| Message ID | 20221118155335.635430-1-matthew.auld@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [i-g-t] tests/i915/gem_exec_balancer: exercise dmabuf import | expand |
On 18.11.2022 16:53, Matthew Auld wrote: > With parallel submission it should be easy to get a fence array as the > output fence. Try importing this into dma-buf reservation object, to see > if anything explodes. > > References: https://gitlab.freedesktop.org/drm/intel/-/issues/7532 > Signed-off-by: Matthew Auld <matthew.auld@intel.com> > Cc: Andrzej Hajda <andrzej.hajda@intel.com> > Cc: Nirmoy Das <nirmoy.das@intel.com> > --- > tests/i915/gem_exec_balancer.c | 39 ++++++++++++++++++++++++++++++++++ > 1 file changed, 39 insertions(+) > > diff --git a/tests/i915/gem_exec_balancer.c b/tests/i915/gem_exec_balancer.c > index 4300dbd1..fdae8de5 100644 > --- a/tests/i915/gem_exec_balancer.c > +++ b/tests/i915/gem_exec_balancer.c > @@ -37,6 +37,7 @@ > #include "igt_sysfs.h" > #include "igt_types.h" > #include "sw_sync.h" > +#include <linux/dma-buf.h> > > IGT_TEST_DESCRIPTION("Exercise in-kernel load-balancing"); > > @@ -2856,6 +2857,24 @@ static void logical_sort_siblings(int i915, > #define PARALLEL_SUBMIT_FENCE (0x1 << 3) > #define PARALLEL_CONTEXTS (0x1 << 4) > #define PARALLEL_VIRTUAL (0x1 << 5) > +#define PARALLEL_OUT_FENCE_DMABUF (0x1 << 6) > + > +struct igt_dma_buf_sync_file { > + __u32 flags; > + __s32 fd; > +}; > + > +#define IGT_DMA_BUF_IOCTL_EXPORT_SYNC_FILE _IOWR(DMA_BUF_BASE, 2, struct igt_dma_buf_sync_file) > +#define IGT_DMA_BUF_IOCTL_IMPORT_SYNC_FILE _IOW(DMA_BUF_BASE, 3, struct igt_dma_buf_sync_file) > + > +static void dmabuf_import_sync_file(int dmabuf, uint32_t flags, int sync_fd) > +{ > + struct igt_dma_buf_sync_file arg; > + > + arg.flags = flags; > + arg.fd = sync_fd; > + do_ioctl(dmabuf, IGT_DMA_BUF_IOCTL_IMPORT_SYNC_FILE, &arg); > +} Wouldn't be good to move code above to some common lib? Anyway: Reviewed-by: Andrzej Hajda <andrzej.hajda@intel.com> Regards Andrzej > > static void parallel_thread(int i915, unsigned int flags, > struct i915_engine_class_instance *siblings, > @@ -2871,6 +2890,8 @@ static void parallel_thread(int i915, unsigned int flags, > uint32_t target_bo_idx = 0; > uint32_t first_bb_idx = 1; > intel_ctx_cfg_t cfg; > + uint32_t dmabuf_handle; > + int dmabuf; > > igt_assert(bb_per_execbuf < 32); > > @@ -2924,11 +2945,20 @@ static void parallel_thread(int i915, unsigned int flags, > execbuf.buffers_ptr = to_user_pointer(obj); > execbuf.rsvd1 = ctx->id; > > + if (flags & PARALLEL_OUT_FENCE_DMABUF) { > + dmabuf_handle = gem_create(i915, 4096); > + dmabuf = prime_handle_to_fd(i915, dmabuf_handle); > + } > + > for (n = 0; n < PARALLEL_BB_LOOP_COUNT; ++n) { > execbuf.flags &= ~0x3full; > gem_execbuf_wr(i915, &execbuf); > > if (flags & PARALLEL_OUT_FENCE) { > + if (flags & PARALLEL_OUT_FENCE_DMABUF) > + dmabuf_import_sync_file(dmabuf, DMA_BUF_SYNC_WRITE, > + execbuf.rsvd2 >> 32); > + > igt_assert_eq(sync_fence_wait(execbuf.rsvd2 >> 32, > 1000), 0); > igt_assert_eq(sync_fence_status(execbuf.rsvd2 >> 32), 1); > @@ -2959,6 +2989,11 @@ static void parallel_thread(int i915, unsigned int flags, > if (fence) > close(fence); > > + if (flags & PARALLEL_OUT_FENCE_DMABUF) { > + gem_close(i915, dmabuf_handle); > + close(dmabuf); > + } > + > check_bo(i915, obj[target_bo_idx].handle, > bb_per_execbuf * PARALLEL_BB_LOOP_COUNT, true); > > @@ -3420,6 +3455,10 @@ igt_main > igt_subtest("parallel-out-fence") > parallel(i915, PARALLEL_OUT_FENCE); > > + igt_subtest("parallel-out-fence-import-dmabuf") > + parallel(i915, PARALLEL_OUT_FENCE | > + PARALLEL_OUT_FENCE_DMABUF); > + > igt_subtest("parallel-keep-in-fence") > parallel(i915, PARALLEL_OUT_FENCE | PARALLEL_IN_FENCE); >
Hi Matthew, few nits, see below. On 2022-11-18 at 15:53:35 +0000, Matthew Auld wrote: > With parallel submission it should be easy to get a fence array as the > output fence. Try importing this into dma-buf reservation object, to see > if anything explodes. > > References: https://gitlab.freedesktop.org/drm/intel/-/issues/7532 > Signed-off-by: Matthew Auld <matthew.auld@intel.com> > Cc: Andrzej Hajda <andrzej.hajda@intel.com> > Cc: Nirmoy Das <nirmoy.das@intel.com> > --- > tests/i915/gem_exec_balancer.c | 39 ++++++++++++++++++++++++++++++++++ > 1 file changed, 39 insertions(+) > > diff --git a/tests/i915/gem_exec_balancer.c b/tests/i915/gem_exec_balancer.c > index 4300dbd1..fdae8de5 100644 > --- a/tests/i915/gem_exec_balancer.c > +++ b/tests/i915/gem_exec_balancer.c > @@ -37,6 +37,7 @@ > #include "igt_sysfs.h" > #include "igt_types.h" > #include "sw_sync.h" > +#include <linux/dma-buf.h> > > IGT_TEST_DESCRIPTION("Exercise in-kernel load-balancing"); > > @@ -2856,6 +2857,24 @@ static void logical_sort_siblings(int i915, > #define PARALLEL_SUBMIT_FENCE (0x1 << 3) > #define PARALLEL_CONTEXTS (0x1 << 4) > #define PARALLEL_VIRTUAL (0x1 << 5) > +#define PARALLEL_OUT_FENCE_DMABUF (0x1 << 6) > + > +struct igt_dma_buf_sync_file { > + __u32 flags; > + __s32 fd; > +}; > + > +#define IGT_DMA_BUF_IOCTL_EXPORT_SYNC_FILE _IOWR(DMA_BUF_BASE, 2, struct igt_dma_buf_sync_file) > +#define IGT_DMA_BUF_IOCTL_IMPORT_SYNC_FILE _IOW(DMA_BUF_BASE, 3, struct igt_dma_buf_sync_file) > + > +static void dmabuf_import_sync_file(int dmabuf, uint32_t flags, int sync_fd) > +{ > + struct igt_dma_buf_sync_file arg; > + > + arg.flags = flags; > + arg.fd = sync_fd; > + do_ioctl(dmabuf, IGT_DMA_BUF_IOCTL_IMPORT_SYNC_FILE, &arg); > +} You did not check for error here, so either add assert do_ioctl ... == 0 or change function name, add __ before like: static int __dmabuf_import_sync_file(int dmabuf, uint32_t flags, int sync_fd) > > static void parallel_thread(int i915, unsigned int flags, > struct