Re: [Y2038] [PATCH 15/16] drm/etnaviv: use ktime_t for timeouts
On Mon, Nov 11, 2019 at 10:55 AM Lucas Stach wrote:
> >
> > > If that's the case then we should never encounter a genuine 0 timeout
> > > and this change would be okay.
> >
> > That's quite likely, I'd say any program passing {0,0} as a timeout without
> > ETNA_WAIT_NONBLOCK is already broken, but if we leave it like that,
> > it would be best to describe the reasoning in the changelog.
> >
> > Should I change the changelog, or change the patch to restore the
> > current behavior instead?
> >
> > I guess I could fold the change below into my patch to make it transparent
> > to the application again.
>
> If we assume 0 to never be a valid timeout, due to monotonic clock
> starting at 0 and never wrapping then I think we shouldn't introduce
> any additional code complexity to fix up the return value for this
> specific case. I'm not aware of any etnaviv userspace being broken in
> this way to rely on the return value for an invalid timeout input.
>
> Please just amend the commit message to mention the change in behavior
> and why we think it is safe to do.
Russell had some additional concerns that he raised on IRC,
and I did a new simpler implementation of the patch, plus a related
bugfix.
Please have a look at those.
Arnd
___
Y2038 mailing list
Y2038@lists.linaro.org
https://lists.linaro.org/mailman/listinfo/y2038
Re: [Y2038] [PATCH 15/16] drm/etnaviv: use ktime_t for timeouts
On Sa, 2019-11-09 at 13:12 +0100, Arnd Bergmann wrote:
> On Sat, Nov 9, 2019 at 12:03 AM Lucas Stach wrote:
> > Am Freitag, den 08.11.2019, 22:32 +0100 schrieb Arnd Bergmann:
> > > struct timespec is being removed from the kernel because it often leads
> > > to code that is not y2038-safe.
> > >
> > > In the etnaviv driver, monotonic timestamps are used, which do not suffer
> > > from overflow, but using ktime_t still leads to better code overall.
> > >
> > > The conversion is straightforward for the most part, except for
> > > etnaviv_timeout_to_jiffies(), which needs to handle arguments larger
> > > than MAX_JIFFY_OFFSET on 32-bit architectures.
> > >
> > > Signed-off-by: Arnd Bergmann
> > > @@ -368,7 +366,7 @@ static int etnaviv_ioctl_wait_fence(struct drm_device
> > > *dev, void *data,
> > > return -ENXIO;
> > >
> > > if (args->flags & ETNA_WAIT_NONBLOCK)
> > > - timeout = NULL;
> > > + timeout = ktime_set(0, 0);
> >
> > This is a change in behavior, as far as I can see. After this change
> > the called internal function is not able to differentiate between a
> > NONBLOCK call and a blocking call with 0 timeout. The difference being
> > that on a busy object the NONBLOCK call will return -EBUSY, while a
> > blocking call will return -ETIMEDOUT.
>
> Ah, good point. I created this patch a long time ago (cherry-picked it out
> of an older branch I had done), so I don't remember how I concluded this
> was a safe conversion, of if I missed that difference originally.
>
> > But then CLOCK_MONOTONIC starts at 0 and should not never wrap, right?
>
> Yes, that is correct.
>
> > If that's the case then we should never encounter a genuine 0 timeout
> > and this change would be okay.
>
> That's quite likely, I'd say any program passing {0,0} as a timeout without
> ETNA_WAIT_NONBLOCK is already broken, but if we leave it like that,
> it would be best to describe the reasoning in the changelog.
>
> Should I change the changelog, or change the patch to restore the
> current behavior instead?
>
> I guess I could fold the change below into my patch to make it transparent
> to the application again.
If we assume 0 to never be a valid timeout, due to monotonic clock
starting at 0 and never wrapping then I think we shouldn't introduce
any additional code complexity to fix up the return value for this
specific case. I'm not aware of any etnaviv userspace being broken in
this way to rely on the return value for an invalid timeout input.
Please just amend the commit message to mention the change in behavior
and why we think it is safe to do.
