[i-g-t v3 2/4] lib/xe/xe_util: Introduce helper functions
Zeng, Oak
oak.zeng at intel.com
Thu Feb 20 23:29:18 UTC 2025
> -----Original Message-----
> From: Kempczynski, Zbigniew <zbigniew.kempczynski at intel.com>
> Sent: February 20, 2025 2:15 AM
> To: Zeng, Oak <oak.zeng at intel.com>
> Cc: igt-dev at lists.freedesktop.org; Thomas.Hellstrom at linux.intel.com;
> Brost, Matthew <matthew.brost at intel.com>; Konieczny, Kamil
> <kamil.konieczny at intel.com>; Dixit, Ashutosh
> <ashutosh.dixit at intel.com>; Heikkila, Juha-pekka <juha-
> pekka.heikkila at intel.com>; Vivi, Rodrigo <rodrigo.vivi at intel.com>
> Subject: Re: [i-g-t v3 2/4] lib/xe/xe_util: Introduce helper functions
>
> On Tue, Feb 18, 2025 at 06:57:53PM -0500, Oak Zeng wrote:
> > From: Bommu Krishnaiah <krishnaiah.bommu at intel.com>
> >
> > Introduce helper functions for buffer creation, binding,
> > destruction and command submission etc. With those helpers,
> > writing a xe igt test will be much easier, which will be
> > showed in a coming example.
> >
> > v2: use to_user_pointer to cast a pointer (Kamil)
> > s/insert_store/xe_insert_store (Kamil)
> > s/cmdbuf_fill_func_t/xe_cmdbuf_fill_func_t (Kamil)
> > v3: refactor command buffer fill interface (Zbigniew)
> >
> > Signed-off-by: Bommu Krishnaiah <krishnaiah.bommu at intel.com>
> > Signed-off-by: Oak Zeng <oak.zeng at intel.com>
> > Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray at intel.com>
> > ---
> > lib/xe/xe_util.c | 239
> +++++++++++++++++++++++++++++++++++++++++++++++
> > lib/xe/xe_util.h | 47 ++++++++++
> > 2 files changed, 286 insertions(+)
> >
> > diff --git a/lib/xe/xe_util.c b/lib/xe/xe_util.c
> > index 06b378ce0..891316ecd 100644
> > --- a/lib/xe/xe_util.c
> > +++ b/lib/xe/xe_util.c
> > @@ -13,6 +13,245 @@
> > #include "xe/xe_query.h"
> > #include "xe/xe_util.h"
> >
> > +/**
> > + * xe_cmdbuf_exec_ufence_gpuva:
> > + * @cmdbuf Pointer to the xe_cmdbuf structure representing the
> command buffer.
> > + *
> > + * Returns the GPU virtual address of the execution user fence
> located at the
> > + * end of the command buffer.
> > + */
> > +static uint64_t xe_cmdbuf_exec_ufence_gpuva(struct xe_cmdbuf
> *cmdbuf)
> > +{
> > + /* the last 8 bytes of the cmd buffer is used as ufence */
> > + return cmdbuf->buf.gpu_addr + cmdbuf->buf.size -
> UFENCE_LENGTH;
>
> I think instead of hardcoding 8 as length of the fence you should use
> sizeof() of timeline_value from drm_xe_sync.
>
> > +}
> > +
> > +/**
> > + * xe_cmdbuf_exec_ufence_cpuva:
> > + * @cmdbuf Pointer to the xe_cmdbuf structure representing the
> command buffer.
> > + *
> > + * Returns the CPU virtual address of the execution user fence
> located at the
> > + * end of the command buffer.