i915_engine_class_instance *siblings, > @@ -2871,6 +2890,8 @@ static void parallel_thread(int i915, unsigned int flags, > uint32_t target_bo_idx = 0; > uint32_t first_bb_idx = 1; > intel_ctx_cfg_t cfg; > + uint32_t dmabuf_handle; > + int dmabuf; > > igt_assert(bb_per_execbuf < 32); > > @@ -2924,11 +2945,20 @@ static void parallel_thread(int i915, unsigned int flags, > execbuf.buffers_ptr = to_user_pointer(obj); > execbuf.rsvd1 = ctx->id; > > + if (flags & PARALLEL_OUT_FENCE_DMABUF) { > + dmabuf_handle = gem_create(i915, 4096); > + dmabuf = prime_handle_to_fd(i915, dmabuf_handle); > + } > + > for (n = 0; n < PARALLEL_BB_LOOP_COUNT; ++n) { > execbuf.flags &= ~0x3full; > gem_execbuf_wr(i915, &execbuf); > > if (flags & PARALLEL_OUT_FENCE) { > + if (flags & PARALLEL_OUT_FENCE_DMABUF) > + dmabuf_import_sync_file(dmabuf, DMA_BUF_SYNC_WRITE, > + execbuf.rsvd2 >> 32); > + > igt_assert_eq(sync_fence_wait(execbuf.rsvd2 >> 32, > 1000), 0); > igt_assert_eq(sync_fence_status(execbuf.rsvd2 >> 32), 1); > @@ -2959,6 +2989,11 @@ static void parallel_thread(int i915, unsigned int flags, > if (fence) > close(fence); > > + if (flags & PARALLEL_OUT_FENCE_DMABUF) { > + gem_close(i915, dmabuf_handle); > + close(dmabuf); > + } > + > check_bo(i915, obj[target_bo_idx].handle, > bb_per_execbuf * PARALLEL_BB_LOOP_COUNT, true); > > @@ -3420,6 +3455,10 @@ igt_main > igt_subtest("parallel-out-fence") > parallel(i915, PARALLEL_OUT_FENCE); > Please put description here. Regards, Kamil > + igt_subtest("parallel-out-fence-import-dmabuf") > + parallel(i915, PARALLEL_OUT_FENCE | > + PARALLEL_OUT_FENCE_DMABUF); > + > igt_subtest("parallel-keep-in-fence") > parallel(i915, PARALLEL_OUT_FENCE | PARALLEL_IN_FENCE); > > -- > 2.38.1 >
Hi Matthew, one more nit, see below. On 2022-11-18 at 15:53:35 +0000, Matthew Auld wrote: > With parallel submission it should be easy to get a fence array as the > output fence. Try importing this into dma-buf reservation object, to see > if anything explodes. > > References: https://gitlab.freedesktop.org/drm/intel/-/issues/7532 > Signed-off-by: Matthew Auld <matthew.auld@intel.com> > Cc: Andrzej Hajda <andrzej.hajda@intel.com> > Cc: Nirmoy Das <nirmoy.das@intel.com> > --- > tests/i915/gem_exec_balancer.c | 39 ++++++++++++++++++++++++++++++++++ > 1 file changed, 39 insertions(+) > > diff --git a/tests/i915/gem_exec_balancer.c b/tests/i915/gem_exec_balancer.c > index 4300dbd1..fdae8de5 100644 > --- a/tests/i915/gem_exec_balancer.c > +++ b/tests/i915/gem_exec_balancer.c > @@ -37,6 +37,7 @@ > #include "igt_sysfs.h" > #include "igt_types.h" > #include "sw_sync.h" > +#include <linux/dma-buf.h> - ^^^^^^^^^^^^^^^^^^^^^^^^^^ This should be above with other <includes>. Also it is linux specific, so please put it with #ifdef __linux__ #include <linux/dma-buf.h> #endif Regards, Kamil > > IGT_TEST_DESCRIPTION("Exercise