You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

323 lines
10 KiB

/*
* Copyright © 2015 Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*
* Authors:
* Tiago Vignatti
*/
/** @file prime_mmap_coherency.c
*
* TODO: need to show the need for prime_sync_end().
*/
#include "igt.h"
IGT_TEST_DESCRIPTION("Test dma-buf mmap on !llc platforms mostly and provoke"
" coherency bugs so we know for sure where we need the sync ioctls.");
int fd;
static drm_intel_bufmgr *bufmgr;
struct intel_batchbuffer *batch;
static int width = 1024, height = 1024;
/*
* Exercises the need for read flush:
* 1. create a BO and write '0's, in GTT domain.
* 2. read BO using the dma-buf CPU mmap.
* 3. write '1's, in GTT domain.
* 4. read again through the mapped dma-buf.
*/
static int test_read_flush(void)
{
drm_intel_bo *bo_1;
drm_intel_bo *bo_2;
uint32_t *ptr_cpu;
uint32_t *ptr_gtt;
int dma_buf_fd, i;
int stale = 0;
bo_1 = drm_intel_bo_alloc(bufmgr, "BO 1", width * height * 4, 4096);
/* STEP #1: put the BO 1 in GTT domain. We use the blitter to copy and fill
* zeros to BO 1, so commands will be submitted and likely to place BO 1 in
* the GTT domain. */
bo_2 = drm_intel_bo_alloc(bufmgr, "BO 2", width * height * 4, 4096);
intel_copy_bo(batch, bo_1, bo_2, width * height);
drm_intel_bo_unreference(bo_2);
/* STEP #2: read BO 1 using the dma-buf CPU mmap. This dirties the CPU caches. */
dma_buf_fd = prime_handle_to_fd_for_mmap(fd, bo_1->handle);
/* STEP #3: write 0x11 into BO 1. */
bo_2 = drm_intel_bo_alloc(bufmgr, "BO 2", width * height * 4, 4096);
ptr_gtt = gem_mmap__gtt(fd, bo_2->handle, width * height, PROT_READ | PROT_WRITE);
gem_set_domain(fd, bo_2->handle,
I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
memset(ptr_gtt, 0xc5, width * height);
munmap(ptr_gtt, width * height);
ptr_cpu = mmap(NULL, width * height, PROT_READ,
MAP_SHARED, dma_buf_fd, 0);
igt_assert(ptr_cpu != MAP_FAILED);
prime_sync_start(dma_buf_fd, false);
for (i = 0; i < (width * height) / 4; i++)
igt_assert_eq(ptr_cpu[i], 0);
prime_sync_end(dma_buf_fd, false);
intel_copy_bo(batch, bo_1, bo_2, width * height);
drm_intel_bo_unreference(bo_2);
/* STEP #4: read again using the CPU mmap. Doing #1 before #3 makes sure we
* don't do a full CPU cache flush in step #3 again. That makes sure all the
* stale cachelines from step #2 survive (mostly, a few will be evicted)
* until we try to read them again in step #4. This behavior could be fixed
* by flush CPU read right before accessing the CPU pointer */
prime_sync_start(dma_buf_fd, false);
for (i = 0; i < (width * height) / 4; i++)
if (ptr_cpu[i] != 0xc5c5c5c5)
stale++;
prime_sync_end(dma_buf_fd, false);
drm_intel_bo_unreference(bo_1);
munmap(ptr_cpu, width * height);
close(dma_buf_fd);
return stale;
}
/*
* Exercises the need for write flush:
* 1. create BO 1 and write '0's, in GTT domain.
* 2. write '1's into BO 1 using the dma-buf CPU mmap.
* 3. copy BO 1 to new BO 2, in GTT domain.
* 4. read via dma-buf mmap BO 2.
