[Intel-gfx] [PATCH i-g-t] tests/gem_softpin: New tests for softpin feature
Tvrtko Ursulin
tvrtko.ursulin at linux.intel.com
Fri Nov 20 08:20:15 PST 2015
Hi,
On 19/11/15 22:29, Vinay Belgaumkar wrote:
> These tests exercise the userptr ioctl to create shared buffers
> between CPU and GPU. They contain error and normal usage scenarios.
> They also contain a couple of stress tests which copy buffers between
> CPU and GPU. These tests rely on the softpin patch in order to pin buffers
> to a certain VA.
>
> Caveat: These tests were designed to run on 64-bit system. Future work
> includes adding logic to ensure these tests can run on 32-bit systems with
> PPGTT support. Some tests are currently disabled for 32-bit systems for that
> reason.
>
> v2: Added cc and signed-off-by fields
>
> v3: Fixed review comments, added helper functions. Removed userptr error
> scenarios covered by existing userptr tests. Modified stress test to have
> 100K buffers, it now runs for ~30 mins, checks every element has been written
> to correctly, and pins buffers at different VMAs.
>
> v4: Changed name to gem_softpin
>
> Cc: Michel Thierry <michel.thierry at intel.com>
Copy your reviewers as well please. :)
> Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar at intel.com>
> ---
> tests/.gitignore | 1 +
> tests/Makefile.sources | 1 +
> tests/gem_softpin.c | 1252 ++++++++++++++++++++++++++++++++++++++++++++++++
> 3 files changed, 1254 insertions(+)
> create mode 100644 tests/gem_softpin.c
>
> diff --git a/tests/.gitignore b/tests/.gitignore
> index 80af9a7..424870b 100644
> --- a/tests/.gitignore
> +++ b/tests/.gitignore
> @@ -21,6 +21,7 @@ gem_bad_blit
> gem_bad_length
> gem_bad_reloc
> gem_basic
> +gem_softpin
> gem_caching
> gem_close_race
> gem_concurrent_all
> diff --git a/tests/Makefile.sources b/tests/Makefile.sources
> index 8fb2de8..2008d4a 100644
> --- a/tests/Makefile.sources
> +++ b/tests/Makefile.sources
> @@ -11,6 +11,7 @@ TESTS_progs_M = \
> drv_hangman \
> gem_bad_reloc \
> gem_basic \
> + gem_softpin \
> gem_caching \
> gem_close_race \
> gem_concurrent_blit \
> diff --git a/tests/gem_softpin.c b/tests/gem_softpin.c
> new file mode 100644
> index 0000000..aed607c
> --- /dev/null
> +++ b/tests/gem_softpin.c
> @@ -0,0 +1,1252 @@
> +/*
> + * 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:
> + * Vinay Belgaumkar <vinay.belgaumkar at intel.com>
> + Thomas Daniel <thomas.daniel at intel.com>
> + *
> + */
> +
> +#include <unistd.h>
> +#include <stdlib.h>
> +#include <stdint.h>
> +#include <stdio.h>
> +#include <string.h>
> +#include <fcntl.h>
> +#include <inttypes.h>
> +#include <errno.h>
> +#include <sys/stat.h>
> +#include <sys/ioctl.h>
> +#include <sys/time.h>
> +#include <malloc.h>
> +#include "drm.h"
> +#include "ioctl_wrappers.h"
> +#include "drmtest.h"
> +#include "intel_chipset.h"
> +#include "intel_io.h"
> +#include "i915_drm.h"
> +#include <assert.h>
> +#include <sys/wait.h>
> +#include <sys/ipc.h>
> +#include <sys/shm.h>
> +#include "igt_kms.h"
> +#include <inttypes.h>
> +#include <sys/types.h>
> +#include <sys/stat.h>
> +
> +#define OBJECT_SIZE 16384
Unused.
> +#define BO_SIZE 4 * 4096
Could it be just one page to save resources?
> +#define MULTIPAGE_BO_SIZE 4 * BO_SIZE
Need more than two pages?
> +#define STORE_BATCH_BUFFER_SIZE 6
> +#define STRESS_BATCH_BUFFER_SIZE 5
> +#define EXEC_OBJECT_PINNED (1<<4)
> +#define EXEC_OBJECT_SUPPORTS_48B_ADDRESS (1<<3)
> +#define SHARED_BUFFER_SIZE 4096
> +#define NUM_EXEC_OBJECTS 2
> +
> +typedef struct drm_i915_gem_userptr i915_gem_userptr;
> +
> +static void gem_create_userptr_struct(i915_gem_userptr*, void* ptr, __u64 size, bool read_only);
Okay I commented on mixing of __u and stdint types and now you have
opted for the former which is slightly unusual for userspace. I would
recommend you use stdint, should be easy with search & replace in any
editor.
Hm actually on a closer look you still use both. I would really
recommend to stick with one for readability.
> +static void *create_mem_buffer(__u64 size);
> +static int gem_call_userptr_ioctl(int fd, i915_gem_userptr *userptr);
> +static void gem_basic_test(void);
> +static void gem_pin_invalid_vma_test(void);
> +static void gem_pin_overlap_test(void);
> +static void gem_shmem_test(void);
> +static void gem_pin_high_address_test(void);
> +static void gem_pin_mmap_anonymous_test(void);
> +static void gem_pin_mmap_file_test(void);
> +
> +static int gem_call_userptr_ioctl(int fd, i915_gem_userptr* userptr)
Coding style, i915_gem_userptr *userptr, probably some more of this around.
> +{
> + int ret;
> +
> + ret = drmIoctl(fd, DRM_IOCTL_I915_GEM_USERPTR, userptr);
> +
> + if (ret)
> + ret = errno;
> +
> + return ret;
> +}
> +
> +static void gem_create_userptr_struct(i915_gem_userptr *userptr, void* ptr, __u64 size, bool read_only)
a) looks by the eye to be over 80 char long. But I am not sure if IGT
allows this, probably better to take care of those.
b) maybe change the name somehow so it does't have create in the name
since it is not creating the struct any more.
Actually.. on a closer look all calls to this are immediately followed
by gem_call_userptr_ioctl so It may be best for readability of test
cases to merge the two.
> +{
> + memset((void*)userptr, 0, sizeof(i915_gem_userptr));
> +
> + userptr->user_ptr = (uintptr_t)ptr;
Bad cast and not even needed.
