* [PATCH] gpu/cuda: GPU_REGISTERED to distinguish GPU memory CPU mapped
@ 2022-04-29 13:52 eagostini
2022-04-29 14:14 ` eagostini
0 siblings, 1 reply; 3+ messages in thread
From: eagostini @ 2022-04-29 13:52 UTC (permalink / raw)
To: dev; +Cc: Elena Agostini
From: Elena Agostini <eagostini@nvidia.com>
Enable GPU_REGISTERED flag in gpu/cuda driver in the memory list.
If a GPU memory address CPU mapped is freed before being
unmapped, CUDA driver unmaps it before freeing the memory.
Signed-off-by: Elena Agostini <eagostini@nvidia.com>
---
drivers/gpu/cuda/cuda.c | 78 +++++++++++++++++++++++++----------------
1 file changed, 47 insertions(+), 31 deletions(-)
diff --git a/drivers/gpu/cuda/cuda.c b/drivers/gpu/cuda/cuda.c
index 5b8476ac20..c6bf54c130 100644
--- a/drivers/gpu/cuda/cuda.c
+++ b/drivers/gpu/cuda/cuda.c
@@ -257,7 +257,7 @@ struct cuda_info {
enum mem_type {
GPU_MEM = 0,
CPU_REGISTERED,
- GPU_REGISTERED /* Not used yet */
+ GPU_REGISTERED
};
/* key associated to a memory address */
@@ -953,13 +953,14 @@ cuda_mem_cpu_map(struct rte_gpu *dev, __rte_unused size_t size, void *ptr_in, vo
return -rte_errno;
}
+ mem_item->mtype = GPU_REGISTERED;
*ptr_out = mem_item->ptr_h;
return 0;
}
static int
-cuda_mem_free(struct rte_gpu *dev, void *ptr)
+cuda_mem_unregister(struct rte_gpu *dev, void *ptr)
{
CUresult res;
struct mem_entry *mem_item;
@@ -978,11 +979,11 @@ cuda_mem_free(struct rte_gpu *dev, void *ptr)
return -rte_errno;
}
- if (mem_item->mtype == GPU_MEM) {
- res = pfn_cuMemFree(mem_item->ptr_orig_d);
+ if (mem_item->mtype == CPU_REGISTERED) {
+ res = pfn_cuMemHostUnregister(ptr);
if (res != 0) {
pfn_cuGetErrorString(res, &(err_string));
- rte_cuda_log(ERR, "cuMemFree current failed with %s",
+ rte_cuda_log(ERR, "cuMemHostUnregister current failed with %s",
err_string);
rte_errno = EPERM;
return -rte_errno;
@@ -993,74 +994,89 @@ cuda_mem_free(struct rte_gpu *dev, void *ptr)
rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
- return -EPERM;
+ rte_errno = EPERM;
+ return -rte_errno;
}
static int
-cuda_mem_unregister(struct rte_gpu *dev, void *ptr)
+cuda_mem_cpu_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 (mem_item->mtype == CPU_REGISTERED) {
- res = pfn_cuMemHostUnregister(ptr);
- if (res != 0) {
- pfn_cuGetErrorString(res, &(err_string));
- rte_cuda_log(ERR, "cuMemHostUnregister 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);
-
- rte_errno = EPERM;
- return -rte_errno;
+ return 0;
}
static int
-cuda_mem_cpu_unmap(struct rte_gpu *dev, void *ptr_in)
+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_in);
+ 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_in);
+ rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory", ptr);
rte_errno = EPERM;
return -rte_errno;
}
- 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;
+ /*
+ * 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);
}
- return 0;
+ 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
--
2.25.1
^ permalink raw reply [flat|nested] 3+ messages in thread
* [PATCH] gpu/cuda: GPU_REGISTERED to distinguish GPU memory CPU mapped
2022-04-29 13:52 [PATCH] gpu/cuda: GPU_REGISTERED to distinguish GPU memory CPU mapped eagostini
@ 2022-04-29 14:14 ` eagostini
2022-05-24 21:11 ` Thomas Monjalon
0 siblings, 1 reply; 3+ messages in thread
From: eagostini @ 2022-04-29 14:14 UTC (permalink / raw)
To: dev; +Cc: Elena Agostini
From: Elena Agostini <eagostini@nvidia.com>
Enable GPU_REGISTERED flag in gpu/cuda driver in the memory list.
If a GPU memory address CPU mapped is freed before being
unmapped, CUDA driver unmaps it before freeing the memory.