in-kernel load-balancing"); > > @@ -2856,6 +2857,24 @@ static void logical_sort_siblings(int i915, > #define PARALLEL_SUBMIT_FENCE (0x1 << 3) > #define PARALLEL_CONTEXTS (0x1 << 4) > #define PARALLEL_VIRTUAL (0x1 << 5) > +#define PARALLEL_OUT_FENCE_DMABUF (0x1 << 6) > + > +struct igt_dma_buf_sync_file { > + __u32 flags; > + __s32 fd; > +}; > + > +#define IGT_DMA_BUF_IOCTL_EXPORT_SYNC_FILE _IOWR(DMA_BUF_BASE, 2, struct igt_dma_buf_sync_file) > +#define IGT_DMA_BUF_IOCTL_IMPORT_SYNC_FILE _IOW(DMA_BUF_BASE, 3, struct igt_dma_buf_sync_file) > + > +static void dmabuf_import_sync_file(int dmabuf, uint32_t flags, int sync_fd) > +{ > + struct igt_dma_buf_sync_file arg; > + > + arg.flags = flags; > + arg.fd = sync_fd; > + do_ioctl(dmabuf, IGT_DMA_BUF_IOCTL_IMPORT_SYNC_FILE, &arg); > +} > > static void parallel_thread(int i915, unsigned int flags, > struct i915_engine_class_instance *siblings, > @@ -2871,6 +2890,8 @@ static void parallel_thread(int i915, unsigned int flags, > uint32_t target_bo_idx = 0; > uint32_t first_bb_idx = 1; > intel_ctx_cfg_t cfg; > + uint32_t dmabuf_handle; > + int dmabuf; > > igt_assert(bb_per_execbuf < 32); > > @@ -2924,11 +2945,20 @@ static void parallel_thread(int i915, unsigned int flags, > execbuf.buffers_ptr = to_user_pointer(obj); > execbuf.rsvd1 = ctx->id; > > + if (flags & PARALLEL_OUT_FENCE_DMABUF) { > + dmabuf_handle = gem_create(i915, 4096); > + dmabuf = prime_handle_to_fd(i915, dmabuf_handle); > + } > + > for (n = 0; n < PARALLEL_BB_LOOP_COUNT; ++n) { > execbuf.flags &= ~0x3full; > gem_execbuf_wr(i915, &execbuf); > > if (flags & PARALLEL_OUT_FENCE) { > + if (flags & PARALLEL_OUT_FENCE_DMABUF) > + dmabuf_import_sync_file(dmabuf, DMA_BUF_SYNC_WRITE, > + execbuf.rsvd2 >> 32); > + > igt_assert_eq(sync_fence_wait(execbuf.rsvd2 >> 32, > 1000), 0); > igt_assert_eq(sync_fence_status(execbuf.rsvd2 >> 32), 1); > @@ -2959,6 +2989,11 @@ static void parallel_thread(int i915, unsigned int flags, > if (fence) > close(fence); > > + if (flags & PARALLEL_OUT_FENCE_DMABUF) { > + gem_close(i915, dmabuf_handle); > + close(dmabuf); > + } > + > check_bo(i915, obj[target_bo_idx].handle, > bb_per_execbuf * PARALLEL_BB_LOOP_COUNT, true); > > @@ -3420,6 +3455,10 @@ igt_main > igt_subtest("parallel-out-fence") > parallel(i915, PARALLEL_OUT_FENCE); > > + igt_subtest("parallel-out-fence-import-dmabuf") > + parallel(i915, PARALLEL_OUT_FENCE | > + PARALLEL_OUT_FENCE_DMABUF); > + > igt_subtest("parallel-keep-in-fence") > parallel(i915, PARALLEL_OUT_FENCE | PARALLEL_IN_FENCE); > > -- > 2.38.1 >