> + userptr->user_size = size;
> + userptr->flags = I915_USERPTR_UNSYNCHRONIZED;
> +
> + if (read_only)
> + userptr->flags |= I915_USERPTR_READ_ONLY;
> +}
> +
> +/* Creates a 4K aligned CPU buffer
> + * @size - size of buffer
> + * RETURNS pointer to buffer of @size
> +*/
Unusual comment formatting.
> +
> +static void* create_mem_buffer(__u64 size)
> +{
> + void* addr;
> + int ret;
> +
> + ret = posix_memalign(&addr, 4096, size);
> + igt_assert(ret == 0);
> +
> + return addr;
> +}
> +
> +/* setup_exec_obj - populate exec object
> + * @exec - exec object
> + * @handle - handle to gem buffer
> + * flags - any flags
> + * offset - requested VMA
> + */
> +static void setup_exec_obj(struct drm_i915_gem_exec_object2 *exec, __u32 handle, __u32 flags, __u64 offset)
> +{
> + memset(exec, 0, sizeof(struct drm_i915_gem_exec_object2));
> + exec->handle = handle;
> + exec->flags = flags;
> + exec->offset = offset;
> +}
> +
> +/*
> + * gem_store_data_svm - populate batch buffer with MI_STORE_DWORD_IMM command
> + * @fd: drm file descriptor
> + * @buf: batch buffer
> + * @buffer_size: size of buffer
> + * @addr: destination Virtual address
> + * @data: data to be store at destination
> + * @end: whether to end batch buffer or not
> + */
> +static int gem_store_data_svm(int fd, uint32_t* cmd_buf, uint64_t vaddr,
> + uint32_t data, bool end)
> +{
> + int i = 0;
> +
> + cmd_buf[i++] = MI_STORE_DWORD_IMM;
> + cmd_buf[i++] = vaddr & 0xFFFFFFFC;
> + cmd_buf[i++] = (vaddr >> 32) & 0xFFFF; /* bits 32:47 */
> +
> + cmd_buf[i++] = data;
> + if (end){
> + cmd_buf[i++] = MI_BATCH_BUFFER_END;
> + cmd_buf[i++] = 0;
> + }
> +
> + return (i * sizeof(uint32_t));
> +}
> +
> +/*
> + * gem_store_data - populate batch buffer with MI_STORE_DWORD_IMM command
> + * This one fills up reloc buffer as well
> + * @fd: drm file descriptor
> + * @buf: batch buffer
> + * @buffer_size: size of buffer
> + * @addr: destination Virtual address
> + * @data: data to be store at destination
> + * @reloc - relocation entry
> + * @end: whether to end batch buffer or not
> + */
> +
> +static int gem_store_data(int fd, uint32_t* cmd_buf,
> + uint32_t handle, uint32_t data,
> + struct drm_i915_gem_relocation_entry *reloc,
> + bool end)
> +{
> + int i = 0;
> +
> + cmd_buf[i++] = MI_STORE_DWORD_IMM;
> + cmd_buf[i++] = 0; /* lower 31 bits of 48 bit address - 0 because reloc is needed */
> + cmd_buf[i++] = 0; /* upper 15 bits of 48 bit address - 0 because reloc is needed */
> + reloc->offset = 1 * sizeof(uint32_t);
> + reloc->delta = 0;
> + reloc->target_handle = handle;
> + reloc->read_domains = I915_GEM_DOMAIN_RENDER;
> + reloc->write_domain = I915_GEM_DOMAIN_RENDER;
> + reloc->presumed_offset = 0;
> + reloc++;
Pointless line.
> + cmd_buf[i++] = data;
> + if (end){
> + cmd_buf[i++] = MI_BATCH_BUFFER_END;
> + cmd_buf[i++] = 0;
> + }
> +
> + return (i * sizeof(uint32_t));
> +}
> +
> +/* Helper function for filling execbuffer struct */
> +static void setup_execbuffer(struct drm_i915_gem_execbuffer2 *execbuf, struct drm_i915_gem_exec_object2 *exec_object,
> + int ring, int buffer_count, int batch_length)
> +{
> + execbuf->buffers_ptr = (uintptr_t)exec_object;
> + execbuf->buffer_count = buffer_count;
> + execbuf->batch_start_offset = 0;
> + execbuf->batch_len = batch_length;
> + execbuf->cliprects_ptr = 0;
> + execbuf->num_cliprects = 0;
> + execbuf->DR1 = 0;
> + execbuf->DR4 = 0;
> + execbuf->flags = ring;
> + i915_execbuffer2_set_context_id(*execbuf, 0);
> + execbuf->rsvd2 = 0;
> +}
> +
> +/* Helper function for exec and sync functions */
> +static void submit_and_sync(int fd, struct drm_i915_gem_execbuffer2 *execbuf, uint32_t batch_buf_handle)
> +{
> + gem_execbuf(fd, execbuf);
> + gem_sync(fd, batch_buf_handle);
> +}
> +
> +/* gem_basic_test - This test will create a shared buffer, and create a command
> + * for GPU to write data in it
> + * CPU will read and make sure expected value is obtained
> + * @valid_shared_buffer - whether test with valid malloc'd buffer or not
> +
> + if (valid_shared_buffer == true)
> + Malloc a 4K buffer
> + Share buffer with with GPU by using userptr ioctl
> + Create batch buffer to write DATA to first dword of buffer
> + Use 0x1000 address as destination address in batch buffer
> + Set EXEC_OBJECT_PINNED flag in exec object
> + Set 'offset' in exec object to 0x1000
> + Submit execbuffer
> + Verify value of first DWORD in shared buffer matches DATA
> +
> + if (valid_shared_buffer == false)
> + Declare null buffer
> + Call Userptr ioctl with null buffer
> + Run Basic Test
> + Test should fail at submit execbuffer
> +*/
> +static void gem_basic_test(void)
> +{
It would be better in my opinion if the basic tests didn't use userptr
but a normal bo.
Not saying you can't have both, on the contrary, I even think we need both.