> > + */
> > +static void *xe_cmdbuf_exec_ufence_cpuva(struct xe_cmdbuf
> *cmdbuf)
> > +{
> > + /* the last 8 bytes of the cmd buffer is used as ufence */
> > + return (char *)cmdbuf->buf.cpu_addr + cmdbuf->buf.size -
> UFENCE_LENGTH;
> > +}
> > +
> > +
> > +/**
> > + * __xe_submit_cmd:
> > + * @cmdbuf Pointer to the command buffer structure
> > + *
> > + * Submits a command buffer to the GPU, waits for its completion,
> and verifies
> > + * the user fence value
> > + *
> > + * Return: The result of waiting for the user fence value
> > + */
> > +int64_t __xe_submit_cmd(struct xe_cmdbuf *cmdbuf)
> > +{
> > + int64_t timeout = NSEC_PER_SEC;
> > + int ret;
> > +
>
> You should assert if cmdbuf is null.
>
> > + struct drm_xe_sync sync[1] = {
> > + { .type = DRM_XE_SYNC_TYPE_USER_FENCE,
> > + .flags = DRM_XE_SYNC_FLAG_SIGNAL,
> > + .timeline_value = USER_FENCE_VALUE,
> > + .addr =
> xe_cmdbuf_exec_ufence_gpuva(cmdbuf),},
> > + };
> > + struct drm_xe_exec exec = {
> > + .num_batch_buffer = 1,
> > + .num_syncs = 1,
> > + .syncs = to_user_pointer(sync),
> > + .exec_queue_id = cmdbuf->exec_queue,
> > + .address = cmdbuf->buf.gpu_addr,
> > + };
> > +
> > + xe_exec(cmdbuf->buf.fd, &exec);
>
> This might assert, shouldn't it be sth like:
>
> ret = __xe_exec(...);
> if (ret)
> return ret;
>
> xe_submit_cmd() is asserting, so it might assert from both
> reasons - unsuccessful exec or unsuccessful wait.
>
> My personal preference is to not assert in underscore functions
> giving caller chance to handle errors. Only exception is to
> checking input parameters.
>
> > + ret = __xe_wait_ufence(cmdbuf->buf.fd,
> > + (uint64_t
> *)xe_cmdbuf_exec_ufence_cpuva(cmdbuf),
> > + USER_FENCE_VALUE, cmdbuf-
> >exec_queue, &timeout);
> > + /* Reset the fence so the exec ufence can be reused */
> > + memset((char *)xe_cmdbuf_exec_ufence_cpuva(cmdbuf), 0,
> UFENCE_LENGTH);
>
> Cosmetic, please add line before return.
>
> > + return ret;
> > +}
> > +
> > +/**
> > + * xe_submit_cmd:
> > + * @cmdbuf Pointer to the command buffer structure
> > + *
> > + * Wrapper function to submit a command buffer and assert its
> successful
> > + * execution.
> > + */
> > +void xe_submit_cmd(struct xe_cmdbuf *cmdbuf)
> > +{
> > + int64_t ret;
> > +
>
> Assert if cmdbuf is not null.
>
> > + ret = __xe_submit_cmd(cmdbuf);
> > + igt_assert_eq(ret, 0);
> > +}
> > +
> > +/**
> > + *xe_create_buffer:
> > + * @buffer Pointer to the xe_buf structure containing buffer
> details.
> > + *
> > + * Creates a buffer, maps it to both CPU and GPU address spaces,
> and binds it
> > + * to a virtual memory (VM) space.
> > + */
> > +void xe_create_buffer(struct xe_buf *buffer)
> > +{
> > + struct drm_xe_sync sync[1] = {
> > + { .type = DRM_XE_SYNC_TYPE_USER_FENCE,
> > + .flags = DRM_XE_SYNC_FLAG_SIGNAL,
> > + .timeline_value = USER_FENCE_VALUE },
> > + };
> > +
>
> Ditto.
>
> This function should check if all necessary fields are set in buffer
> structure before proceed - fd, vm, size, placement.
> Calling this function on not properly initialized buffer likely will
> lead to segfault.
I will check fd, size and placement.
I didn't find a good way to check vm. I assume vm==0 is legal.
I will leave vm check for now.
