Add code to get the physical address of the GPU memory.
This should allow some NIC hardware to directly use the
physical address for DMA instead of the CUDA Unified Memory
provided by cuda_mem_map.
Signed-off-by: Ajit Khaparde <ajit.khapa...@broadcom.com>
---
drivers/gpu/cuda/common.h | 2 +
drivers/gpu/cuda/cuda.c | 117 ++++++++++++++++++++++++++++++-------
drivers/gpu/cuda/gdrcopy.c | 50 ++++++++++++++++
3 files changed, 147 insertions(+), 22 deletions(-)
diff --git a/drivers/gpu/cuda/common.h b/drivers/gpu/cuda/common.h
index e809988503..5da5e7d176 100644
--- a/drivers/gpu/cuda/common.h
+++ b/drivers/gpu/cuda/common.h
@@ -33,6 +33,8 @@ extern int cuda_logtype;
int gdrcopy_pin(gdr_t *gdrc_h, __rte_unused gdr_mh_t *mh,
uint64_t d_addr, size_t size, void **h_addr);
+int gdrcopy_dma(gdr_t *gdrc_h, __rte_unused gdr_mh_t *mh,
+ uint64_t d_addr, size_t size, void **h_addr, uint64_t *paddr);
int gdrcopy_unpin(gdr_t gdrc_h, __rte_unused gdr_mh_t mh,
void *d_addr, size_t size);
diff --git a/drivers/gpu/cuda/cuda.c b/drivers/gpu/cuda/cuda.c
index a552aabeb8..36cff7b3a5 100644
--- a/drivers/gpu/cuda/cuda.c
+++ b/drivers/gpu/cuda/cuda.c
@@ -959,6 +959,87 @@ cuda_mem_cpu_map(struct rte_gpu *dev, __rte_unused size_t
size, void *ptr_in, vo
return 0;
}
+static int
+cuda_mem_dma_map(struct rte_gpu *dev, __rte_unused size_t size, void *ptr_in,
void **ptr_out)
+{
+ struct mem_entry *mem_item;
+ cuda_ptr_key hk;
+ uint64_t paddr;
+
+ if (dev == NULL)
+ return -ENODEV;
+
+ hk = get_hash_from_ptr((void *)ptr_in);
+
+ mem_item = mem_list_find_item(hk);
+ if (mem_item == NULL) {
+ rte_cuda_log(ERR, "Memory address 0x%p not found in driver
memory.", ptr_in);
+ rte_errno = EPERM;
+ return -rte_errno;
+ }
+
+ if (mem_item->mtype != GPU_MEM) {
+ rte_cuda_log(ERR, "Memory address 0x%p is not GPU memory
type.", ptr_in);
+ rte_errno = EPERM;
+ return -rte_errno;
+ }
+
+ if (mem_item->size != size)
+ rte_cuda_log(WARNING,
+ "Can't expose memory area with size (%zd)
different from original size (%zd).",
+ size, mem_item->size);
+
+ if (gdrcopy_dma(&gdrc_h, &(mem_item->mh), (uint64_t)mem_item->ptr_d,
+ mem_item->size, &(mem_item->ptr_h),
&paddr)) {
+ rte_cuda_log(ERR, "Error exposing GPU memory address 0x%p.",
ptr_in);
+ rte_errno = EPERM;
+ return -rte_errno;
+ }
+
+ mem_item->mtype = GPU_REGISTERED;
+ *ptr_out = (void *)paddr;
+
+ return 0;
+}
+
+static int
+cuda_mem_free(struct rte_gpu *dev, void *ptr)
+{
+ CUresult res;
+ struct mem_entry *mem_item;
+ const char *err_string;
+ cuda_ptr_key hk;
+
+ if (dev == NULL)
+ return -ENODEV;
+
+ hk = get_hash_from_ptr((void *)ptr);
+
+ mem_item = mem_list_find_item(hk);
+ if (mem_item == NULL) {
+ rte_cuda_log(ERR, "Memory address 0x%p not found in driver
memory", ptr);
+ rte_errno = EPERM;
+ return -rte_errno;
+ }
+
+ if (mem_item->mtype == GPU_MEM) {
+ res = pfn_cuMemFree(mem_item->ptr_orig_d);
+ if (res != 0) {
+ pfn_cuGetErrorString(res, &(err_string));
+ rte_cuda_log(ERR, "cuMemFree current failed with %s",
+ err_string);
+ rte_errno = EPERM;
+ return -rte_errno;
+ }
+
+ return mem_list_del_item(hk);
+ }
+
+ rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
+
+ return -EPERM;
+}
+
static int
cuda_mem_unregister(struct rte_gpu *dev, void *ptr)
{
@@ -1034,48 +1115,38 @@ cuda_mem_cpu_unmap(struct rte_gpu *dev, void *ptr_in)
}
static int
-cuda_mem_free(struct rte_gpu *dev, void *ptr)
+cuda_mem_dma_unmap(struct rte_gpu *dev, void *ptr_in)
{
- CUresult res;
struct mem_entry *mem_item;
- const char *err_string;
cuda_ptr_key hk;
if (dev == NULL)
return -ENODEV;
- hk = get_hash_from_ptr((void *)ptr);
+ hk = get_hash_from_ptr((void *)ptr_in);
mem_item = mem_list_find_item(hk);
if (mem_item == NULL) {
- rte_cuda_log(ERR, "Memory address 0x%p not found in driver
memory", ptr);
+ rte_cuda_log(ERR, "Memory address 0x%p not found in driver
memory.", ptr_in);
rte_errno = EPERM;
return -rte_errno;
}
- /*
- * If a GPU memory area that's CPU mapped is being freed
- * without calling cpu_unmap, force the unmapping.