> + i915_gem_userptr userptr;
> + int fd, ret;
> + uint32_t* shared_buffer = NULL;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS];
> + uint32_t batch_buffer[STORE_BATCH_BUFFER_SIZE];
> + uint32_t batch_buf_handle, shared_buf_handle;
> + int ring, len;
> + const uint32_t data = 0x12345678;
> + uint64_t pinning_offset = 0x1000;
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, BO_SIZE);
> +
> + /* create cpu buffer, set to all 0xF's */
> + shared_buffer = create_mem_buffer(BO_SIZE);
Comment seems wrong, or something, I don't see setting to 0xFs.
> +
> + /* share with GPU */
> + gem_create_userptr_struct(&userptr, shared_buffer, BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
I recommend moving this assert into the userptr helper and at the same
time...
> +
> + /* Get handle for shared buffer */
> + shared_buf_handle = userptr.handle;
... make the helper return the handle. Save some lines of code and make
the test cases more readable.
> + /* create command buffer with write command */
> + len = gem_store_data_svm(fd, batch_buffer, pinning_offset, data, true);
> + igt_assert_lte(len, STORE_BATCH_BUFFER_SIZE * 4);
Range checking should happen in gem_store_data_svm since it is pretty
weird to do it after the fact.
> +
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + setup_exec_obj(&exec_object2[0], shared_buf_handle, EXEC_OBJECT_PINNED, pinning_offset);
> + setup_exec_obj(&exec_object2[1], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
I've asked about this in the previous round - why it is interesting to
use the blitter ring if available and not always the render ring?
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS, len);
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
In the previous round I suggested that you check the offset and I stand
by that. It is stronger to check the offset as returned by the kernel
and not just the written data.
> + /* check on CPU to see if value changes */
> + igt_fail_on_f(shared_buffer[0] != data,
> + "\nCPU read does not match GPU write, expected: 0x%x, got: 0x%x\n", data, shared_buffer[0]);
Why don't you need to move the shared_buf_handle to the CPU domain
before accessing it? I thought that is required to ensure GPU as
finished writing to it and you don't hit cache coherency issues?
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> +
> + free(shared_buffer);
> +}
> +
> +/* gem_multiple_process_test - Run basic test simultaneously with multiple processes
> +* This will test pinning same VA separately in each process
> +
> + fork();
> + Execute basic test in parent/child processes
> +
> +*/
> +
> +#define MAX_NUM_PROCESSES 10
> +
> +static void gem_multiple_process_test(void)
> +{
> + igt_fork(child, MAX_NUM_PROCESSES) {
> + gem_basic_test();
> + }
> + igt_waitchildren();
> +}
> +
> +
> +/* gem_repin_test
> + * This test tries to repin a buffer at a previously pinned vma
> + * from a different execbuf.
> +
> + * Malloc a 4K buffer
> + * Share buffer with with GPU by using userptr ioctl
> + * Create batch buffer to write DATA to first dword of buffer
> + * Use 0x1000 address as destination address in batch buffer
> + * Set EXEC_OBJECT_PINNED flag in exec object
> + * Set 'offset' in exec object to 0x1000 VMA
> + * Submit execbuffer
> + * Verify value of first DWORD in shared buffer matches DATA
> +
> + * Create second shared buffer
> + * Follow all steps above
> + * Execpt, for offset, use VMA of first buffer above
> + * Submit execbuffer
> + * Verify value of first DWORD in second shared buffer matches DATA
> +*/
I hate it how you have all possible comment styles in one file.
/**
* Blah
/* Blah
/*
* blah
/*
@code
description of steps
/*
* description of steps
/*
description of steps
...
It is just too much..
> +static void gem_repin_test(void)
> +{
> + i915_gem_userptr userptr;
> + i915_gem_userptr userptr1;
> + int fd, ret;
> + uint32_t* shared_buffer = NULL;
> + uint32_t* shared_buffer1 = NULL;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS];
> + uint32_t batch_buffer[STORE_BATCH_BUFFER_SIZE];
> + uint32_t batch_buf_handle, shared_buf_handle, shared_buf_handle1;
> + int ring, len;
> + const uint32_t data = 0x12345678;
> + uint64_t pinning_offset = 0x1000;
> +
> + /* Create gem object */
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, BO_SIZE);
> +
> + /* create cpu buffer, set to all 0xF's */
> + shared_buffer = create_mem_buffer(BO_SIZE);
> + shared_buffer1 = create_mem_buffer(BO_SIZE);
> + shared_buffer[0] = 0x0;
> + shared_buffer1[0] = 0x0;
Oops.
> +
> + /* share with GPU */
> + gem_create_userptr_struct(&userptr, shared_buffer, BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> +
> + //gem_create_userptr_struct(&userptr1, shared_buffer1, BO_SIZE * 2, false);
Remove?