>
> > + buffer->bind_queue = xe_bind_exec_queue_create(buffer-
> >fd,
> > + buffer->vm, 0);
> > + buffer->bind_ufence =
> aligned_alloc(xe_get_default_alignment(buffer->fd),
> > + PAGE_ALIGN_UFENCE);
> > + sync->addr = to_user_pointer(buffer->bind_ufence);
> > +
> > +
>
> Unnecessary double newline.
>
> > + /* create and bind the buffer->bo */
> > + buffer->bo = xe_bo_create(buffer->fd, 0, buffer->size,
> > + buffer->placement, buffer->flag);
> > + buffer->cpu_addr = xe_bo_map(buffer->fd, buffer->bo,
> buffer->size);
> > + xe_vm_bind_async(buffer->fd, buffer->vm, buffer-
> >bind_queue,
> > + buffer->bo, 0, buffer->gpu_addr,
> > + buffer->size, sync, 1);
> > +
> > + xe_wait_ufence(buffer->fd, buffer->bind_ufence,
> > + USER_FENCE_VALUE, buffer->bind_queue,
> NSEC_PER_SEC);
> > + memset(buffer->bind_ufence, 0, PAGE_ALIGN_UFENCE);
> > +}
> > +
> > +/**
> > + * xe_destroy_buffer:
> > + * @buffer Pointer to the xe_buf structure containing buffer
> details
> > + *
> > + * Destroys a buffer created by xe_create_buffer and releases
> associated
> > + * resources.
> > + */
> > +void xe_destroy_buffer(struct xe_buf *buffer)
> > +{
> > + struct drm_xe_sync sync[1] = {
> > + { .type = DRM_XE_SYNC_TYPE_USER_FENCE,
> > + .flags = DRM_XE_SYNC_FLAG_SIGNAL,
> > + .timeline_value = USER_FENCE_VALUE },
> > + };
>
> Assert if buffer is null
>
> > + sync->addr = to_user_pointer(buffer->bind_ufence);
> > +
> > + xe_vm_unbind_async(buffer->fd, buffer->vm, buffer-
> >bind_queue,
> > + 0, buffer->gpu_addr, buffer->size, sync, 1);
> > + xe_wait_ufence(buffer->fd, buffer->bind_ufence,
> > + USER_FENCE_VALUE, buffer->bind_queue,
> NSEC_PER_SEC);
> > + memset(buffer->bind_ufence, 0, PAGE_ALIGN_UFENCE);
> > +
> > + munmap(buffer->cpu_addr, buffer->size);
> > + gem_close(buffer->fd, buffer->bo);
> > +
> > + free(buffer->bind_ufence);
> > + xe_exec_queue_destroy(buffer->fd, buffer->bind_queue);
> > +}
> > +
> > +/**
> > + * xe_cmdbuf_insert_store:
> > + * @cmdbuf: command buffer where commands will be inserted.
> > + * @dst_va Destination virtual address to store the value.
> > + * @val Value to be stored.
> > + *
> > + * Inserts a MI_STORE_DWORD_IMM_GEN4 command into a
> command buffer, which stores
> > + * an immediate value to a given destination virtual address.
> > + */
> > +void xe_cmdbuf_insert_store(struct xe_cmdbuf *cmdbuf,
> > + uint64_t dst_va, uint32_t val)
> > +{
> > + uint32_t *batch = cmdbuf->buf.cpu_addr;
> > +
> > + igt_assert(!cmdbuf->closed);
>
> You may check if cmdbuf->write_index + length of store-dword (4
> dwords).
> As xe_bb_size() adds additional page you won't overwrite memory,
> just
> assert due to lack of space for BBE.
I think the last line of this function serves this purpose.
I will move the assert up.
>
> > +
> > + batch[cmdbuf->write_index++] =
> MI_STORE_DWORD_IMM_GEN4;
> > + batch[cmdbuf->write_index++] = dst_va;
> > + batch[cmdbuf->write_index++] = dst_va >> 32;
> > + batch[cmdbuf->write_index++] = val;
> > +
> > + /* Leaves at least one dword for MI_BATCH_BUFFER_END */
> > + igt_assert(cmdbuf->write_index <= cmdbuf-
> >cmd_size/sizeof(uint32_t) - 2);
> > +}
> > +
> > +/**
> > + * xe_create_cmdbuf:
> > + * @cmdbuf Pointer to the xe_cmdbuf structure representing the
> command buffer.