- */
- if (mem_item->mtype == GPU_REGISTERED)
- cuda_mem_cpu_unmap(dev, ptr);
-
- if (mem_item->mtype == GPU_MEM) {
- res = pfn_cuMemFree(mem_item->ptr_orig_d);
- if (res != 0) {
- pfn_cuGetErrorString(res, &(err_string));
- rte_cuda_log(ERR, "cuMemFree current failed with %s",
- err_string);
+ if (mem_item->mtype == GPU_REGISTERED) {
+ if (gdrcopy_unpin(gdrc_h, mem_item->mh, (void *)mem_item->ptr_d,
+ mem_item->size)) {
+ rte_cuda_log(ERR, "Error unexposing GPU memory address
0x%p.", ptr_in);
rte_errno = EPERM;
return -rte_errno;
}
- return mem_list_del_item(hk);
+ mem_item->mtype = GPU_MEM;
+ } else {
+ rte_errno = EPERM;
+ return -rte_errno;
}
- rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
-
- return -EPERM;
+ return 0;
}
static int
@@ -1391,7 +1462,9 @@ cuda_gpu_probe(__rte_unused struct rte_pci_driver
*pci_drv, struct rte_pci_devic
dev->ops.mem_register = cuda_mem_register;
dev->ops.mem_unregister = cuda_mem_unregister;
dev->ops.mem_cpu_map = cuda_mem_cpu_map;
+ dev->ops.mem_dma_map = cuda_mem_dma_map;
dev->ops.mem_cpu_unmap = cuda_mem_cpu_unmap;
+ dev->ops.mem_dma_unmap = cuda_mem_dma_unmap;
dev->ops.wmb = cuda_wmb;
rte_gpu_complete_new(dev);
diff --git a/drivers/gpu/cuda/gdrcopy.c b/drivers/gpu/cuda/gdrcopy.c
index bd56b73ce4..ac6b3d773a 100644
--- a/drivers/gpu/cuda/gdrcopy.c
+++ b/drivers/gpu/cuda/gdrcopy.c
@@ -12,6 +12,8 @@ static void *gdrclib;
static gdr_t (*sym_gdr_open)(void);
static int (*sym_gdr_pin_buffer)(gdr_t g, unsigned long addr, size_t size,
uint64_t p2p_token, uint32_t va_space, gdr_mh_t *handle);
+static int (*sym_gdr_p2p_dma_map_buffer)(gdr_t g, unsigned long addr, size_t
size,
+ uint64_t p2p_token, uint32_t va_space, gdr_mh_t *handle,
uint64_t *paddr);
static int (*sym_gdr_unpin_buffer)(gdr_t g, gdr_mh_t handle);
static int (*sym_gdr_map)(gdr_t g, gdr_mh_t handle, void **va, size_t size);
static int (*sym_gdr_unmap)(gdr_t g, gdr_mh_t handle, void *va, size_t size);
@@ -45,6 +47,13 @@ gdrcopy_loader(void)
return -1;
}
+ sym_gdr_p2p_dma_map_buffer = dlsym(gdrclib, "gdr_p2p_dma_map_buffer");
+ if (sym_gdr_p2p_dma_map_buffer == NULL) {
+ rte_cuda_log(ERR, "Failed to load GDRCopy symbol
gdr_p2p_dma_map_buffer\n");
+ printf("Failed to load GDRCopy symbol
gdr_p2p_dma_map_buffer\n");
+ return -1;
+ }
+
sym_gdr_unpin_buffer = dlsym(gdrclib, "gdr_unpin_buffer");
if (sym_gdr_unpin_buffer == NULL) {
rte_cuda_log(ERR, "Failed to load GDRCopy symbols\n");
@@ -119,6 +128,47 @@ gdrcopy_pin(__rte_unused gdr_t *gdrc_h, __rte_unused
gdr_mh_t *mh,
#endif
}
+int
+gdrcopy_dma(__rte_unused gdr_t *gdrc_h, __rte_unused gdr_mh_t *mh,
+ __rte_unused uint64_t d_addr, __rte_unused size_t size,
+ __rte_unused void **h_addr, __rte_unused uint64_t *paddr)
+{
+#ifdef DRIVERS_GPU_CUDA_GDRCOPY_H
+ uint64_t phys;
+
+ if (*gdrc_h == NULL) {
+ if (gdrcopy_loader())
+ return -ENOTSUP;
+
+ if (gdrcopy_open(gdrc_h)) {
+ rte_cuda_log(ERR,
+ "GDRCopy gdrdrv kernel module not
found. Can't CPU map GPU memory.");
+ return -EPERM;
+ }
+ }
+
+ /* Pin the device buffer */
+ if (sym_gdr_p2p_dma_map_buffer(*gdrc_h, d_addr, size, 0, 0, mh, &phys)
!= 0) {
+ rte_cuda_log(ERR, "GDRCopy p2p dma map buffer error.");
+ return -1;
+ }
+ *paddr = phys;
+
+ /* Map the buffer to user space */
+ if (sym_gdr_map(*gdrc_h, *mh, h_addr, size) != 0) {
+ rte_cuda_log(ERR, "GDRCopy map buffer error.");
+ sym_gdr_unpin_buffer(*gdrc_h, *mh);
+ return -1;
+ }
+
+ return 0;
+#else
+ rte_cuda_log(ERR,
+ "GDRCopy headers not provided at DPDK building time.
Can't CPU map GPU memory.");
+ return -ENOTSUP;
+#endif
+}
+
int
gdrcopy_unpin(gdr_t gdrc_h, __rte_unused gdr_mh_t mh,
__rte_unused void *d_addr, __rte_unused size_t size)
--
2.39.3 (Apple Git-146)