> + gem_create_userptr_struct(&userptr1, shared_buffer1, BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr1);
> + igt_assert_eq(ret, 0);
> +
> + /* Get handle for shared buffer */
> + shared_buf_handle = userptr.handle;
> + shared_buf_handle1 = userptr1.handle;
> +
> + /* create command buffer with write command */
> + len = gem_store_data_svm(fd, batch_buffer, pinning_offset, data, true);
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + setup_exec_obj(&exec_object2[0], shared_buf_handle, EXEC_OBJECT_PINNED, pinning_offset);
> + setup_exec_obj(&exec_object2[1], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS, len);
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + igt_assert(*shared_buffer == data);
> +
> + /* Second buffer */
> + /* create command buffer with write command */
> + pinning_offset = exec_object2[0].offset;
> + len = gem_store_data_svm(fd, batch_buffer, pinning_offset, data, true);
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + /* Pin at shared_buffer, not shared_buffer1 */
> + setup_exec_obj(&exec_object2[0], shared_buf_handle1, EXEC_OBJECT_PINNED, pinning_offset);
> + setup_exec_obj(&exec_object2[1], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS, len);
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + igt_assert(*shared_buffer1 == data);
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> +
> + free(shared_buffer);
> + free(shared_buffer1);
> +}
> +
> +
> +/** gem_repin_overlap_test
> + * This test will attempt to pin two buffers at the same VMA as part of the same
> + execbuffer object
> +
> + at code
> + Malloc a 4K buffer
> + Share buffer with with GPU by using userptr ioctl
> + Create second shared buffer
> + Create batch buffer to write DATA to first dword of each buffer
> + Use same virtual address as destination addresses in batch buffer
> + Set EXEC_OBJECT_PINNED flag in both exec objects
> + Set 'offset' in both exec objects to same VMA
> + Submit execbuffer
> + Command should return EINVAL, since we are trying to pin to same VMA
> + at endcode
> +**/
> +static void gem_pin_overlap_test(void)
> +{
> + i915_gem_userptr userptr;
> + i915_gem_userptr userptr1;
> + int fd, ret;
> + uint32_t* shared_buffer = NULL;
> + uint32_t* shared_buffer1 = NULL;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS + 1];
> + uint32_t batch_buffer[BO_SIZE];
> + uint32_t batch_buf_handle, shared_buf_handle, shared_buf_handle1;
> + int ring, len;
> + const uint32_t data = 0x12345678;
> + uint64_t pinning_offset = 0x1000;
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, BO_SIZE);
> +
> + shared_buffer = create_mem_buffer(BO_SIZE);
> + shared_buffer1 = create_mem_buffer(BO_SIZE * 2);
> +
> + /* share with GPU */
> + gem_create_userptr_struct(&userptr, shared_buffer, BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> +
> + gem_create_userptr_struct(&userptr1, shared_buffer1, BO_SIZE * 2, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr1);
> + igt_assert_eq(ret, 0);
> +
> + shared_buf_handle = userptr.handle;
> + shared_buf_handle1 = userptr1.handle;
> +
> + len = gem_store_data_svm(fd, batch_buffer, pinning_offset, data, false);
> + len += gem_store_data_svm(fd, (batch_buffer + len/4), pinning_offset, data, true);
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + setup_exec_obj(&exec_object2[0], shared_buf_handle, EXEC_OBJECT_PINNED, pinning_offset);
> + setup_exec_obj(&exec_object2[1], shared_buf_handle1, EXEC_OBJECT_PINNED, pinning_offset);
> + setup_exec_obj(&exec_object2[2], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS+1, len);
> +
> + ret = drmIoctl(fd,
> + DRM_IOCTL_I915_GEM_EXECBUFFER2,
> + &execbuf);
> +
> + /* expect to fail */
> + igt_assert_neq(ret, 0);
> + igt_assert(errno == 22);
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> +
> + free(shared_buffer);
> + free(shared_buffer1);
> +}
> +
> +
> +/* gem_evict_test
> + * create shared buffer, pin it
> + * create normal buffer
> + * try and relocate to shared location
> + * Opens: How to detect eviction occured?
> + * i915_gem_gtt - debugfs api - grep it for the page?
> +
> + Create a gem buffer of 4K
> + Malloc a 4K buffer
> + Share buffer with GPU using userptr ioctl
> + Create a batch buffer to write 0x11111111 and 0x22222222 in above 2 buffers
> + Pin Shared buffer to offset '0' in GTT
Why at zero?
> + Create reloc buffer to ensure gem buffer is relocated to GTT
> + Submit execbuffer
> + Verify shared buffer has 0x22222222 as expected
> + Obtain offset of where gem object has been placed from exec object field
> + Try to pin shared buffer at that address using 'offset' field in exec object
> + Prevent relocation by setting relocation_count = 0
> + Submit execbuffer
> + Shared buffer will be pinned to previous address of gem object
> + Unshared buffer will be evicted, since relocation is not allowed
> + Second batch buffer will write 0x11111111 to shared buffer instead of unshared
> + Verify shared buffer contains 0x11111111
> + Reverse order of instructions in batch buffer to write to unshared first
> +*/
> +
> +static void gem_evict_test(void)
> +{
> + i915_gem_userptr userptr;
> + int fd, ret;
> + uint32_t* shared_buffer = NULL;
> + struct drm_i915_gem_relocation_entry reloc[4];
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS + 1];
> + uint32_t batch_buffer[STORE_BATCH_BUFFER_SIZE * 2];
> + uint32_t batch_buf_handle, shared_buf_handle, unshared_buf_handle;
> + int ring, len;
> + uint32_t data1, data2;
> + uint32_t value;
> + uint64_t pinning_offset = 0;
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, BO_SIZE);
> + unshared_buf_handle = gem_create(fd, BO_SIZE);
> +
> + shared_buffer = create_mem_buffer(BO_SIZE);
> + *shared_buffer = 0xFFFFFFFF;
Here is some setting to ffffs! But this time no comment. I am not saying
it needs a comment just that it is funny. :)
> +
> + /* share with GPU */
> + gem_create_userptr_struct(&userptr, shared_buffer, BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> +
> + /* Get handle for shared buffer */
> + shared_buf_handle = userptr.handle;
> +
> + /* create command buffer with write commands */
> + data1 = 0x11111111;
> + data2 = 0x22222222;
> + len = gem_store_data(fd, batch_buffer, unshared_buf_handle, data1, reloc, false);
> + len += gem_store_data_svm(fd, batch_buffer + (len/4), pinning_offset, data2, true);
> + igt_assert_lte(len, STORE_BATCH_BUFFER_SIZE * 2 * 4);
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + setup_exec_obj(&exec_object2[0], shared_buf_handle, EXEC_OBJECT_PINNED, pinning_offset);
> + setup_exec_obj(&exec_object2[1], unshared_buf_handle, 0, 0);
> + setup_exec_obj(&exec_object2[2], batch_buf_handle, 0, 0);
> +
> + exec_object2[2].relocation_count = 1;
> + exec_object2[2].relocs_ptr = (uintptr_t)reloc;
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS+1, len);
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + igt_assert(*shared_buffer == data2);
> + gem_read(fd, unshared_buf_handle, 0, (void*)&value, 4);
> + igt_assert(value == data1);
> +
> +
> + *shared_buffer = 0xffffffff;
> + /* Now cause eviction of unshared buffer by pinning shared buffer there */
> + exec_object2[0].offset = exec_object2[1].offset;
> + /* Prevent relocation */
> + exec_object2[2].relocation_count = 0;
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
First MI_STORE_DATA_IMM now writes to where? To the previously resolved
address, where unshared buffer was? This is now shared so it writes
data1 there.
Second MI_STORE_DATA_IMM writes data2 to the pinning_offset. And what is
at that address now? Shared is not any more since it has been moved. I
don't get it.
I've asked this in the last round...
