[igt-dev] [PATCH i-g-t v3 1/3] lib/xe_spin: xe_spin_opts for xe_spin initialization

Zbigniew Kempczyński zbigniew.kempczynski at intel.com
Fri Sep 8 08:52:48 UTC 2023


On Tue, Sep 05, 2023 at 03:02:23PM +0000, Marcin Bernatowicz wrote:
> Introduced struct xe_spin_opts for xe_spin initialization,
> adjusted tests to new xe_spin_init signature.
> Added xe_spin_init_opts macro (Zbyszek).
> 
> Signed-off-by: Marcin Bernatowicz <marcin.bernatowicz at linux.intel.com>
> ---
>  lib/xe/xe_spin.c               | 28 ++++++++++------------------
>  lib/xe/xe_spin.h               | 19 ++++++++++++++++++-
>  tests/intel/xe_dma_buf_sync.c  |  6 +++---
>  tests/intel/xe_exec_balancer.c |  9 ++++-----
>  tests/intel/xe_exec_reset.c    | 24 ++++++++++++++----------
>  tests/intel/xe_exec_threads.c  |  7 ++++---
>  tests/intel/xe_vm.c            |  7 ++++---
>  7 files changed, 57 insertions(+), 43 deletions(-)
> 
> diff --git a/lib/xe/xe_spin.c b/lib/xe/xe_spin.c
> index 7113972ee..27f837ef9 100644
> --- a/lib/xe/xe_spin.c
> +++ b/lib/xe/xe_spin.c
> @@ -19,17 +19,13 @@
>  /**
>   * xe_spin_init:
>   * @spin: pointer to mapped bo in which spinner code will be written
> - * @addr: offset of spinner within vm
> - * @preempt: allow spinner to be preempted or not
> + * @opts: pointer to spinner initialization options
>   */
> -void xe_spin_init(struct xe_spin *spin, uint64_t addr, bool preempt)
> +void xe_spin_init(struct xe_spin *spin, struct xe_spin_opts *opts)
>  {
> -	uint64_t batch_offset = (char *)&spin->batch - (char *)spin;
> -	uint64_t batch_addr = addr + batch_offset;
> -	uint64_t start_offset = (char *)&spin->start - (char *)spin;
> -	uint64_t start_addr = addr + start_offset;
> -	uint64_t end_offset = (char *)&spin->end - (char *)spin;
> -	uint64_t end_addr = addr + end_offset;
> +	uint64_t loop_addr = opts->addr + offsetof(struct xe_spin, batch);
> +	uint64_t start_addr = opts->addr + offsetof(struct xe_spin, start);
> +	uint64_t end_addr = opts->addr + offsetof(struct xe_spin, end);
>  	int b = 0;
>  
>  	spin->start = 0;
> @@ -40,7 +36,7 @@ void xe_spin_init(struct xe_spin *spin, uint64_t addr, bool preempt)
>  	spin->batch[b++] = start_addr >> 32;
>  	spin->batch[b++] = 0xc0ffee;
>  
> -	if (preempt)
> +	if (opts->preempt)
>  		spin->batch[b++] = (0x5 << 23);
>  
>  	spin->batch[b++] = MI_COND_BATCH_BUFFER_END | MI_DO_COMPARE | 2;
> @@ -49,8 +45,8 @@ void xe_spin_init(struct xe_spin *spin, uint64_t addr, bool preempt)
>  	spin->batch[b++] = end_addr >> 32;
>  
>  	spin->batch[b++] = MI_BATCH_BUFFER_START | 1 << 8 | 1;
> -	spin->batch[b++] = batch_addr;
> -	spin->batch[b++] = batch_addr >> 32;
> +	spin->batch[b++] = loop_addr;
> +	spin->batch[b++] = loop_addr >> 32;
>  
>  	igt_assert(b <= ARRAY_SIZE(spin->batch));
>  }
> @@ -133,11 +129,7 @@ xe_spin_create(int fd, const struct igt_spin_factory *opt)