Regards,
Lucas
> Arnd
>
> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> index 1250c5e06329..162cedfb7f72 100644
> --- a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> @@ -354,6 +354,7 @@ static int etnaviv_ioctl_wait_fence(struct
> drm_device *dev, void *data,
> ktime_t timeout = ktime_set(args->timeout.tv_sec,
> args->timeout.tv_nsec);
> struct etnaviv_gpu *gpu;
> + int ret;
>
> if (args->flags & ~(ETNA_WAIT_NONBLOCK))
> return -EINVAL;
> @@ -365,8 +366,12 @@ static int etnaviv_ioctl_wait_fence(struct
> drm_device *dev, void *data,
> if (!gpu)
> return -ENXIO;
>
> - if (args->flags & ETNA_WAIT_NONBLOCK)
> - timeout = ktime_set(0, 0);
> + if (args->flags & ETNA_WAIT_NONBLOCK) {
> + ret = etnaviv_gpu_wait_fence_interruptible(gpu, args->fence,
> + ktime_set(0, 0));
> +
> + return (ret == -ETIMEDOUT) ? -EBUSY : ret;
> + }
>
> return etnaviv_gpu_wait_fence_interruptible(gpu, args->fence,
> timeout);
> @@ -421,10 +426,13 @@ static int etnaviv_ioctl_gem_wait(struct
> drm_device *dev, void *data,
> if (!obj)
> return -ENOENT;
>
> - if (args->flags & ETNA_WAIT_NONBLOCK)
> - timeout = ktime_set(0, 0);
> -
> - ret = etnaviv_gem_wait_bo(gpu, obj, timeout);
> + if (args->flags & ETNA_WAIT_NONBLOCK) {
> + ret = etnaviv_gem_wait_bo(gpu, obj, ktime_set(0, 0));
> + if (ret == -ETIMEDOUT)
> + ret = -EBUSY;
> + } else {
> + ret = etnaviv_gem_wait_bo(gpu, obj, timeout);
> + }
>
> drm_gem_object_put_unlocked(obj);
>
> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> index e42b1c4d902c..fa6986c5a5fe 100644
> --- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
> @@ -1135,6 +1135,7 @@ int etnaviv_gpu_wait_fence_interruptible(struct
> etnaviv_gpu *gpu,
> u32 id, ktime_t timeout)
> {
> struct dma_fence *fence;
> + unsigned long
Re: [Y2038] [PATCH 15/16] drm/etnaviv: use ktime_t for timeouts
On Sat, Nov 9, 2019 at 12:03 AM Lucas Stach wrote:
>
> Am Freitag, den 08.11.2019, 22:32 +0100 schrieb Arnd Bergmann:
> > struct timespec is being removed from the kernel because it often leads
> > to code that is not y2038-safe.
> >
> > In the etnaviv driver, monotonic timestamps are used, which do not suffer
> > from overflow, but using ktime_t still leads to better code overall.
> >
> > The conversion is straightforward for the most part, except for
> > etnaviv_timeout_to_jiffies(), which needs to handle arguments larger
> > than MAX_JIFFY_OFFSET on 32-bit architectures.
> >
> > Signed-off-by: Arnd Bergmann
> > @@ -368,7 +366,7 @@ static int etnaviv_ioctl_wait_fence(struct drm_device
> > *dev, void *data,
> > return -ENXIO;
> >
> > if (args->flags & ETNA_WAIT_NONBLOCK)
> > - timeout = NULL;
> > + timeout = ktime_set(0, 0);
>
> This is a change in behavior, as far as I can see. After this change
> the called internal function is not able to differentiate between a
> NONBLOCK call and a blocking call with 0 timeout. The difference being
> that on a busy object the NONBLOCK call will return -EBUSY, while a
> blocking call will return -ETIMEDOUT.
Ah, good point. I created this patch a long time ago (cherry-picked it out
of an older branch I had done), so I don't remember how I concluded this
was a safe conversion, of if I missed that difference originally.
> But then CLOCK_MONOTONIC starts at 0 and should not never wrap, right?
Yes, that is correct.
> If that's the case then we should never encounter a genuine 0 timeout
> and this change would be okay.
That's quite likely, I'd say any program passing {0,0} as a timeout without
ETNA_WAIT_NONBLOCK is already broken, but if we leave it like that,
it would be best to describe the reasoning in the changelog.
Should I change the changelog, or change the patch to restore the
current behavior instead?
I guess I could fold the change below into my patch to make it transparent
to the application again.