> + igt_assert(*shared_buffer == data1);
>
> + igt_assert(exec_object2[0].offset != exec_object2[1].offset);
> + /* Unshared buffer gets evicted, and retains its value, since no batch buffer writes to it */
> + gem_read(fd, unshared_buf_handle, 0, (void*)&value, 4);
> + igt_assert(value == data1);
> +
> + /* Now lets do it again with the objects listed in reverse order... */
> + *shared_buffer = 0xffffffff;
> + setup_exec_obj(&exec_object2[0], unshared_buf_handle, 0, 0);
> + setup_exec_obj(&exec_object2[1], shared_buf_handle, EXEC_OBJECT_PINNED, 0);
> +
> + exec_object2[2].relocation_count = 1;
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + igt_assert(*shared_buffer == data2);
> + *shared_buffer = 0xffffffff;
> + /* Now cause eviction of unshared buffer by pinning shared buffer there */
> + exec_object2[1].offset = exec_object2[0].offset;
> + /* Prevent relocation */
> + exec_object2[2].relocation_count = 0;
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + igt_assert(*shared_buffer == data1);
> + igt_assert(exec_object2[0].offset != exec_object2[1].offset);
> + gem_read(fd, unshared_buf_handle, 0, (void*)&value, 4);
> + igt_assert(value == data1);
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> +
> + free(shared_buffer);
> +}
> +
> +/** gem_softpin_stress_test - Stress test which creates 10K buffers and shares with GPU
> + Create 100K uint32 buffers of size 4K each
> + Share with GPU using userptr ioctl
> + Create batch buffer to write DATA in first element of each buffer
> + Pin each buffer to varying addresses starting from 0x800000000000 going below
> + Execute Batch Buffer on Blit ring STRESS_NUM_LOOPS times
> + Validate every buffer has DATA in first element
> + Rinse and Repeat on Render ring
> +**/
> +#define STRESS_NUM_BUFFERS 100240
> +#define STRESS_NUM_LOOPS 1
> +#define STORE_COMMANDS 4 * STRESS_NUM_BUFFERS
> +
> +static void gem_softpin_stress_test(void)
> +{
> + i915_gem_userptr userptr;
> + int fd, ret;
> + uint32_t* shared_buffer[STRESS_NUM_BUFFERS];
> + uint32_t shared_handle[STRESS_NUM_BUFFERS];
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[STRESS_NUM_BUFFERS + 1];
> + uint32_t batch_buffer[STORE_COMMANDS + 2]; /* 4 dwords per buffer + 2 for the end of batchbuffer */
This test will allocate 1.5GiB of buffers which will fit in GTT on all
GENs and all mem config? Like some Android tablet reference design?
Maybe you need to auto-tune the numbers?
And it will use quite a lot of stack - some people complained when I
allocated 48k of it. :)
Lets see.. you'll need 400k + 400k + 1.6MiB + 5.3Mib = 7.3MiB of stack,
did I get that right?
Easy fix is to allocate blobs for each. For example:
struct drm_i915_gem_exec_object2 *exec_object2 = calloc(
STRESS_NUM_BUFFERS + 1, sizeof(struct ...exec_object2));
Etc for the rest.
Even for share_buffer, you don't need separate pointers but could also
allocate the huge blob.
> + uint32_t batch_buf_handle;
> + int ring, len, i, j;
> + uint64_t pinning_offset = 0x800000000000;
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, sizeof(batch_buffer));
> +
> + /* create command buffer with write commands */
> + len = 0;
> + for(i = 0; i < STRESS_NUM_BUFFERS; i++)
> + {
> + shared_buffer[i] = create_mem_buffer(BO_SIZE);
> + *shared_buffer[i] = 0xFFFFFFFF;
> +
> + /* share with GPU */
> + gem_create_userptr_struct(&userptr, shared_buffer[i], BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> +
> + /* Get handle for shared buffer */
> + shared_handle[i] = userptr.handle;
> +
> + setup_exec_obj(&exec_object2[i], shared_handle[i], EXEC_OBJECT_PINNED, pinning_offset);
> + len += gem_store_data_svm(fd, batch_buffer + (len/4), pinning_offset, i , (i == STRESS_NUM_BUFFERS-1) ? true:false);
> +
> + pinning_offset -= 0x200000; /* incremental 4K aligned address */
ALIGN(BO_SIZE, 4096) instead of hardcoded number?
> + }
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + setup_exec_obj(&exec_object2[STRESS_NUM_BUFFERS], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, STRESS_NUM_BUFFERS + 1, len);
> +
> + for (i = 0; i < STRESS_NUM_LOOPS; i++){
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> + for(j = 0; j < STRESS_NUM_BUFFERS; j++){
> + igt_fail_on_f(*shared_buffer[j] != j,
> + "Mismatch in buffer %d, iteration %d: 0x%08X\n", j, i, *shared_buffer[j]);
> + }
> + }
> +
> + // Now Render Ring
> + ring = I915_EXEC_RENDER;
> + execbuf.flags = ring;
> + for (i = 0; i < STRESS_NUM_LOOPS; i++){
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> + for(j = 0; j < STRESS_NUM_BUFFERS; j++){
> + igt_fail_on_f(*shared_buffer[j] != j,
> + "Mismatch in buffer %d, iteration %d: 0x%08X\n", j, i, *shared_buffer[j]);
> + }
> + }
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> +
> + for(i = 0; i < STRESS_NUM_BUFFERS; i++)
> + {
> + free(shared_buffer[i]);
> + }
Coding style.. or you can just remove the braces.
> +}
> +
> +/* gem_write_multipage_buffer - Create a buffer spanning multiple
> + pages, and share with GPU. Write to every element of the buffer
> + and verify correct contents.
> +
> + Create 16K uint32 buffer
> + Share with GPU using userptr ioctl
> + Create batch buffer to write DATA in all elements of buffer
> + Execute Batch Buffer
> + Validate every element has DATA
> + */
> +static void gem_write_multipage_buffer_test(void)
> +{
> + i915_gem_userptr userptr;
> + int fd, ret;
> + uint32_t* shared_buffer;
> + uint32_t shared_handle;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS];
> + uint32_t batch_buffer[(4 * BO_SIZE) + 2];
It is hard to figure out why that many elements. You have a store bb
size defined somewhere already no? And then you need BO_SIZE /
sizeof(uint32_t) of those plus 2, right?