>  	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);
>  
> -	if (!(opt->flags & IGT_SPIN_NO_PREEMPTION))
> -		xe_spin_init(xe_spin, addr, true);
> -	else
> -		xe_spin_init(xe_spin, addr, false);
> -
> +	xe_spin_init_opts(xe_spin, .addr = addr, .preempt = !(opt->flags & IGT_SPIN_NO_PREEMPTION));
>  	exec.exec_queue_id = spin->engine;
>  	exec.address = addr;
>  	sync.handle = spin->syncobj;
> @@ -219,7 +211,7 @@ void xe_cork_init(int fd, struct drm_xe_engine_class_instance *hwe,
>  	exec_queue = xe_exec_queue_create(fd, vm, hwe, 0);
>  	syncobj = syncobj_create(fd, 0);
>  
> -	xe_spin_init(spin, addr, true);
> +	xe_spin_init_opts(spin, .addr = addr, .preempt = true);
>  	exec.exec_queue_id = exec_queue;
>  	exec.address = addr;
>  	sync.handle = syncobj;
> diff --git a/lib/xe/xe_spin.h b/lib/xe/xe_spin.h
> index c84db175d..9f1d33294 100644
> --- a/lib/xe/xe_spin.h
> +++ b/lib/xe/xe_spin.h
> @@ -15,6 +15,18 @@
>  #include "xe_query.h"
>  #include "lib/igt_dummyload.h"
>  
> +/** struct xe_spin_opts
> + *
> + * @addr: offset of spinner within vm
> + * @preempt: allow spinner to be preempted or not
> + *
> + * Used to initialize struct xe_spin spinner behavior.
> + */
> +struct xe_spin_opts {
> +	uint64_t addr;
> +	bool preempt;
> +};
> +
>  /* Mapped GPU object */
>  struct xe_spin {
>  	uint32_t batch[16];
> @@ -22,8 +34,13 @@ struct xe_spin {
>  	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);
> +void xe_spin_init(struct xe_spin *spin, struct xe_spin_opts *opts);
> +
> +#define xe_spin_init_opts(fd, ...) \
> +	xe_spin_init(fd, &((struct xe_spin_opts){__VA_ARGS__}))
> +
>  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);
> diff --git a/tests/intel/xe_dma_buf_sync.c b/tests/intel/xe_dma_buf_sync.c
> index 29d675154..627f4c1e5 100644
> --- a/tests/intel/xe_dma_buf_sync.c
> +++ b/tests/intel/xe_dma_buf_sync.c
> @@ -144,7 +144,6 @@ test_export_dma_buf(struct drm_xe_engine_class_instance *hwe0,
>  		uint64_t sdi_offset = (char *)&data[i]->data - (char *)data[i];
>  		uint64_t sdi_addr = addr + sdi_offset;
>  		uint64_t spin_offset = (char *)&data[i]->spin - (char *)data[i];
> -		uint64_t spin_addr = addr + spin_offset;
>  		struct drm_xe_sync sync[2] = {
>  			{ .flags = DRM_XE_SYNC_SYNCOBJ, },
>  			{ .flags = DRM_XE_SYNC_SYNCOBJ | DRM_XE_SYNC_SIGNAL, },
> @@ -153,14 +152,15 @@ test_export_dma_buf(struct drm_xe_engine_class_instance *hwe0,
>  			.num_batch_buffer = 1,
>  			.syncs = to_user_pointer(sync),
>  		};
> +		struct xe_spin_opts spin_opts = { .addr = addr + spin_offset, .preempt = true };
>  		uint32_t syncobj;
>  		int b = 0;
>  		int sync_fd;
>  
>  		/* Write spinner on FD[0] */
> -		xe_spin_init(&data[i]->spin, spin_addr, true);
> +		xe_spin_init(&data[i]->spin, &spin_opts);
>  		exec.exec_queue_id = exec_queue[0];
> -		exec.address = spin_addr;
