[PATCH i-g-t 01/29] lib/igt_dummyload: Add support of allocator use in igt spinner
Zbigniew Kempczyński
zbigniew.kempczynski at intel.com
Mon Jun 14 04:57:36 UTC 2021
For gens without relocations we need to use softpin with valid offsets
which do not overlap other execbuf objects. As spinner during creation
knows nothing about vm it has to run into allocator handle must be
passed to properly acquire offsets from allocator instance.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
Cc: Petri Latvala <petri.latvala at intel.com>
---
lib/igt_dummyload.c | 135 ++++++++++++++++++++++++++++++--------------
lib/igt_dummyload.h | 3 +
2 files changed, 96 insertions(+), 42 deletions(-)
diff --git a/lib/igt_dummyload.c b/lib/igt_dummyload.c
index 82459e524..45be93d30 100644
--- a/lib/igt_dummyload.c
+++ b/lib/igt_dummyload.c
@@ -35,11 +35,13 @@
#include "i915/gem_engine_topology.h"
#include "i915/gem_mman.h"
#include "i915/gem_submission.h"
+#include "igt_aux.h"
#include "igt_core.h"
#include "igt_device.h"
#include "igt_dummyload.h"
#include "igt_gt.h"
#include "igt_vgem.h"
+#include "intel_allocator.h"
#include "intel_chipset.h"
#include "intel_reg.h"
#include "ioctl_wrappers.h"
@@ -99,8 +101,8 @@ emit_recursive_batch(igt_spin_t *spin,
unsigned int flags[GEM_MAX_ENGINES];
unsigned int nengine;
int fence_fd = -1;
- uint64_t addr;
- uint32_t *cs;
+ uint64_t addr, addr_scratch, ahnd = opts->ahnd, objflags = 0;
+ uint32_t *cs, delta;
int i;
/*
@@ -117,11 +119,16 @@ emit_recursive_batch(igt_spin_t *spin,
* are not allowed in the first 256KiB, for fear of negative relocations
* that wrap.
*/
- addr = gem_aperture_size(fd) / 2;
- if (addr >> 31)
- addr = 1 << 31;
- addr += random() % addr / 2;
- addr &= -4096;
+ if (!ahnd) {
+ addr = gem_aperture_size(fd) / 2;
+ if (addr >> 31)
+ addr = 1 << 31;
+ addr += random() % addr / 2;
+ addr &= -4096;
+ } else {
+ spin->ahnd = ahnd;
+ objflags |= EXEC_OBJECT_PINNED;
+ }
igt_assert(!(opts->ctx && opts->ctx_id));
@@ -171,15 +178,31 @@ emit_recursive_batch(igt_spin_t *spin,
execbuf->buffer_count++;
cs = spin->batch;
- obj[BATCH].offset = addr;
- addr += BATCH_SIZE;
+ if (ahnd)
+ addr = intel_allocator_alloc_with_strategy(ahnd, obj[BATCH].handle,
+ BATCH_SIZE, 0,
+ ALLOC_STRATEGY_LOW_TO_HIGH);
+ obj[BATCH].offset = CANONICAL(addr);
+ obj[BATCH].flags |= objflags;
+ if (obj[BATCH].offset >= (1ull << 32))
+ obj[BATCH].flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
if (opts->dependency) {