> + uint32_t batch_buf_handle;
> + int ring, len, j;
> + uint64_t pinning_offset=0x1000;
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, sizeof(batch_buffer));
> +
> + /* create command buffer with write commands */
> + len = 0;
> + shared_buffer = create_mem_buffer(MULTIPAGE_BO_SIZE);
> + memset(batch_buffer, 0, sizeof(batch_buffer));
> +
> + memset(shared_buffer, 0, MULTIPAGE_BO_SIZE);
Maybe group the buffers or logical operations more clearly.
> +
> + /* share with GPU */
> + gem_create_userptr_struct(&userptr, shared_buffer, MULTIPAGE_BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> +
> + /* Get handle for shared buffer */
> + shared_handle = userptr.handle;
> +
> + setup_exec_obj(&exec_object2[0], shared_handle, EXEC_OBJECT_PINNED, pinning_offset);
> +
> + /* Every element of buffer */
> + for(j=0; j< (BO_SIZE); j++) /* BO_SIZE because it is 16K 4 byte entries */
MULTIPAGE_BO_SIZE / sizeof(uint32_t) for clarity ?
> + {
> + len += gem_store_data_svm(fd, batch_buffer + (len/4), pinning_offset, j, (j == ((BO_SIZE)-1)) ? true:false);
This is confusing since it doesn't use SVM, I mean doesn't attempt to
pin the buffer to share_buffer address.
> + pinning_offset += sizeof(shared_buffer[0]); /* 4 bytes */
> + }
> +
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + // submit command buffer
C++ comment.
> + setup_exec_obj(&exec_object2[1], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS, len);
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + for(j = 0; j < (BO_SIZE); j++)
> + {
> + igt_fail_on_f(shared_buffer[j] != j,
> + "Mismatch in index %d: 0x%08X\n", j, shared_buffer[j]);
> + }
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> +
> + free(shared_buffer);
> +}
> +
> +/** This test will request to pin a shared buffer to an invalid
> + VMA > 48-bit address
> +
> + Create shared buffer of size 4K
> + Try and Pin object to address 0x9000000000000
> +**/
> +static void gem_pin_invalid_vma_test(void)
> +{
> + i915_gem_userptr userptr;
> + int fd, ret;
> + uint32_t* shared_buffer = NULL;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS];
> + uint32_t batch_buffer[BO_SIZE];
> + uint32_t batch_buf_handle, shared_buf_handle;
> + int ring, len;
> + const uint32_t data = 0x12345678;
> + uint64_t invalid_address = 0x9000000000000; /* 52 bit address */
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, BO_SIZE);
> +
> + shared_buffer = create_mem_buffer(BO_SIZE);
> + *shared_buffer = 0xFFFFFFFF;
> +
> + // share with GPU
> + gem_create_userptr_struct(&userptr, shared_buffer, BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> +
> + shared_buf_handle = userptr.handle;
> +
> + len = gem_store_data_svm(fd, batch_buffer, invalid_address, data, true);
You don't even need anything in the batch in this test so could further
simplify it.
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + setup_exec_obj(&exec_object2[0], shared_buf_handle, EXEC_OBJECT_PINNED, invalid_address);
> + setup_exec_obj(&exec_object2[1], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS, len);
> +
> + /* Expect execbuf to fail */
> + ret = drmIoctl(fd,
> + DRM_IOCTL_I915_GEM_EXECBUFFER2,
> + &execbuf);
> +
> + igt_assert_neq(ret, 0);
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> +
> + free(shared_buffer);
> +}
> +
> +#define SHM_KEY 56789
> +#define SHMEM_SIZE 4096
> +/* gem_shmem_svm_test - Test userptr ioctl with shared memory
> + * This test creates a sysV IPC buffer and shares with GPU.
> + * It will send GPU commands to write DATA in the buffer and
> + * validate it on the CPU side when the command completes.
> +
> + * Create arbitrary shmem id
> + * Use shmat to attach a 4K uint32 buffer to above id
> + * Share buffer with GPU using userptr ioctl
> + * Create Batch buffer to write DATA in the first element
> + * submit execbuffer
> + * Validate on CPU side that DATA was indeed written
> + */
> +static void gem_shmem_test(void)
Move to gem_userptr_blits, without the softpin part.
> +{
> + int shmid;
> + i915_gem_userptr userptr;
> + int fd, ret;
> + uint32_t* shared_buffer = NULL;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS];
> + uint32_t batch_buffer[BO_SIZE];
> + uint32_t batch_buf_handle, shared_buf_handle;
> + int ring, len;
> + const uint32_t data = 0x12345678;
> + uint64_t pinning_offset = 0x1000;
> +
> + shmid = shmget(SHM_KEY, SHMEM_SIZE, IPC_CREAT);
> + igt_assert_neq(shmid, -1);
> +
> + shared_buffer = shmat(shmid, NULL, 0);
> + igt_assert(shared_buffer != (void*)-1);
> +
> + memset(shared_buffer, 0, SHMEM_SIZE);
> + shared_buffer[0] = 0xFFFFFFFF;
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, BO_SIZE);
> +
> + gem_create_userptr_struct(&userptr, shared_buffer, SHMEM_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> +
> + shared_buf_handle = userptr.handle;
> +
> + /* create command buffer with write command */
> + len = gem_store_data_svm(fd, batch_buffer, pinning_offset, data, true);
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + setup_exec_obj(&exec_object2[0], shared_buf_handle, EXEC_OBJECT_PINNED, pinning_offset);
> + setup_exec_obj(&exec_object2[1], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS, len);
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> +
> + /* check on CPU to see if value changes */
> + igt_fail_on_f(shared_buffer[0] != data,
> + "\nCPU read does not match GPU write, expected: 0x%x, got: 0x%x\n", data, shared_buffer[0]);
> +
> + ret = shmdt(shared_buffer);
> + igt_assert_eq(ret, 0);
> +}
> +
> +/* gem_pin_high_address_test - This test will create a shared buffer, and create a command
> + * for GPU to write data in it. It will attempt to pin the buffer at address > 32 bits.
> + * CPU will read and make sure expected value is obtained
> +
> + Malloc a 4K buffer
> + Share buffer with with GPU by using userptr ioctl
> + Create batch buffer to write DATA to first dword of buffer
> + Use virtual address of buffer as 0x1100000000 (> 32 bit)
> + Set EXEC_OBJECT_PINNED flag in exec object
> + Set 'offset' in exec object to shared buffer VMA
> + Submit execbuffer
> + Verify value of first DWORD in shared buffer matches DATA
> +*/
> +
> +static void gem_pin_high_address_test(void)
> +{
You also need a variant of this test with a normal BO, not userptr.