diff --git a/tests/i915/gem_exec_balancer.c b/tests/i915/gem_exec_balancer.c index 4300dbd1..fdae8de5 100644 --- a/tests/i915/gem_exec_balancer.c +++ b/tests/i915/gem_exec_balancer.c @@ -37,6 +37,7 @@ #include "igt_sysfs.h" #include "igt_types.h" #include "sw_sync.h" +#include <linux/dma-buf.h> IGT_TEST_DESCRIPTION("Exercise in-kernel load-balancing"); @@ -2856,6 +2857,24 @@ static void logical_sort_siblings(int i915, #define PARALLEL_SUBMIT_FENCE (0x1 << 3) #define PARALLEL_CONTEXTS (0x1 << 4) #define PARALLEL_VIRTUAL (0x1 << 5) +#define PARALLEL_OUT_FENCE_DMABUF (0x1 << 6) + +struct igt_dma_buf_sync_file { + __u32 flags; + __s32 fd; +}; + +#define IGT_DMA_BUF_IOCTL_EXPORT_SYNC_FILE _IOWR(DMA_BUF_BASE, 2, struct igt_dma_buf_sync_file) +#define IGT_DMA_BUF_IOCTL_IMPORT_SYNC_FILE _IOW(DMA_BUF_BASE, 3, struct igt_dma_buf_sync_file) + +static void dmabuf_import_sync_file(int dmabuf, uint32_t flags, int sync_fd) +{ + struct igt_dma_buf_sync_file arg; + + arg.flags = flags; + arg.fd = sync_fd; + do_ioctl(dmabuf, IGT_DMA_BUF_IOCTL_IMPORT_SYNC_FILE, &arg); +} static void parallel_thread(int i915, unsigned int flags, struct i915_engine_class_instance *siblings, @@ -2871,6 +2890,8 @@ static void parallel_thread(int i915, unsigned int flags, uint32_t target_bo_idx = 0; uint32_t first_bb_idx = 1; intel_ctx_cfg_t cfg; + uint32_t dmabuf_handle; + int dmabuf; igt_assert(bb_per_execbuf < 32); @@ -2924,11 +2945,20 @@ static void parallel_thread(int i915, unsigned int flags, execbuf.buffers_ptr = to_user_pointer(obj); execbuf.rsvd1 = ctx->id; + if (flags & PARALLEL_OUT_FENCE_DMABUF) { + dmabuf_handle = gem_create(i915, 4096); + dmabuf = prime_handle_to_fd(i915, dmabuf_handle); + } + for (n = 0; n < PARALLEL_BB_LOOP_COUNT; ++n) { execbuf.flags &= ~0x3full; gem_execbuf_wr(i915, &execbuf); if (flags & PARALLEL_OUT_FENCE) { + if (flags & PARALLEL_OUT_FENCE_DMABUF) + dmabuf_import_sync_file(dmabuf, DMA_BUF_SYNC_WRITE, + execbuf.rsvd2 >> 32); + igt_assert_eq(sync_fence_wait(execbuf.rsvd2 >> 32, 1000), 0); igt_assert_eq(sync_fence_status(execbuf.rsvd2 >> 32), 1); @@ -2959,6 +2989,11 @@ static void parallel_thread(int i915, unsigned int flags, if (fence) close(fence); + if (flags & PARALLEL_OUT_FENCE_DMABUF) { + gem_close(i915, dmabuf_handle); + close(dmabuf); + } + check_bo(i915, obj[target_bo_idx].handle, bb_per_execbuf * PARALLEL_BB_LOOP_COUNT, true); @@ -3420,6 +3455,10 @@ igt_main igt_subtest("parallel-out-fence") parallel(i915, PARALLEL_OUT_FENCE); + igt_subtest("parallel-out-fence-import-dmabuf") + parallel(i915, PARALLEL_OUT_FENCE | + PARALLEL_OUT_FENCE_DMABUF); + igt_subtest("parallel-keep-in-fence") parallel(i915, PARALLEL_OUT_FENCE | PARALLEL_IN_FENCE);
With parallel submission it should be easy to get a fence array as the output fence. Try importing this into dma-buf reservation object, to see if anything explodes. References: https://gitlab.freedesktop.org/drm/intel/-/issues/7532 Signed-off-by: Matthew Auld <matthew.auld@intel.com> Cc: Andrzej Hajda <andrzej.hajda@intel.com> Cc: Nirmoy Das <nirmoy.das@intel.com> --- tests/i915/gem_exec_balancer.c | 39 ++++++++++++++++++++++++++++++++++ 1 file changed, 39 insertions(+)