> > + * @eci Pointer to the engine class instance for execution.
> > + *
> > + * Creates a command buffer, fills it with commands using the
> provided fill
> > + * function, and sets up the execution queue for submission.
> > + */
> > +void xe_create_cmdbuf(struct xe_cmdbuf *cmdbuf,
> > + struct drm_xe_engine_class_instance *eci)
> > +{
> > + struct xe_buf *buf = &cmdbuf->buf;
> > + /*
> > + * make some room for a exec_ufence, which will be used to
> sync the
> > + * submission of this command....
> > + */
> > + buf->size = xe_bb_size(buf->fd,
> > + cmdbuf->cmd_size +
> PAGE_ALIGN_UFENCE);
> > + xe_create_buffer(buf);
> > + cmdbuf->exec_queue = xe_exec_queue_create(buf->fd,
> buf->vm, eci, 0);
> > + cmdbuf->write_index = 0;
> > + cmdbuf->closed = false;
> > +}
> > +
> > +/**
> > + * xe_fill_cmdbuf:
> > + * @cmdbuf Pointer to the xe_cmdbuf structure representing the
> command buffer.
> > + * @fill_func Pointer to the function that fills the command buffer.
> > + * @dst_gpu_va Destination GPU virtual address to store the value.
> > + * @val Value to be stored.
> > + *
> > + * Fills the command buffer with commands using the provided fill
> function.
> > + */
> > +void xe_fill_cmdbuf(struct xe_cmdbuf *cmdbuf,
> > + xe_cmdbuf_fill_func_t fill_func,
> > + uint64_t dst_gpu_va, uint32_t val)
> > +{
> > + igt_assert(!cmdbuf->closed);
> > + fill_func(cmdbuf, dst_gpu_va, val);
>
> I don't like this function and I vote for it's removal. You should
> just operate directly on cmdbuf (it already have everything you need).
> xe_cmdbuf_fill_func_t is not generic (limited to gpu_va/val) so what's
> for defining it?
>
> So instead:
>
> xe_create_cmdbuf(&cmdbuf, eci);
> xe_fill_cmdbuf(&cmdbuf, xe_cmdbuf_insert_store,
> dst_buf.gpu_addr, 0xc0ffee);
> xe_close_cmdbuf(&cmdbuf);
> xe_submit_cmd(&cmdbuf);
>
> I would imagine:
>
> xe_create_cmdbuf(&cmdbuf, eci);
> xe_cmdbuf_insert_store(&cmdbuf, dst_buf.gpu_addr,
> 0xc0ffee);
> xe_cmdbuf_insert_bbe(&cmdbuf, ...)
> xe_submit_cmd(&cmdbuf);
>
> Any additional commands might be added by generic calls:
>
> xe_cmdbuf_insert_dw(&cmdbuf, 0xc0de);
> xe_cmdbuf_insert_dw(&cmdbuf, 0xc0de2);
>
> You may even drop xe_cmdbuf_insert_bbe() by:
>
> xe_cmdbuf_insert_dw(&cmdbuf, MI_BATCH_BUFFER_END);
>
It is a fair argument. I will do the interface you laid out above.
>
> > +}
> > +
> > +/**
> > + * xe_close_cmdbuf:
> > + * @cmdbuf command buffer to be closed.
> > + *
> > + * After this function is called, no more commands can be added to
> the
> > + * command buffer.
> > + */
> > +void xe_close_cmdbuf(struct xe_cmdbuf *cmdbuf)
> > +{
> > + uint32_t *batch = cmdbuf->buf.cpu_addr;
> > +
> > + igt_assert(!cmdbuf->closed);
> > + batch[cmdbuf->write_index++] = MI_BATCH_BUFFER_END;
> > + cmdbuf->closed = true;
>
> Please drop this function. You don't need to keep any logic
> regarding if buffer has bbe or not.