> +		exec.address = spin_opts.addr;
>  		xe_exec(fd[0], &exec);
>  
>  		/* Export prime BO as sync file and veify business */
> diff --git a/tests/intel/xe_exec_balancer.c b/tests/intel/xe_exec_balancer.c
> index f364a4b7a..d7d8dd8fb 100644
> --- a/tests/intel/xe_exec_balancer.c
> +++ b/tests/intel/xe_exec_balancer.c
> @@ -52,6 +52,7 @@ static void test_all_active(int fd, int gt, int class)
>  	struct {
>  		struct xe_spin spin;
>  	} *data;
> +	struct xe_spin_opts spin_opts = { .preempt = false };
>  	struct drm_xe_engine_class_instance *hwe;
>  	struct drm_xe_engine_class_instance eci[MAX_INSTANCE];
>  	int i, num_placements = 0;
> @@ -90,16 +91,14 @@ static void test_all_active(int fd, int gt, int class)
>  	xe_vm_bind_async(fd, vm, 0, bo, 0, addr, bo_size, sync, 1);
>  
>  	for (i = 0; i < num_placements; i++) {
> -		uint64_t spin_offset = (char *)&data[i].spin - (char *)data;
> -		uint64_t spin_addr = addr + spin_offset;
> -
> -		xe_spin_init(&data[i].spin, spin_addr, false);
> +		spin_opts.addr = addr + (char *)&data[i].spin - (char *)data;
> +		xe_spin_init(&data[i].spin, &spin_opts);
>  		sync[0].flags &= ~DRM_XE_SYNC_SIGNAL;
>  		sync[1].flags |= DRM_XE_SYNC_SIGNAL;
>  		sync[1].handle = syncobjs[i];
>  
>  		exec.exec_queue_id = exec_queues[i];
> -		exec.address = spin_addr;
> +		exec.address = spin_opts.addr;
>  		xe_exec(fd, &exec);
>  		xe_spin_wait_started(&data[i].spin);
>  	}
> diff --git a/tests/intel/xe_exec_reset.c b/tests/intel/xe_exec_reset.c
> index a2d33baf1..be6bbada6 100644
> --- a/tests/intel/xe_exec_reset.c
> +++ b/tests/intel/xe_exec_reset.c
> @@ -44,6 +44,7 @@ static void test_spin(int fd, struct drm_xe_engine_class_instance *eci)
>  	size_t bo_size;
>  	uint32_t bo = 0;
>  	struct xe_spin *spin;
> +	struct xe_spin_opts spin_opts = { .addr = addr, .preempt = false };
>  
>  	vm = xe_vm_create(fd, DRM_XE_VM_CREATE_ASYNC_BIND_OPS, 0);
>  	bo_size = sizeof(*spin);
> @@ -60,7 +61,7 @@ static void test_spin(int fd, struct drm_xe_engine_class_instance *eci)
>  	sync[0].handle = syncobj_create(fd, 0);
>  	xe_vm_bind_async(fd, vm, 0, bo, 0, addr, bo_size, sync, 1);
>  
> -	xe_spin_init(spin, addr, false);
> +	xe_spin_init(spin, &spin_opts);
>  
>  	sync[0].flags &= ~DRM_XE_SYNC_SIGNAL;
>  	sync[1].flags |= DRM_XE_SYNC_SIGNAL;
> @@ -165,6 +166,7 @@ test_balancer(int fd, int gt, int class, int n_exec_queues, int n_execs,
>  		uint64_t pad;
>  		uint32_t data;
>  	} *data;
> +	struct xe_spin_opts spin_opts = { .preempt = false };
>  	struct drm_xe_engine_class_instance *hwe;
>  	struct drm_xe_engine_class_instance eci[MAX_INSTANCE];
>  	int i, j, b, num_placements = 0, bad_batches = 1;
> @@ -236,7 +238,6 @@ test_balancer(int fd, int gt, int class, int n_exec_queues, int n_execs,
>  		uint64_t batch_offset = (char *)&data[i].batch - (char *)data;
>  		uint64_t batch_addr = base_addr + batch_offset;
>  		uint64_t spin_offset = (char *)&data[i].spin - (char *)data;
> -		uint64_t spin_addr = base_addr + spin_offset;
>  		uint64_t sdi_offset = (char *)&data[i].data - (char *)data;
>  		uint64_t sdi_addr = base_addr + sdi_offset;
>  		uint64_t exec_addr;
> @@ -247,8 +248,9 @@ test_balancer(int fd, int gt, int class, int n_exec_queues, int n_execs,
>  			batches[j] = batch_addr;
>  
>  		if (i < bad_batches) {
> -			xe_spin_init(&data[i].spin, spin_addr, false);
> -			exec_addr = spin_addr;
> +			spin_opts.addr = base_addr + spin_offset;
> +			xe_spin_init(&data[i].spin, &spin_opts);
> +			exec_addr = spin_opts.addr;
>  		} else {
>  			b = 0;
>  			data[i].batch[b++] = MI_STORE_DWORD_IMM_GEN4;
> @@ -368,6 +370,7 @@ test_legacy_mode(int fd, struct drm_xe_engine_class_instance *eci,
>  		uint64_t pad;
>  		uint32_t data;
>  	} *data;
> +	struct xe_spin_opts spin_opts = { .preempt = false };
>  	int i, b;
>  
>  	igt_assert(n_exec_queues <= MAX_N_EXECQUEUES);
> @@ -417,15 +420,15 @@ test_legacy_mode(int fd, struct drm_xe_engine_class_instance *eci,
>  		uint64_t batch_offset = (char *)&data[i].batch - (char *)data;
>  		uint64_t batch_addr = base_addr + batch_offset;
>  		uint64_t spin_offset = (char *)&data[i].spin - (char *)data;
> -		uint64_t spin_addr = base_addr + spin_offset;
>  		uint64_t sdi_offset = (char *)&data[i].data - (char *)data;
>  		uint64_t sdi_addr = base_addr + sdi_offset;
>  		uint64_t exec_addr;
>  		int e = i % n_exec_queues;
>  
>  		if (!i) {
> -			xe_spin_init(&data[i].spin, spin_addr, false);
> -			exec_addr = spin_addr;
> +			spin_opts.addr = base_addr + spin_offset;
> +			xe_spin_init(&data[i].spin, &spin_opts);
> +			exec_addr = spin_opts.addr;
>  		} else {
>  			b = 0;
>  			data[i].batch[b++] = MI_STORE_DWORD_IMM_GEN4;
> @@ -539,6 +542,7 @@ test_compute_mode(int fd, struct drm_xe_engine_class_instance *eci,
>  		uint64_t exec_sync;
>  		uint32_t data;
>  	} *data;
> +	struct xe_spin_opts spin_opts = { .preempt = false };
>  	int i, b;
>  
>  	igt_assert(n_exec_queues <= MAX_N_EXECQUEUES);
> @@ -593,15 +597,15 @@ test_compute_mode(int fd, struct drm_xe_engine_class_instance *eci,
>  		uint64_t batch_offset = (char *)&data[i].batch - (char *)data;
>  		uint64_t batch_addr = base_addr + batch_offset;
>  		uint64_t spin_offset = (char *)&data[i].spin - (char *)data;
> -		uint64_t spin_addr = base_addr + spin_offset;
>  		uint64_t sdi_offset = (char *)&data[i].data - (char *)data;
>  		uint64_t sdi_addr = base_addr + sdi_offset;
>  		uint64_t exec_addr;
>  		int e = i % n_exec_queues;
>  
>  		if (!i) {
> -			xe_spin_init(&data[i].spin, spin_addr, false);
> -			exec_addr = spin_addr;
> +			spin_opts.addr = base_addr + spin_offset;
> +			xe_spin_init(&data[i].spin, &spin_opts);
> +			exec_addr = spin_opts.addr;
>  		} else {
