On Sat, Nov 27, 2021 at 02:39:52PM +0100, Heiner Kallweit wrote:
> When discussing whether pm_runtime_resume_and_get() should be annotated as
> __must_check, I became aware of RPM usage in crypto/stm32.
>
> Following two patches replace usage of pm_runtime_get_sync() with
> pm_runtime_resume_and_get() w/o checking the return code.
>
> 747bf30fd944 ("crypto: stm32/cryp - Fix PM reference leak on stm32-cryp.c")
> 1cb3ad701970 ("crypto: stm32/hash - Fix PM reference leak on stm32-hash.c")
>
> This results in RPM usage like the following in stm32_hash_export():
>
> pm_runtime_resume_and_get(hdev->dev);
> ...
> pm_runtime_mark_last_busy(hdev->dev);
> pm_runtime_put_autosuspend(hdev->dev);
>
> This is broken. After pm_runtime_resume_and_get() the usage count may be
> incremented or not. If not, then the call to pm_runtime_put_autosuspend()
> results in exactly the imbalance that the patch claims to fix.
>
> Therefore I think both patches should be reverted, or the return code
> of pm_runtime_resume_and_get() has to be checked and properly handled
> in the driver logic.
I agree. But we can't revert them completely because it does
fix some genuine issues with the ones where we do check the error
code. What about this patch?
---8<---
We should not call pm_runtime_resume_and_get where the reference
count is expected to be incremented unconditionally. This patch
reverts these calls to the original unconditional get_sync call.
Reported-by: Heiner Kallweit <[email protected]>
Fixes: 747bf30fd944 ("crypto: stm32/cryp - Fix PM reference leak...")
Fixes: 1cb3ad701970 ("crypto: stm32/hash - Fix PM reference leak...")
Signed-off-by: Herbert Xu <[email protected]>
diff --git a/drivers/crypto/stm32/stm32-cryp.c b/drivers/crypto/stm32/stm32-cryp.c
index 7389a0536ff0..05087831e199 100644
--- a/drivers/crypto/stm32/stm32-cryp.c
+++ b/drivers/crypto/stm32/stm32-cryp.c
@@ -542,7 +542,7 @@ static int stm32_cryp_hw_init(struct stm32_cryp *cryp)
int ret;
u32 cfg, hw_mode;
- pm_runtime_resume_and_get(cryp->dev);
+ pm_runtime_get_sync(cryp->dev);
/* Disable interrupt */
stm32_cryp_write(cryp, CRYP_IMSCR, 0);
diff --git a/drivers/crypto/stm32/stm32-hash.c b/drivers/crypto/stm32/stm32-hash.c
index 389de9e3302d..d33006d43f76 100644
--- a/drivers/crypto/stm32/stm32-hash.c
+++ b/drivers/crypto/stm32/stm32-hash.c
@@ -813,7 +813,7 @@ static void stm32_hash_finish_req(struct ahash_request *req, int err)
static int stm32_hash_hw_init(struct stm32_hash_dev *hdev,
struct stm32_hash_request_ctx *rctx)
{
- pm_runtime_resume_and_get(hdev->dev);
+ pm_runtime_get_sync(hdev->dev);
if (!(HASH_FLAGS_INIT & hdev->flags)) {
stm32_hash_write(hdev, HASH_CR, HASH_CR_INIT);
@@ -962,7 +962,7 @@ static int stm32_hash_export(struct ahash_request *req, void *out)
u32 *preg;
unsigned int i;
- pm_runtime_resume_and_get(hdev->dev);
+ pm_runtime_get_sync(hdev->dev);
while ((stm32_hash_read(hdev, HASH_SR) & HASH_SR_BUSY))
cpu_relax();
@@ -1000,7 +1000,7 @@ static int stm32_hash_import(struct ahash_request *req, const void *in)
preg = rctx->hw_context;
- pm_runtime_resume_and_get(hdev->dev);
+ pm_runtime_get_sync(hdev->dev);
stm32_hash_write(hdev, HASH_IMR, *preg++);
stm32_hash_write(hdev, HASH_STR, *preg++);
--
Email: Herbert Xu <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
On 01.12.2021 07:30, Herbert Xu wrote:
> On Sat, Nov 27, 2021 at 02:39:52PM +0100, Heiner Kallweit wrote:
>> When discussing whether pm_runtime_resume_and_get() should be annotated as
>> __must_check, I became aware of RPM usage in crypto/stm32.