*/
static int test_write_flush(void)
{
drm_intel_bo *bo_1;
drm_intel_bo *bo_2;
uint32_t *ptr_cpu;
uint32_t *ptr2_cpu;
int dma_buf_fd, dma_buf2_fd, i;
int stale = 0;
bo_1 = drm_intel_bo_alloc(bufmgr, "BO 1", width * height * 4, 4096);
/* STEP #1: Put the BO 1 in GTT domain. We use the blitter to copy and fill
* zeros to BO 1, so commands will be submitted and likely to place BO 1 in
* the GTT domain. */
bo_2 = drm_intel_bo_alloc(bufmgr, "BO 2", width * height * 4, 4096);
intel_copy_bo(batch, bo_1, bo_2, width * height);
drm_intel_bo_unreference(bo_2);
/* STEP #2: Write '1's into BO 1 using the dma-buf CPU mmap. */
dma_buf_fd = prime_handle_to_fd_for_mmap(fd, bo_1->handle);
igt_skip_on(errno == EINVAL);
ptr_cpu = mmap(NULL, width * height, PROT_READ | PROT_WRITE,
MAP_SHARED, dma_buf_fd, 0);
igt_assert(ptr_cpu != MAP_FAILED);
/* This is the main point of this test: !llc hw requires a cache write
* flush right here (explained in step #4). */
prime_sync_start(dma_buf_fd, true);
memset(ptr_cpu, 0x11, width * height);
prime_sync_end(dma_buf_fd, true);
/* STEP #3: Copy BO 1 into BO 2, using blitter. */
bo_2 = drm_intel_bo_alloc(bufmgr, "BO 2", width * height * 4, 4096);
intel_copy_bo(batch, bo_2, bo_1, width * height);
/* STEP #4: compare BO 2 against written BO 1. In !llc hardware, there
* should be some cache lines that didn't get flushed out and are still 0,
* requiring cache flush before the write in step 2. */
dma_buf2_fd = prime_handle_to_fd_for_mmap(fd, bo_2->handle);
igt_skip_on(errno == EINVAL);
ptr2_cpu = mmap(NULL, width * height, PROT_READ | PROT_WRITE,
MAP_SHARED, dma_buf2_fd, 0);
igt_assert(ptr2_cpu != MAP_FAILED);
prime_sync_start(dma_buf2_fd, false);
for (i = 0; i < (width * height) / 4; i++)
if (ptr2_cpu[i] != 0x11111111)
stale++;
prime_sync_end(dma_buf2_fd, false);
drm_intel_bo_unreference(bo_1);
drm_intel_bo_unreference(bo_2);
munmap(ptr_cpu, width * height);
close(dma_buf2_fd);
close(dma_buf_fd);
return stale;
}
static void blit_and_cmp(void)
{
drm_intel_bo *bo_1;
drm_intel_bo *bo_2;
uint32_t *ptr_cpu;
uint32_t *ptr2_cpu;
int dma_buf_fd, dma_buf2_fd, i;
int local_fd;
drm_intel_bufmgr *local_bufmgr;
struct intel_batchbuffer *local_batch;
/* recreate process local variables */
local_fd = drm_open_driver(DRIVER_INTEL);
local_bufmgr = drm_intel_bufmgr_gem_init(local_fd, 4096);
igt_assert(local_bufmgr);
local_batch = intel_batchbuffer_alloc(local_bufmgr, intel_get_drm_devid(local_fd));
igt_assert(local_batch);
bo_1 = drm_intel_bo_alloc(local_bufmgr, "BO 1", width * height * 4, 4096);
dma_buf_fd = prime_handle_to_fd_for_mmap(local_fd, bo_1->handle);
igt_skip_on(errno == EINVAL);
ptr_cpu = mmap(NULL, width * height, PROT_READ | PROT_WRITE,
MAP_SHARED, dma_buf_fd, 0);
igt_assert(ptr_cpu != MAP_FAILED);
bo_2 = drm_intel_bo_alloc(local_bufmgr, "BO 2", width * height * 4, 4096);
dma_buf2_fd = prime_handle_to_fd_for_mmap(local_fd, bo_2->handle);
ptr2_cpu = mmap(NULL, width * height, PROT_READ | PROT_WRITE,