igt_assert(!(opts->flags & IGT_SPIN_POLL_RUN));
+ if (ahnd)
+ addr_scratch = intel_allocator_alloc_with_strategy(ahnd, opts->dependency,
+ BATCH_SIZE, 0,
+ ALLOC_STRATEGY_LOW_TO_HIGH);
+ else
+ addr_scratch = addr + BATCH_SIZE;
obj[SCRATCH].handle = opts->dependency;
- obj[SCRATCH].offset = addr;
- if (!(opts->flags & IGT_SPIN_SOFTDEP)) {
+ obj[SCRATCH].offset = CANONICAL(addr_scratch);
+ obj[SCRATCH].flags |= objflags | EXEC_OBJECT_WRITE;
+ if (obj[SCRATCH].offset >= (1ull << 32))
+ obj[SCRATCH].flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+
+ if (!(opts->flags & IGT_SPIN_SOFTDEP) && !ahnd) {
obj[SCRATCH].flags = EXEC_OBJECT_WRITE;
/* dummy write to dependency */
@@ -194,7 +217,8 @@ emit_recursive_batch(igt_spin_t *spin,
execbuf->buffer_count++;
} else if (opts->flags & IGT_SPIN_POLL_RUN) {
- r = &relocs[obj[BATCH].relocation_count++];
+ if (!ahnd)
+ r = &relocs[obj[BATCH].relocation_count++];
igt_assert(!opts->dependency);
@@ -219,29 +243,40 @@ emit_recursive_batch(igt_spin_t *spin,
0, 4096,
PROT_READ | PROT_WRITE);
}
+
+ if (ahnd)
+ addr = intel_allocator_alloc_with_strategy(ahnd, opts->dependency,
+ BATCH_SIZE * 3, 0,
+ ALLOC_STRATEGY_LOW_TO_HIGH);
addr += 4096; /* guard page */
- obj[SCRATCH].offset = addr;
+ obj[SCRATCH].offset = CANONICAL(addr);
addr += 4096;
igt_assert_eq(spin->poll[SPIN_POLL_START_IDX], 0);
- r->presumed_offset = obj[SCRATCH].offset;
- r->target_handle = obj[SCRATCH].handle;
- r->offset = sizeof(uint32_t) * 1;
- r->delta = sizeof(uint32_t) * SPIN_POLL_START_IDX;
+ delta = sizeof(uint32_t) * SPIN_POLL_START_IDX;
+ if (!ahnd) {
+ r->presumed_offset = obj[SCRATCH].offset;
+ r->target_handle = obj[SCRATCH].handle;
+ r->offset = sizeof(uint32_t) * 1;
+ r->delta = delta;
+ }
+ obj[SCRATCH].flags |= objflags;
+ if (obj[SCRATCH].offset >= (1ull << 32))
+ obj[SCRATCH].flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
*cs++ = MI_STORE_DWORD_IMM | (gen < 6 ? 1 << 22 : 0);
if (gen >= 8) {
- *cs++ = r->presumed_offset + r->delta;
- *cs++ = 0;
+ *cs++ = obj[SCRATCH].offset + delta;
+ *cs++ = obj[SCRATCH].offset >> 32;
} else if (gen >= 4) {
*cs++ = 0;
- *cs++ = r->presumed_offset + r->delta;
+ *cs++ = obj[SCRATCH].offset + delta;
r->offset += sizeof(uint32_t);
} else {
cs[-1]--;
- *cs++ = r->presumed_offset + r->delta;
+ *cs++ = obj[SCRATCH].offset + delta;
}
*cs++ = 1;
@@ -285,7 +320,8 @@ emit_recursive_batch(igt_spin_t *spin,
* no matter how they modify it (from either the GPU or CPU).