>>
>> Following two patches replace usage of pm_runtime_get_sync() with
>> pm_runtime_resume_and_get() w/o checking the return code.
>>
>> 747bf30fd944 ("crypto: stm32/cryp - Fix PM reference leak on stm32-cryp.c")
>> 1cb3ad701970 ("crypto: stm32/hash - Fix PM reference leak on stm32-hash.c")
>>
>> This results in RPM usage like the following in stm32_hash_export():
>>
>> pm_runtime_resume_and_get(hdev->dev);
>> ...
>> pm_runtime_mark_last_busy(hdev->dev);
>> pm_runtime_put_autosuspend(hdev->dev);
>>
>> This is broken. After pm_runtime_resume_and_get() the usage count may be
>> incremented or not. If not, then the call to pm_runtime_put_autosuspend()
>> results in exactly the imbalance that the patch claims to fix.
>>
>> Therefore I think both patches should be reverted, or the return code
>> of pm_runtime_resume_and_get() has to be checked and properly handled
>> in the driver logic.
>
> I agree. But we can't revert them completely because it does
> fix some genuine issues with the ones where we do check the error
> code. What about this patch?
Looks good to me, thanks!
>
> ---8<---
> We should not call pm_runtime_resume_and_get where the reference
> count is expected to be incremented unconditionally. This patch
> reverts these calls to the original unconditional get_sync call.
>
> Reported-by: Heiner Kallweit <[email protected]>
> Fixes: 747bf30fd944 ("crypto: stm32/cryp - Fix PM reference leak...")
> Fixes: 1cb3ad701970 ("crypto: stm32/hash - Fix PM reference leak...")
> Signed-off-by: Herbert Xu <[email protected]>
>
> diff --git a/drivers/crypto/stm32/stm32-cryp.c b/drivers/crypto/stm32/stm32-cryp.c
> index 7389a0536ff0..05087831e199 100644
> --- a/drivers/crypto/stm32/stm32-cryp.c
> +++ b/drivers/crypto/stm32/stm32-cryp.c
> @@ -542,7 +542,7 @@ static int stm32_cryp_hw_init(struct stm32_cryp *cryp)
> int ret;
> u32 cfg, hw_mode;
>
> - pm_runtime_resume_and_get(cryp->dev);
> + pm_runtime_get_sync(cryp->dev);
>
> /* Disable interrupt */
> stm32_cryp_write(cryp, CRYP_IMSCR, 0);
> diff --git a/drivers/crypto/stm32/stm32-hash.c b/drivers/crypto/stm32/stm32-hash.c
> index 389de9e3302d..d33006d43f76 100644
> --- a/drivers/crypto/stm32/stm32-hash.c
> +++ b/drivers/crypto/stm32/stm32-hash.c
> @@ -813,7 +813,7 @@ static void stm32_hash_finish_req(struct ahash_request *req, int err)
> static int stm32_hash_hw_init(struct stm32_hash_dev *hdev,
> struct stm32_hash_request_ctx *rctx)
> {
> - pm_runtime_resume_and_get(hdev->dev);
> + pm_runtime_get_sync(hdev->dev);
>
> if (!(HASH_FLAGS_INIT & hdev->flags)) {
> stm32_hash_write(hdev, HASH_CR, HASH_CR_INIT);
> @@ -962,7 +962,7 @@ static int stm32_hash_export(struct ahash_request *req, void *out)
> u32 *preg;
> unsigned int i;
>
> - pm_runtime_resume_and_get(hdev->dev);
> + pm_runtime_get_sync(hdev->dev);
>
> while ((stm32_hash_read(hdev, HASH_SR) & HASH_SR_BUSY))
> cpu_relax();
> @@ -1000,7 +1000,7 @@ static int stm32_hash_import(struct ahash_request *req, const void *in)
>
> preg = rctx->hw_context;
>
> - pm_runtime_resume_and_get(hdev->dev);
> + pm_runtime_get_sync(hdev->dev);
>
> stm32_hash_write(hdev, HASH_IMR, *preg++);
> stm32_hash_write(hdev, HASH_STR, *preg++);
>
On Wed, Dec 1, 2021 at 7:30 AM Herbert Xu <[email protected]> wrote:
>
> On Sat, Nov 27, 2021 at 02:39:52PM +0100, Heiner Kallweit wrote:
> > When discussing whether pm_runtime_resume_and_get() should be annotated as
> > __must_check, I became aware of RPM usage in crypto/stm32.
