[v3] gpudev: manage NULL pointer
Checks
Commit Message
From: Elena Agostini <eagostini@nvidia.com>
gpudev free and unregister functions return gracefully if input pointer is NULL
Fixes: 818a067baf90 ("gpudev: manage NULL pointer")
Signed-off-by: Elena Agostini <eagostini@nvidia.com>
---
drivers/gpu/cuda/cuda.c | 6 ------
lib/gpudev/gpudev.c | 6 ++++++
2 files changed, 6 insertions(+), 6 deletions(-)
Comments
22/11/2021 19:24, eagostini@nvidia.com:
> From: Elena Agostini <eagostini@nvidia.com>
>
> gpudev free and unregister functions return gracefully if input pointer is NULL
We could add that the API doc was indicating NULL as a no-op accepted value.
Another explanation to add: cuda driver checks are removed because redundant
with the checks added in gpudev library.
> Fixes: 818a067baf90 ("gpudev: manage NULL pointer")
It should be:
Fixes: e818c4e2bf50 ("gpudev: add memory API")
>
> Signed-off-by: Elena Agostini <eagostini@nvidia.com>
> ---
> drivers/gpu/cuda/cuda.c | 6 ------
> lib/gpudev/gpudev.c | 6 ++++++
> 2 files changed, 6 insertions(+), 6 deletions(-)
[...]
> --- a/lib/gpudev/gpudev.c
> +++ b/lib/gpudev/gpudev.c
> @@ -569,6 +569,9 @@ rte_gpu_mem_free(int16_t dev_id, void *ptr)
> {
> struct rte_gpu *dev;
>
> + if (ptr == NULL)
> + return 0;
> +
> dev = gpu_get_by_id(dev_id);
> if (dev == NULL) {
> GPU_LOG(ERR, "free mem for invalid device ID %d", dev_id);
I think we should keep this check first.
> @@ -612,6 +615,9 @@ rte_gpu_mem_unregister(int16_t dev_id, void *ptr)
> {
> struct rte_gpu *dev;
>
> + if (ptr == NULL)
> + return 0;
> +
> dev = gpu_get_by_id(dev_id);
> if (dev == NULL) {
> GPU_LOG(ERR, "unregister mem for invalid device ID %d", dev_id);
Same here.
There is third function where NULL should be accepted: rte_gpu_mem_register
> From: Thomas Monjalon <thomas@monjalon.net>
> Date: Monday, 22 November 2021 at 12:23
> To: Elena Agostini <eagostini@nvidia.com>
> Cc: dev@dpdk.org <dev@dpdk.org>
> Subject: Re: [PATCH v3] gpudev: manage NULL pointer
> External email: Use caution opening links or attachments>
>
> 22/11/2021 19:24, eagostini@nvidia.com:
> > From: Elena Agostini <eagostini@nvidia.com>
> >
> > gpudev free and unregister functions return gracefully if input pointer is NULL>
> We could add that the API doc was indicating NULL as a no-op accepted value.>
> Another explanation to add: cuda driver checks are removed because redundant
> with the checks added in gpudev library.>
> > Fixes: 818a067baf90 ("gpudev: manage NULL pointer")>
> It should be:
> Fixes: e818c4e2bf50 ("gpudev: add memory API")>
> >
> > Signed-off-by: Elena Agostini <eagostini@nvidia.com>
> > ---
> > drivers/gpu/cuda/cuda.c | 6 ------
> > lib/gpudev/gpudev.c | 6 ++++++
> > 2 files changed, 6 insertions(+), 6 deletions(-)
> [...]
> > --- a/lib/gpudev/gpudev.c
> > +++ b/lib/gpudev/gpudev.c
> > @@ -569,6 +569,9 @@ rte_gpu_mem_free(int16_t dev_id, void *ptr)
> > {
> > struct rte_gpu *dev;
> >
> > + if (ptr == NULL)
> > + return 0;
> > +
> > dev = gpu_get_by_id(dev_id);
> > if (dev == NULL) {
> > GPU_LOG(ERR, "free mem for invalid device ID %d", dev_id);>
> I think we should keep this check first.
Why should gpudev waste more latency in looking for the device if the ptr is NULL?
> > @@ -612,6 +615,9 @@ rte_gpu_mem_unregister(int16_t dev_id, void *ptr)
> > {
> > struct rte_gpu *dev;
> >
> > + if (ptr == NULL)
> > + return 0;
> > +
> > dev = gpu_get_by_id(dev_id);
> > if (dev == NULL) {
> > GPU_LOG(ERR, "unregister mem for invalid device ID %d", dev_id);>
> Same here.
> There is third function where NULL should be accepted: rte_gpu_mem_register>
Thanks, let me add it
22/11/2021 12:34, Elena Agostini:
> From: Thomas Monjalon <thomas@monjalon.net>
> > 22/11/2021 19:24, eagostini@nvidia.com:
> > > --- a/lib/gpudev/gpudev.c
> > > +++ b/lib/gpudev/gpudev.c
> > > @@ -569,6 +569,9 @@ rte_gpu_mem_free(int16_t dev_id, void *ptr)
> > > {
> > > struct rte_gpu *dev;
> > >
> > > + if (ptr == NULL)
> > > + return 0;
> > > +
> > > dev = gpu_get_by_id(dev_id);
> > > if (dev == NULL) {
> > > GPU_LOG(ERR, "free mem for invalid device ID %d", dev_id);>
>
> > I think we should keep this check first.
>
> Why should gpudev waste more latency in looking for the device if the ptr is NULL?
Freeing with NULL pointer is not in the datapath I think,
probably just a failure cleanup case.
Having the dev_id check first allows to catch more bugs.
Returning 0 without checking the id looks weird to me.
@@ -764,9 +764,6 @@ cuda_mem_free(struct rte_gpu *dev, void *ptr)
if (dev == NULL)
return -ENODEV;
- if (ptr == NULL)
- return -EINVAL;
-
hk = get_hash_from_ptr((void *)ptr);
mem_item = mem_list_find_item(hk);
@@ -803,9 +800,6 @@ cuda_mem_unregister(struct rte_gpu *dev, void *ptr)
if (dev == NULL)
return -ENODEV;
- if (ptr == NULL)
- return -EINVAL;
-
hk = get_hash_from_ptr((void *)ptr);
mem_item = mem_list_find_item(hk);
@@ -569,6 +569,9 @@ rte_gpu_mem_free(int16_t dev_id, void *ptr)
{
struct rte_gpu *dev;
+ if (ptr == NULL)
+ return 0;
+
dev = gpu_get_by_id(dev_id);
if (dev == NULL) {
GPU_LOG(ERR, "free mem for invalid device ID %d", dev_id);
@@ -612,6 +615,9 @@ rte_gpu_mem_unregister(int16_t dev_id, void *ptr)
{
struct rte_gpu *dev;
+ if (ptr == NULL)
+ return 0;
+
dev = gpu_get_by_id(dev_id);
if (dev == NULL) {
GPU_LOG(ERR, "unregister mem for invalid device ID %d", dev_id);