> + i915_gem_userptr userptr;
> + int fd, ret;
> + uint32_t* shared_buffer = NULL;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS];
> + uint32_t batch_buffer[BO_SIZE];
> + uint32_t batch_buf_handle, shared_buf_handle;
> + int ring, len;
> + const uint32_t data = 0x12345678;
> + uint64_t high_address = 0x1111FFFF000; /* 44 bit address */
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, BO_SIZE);
> +
> + /* create cpu buffer, set to all 0xF's */
> + shared_buffer = create_mem_buffer(BO_SIZE);
> + *shared_buffer = 0xFFFFFFFF;
> +
> + /* share with GPU */
> + gem_create_userptr_struct(&userptr, shared_buffer, BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> +
> + /* Get handle for shared buffer */
> + shared_buf_handle = userptr.handle;
> +
> + /* create command buffer with write command */
> + len = gem_store_data_svm(fd, batch_buffer, high_address, data, true);
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + setup_exec_obj(&exec_object2[0], shared_buf_handle, EXEC_OBJECT_PINNED | EXEC_OBJECT_SUPPORTS_48B_ADDRESS, high_address);
> + setup_exec_obj(&exec_object2[1], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS, len);
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + /* check on CPU to see if value changes */
> + igt_fail_on_f(shared_buffer[0] != data,
> + "\nCPU read does not match GPU write, expected: 0x%x, got: 0x%x\n", data, shared_buffer[0]);
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> + free(shared_buffer);
> +}
> +
> +/* gem_pin_near_48Bit_test - This test will create a shared buffer, and create a command
> + * for GPU to write data in it. It will attempt to pin the buffer at address > 47 bits <= 48-bit.
> + * CPU will read and make sure expected value is obtained
> +
> + Malloc a 4K buffer
> + Share buffer with with GPU by using userptr ioctl
> + Create batch buffer to write DATA to first dword of buffer
> + Use virtual address of buffer as range between 47-bit and 48-bit
> + Set EXEC_OBJECT_PINNED flag in exec object
> + Set 'offset' in exec object to shared buffer VMA
> + Submit execbuffer
> + Verify value of first DWORD in shared buffer matches DATA
> +*/
> +#define BEGIN_HIGH_ADDRESS 0x7FFFFFFFF000
> +#define END_HIGH_ADDRESS 0xFFFFFFFFC000
> +#define ADDRESS_INCREMENT 0x2000000000
> +static void gem_pin_near_48Bit_test(void)
> +{
> + i915_gem_userptr userptr;
> + int fd, ret;
> + uint32_t* shared_buffer = NULL;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS];
> + uint32_t batch_buffer[BO_SIZE];
> + uint32_t batch_buf_handle, shared_buf_handle;
> + int ring, len;
> + const uint32_t data = 0x12345678;
> + uint64_t high_address;
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, BO_SIZE);
> +
> + /* create cpu buffer, set to all 0xF's */
> + shared_buffer = create_mem_buffer(BO_SIZE);
> + *shared_buffer = 0xFFFFFFFF;
> +
> + /* share with GPU */
> + gem_create_userptr_struct(&userptr, shared_buffer, BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> +
> + /* Get handle for shared buffer */
> + shared_buf_handle = userptr.handle;
> +
> + for (high_address = BEGIN_HIGH_ADDRESS; high_address <= END_HIGH_ADDRESS;
> + high_address+=ADDRESS_INCREMENT){
> + /* create command buffer with write command */
> + len = gem_store_data_svm(fd, batch_buffer, high_address,
> + data, true);
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + setup_exec_obj(&exec_object2[0], shared_buf_handle,
> + EXEC_OBJECT_PINNED | EXEC_OBJECT_SUPPORTS_48B_ADDRESS,
> + high_address);
> + setup_exec_obj(&exec_object2[1], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS, len);
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + /* check on CPU to see if value changes */
> + igt_fail_on_f(shared_buffer[0] != data,
> + "\nCPU read does not match GPU write, expected: 0x%x, got: 0x%x\n, 0x%"PRIx64"", data, shared_buffer[0], high_address);
> + }
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> + free(shared_buffer);
> +}
> +
> +/* gem_pin_mmap_anonymous_test - This test will create a mmap anonymous buffer and
> + * share with GPU. It will run basic test on this buffer.
> +
> + Create a anonymous mmap buffer
> + Share buffer with with GPU by using userptr ioctl
> + Create batch buffer to write DATA to first dword of buffer
> + Set EXEC_OBJECT_PINNED flag in exec object
> + Set 'offset' in exec object to pinning_offset VMA
> + Submit execbuffer
> + Verify value of first DWORD in shared buffer matches DATA
> +*/
> +void gem_pin_mmap_anonymous_test(void)
> +{
Please move to gem_userptr_blits, without the softpin parts.
> + i915_gem_userptr userptr;
> + int fd, ret;
> + uint32_t* shared_buffer = NULL;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS];
> + uint32_t batch_buffer[BO_SIZE];
> + uint32_t batch_buf_handle, shared_buf_handle;
> + int ring, len;
> + const uint32_t data = 0x12345678;
> + uint64_t pinning_offset = 0x1000;
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, BO_SIZE);
> +
> + /* create anonymus mmap buffer, set to all 0xF's */
> + shared_buffer = mmap(NULL, BO_SIZE, PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
> + igt_fail_on_f(shared_buffer == (void *)-1,
> + "mmap call failed with %s\n", strerror(errno));
MAP_FAILED, not -1.