Arnd
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
index 1250c5e06329..162cedfb7f72 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
@@ -354,6 +354,7 @@ static int etnaviv_ioctl_wait_fence(struct
drm_device *dev, void *data,
ktime_t timeout = ktime_set(args->timeout.tv_sec,
args->timeout.tv_nsec);
struct etnaviv_gpu *gpu;
+ int ret;
if (args->flags & ~(ETNA_WAIT_NONBLOCK))
return -EINVAL;
@@ -365,8 +366,12 @@ static int etnaviv_ioctl_wait_fence(struct
drm_device *dev, void *data,
if (!gpu)
return -ENXIO;
- if (args->flags & ETNA_WAIT_NONBLOCK)
- timeout = ktime_set(0, 0);
+ if (args->flags & ETNA_WAIT_NONBLOCK) {
+ ret = etnaviv_gpu_wait_fence_interruptible(gpu, args->fence,
+ ktime_set(0, 0));
+
+ return (ret == -ETIMEDOUT) ? -EBUSY : ret;
+ }
return etnaviv_gpu_wait_fence_interruptible(gpu, args->fence,
timeout);
@@ -421,10 +426,13 @@ static int etnaviv_ioctl_gem_wait(struct
drm_device *dev, void *data,
if (!obj)
return -ENOENT;
- if (args->flags & ETNA_WAIT_NONBLOCK)
- timeout = ktime_set(0, 0);
-
- ret = etnaviv_gem_wait_bo(gpu, obj, timeout);
+ if (args->flags & ETNA_WAIT_NONBLOCK) {
+ ret = etnaviv_gem_wait_bo(gpu, obj, ktime_set(0, 0));
+ if (ret == -ETIMEDOUT)
+ ret = -EBUSY;
+ } else {
+ ret = etnaviv_gem_wait_bo(gpu, obj, timeout);
+ }
drm_gem_object_put_unlocked(obj);
diff --git a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
index e42b1c4d902c..fa6986c5a5fe 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_gpu.c
@@ -1135,6 +1135,7 @@ int etnaviv_gpu_wait_fence_interruptible(struct
etnaviv_gpu *gpu,
u32 id, ktime_t timeout)
{
struct dma_fence *fence;
+ unsigned long remaining;
int ret;
/*
@@ -1151,12 +1152,12 @@ int
etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
if (!fence)
return 0;
- if (!timeout) {
- /* No timeout was requested: just test for completion */
- ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
+ if (!timeout ||
+ (remaining = etnaviv_timeout_to_jiffies(timeout)) == 0) {
+ /* No timeout was requested, or timeout is already expired,
+* just test for completion */
+ ret = dma_fence_is_signaled(fence) ? 0 : -ETIMEDOUT;
} else {
- unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
-
Re: [Y2038] [PATCH 15/16] drm/etnaviv: use ktime_t for timeouts
Am Freitag, den 08.11.2019, 22:32 +0100 schrieb Arnd Bergmann:
> struct timespec is being removed from the kernel because it often leads
> to code that is not y2038-safe.
>
> In the etnaviv driver, monotonic timestamps are used, which do not suffer
> from overflow, but using ktime_t still leads to better code overall.
>
> The conversion is straightforward for the most part, except for
> etnaviv_timeout_to_jiffies(), which needs to handle arguments larger
> than MAX_JIFFY_OFFSET on 32-bit architectures.