*/
if (gen >= 8) { /* arbitrary cutoff between ring/execlists submission */
- r = &relocs[obj[BATCH].relocation_count++];
+ if (!ahnd)
+ r = &relocs[obj[BATCH].relocation_count++];
/*
* On Sandybridge+ the comparison is a strict greater-than:
@@ -300,37 +336,46 @@ emit_recursive_batch(igt_spin_t *spin,
spin->condition[0] = 0xffffffff;
spin->condition[1] = 0xffffffff;
- r->presumed_offset = obj[BATCH].offset;
- r->target_handle = obj[BATCH].handle;
- r->offset = (cs + 2 - spin->batch) * sizeof(*cs);
- r->read_domains = I915_GEM_DOMAIN_COMMAND;
- r->delta = (spin->condition - spin->batch) * sizeof(*cs);
+ delta = (spin->condition - spin->batch) * sizeof(*cs);
+ if (!ahnd) {
+ r->presumed_offset = obj[BATCH].offset;
+ r->target_handle = obj[BATCH].handle;
+ r->offset = (cs + 2 - spin->batch) * sizeof(*cs);
+ r->read_domains = I915_GEM_DOMAIN_COMMAND;
+ r->delta = delta;
+ }
*cs++ = MI_COND_BATCH_BUFFER_END | MI_DO_COMPARE | 2;
*cs++ = MI_BATCH_BUFFER_END;
- *cs++ = r->presumed_offset + r->delta;
+ *cs++ = obj[BATCH].offset + delta;
*cs++ = 0;
}
/* recurse */
- r = &relocs[obj[BATCH].relocation_count++];
- r->target_handle = obj[BATCH].handle;
- r->presumed_offset = obj[BATCH].offset;
- r->offset = (cs + 1 - spin->batch) * sizeof(*cs);
- r->read_domains = I915_GEM_DOMAIN_COMMAND;
- r->delta = LOOP_START_OFFSET;
+ delta = LOOP_START_OFFSET;
+ if (!ahnd) {
+ r = &relocs[obj[BATCH].relocation_count++];
+ r->target_handle = obj[BATCH].handle;
+ r->presumed_offset = obj[BATCH].offset;
+ r->offset = (cs + 1 - spin->batch) * sizeof(*cs);
+ r->read_domains = I915_GEM_DOMAIN_COMMAND;
+ r->delta = delta;
+ }
+
if (gen >= 8) {
*cs++ = MI_BATCH_BUFFER_START | 1 << 8 | 1;
- *cs++ = r->presumed_offset + r->delta;
- *cs++ = 0;
+ *cs++ = obj[BATCH].offset + delta;
+ *cs++ = obj[BATCH].offset >> 32;
} else if (gen >= 6) {
*cs++ = MI_BATCH_BUFFER_START | 1 << 8;
- *cs++ = r->presumed_offset + r->delta;
+ *cs++ = obj[BATCH].offset + delta;
} else {
*cs++ = MI_BATCH_BUFFER_START | 2 << 6;
- if (gen < 4)
- r->delta |= 1;
- *cs = r->presumed_offset + r->delta;
+ if (gen < 4) {
+ delta |= 1;
+ r->delta = delta;
+ }
+ *cs++ = obj[BATCH].offset + delta;
cs++;
}
obj[BATCH].relocs_ptr = to_user_pointer(relocs);
@@ -580,11 +625,17 @@ static void __igt_spin_free(int fd, igt_spin_t *spin)
if (spin->batch)
gem_munmap(spin->batch, BATCH_SIZE);
- if (spin->poll_handle)
+ if (spin->poll_handle) {
gem_close(fd, spin->poll_handle);
+ if (spin->ahnd)
+ intel_allocator_free(spin->ahnd, spin->poll_handle);
+ }
- if (spin->handle)
+ if (spin->handle) {
gem_close(fd, spin->handle);
+ if (spin->ahnd)
+ intel_allocator_free(spin->ahnd, spin->handle);
+ }
if (spin->out_fence >= 0)
close(spin->out_fence);
diff --git a/lib/igt_dummyload.h b/lib/igt_dummyload.h
index 67e1a08ed..f02052614 100644
--- a/lib/igt_dummyload.h
+++ b/lib/igt_dummyload.h
@@ -58,6 +58,8 @@ typedef struct igt_spin {
unsigned int flags;
#define SPIN_CLFLUSH (1 << 0)
+
+ uint64_t ahnd;
} igt_spin_t;
/**
@@ -78,6 +80,7 @@ typedef struct igt_spin_factory {
unsigned int engine;
unsigned int flags;
int fence;
+ uint64_t ahnd;
} igt_spin_factory_t;
#define IGT_SPIN_FENCE_IN (1 << 0)
--
2.26.0
More information about the Intel-gfx-trybot
mailing list