> +
> + *shared_buffer = 0xFFFFFFFF;
> +
> + /* share with GPU */
> + gem_create_userptr_struct(&userptr, shared_buffer, BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> +
> + /* Get handle for shared buffer */
> + shared_buf_handle = userptr.handle;
> +
> + /* create command buffer with write command */
> + len = gem_store_data_svm(fd, batch_buffer, pinning_offset, data, true);
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + setup_exec_obj(&exec_object2[0], shared_buf_handle, EXEC_OBJECT_PINNED, pinning_offset);
> + setup_exec_obj(&exec_object2[1], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_BLT;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS, len);
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + // check on CPU to see if value changes
> + igt_fail_on_f(shared_buffer[0] != data,
> + "\nCPU read does not match GPU write, expected: 0x%x, got: 0x%x\n", data, shared_buffer[0]);
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> + igt_fail_on_f(munmap(shared_buffer, BO_SIZE) != 0,
> + "munmap failed with: %s", strerror(errno));
> +}
> +
> +/* gem_pin_mmap_file_test - This test will use mmap command to map
> + * a file in memory. It will then attempt to share the buffer with GPU
> + * using the userptr ioctl. It will verify if CPU/GPU writes are consistent
> +
> + open/create a file
> + lseek into the file and write some arbitrary data
> + this allows the mmap'd page to become resident
> + use mmap command to map the file into memory
> + Share buffer with with GPU by using userptr ioctl
> + Create batch buffer to write DATA to first dword of buffer
> + Set EXEC_OBJECT_PINNED flag in exec object
> + Set 'offset' in exec object to pinning_offest VMA
> + Submit execbuffer
> + Verify value of first DWORD in shared buffer matches DATA
> + Close file
> +*/
> +void gem_pin_mmap_file_test(void)
> +{
Please move to gem_userptr_blits, without the softpin parts.
> + i915_gem_userptr userptr;
> + int fd, ret;
> + struct drm_i915_gem_execbuffer2 execbuf;
> + struct drm_i915_gem_exec_object2 exec_object2[NUM_EXEC_OBJECTS];
> + uint32_t batch_buffer[BO_SIZE];
> + uint32_t batch_buf_handle, dest_buf_handle;
> + int ring, len;
> + const uint32_t data = 0x12345678;
> + int fdout;
> + uint32_t *dest;
> + const char filename[] = "/tmp/svm_mmap.txt";
> + uint64_t pinning_offset = 0x1000;
> +
> + fdout = open(filename, O_RDWR | O_CREAT | O_TRUNC, 0640);
> + igt_fail_on_f(fdout < 0, "Cannot open output file\n");
> +
> + /* Do this to ensure backing physical memory for the file */
> + /* go to the location corresponding to the last byte */
> + if (lseek (fdout, BO_SIZE, SEEK_SET) == -1)
> + igt_info("lseek error");
> +
> + /* write a dummy byte at the last location */
> + if (write (fdout, "", 1) != 1)
> + igt_info("write error");
> +
> + fd = drm_open_driver(DRIVER_INTEL);
> + batch_buf_handle = gem_create(fd, BO_SIZE);
> +
> + /* create anonymus mmap buffer, set to all 0xF's */
> + dest = mmap(0, BO_SIZE, PROT_WRITE, MAP_SHARED, fdout, 0);
> + igt_fail_on_f(dest == (void *)-1,
> + "mmap call failed with %s\n", strerror(errno));
> + *dest = 0x11111111;
> +
> + gem_create_userptr_struct(&userptr, dest, BO_SIZE, false);
> + ret = gem_call_userptr_ioctl(fd, &userptr);
> + igt_assert_eq(ret, 0);
> + dest_buf_handle = userptr.handle;
> +
> + /* create command buffer with write command */
> + len = gem_store_data_svm(fd, batch_buffer, pinning_offset, data, true);
> + gem_write(fd, batch_buf_handle, 0, batch_buffer, len);
> +
> + /* submit command buffer */
> + setup_exec_obj(&exec_object2[0], dest_buf_handle, EXEC_OBJECT_PINNED, pinning_offset);
> + setup_exec_obj(&exec_object2[1], batch_buf_handle, 0, 0);
> +
> + ring = 0;
> + if (HAS_BLT_RING(intel_get_drm_devid(fd)))
> + ring = I915_EXEC_RENDER;
> +
> + setup_execbuffer(&execbuf, exec_object2, ring, NUM_EXEC_OBJECTS, len);
> + submit_and_sync(fd, &execbuf, batch_buf_handle);
> +
> + /* check on CPU to see if value changes */
> + igt_fail_on_f(*dest != data,
> + "\nCPU read does not match GPU write, expected: 0x%x, got: 0x%x\n", data, dest[0]);
> +
> + gem_close(fd, batch_buf_handle);
> + close(fd);
> + igt_fail_on_f(munmap(dest, BO_SIZE) != 0,
> + "munmap failed with: %s", strerror(errno));
> + close(fdout);
> +}
> +
> +
> +int main(int argc, char* argv[])
> +{
> + igt_subtest_init(argc, argv);
> + igt_skip_on_simulation();
> +
> + igt_subtest("gem_basic"){
> + gem_basic_test();
> + }
> + igt_subtest("gem_multiple_process"){
> + gem_multiple_process_test();
> + }
> + igt_subtest("gem_repin"){
> + gem_repin_test();
> + }
> + igt_subtest("gem_evict"){
> + gem_evict_test();
> + }
> + igt_subtest("gem_softpin_stress"){
> + gem_softpin_stress_test();
> + }
> + igt_subtest("gem_pin_overlap"){
> + gem_pin_overlap_test();
> + }
> + igt_subtest("gem_shmem"){
> + gem_shmem_test();
> + }
> + igt_subtest("gem_write_multipage_buffer"){
> + gem_write_multipage_buffer_test();
> + }
> + igt_subtest("gem_pin_high_address"){
> + gem_pin_high_address_test();
> + }
> + igt_subtest("gem_pin_near_48Bit"){
> + gem_pin_near_48Bit_test();
> + }
> + igt_subtest("gem_pin_invalid_vma"){
> + gem_pin_invalid_vma_test();
> + }
> + igt_subtest("gem_pin_mmap_anon"){
> + gem_pin_mmap_anonymous_test();
> + }
> + igt_subtest("gem_pin_mmap_file"){
> + gem_pin_mmap_file_test();
> + }
> +
> + igt_exit();
> +}
Same comments as before :(
1. More coding style cleanup (including comment styles).
2. More tests on normal BOs, no need to use userptr for everything.
3. Move "can we create an userptr from this backing store" tests to
gem_userptr_blits.
4. And probably some more intersting negative tests can be added about
when softpin can/will fail. Need to brainstorm it a bit.
Regards,
Tvrtko
More information about the Intel-gfx
mailing list