Will do.
Oak
>
> --
> Zbigniew
>
> > + igt_assert(cmdbuf->write_index <= cmdbuf-
> >cmd_size/sizeof(uint32_t) - 1);
> > +}
> > +
> > +/**
> > + * xe_destroy_cmdbuf:
> > + * @cmdbuf Pointer to the xe_buf structure representing the
> command buffer.
> > + *
> > + * Destroys a command buffer created by xe_create_cmdbuf and
> releases
> > + * associated resources.
> > + */
> > +void xe_destroy_cmdbuf(struct xe_cmdbuf *cmdbuf)
> > +{
> > + xe_exec_queue_destroy(cmdbuf->buf.fd, cmdbuf-
> >exec_queue);
> > + xe_destroy_buffer(&cmdbuf->buf);
> > +}
> > +
> > static bool __region_belongs_to_regions_type(struct
> drm_xe_mem_region *region,
> > uint32_t
> *mem_regions_type,
> > int num_regions)
> > diff --git a/lib/xe/xe_util.h b/lib/xe/xe_util.h
> > index 06ebd3c2a..26321f957 100644
> > --- a/lib/xe/xe_util.h
> > +++ b/lib/xe/xe_util.h
> > @@ -14,6 +14,53 @@
> >
> > #include "xe_query.h"
> >
> > +#define USER_FENCE_VALUE 0xdeadbeefdeadbeefull
> > +#define PAGE_ALIGN_UFENCE 4096
> > +#define UFENCE_LENGTH 8
> > +
> > +struct xe_buf {
> > + void *cpu_addr;
> > + uint64_t gpu_addr;
> > + /*the user fence used to vm bind this buffer*/
> > + uint64_t *bind_ufence;
> > + uint64_t size;
> > + uint32_t flag;
> > + uint32_t vm;
> > + uint32_t bo;
> > + uint32_t placement;
> > + uint32_t bind_queue;
> > + int fd;
> > + bool is_userptr;
> > +};
> > +
> > +struct xe_cmdbuf {
> > + struct xe_buf buf;
> > + /* command size in bytes, not including exec_ufence */
> > + uint64_t cmd_size;
> > + uint32_t exec_queue;
> > + /* Dword index to writ to command buffer */
> > + uint32_t write_index;
> > + bool closed;
> > +};
> > +
> > +typedef void (*xe_cmdbuf_fill_func_t) (struct xe_cmdbuf
> *cmdbuf,
> > + uint64_t dst_gpu_va, uint32_t val);
> > +
> > +void xe_create_buffer(struct xe_buf *buffer);
> > +void xe_destroy_buffer(struct xe_buf *buffer);
> > +
> > +void xe_create_cmdbuf(struct xe_cmdbuf *cmdbuf,
> > + struct drm_xe_engine_class_instance *eci);
> > +void xe_fill_cmdbuf(struct xe_cmdbuf *cmdbuf,
> > + xe_cmdbuf_fill_func_t fill_func,
> > + uint64_t dst_gpu_va, uint32_t val);
> > +void xe_close_cmdbuf(struct xe_cmdbuf *cmdbuf);
> > +void xe_cmdbuf_insert_store(struct xe_cmdbuf *cmdbuf,
> uint64_t dst_va,
> > + uint32_t val);
> > +void xe_submit_cmd(struct xe_cmdbuf *cmdbuf);
> > +int64_t __xe_submit_cmd(struct xe_cmdbuf *cmdbuf);
> > +void xe_destroy_cmdbuf(struct xe_cmdbuf *cmdbuf);
> > +
> > #define XE_IS_SYSMEM_MEMORY_REGION(fd, region) \
> > (xe_region_class(fd, region) ==
> DRM_XE_MEM_REGION_CLASS_SYSMEM)
> > #define XE_IS_VRAM_MEMORY_REGION(fd, region) \
> > --
> > 2.26.3
> >
More information about the igt-dev
mailing list