Signed-off-by: Elena Agostini <eagostini@nvidia.com>
---
drivers/gpu/cuda/cuda.c | 77 ++++++++++++++++++++++++-----------------
1 file changed, 46 insertions(+), 31 deletions(-)
diff --git a/drivers/gpu/cuda/cuda.c b/drivers/gpu/cuda/cuda.c
index 5b8476ac20..664605d9fb 100644
--- a/drivers/gpu/cuda/cuda.c
+++ b/drivers/gpu/cuda/cuda.c
@@ -257,7 +257,7 @@ struct cuda_info {
enum mem_type {
GPU_MEM = 0,
CPU_REGISTERED,
- GPU_REGISTERED /* Not used yet */
+ GPU_REGISTERED
};
/* key associated to a memory address */
@@ -953,13 +953,14 @@ cuda_mem_cpu_map(struct rte_gpu *dev, __rte_unused size_t size, void *ptr_in, vo
return -rte_errno;
}
+ mem_item->mtype = GPU_REGISTERED;
*ptr_out = mem_item->ptr_h;
return 0;
}
static int
-cuda_mem_free(struct rte_gpu *dev, void *ptr)
+cuda_mem_unregister(struct rte_gpu *dev, void *ptr)
{
CUresult res;
struct mem_entry *mem_item;
@@ -978,11 +979,11 @@ cuda_mem_free(struct rte_gpu *dev, void *ptr)
return -rte_errno;
}
- if (mem_item->mtype == GPU_MEM) {
- res = pfn_cuMemFree(mem_item->ptr_orig_d);
+ if (mem_item->mtype == CPU_REGISTERED) {
+ res = pfn_cuMemHostUnregister(ptr);
if (res != 0) {
pfn_cuGetErrorString(res, &(err_string));
- rte_cuda_log(ERR, "cuMemFree current failed with %s",
+ rte_cuda_log(ERR, "cuMemHostUnregister current failed with %s",
err_string);
rte_errno = EPERM;
return -rte_errno;
@@ -993,74 +994,88 @@ cuda_mem_free(struct rte_gpu *dev, void *ptr)
rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
- return -EPERM;
+ rte_errno = EPERM;
+ return -rte_errno;
}
static int
-cuda_mem_unregister(struct rte_gpu *dev, void *ptr)
+cuda_mem_cpu_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 (mem_item->mtype == CPU_REGISTERED) {
- res = pfn_cuMemHostUnregister(ptr);
- if (res != 0) {
- pfn_cuGetErrorString(res, &(err_string));
- rte_cuda_log(ERR, "cuMemHostUnregister 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);
-
- rte_errno = EPERM;
- return -rte_errno;
+ return 0;
}
static int
-cuda_mem_cpu_unmap(struct rte_gpu *dev, void *ptr_in)
+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_in);
+ 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_in);
+ rte_cuda_log(ERR, "Memory address 0x%p not found in driver memory", ptr);
rte_errno = EPERM;
return -rte_errno;
}
- 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;
+ /*
+ * 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);
+ rte_errno = EPERM;
+ return -rte_errno;
+ }
+
+ return mem_list_del_item(hk);
}
- return 0;
+ rte_cuda_log(ERR, "Memory type %d not supported", mem_item->mtype);
+
+ return -EPERM;
}
static int
--
2.25.1
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] gpu/cuda: GPU_REGISTERED to distinguish GPU memory CPU mapped
2022-04-29 14:14 ` eagostini
@ 2022-05-24 21:11 ` Thomas Monjalon
0 siblings, 0 replies; 3+ messages in thread
From: Thomas Monjalon @ 2022-05-24 21:11 UTC (permalink / raw)
To: Elena Agostini; +Cc: dev
29/04/2022 16:14, eagostini@nvidia.com:
> From: Elena Agostini <eagostini@nvidia.com>
>
> Enable GPU_REGISTERED flag in gpu/cuda driver in the memory list.
> If a GPU memory address CPU mapped is freed before being
> unmapped, CUDA driver unmaps it before freeing the memory.
>
> Signed-off-by: Elena Agostini <eagostini@nvidia.com>
Applied, thanks.
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2022-05-24 21:11 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-04-29 13:52 [PATCH] gpu/cuda: GPU_REGISTERED to distinguish GPU memory CPU mapped eagostini
2022-04-29 14:14 ` eagostini
2022-05-24 21:11 ` Thomas Monjalon
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).