> >
> > Following two patches replace usage of pm_runtime_get_sync() with
> > pm_runtime_resume_and_get() w/o checking the return code.
> >
> > 747bf30fd944 ("crypto: stm32/cryp - Fix PM reference leak on stm32-cryp.c")
> > 1cb3ad701970 ("crypto: stm32/hash - Fix PM reference leak on stm32-hash.c")
> >
> > This results in RPM usage like the following in stm32_hash_export():
> >
> > pm_runtime_resume_and_get(hdev->dev);
> > ...
> > pm_runtime_mark_last_busy(hdev->dev);
> > pm_runtime_put_autosuspend(hdev->dev);
> >
> > This is broken. After pm_runtime_resume_and_get() the usage count may be
> > incremented or not. If not, then the call to pm_runtime_put_autosuspend()
> > results in exactly the imbalance that the patch claims to fix.
> >
> > Therefore I think both patches should be reverted, or the return code
> > of pm_runtime_resume_and_get() has to be checked and properly handled
> > in the driver logic.
>
> I agree. But we can't revert them completely because it does
> fix some genuine issues with the ones where we do check the error
> code. What about this patch?
>
> ---8<---
> We should not call pm_runtime_resume_and_get where the reference
> count is expected to be incremented unconditionally. This patch
> reverts these calls to the original unconditional get_sync call.
Acked-by: Rafael J. Wysocki <[email protected]>
The rule of thumb is to use pm_runtime_resume_and_get() if you want to
check the return value (and act on it) and pm_runtime_get_sync()
otherwise.
> Reported-by: Heiner Kallweit <[email protected]>
> Fixes: 747bf30fd944 ("crypto: stm32/cryp - Fix PM reference leak...")
> Fixes: 1cb3ad701970 ("crypto: stm32/hash - Fix PM reference leak...")
> Signed-off-by: Herbert Xu <[email protected]>
>
> diff --git a/drivers/crypto/stm32/stm32-cryp.c b/drivers/crypto/stm32/stm32-cryp.c
> index 7389a0536ff0..05087831e199 100644
> --- a/drivers/crypto/stm32/stm32-cryp.c
> +++ b/drivers/crypto/stm32/stm32-cryp.c
> @@ -542,7 +542,7 @@ static int stm32_cryp_hw_init(struct stm32_cryp *cryp)
> int ret;
> u32 cfg, hw_mode;
>
> - pm_runtime_resume_and_get(cryp->dev);
> + pm_runtime_get_sync(cryp->dev);
>
> /* Disable interrupt */
> stm32_cryp_write(cryp, CRYP_IMSCR, 0);
> diff --git a/drivers/crypto/stm32/stm32-hash.c b/drivers/crypto/stm32/stm32-hash.c
> index 389de9e3302d..d33006d43f76 100644
> --- a/drivers/crypto/stm32/stm32-hash.c
> +++ b/drivers/crypto/stm32/stm32-hash.c
> @@ -813,7 +813,7 @@ static void stm32_hash_finish_req(struct ahash_request *req, int err)
> static int stm32_hash_hw_init(struct stm32_hash_dev *hdev,
> struct stm32_hash_request_ctx *rctx)
> {
> - pm_runtime_resume_and_get(hdev->dev);
> + pm_runtime_get_sync(hdev->dev);
>
> if (!(HASH_FLAGS_INIT & hdev->flags)) {
> stm32_hash_write(hdev, HASH_CR, HASH_CR_INIT);
> @@ -962,7 +962,7 @@ static int stm32_hash_export(struct ahash_request *req, void *out)
> u32 *preg;
> unsigned int i;
>
> - pm_runtime_resume_and_get(hdev->dev);
> + pm_runtime_get_sync(hdev->dev);
>
> while ((stm32_hash_read(hdev, HASH_SR) & HASH_SR_BUSY))
> cpu_relax();
> @@ -1000,7 +1000,7 @@ static int stm32_hash_import(struct ahash_request *req, const void *in)
>
> preg = rctx->hw_context;
>
> - pm_runtime_resume_and_get(hdev->dev);
> + pm_runtime_get_sync(hdev->dev);
>
> stm32_hash_write(hdev, HASH_IMR, *preg++);
> stm32_hash_write(hdev, HASH_STR, *preg++);
> --
> Email: Herbert Xu <[email protected]>
> Home Page: http://gondor.apana.org.au/~herbert/
> PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt