fd, spin->vm, opt->hwe, 0);
opt->engine == 0 and opt->hwe == NULL -> SIGSEGV.
> +
> + bo = xe_bo_create(fd, 0, spin->vm, bo_size);
> + spin->handle = bo;
> + xe_spin = xe_bo_map(fd, spin->handle, bo_size);
> + addr = intel_allocator_alloc_with_strategy(ahnd, spin->handle, bo_size, 0, ALLOC_STRATEGY_LOW_TO_HIGH);
> + xe_vm_bind_sync(fd, spin->vm, spin->handle, 0, addr, bo_size);
> +
> + xe_spin_init(xe_spin, addr, true);
> + exec.engine_id = spin->engine;
> + exec.address = addr;
A lot of code is duplicated in if/else. Avoid this duplication.
> + }
> + sync.handle = spin->syncobj;
> + igt_assert_eq(igt_ioctl(fd, DRM_IOCTL_XE_EXEC, &exec), 0);
> + xe_spin_wait_started(xe_spin);
> + igt_info("Spinner started\n");
Unnecessary noise.
> +
> + spin->bo_size = bo_size;
> + spin->address = addr;
> + spin->xe_spin = xe_spin;
> +
> + return spin;
> +}
> +
> +void xe_spin_sync_wait(int fd, struct igt_spin *spin)
> +{
> + igt_assert(syncobj_wait(fd, &spin->syncobj, 1, INT64_MAX, 0,
> + NULL));
> +}
> +
> +void xe_spin_free(int fd, struct igt_spin *spin)
> +{
> + xe_spin_end(spin->xe_spin);
> + xe_spin_sync_wait(fd, spin);
> +
> + if (!spin->opts.vm)
> + xe_vm_unbind_sync(fd, spin->vm, 0, spin->address, spin->bo_size);
> + else
> + xe_vm_unbind_sync(fd, spin->opts.vm, 0, spin->address, spin->bo_size);
Hint: Assing spin->vm = spin->opts.vm always. If it will be 0 just create it then
use hwe to create on top of it. And as you still have spin->opts available
(I assume you'll copy it on spinner creation) you can use it when vm should
be destroyed in spinner code. And above if/else may be totally dropped to
single xe_vm_unbind_sync() line.
+ @Bhanu
Take a look to kms usecases Bhanu mentioned, I assume there're:
kms_busy.c:
t = igt_spin_new(dpy->drm_fd,
.ahnd = ahnd,
.fence = fence,
.dependency = fb->gem_handle,
.flags = IGT_SPIN_FENCE_IN);
igt_spin_t *t = igt_spin_new(dpy->drm_fd,
.ahnd = ahnd,
.dependency = busy_fb->gem_handle,
.flags = IGT_SPIN_NO_PREEMPTION);
or in kms_cursor_legacy.c:
spin = igt_spin_new(display->drm_fd,
.ahnd = ahnd,
.dependency = fb_info[1].gem_handle,
.dependency_size = fb_info[1].size);
I'm not sure about prime_busy.c.
I think above usecases should be handled by xe spinner.
Bhanu - may you confirm my concerns?
--
Zbigniew
> +
> + syncobj_destroy(fd, spin->syncobj);
> + gem_munmap(spin->xe_spin, spin->bo_size);
> + gem_close(fd, spin->handle);
> +
> + if (!spin->opts.engine) {
> + xe_engine_destroy(fd, spin->engine);
> + xe_vm_destroy(fd, spin->vm);
> + }
> + free(spin);
> +}
> +
> void xe_cork_init(int fd, struct drm_xe_engine_class_instance *hwe,
> struct xe_cork *cork)
> {
> diff --git a/lib/xe/xe_spin.h b/lib/xe/xe_spin.h
> index 73f9a026..48867eb8 100644
> --- a/lib/xe/xe_spin.h
> +++ b/lib/xe/xe_spin.h
> @@ -13,19 +13,26 @@
> #include <stdbool.h>
>
> #include "xe_query.h"
> +#include "lib/igt_dummyload.h"
>
> /* Mapped GPU object */
> +
> struct xe_spin {
> uint32_t batch[16];
> uint64_t pad;
> uint32_t start;
> uint32_t end;
> +
> };
>
> +igt_spin_t *
> +xe_spin_create(int fd, const struct igt_spin_factory *opt);
> void xe_spin_init(struct xe_spin *spin, uint64_t addr, bool preempt);
> bool xe_spin_started(struct xe_spin *spin);
> +void xe_spin_sync_wait(int fd, struct igt_spin *spin);
> void xe_spin_wait_started(struct xe_spin *spin);
> void xe_spin_end(struct xe_spin *spin);
> +void xe_spin_free(int fd, struct igt_spin *spin);
>
> struct xe_cork {
> struct xe_spin *spin;
> --
> 2.39.1
>
More information about the igt-dev
mailing list