AI/rocm-systems
2
0
Forka 0
Codice Problemi Pull Request Actions Pacchetti Progetti Rilasci Wiki Attività

Use drm render device to map kfd BOs

Sfoglia il codice sorgente 
Previously kfd device is used to map memory for CPU access.
However this is not compatible with how TTM handles CPU mapping
on eviction - memory won't be unmapped and remapped on restore.
This fixes the issue by mmapping memory using DRM render device.

This patch requires a coordinated kernel driver change to work.
To make it compatible with old kernel driver, some temporary codes
are included. Once the coordinated kernel driver is checked in,
the temporary codes can be removed.



Change-Id: Ie7b304c4a82b7e8d5ab703acb81d66430af4f0bc
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
This commit is contained in:
Oak Zeng
2017-10-27 16:17:51 -04:00
committato da Felix Kuehling
parent 55fc06dac3
commit 68a2d286ca
4 ha cambiato i file con 47 aggiunte e 9 eliminazioni
Scarica il file Patch Scarica il file Diff Expand all files Collapse all files
src/fmm.c
+14 -7
Vedi File
@@ -1056,9 +1056,12 @@ void *fmm_allocate_device(uint32_t gpu_id, uint64_t MemorySizeInBytes, HsaMemFla
}
if (mem && flags.ui32.HostAccess) {
int map_fd = mmap_offset >= (1ULL<<40) ? kfd_fd :
get_drm_render_fd_by_gpu_id(gpu_id);
void *ret = mmap(mem, MemorySizeInBytes,
PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_FIXED, kfd_fd, mmap_offset);
MAP_SHARED | MAP_FIXED,
map_fd, mmap_offset);
if (ret == MAP_FAILED) {
__fmm_release(mem, aperture);
return NULL;
@@ -1245,9 +1248,11 @@ static void *fmm_allocate_host_gpu(uint32_t node_id, uint64_t MemorySizeInBytes,
ioc_flags, &vm_obj);
if (mem && flags.ui32.HostAccess) {
int map_fd = mmap_offset >= (1ULL<<40) ? kfd_fd :
get_drm_render_fd_by_gpu_id(gpu_id);
void *ret = mmap(mem, MemorySizeInBytes,
PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_FIXED, kfd_fd, mmap_offset);
MAP_SHARED | MAP_FIXED, map_fd, mmap_offset);
if (ret == MAP_FAILED) {
__fmm_release(mem, aperture);
return NULL;
@@ -1259,7 +1264,7 @@ static void *fmm_allocate_host_gpu(uint32_t node_id, uint64_t MemorySizeInBytes,
memset(ret, 0, MemorySizeInBytes);
mmap(VOID_PTR_ADD(mem, my_buf_size), MemorySizeInBytes,
PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_FIXED, kfd_fd, mmap_offset);
MAP_SHARED | MAP_FIXED, map_fd, mmap_offset);
}
}
}
@@ -1827,6 +1832,8 @@ static int _fmm_map_to_gpu_scratch(uint32_t gpu_id, manageable_aperture_t *apert
if (!obj)
return -1;
} else {
int map_fd = mmap_offset >= (1ULL<<40) ? kfd_fd :
get_drm_render_fd_by_gpu_id(gpu_id);
fmm_allocate_memory_in_device(gpu_id,
address,
size,
@@ -1835,8 +1842,7 @@ static int _fmm_map_to_gpu_scratch(uint32_t gpu_id, manageable_aperture_t *apert
KFD_IOC_ALLOC_MEM_FLAGS_GTT);
mmap_ret = mmap(address, size,
PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_FIXED,
kfd_fd, mmap_offset);
MAP_SHARED | MAP_FIXED, map_fd, mmap_offset);
if (mmap_ret == MAP_FAILED) {
__fmm_release(mem, aperture);
return -1;
@@ -2753,10 +2759,11 @@ HSAKMT_STATUS fmm_register_shared_memory(const HsaSharedMemoryHandle *SharedMemo
pthread_mutex_unlock(&aperture->fmm_mutex);
if (importArgs.mmap_offset) {
int map_fd = importArgs.mmap_offset >= (1ULL<<40) ? kfd_fd :
get_drm_render_fd_by_gpu_id(importArgs.gpu_id);
void *ret = mmap(reservedMem, (SharedMemoryStruct->SizeInPages << PAGE_SHIFT),
PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_FIXED, kfd_fd,
importArgs.mmap_offset);
MAP_SHARED | MAP_FIXED, map_fd, importArgs.mmap_offset);
if (ret == MAP_FAILED) {
err = HSAKMT_STATUS_ERROR;
goto err_free_obj;