MAP_SHARED, dma_buf2_fd, 0);
igt_assert(ptr2_cpu != MAP_FAILED);
/* Fill up BO 1 with '1's and BO 2 with '0's */
prime_sync_start(dma_buf_fd, true);
memset(ptr_cpu, 0x11, width * height);
prime_sync_end(dma_buf_fd, true);
prime_sync_start(dma_buf2_fd, true);
memset(ptr2_cpu, 0x00, width * height);
prime_sync_end(dma_buf2_fd, true);
/* Copy BO 1 into BO 2, using blitter. */
intel_copy_bo(local_batch, bo_2, bo_1, width * height);
usleep(0); /* let someone else claim the mutex */
/* Compare BOs. If prime_sync_* were executed properly, the caches
* should be synced. */
prime_sync_start(dma_buf2_fd, false);
for (i = 0; i < (width * height) / 4; i++)
igt_fail_on_f(ptr2_cpu[i] != 0x11111111, "Found 0x%08x at offset 0x%08x\n", ptr2_cpu[i], i);
prime_sync_end(dma_buf2_fd, false);
drm_intel_bo_unreference(bo_1);
drm_intel_bo_unreference(bo_2);
munmap(ptr_cpu, width * height);
munmap(ptr2_cpu, width * height);
close(dma_buf_fd);
close(dma_buf2_fd);
intel_batchbuffer_free(local_batch);
drm_intel_bufmgr_destroy(local_bufmgr);
close(local_fd);
}
/*
* Constantly interrupt concurrent blits to stress out prime_sync_* and make
* sure these ioctl errors are handled accordingly.
*
* Important to note that in case of failure (e.g. in a case where the ioctl
* wouldn't try again in a return error) this test does not reliably catch the
* problem with 100% of accuracy.
*/
static void test_ioctl_errors(void)
{
int ncpus = sysconf(_SC_NPROCESSORS_ONLN);
/* Ensure we can do at least one child */
intel_require_memory(2, width*height*4, CHECK_RAM);
for (int num_children = 1; num_children <= 8 *ncpus; num_children <<= 1) {
uint64_t required, total;
igt_info("Spawing %d interruptible children\n", num_children);
if (!__intel_check_memory(2*num_children,
width*height*4,
CHECK_RAM,
&required, &total)) {
igt_debug("Estimated that we need %'lluMiB for test, but only have %'lluMiB\n",
(long long)(required >> 20),
(long long)(total >> 20));
break;
}
igt_fork(child, num_children)
igt_while_interruptible(true) blit_and_cmp();
igt_waitchildren();
}
}
igt_main
{
igt_fixture {
fd = drm_open_driver(DRIVER_INTEL);
igt_require_gem(fd);
bufmgr = drm_intel_bufmgr_gem_init(fd, 4096);
batch = intel_batchbuffer_alloc(bufmgr, intel_get_drm_devid(fd));
}
/* Cache coherency and the eviction are pretty much unpredictable, so
* reproducing boils down to trial and error to hit different scenarios.
* TODO: We may want to improve tests a bit by picking random subranges. */
igt_subtest("read") {
igt_until_timeout(5) {
int stale = test_read_flush();
igt_fail_on_f(stale,
"num of stale cache lines %d\n", stale);
}
}
igt_subtest("write") {
igt_until_timeout(5) {
int stale = test_write_flush();
igt_fail_on_f(stale,
"num of stale cache lines %d\n", stale);
}
}
igt_subtest("ioctl-errors") {
igt_info("exercising concurrent blit to get ioctl errors\n");
test_ioctl_errors();
}
igt_fixture {
intel_batchbuffer_free(batch);
drm_intel_bufmgr_destroy(bufmgr);
close(fd);
}
}