>
> Signed-off-by: Arnd Bergmann
> ---
> drivers/gpu/drm/etnaviv/etnaviv_drv.c | 19 +--
> drivers/gpu/drm/etnaviv/etnaviv_drv.h | 21 +
> drivers/gpu/drm/etnaviv/etnaviv_gem.c | 5 ++---
> drivers/gpu/drm/etnaviv/etnaviv_gem.h | 2 +-
> drivers/gpu/drm/etnaviv/etnaviv_gpu.c | 4 ++--
> drivers/gpu/drm/etnaviv/etnaviv_gpu.h | 4 ++--
> 6 files changed, 25 insertions(+), 30 deletions(-)
>
> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> index 1f9c01be40d7..1250c5e06329 100644
> --- a/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_drv.c
> @@ -282,16 +282,13 @@ static int etnaviv_ioctl_gem_new(struct drm_device
> *dev, void *data,
> args->flags, >handle);
> }
>
> -#define TS(t) ((struct timespec){ \
> - .tv_sec = (t).tv_sec, \
> - .tv_nsec = (t).tv_nsec \
> -})
> -
> static int etnaviv_ioctl_gem_cpu_prep(struct drm_device *dev, void *data,
> struct drm_file *file)
> {
> struct drm_etnaviv_gem_cpu_prep *args = data;
> struct drm_gem_object *obj;
> + ktime_t timeout = ktime_set(args->timeout.tv_sec,
> + args->timeout.tv_nsec);
> int ret;
>
> if (args->op & ~(ETNA_PREP_READ | ETNA_PREP_WRITE | ETNA_PREP_NOSYNC))
> @@ -301,7 +298,7 @@ static int etnaviv_ioctl_gem_cpu_prep(struct drm_device
> *dev, void *data,
> if (!obj)
> return -ENOENT;
>
> - ret = etnaviv_gem_cpu_prep(obj, args->op, (args->timeout));
> + ret = etnaviv_gem_cpu_prep(obj, args->op, timeout);
>
> drm_gem_object_put_unlocked(obj);
>
> @@ -354,7 +351,8 @@ static int etnaviv_ioctl_wait_fence(struct drm_device
> *dev, void *data,
> {
> struct drm_etnaviv_wait_fence *args = data;
> struct etnaviv_drm_private *priv = dev->dev_private;
> - struct timespec *timeout = (args->timeout);
> + ktime_t timeout = ktime_set(args->timeout.tv_sec,
> + args->timeout.tv_nsec);
> struct etnaviv_gpu *gpu;
>
> if (args->flags & ~(ETNA_WAIT_NONBLOCK))
> @@ -368,7 +366,7 @@ static int etnaviv_ioctl_wait_fence(struct drm_device
> *dev, void *data,
> return -ENXIO;
>
> if (args->flags & ETNA_WAIT_NONBLOCK)
> - timeout = NULL;
> + timeout = ktime_set(0, 0);
This is a change in behavior, as far as I can see. After this change
the called internal function is not able to differentiate between a
NONBLOCK call and a blocking call with 0 timeout. The difference being
that on a busy object the NONBLOCK call will return -EBUSY, while a
blocking call will return -ETIMEDOUT.
But then CLOCK_MONOTONIC starts at 0 and should not never wrap, right?
If that's the case then we should never encounter a genuine 0 timeout
and this change would be okay.
Regards,
Lucas
>
> return etnaviv_gpu_wait_fence_interruptible(gpu, args->fence,
> timeout);
> @@ -403,7 +401,8 @@ static int etnaviv_ioctl_gem_wait(struct drm_device *dev,
> void *data,
> {
> struct etnaviv_drm_private *priv = dev->dev_private;
> struct drm_etnaviv_gem_wait *args = data;
> - struct timespec *timeout = (args->timeout);
> + ktime_t timeout = ktime_set(args->timeout.tv_sec,
> + args->timeout.tv_nsec);
> struct drm_gem_object *obj;
> struct etnaviv_gpu *gpu;
> int ret;
> @@ -423,7 +422,7 @@ static int etnaviv_ioctl_gem_wait(struct drm_device *dev,
> void *data,
> return -ENOENT;
>
> if (args->flags & ETNA_WAIT_NONBLOCK)
> - timeout = NULL;
> + timeout = ktime_set(0, 0);
>
> ret = etnaviv_gem_wait_bo(gpu, obj, timeout);
>
> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_drv.h
> b/drivers/gpu/drm/etnaviv/etnaviv_drv.h
> index 32cfa5a48d42..57a4e247bbcf 100644
> --- a/drivers/gpu/drm/etnaviv/etnaviv_drv.h
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_drv.h
> @@ -60,8 +60,7 @@ struct drm_gem_object
> *etnaviv_gem_prime_import_sg_table(struct drm_device *dev,
> int etnaviv_gem_prime_pin(struct drm_gem_object *obj);
> void etnaviv_gem_prime_unpin(struct drm_gem_object *obj);
> void *etnaviv_gem_vmap(struct drm_gem_object *obj);
> -int etnaviv_gem_cpu_prep(struct drm_gem_object *obj, u32 op,
> - struct timespec *timeout);
> +int etnaviv_gem_cpu_prep(struct