>  			b = 0;
>  			data[i].batch[b++] = MI_STORE_DWORD_IMM_GEN4;
> diff --git a/tests/intel/xe_exec_threads.c b/tests/intel/xe_exec_threads.c
> index e64c1639a..ff4ebc280 100644
> --- a/tests/intel/xe_exec_threads.c
> +++ b/tests/intel/xe_exec_threads.c
> @@ -486,6 +486,7 @@ test_legacy_mode(int fd, uint32_t vm, uint64_t addr, uint64_t userptr,
>  		uint64_t pad;
>  		uint32_t data;
>  	} *data;
> +	struct xe_spin_opts spin_opts = { .preempt = false };
>  	int i, j, b, hang_exec_queue = n_exec_queues / 2;
>  	bool owns_vm = false, owns_fd = false;
>  
> @@ -562,15 +563,15 @@ test_legacy_mode(int fd, uint32_t vm, uint64_t addr, uint64_t userptr,
>  		uint64_t batch_offset = (char *)&data[i].batch - (char *)data;
>  		uint64_t batch_addr = addr + batch_offset;
>  		uint64_t spin_offset = (char *)&data[i].spin - (char *)data;
> -		uint64_t spin_addr = addr + spin_offset;
>  		uint64_t sdi_offset = (char *)&data[i].data - (char *)data;
>  		uint64_t sdi_addr = addr + sdi_offset;
>  		uint64_t exec_addr;
>  		int e = i % n_exec_queues;
>  
>  		if (flags & HANG && e == hang_exec_queue && i == e) {
> -			xe_spin_init(&data[i].spin, spin_addr, false);
> -			exec_addr = spin_addr;
> +			spin_opts.addr = addr + spin_offset;
> +			xe_spin_init(&data[i].spin, &spin_opts);
> +			exec_addr = spin_opts.addr;
>  		} else {
>  			b = 0;
>  			data[i].batch[b++] = MI_STORE_DWORD_IMM_GEN4;
> diff --git a/tests/intel/xe_vm.c b/tests/intel/xe_vm.c
> index e42c04e33..dc1850338 100644
> --- a/tests/intel/xe_vm.c
> +++ b/tests/intel/xe_vm.c
> @@ -727,6 +727,7 @@ test_bind_execqueues_independent(int fd, struct drm_xe_engine_class_instance *ec
>  		uint64_t pad;
>  		uint32_t data;
>  	} *data;
> +	struct xe_spin_opts spin_opts = { .preempt = true };
>  	int i, b;
>  
>  	vm = xe_vm_create(fd, DRM_XE_VM_CREATE_ASYNC_BIND_OPS, 0);
> @@ -755,14 +756,14 @@ test_bind_execqueues_independent(int fd, struct drm_xe_engine_class_instance *ec
>  		uint64_t sdi_offset = (char *)&data[i].data - (char *)data;
>  		uint64_t sdi_addr = addr + sdi_offset;
>  		uint64_t spin_offset = (char *)&data[i].spin - (char *)data;
> -		uint64_t spin_addr = addr + spin_offset;
>  		int e = i;
>  
>  		if (i == 0) {
>  			/* Cork 1st exec_queue with a spinner */
> -			xe_spin_init(&data[i].spin, spin_addr, true);
> +			spin_opts.addr = addr + spin_offset;
> +			xe_spin_init(&data[i].spin, &spin_opts);
>  			exec.exec_queue_id = exec_queues[e];
> -			exec.address = spin_addr;
> +			exec.address = spin_opts.addr;
>  			sync[0].flags &= ~DRM_XE_SYNC_SIGNAL;
>  			sync[1].flags |= DRM_XE_SYNC_SIGNAL;
>  			sync[1].handle = syncobjs[e];
> -- 
> 2.30.2
>

Ok, I like your change - opts is neat + offsetof() either.

Reviewed-by: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>

--
Zbigniew


More information about the igt-dev mailing list