On Tue, Jan 21, 2020 at 11:22 AM Lucas Stach l.stach@pengutronix.de wrote:
On Mo, 2020-01-20 at 19:47 +0100, Arnd Bergmann wrote:
On Mon, Jan 20, 2020 at 6:48 PM Lucas Stach l.stach@pengutronix.de wrote:
On Fr, 2020-01-17 at 16:47 +0100, Guido Günther wrote:
This breaks rendering here on arm64/gc7000 due to
ioctl(6, DRM_IOCTL_ETNAVIV_GEM_CPU_PREP or DRM_IOCTL_MSM_GEM_CPU_PREP, 0xfffff7888680) = -1 EINVAL (Invalid argument) ioctl(6, DRM_IOCTL_ETNAVIV_GEM_CPU_FINI or DRM_IOCTL_QXL_CLIENTCAP, 0xfffff78885e0) = 0 ioctl(6, DRM_IOCTL_ETNAVIV_GEM_CPU_PREP or DRM_IOCTL_MSM_GEM_CPU_PREP, 0xfffff7888680) = -1 EINVAL (Invalid argument) ioctl(6, DRM_IOCTL_ETNAVIV_GEM_CPU_FINI or DRM_IOCTL_QXL_CLIENTCAP, 0xfffff78885e0) = 0 ioctl(6, DRM_IOCTL_ETNAVIV_GEM_CPU_PREP or DRM_IOCTL_MSM_GEM_CPU_PREP, 0xfffff7888680) = -1 EINVAL (Invalid argument) ioctl(6, DRM_IOCTL_ETNAVIV_GEM_CPU_FINI or DRM_IOCTL_QXL_CLIENTCAP, 0xfffff78885e0) = 0
This is due to
get_abs_timeout(&req.timeout, 5000000000);in etna_bo_cpu_prep which can exceed NSEC_PER_SEC.
Should i send a patch to revert that change since it breaks existing userspace?
No need to revert. This patch has not been applied to the etnaviv tree yet, I guess it's just in one of Arnds branches feeding into -next.
That part of userspace is pretty dumb, as it misses to renormalize tv_nsec when it overflows the second boundary. So if what I see is correct it should be enough to allow 2 * NSEC_PER_SEC, which should both reject broken large timeout and keep existing userspace working.
Ah, so it's never more than 2 billion nanoseconds in known user space? I can definitely change my patch (actually add one on top) to allow that and handle it as before, or alternatively accept any 64-bit nanosecond value as arm64 already did, but make it less inefficient to handle.
So the broken userspace code looks like this:
static inline void get_abs_timeout(struct drm_etnaviv_timespec *tv, uint64_t ns) { struct timespec t; uint32_t s = ns / 1000000000; clock_gettime(CLOCK_MONOTONIC, &t); tv->tv_sec = t.tv_sec + s; tv->tv_nsec = t.tv_nsec + ns - (s * 1000000000); }
Which means it _tries_ to do the right thing by putting the billion part into the tv_sec member and only the remaining ns part is added to tv_nsec, but then it fails to propagate a tv_nsec overflow over NSEC_PER_SEC into tv_sec.
Which means the tv_nsec should never be more than 2 * NSEC_PER_SEC in known userspace. I would prefer if we could make the interface as strict as possible (i.e. no arbitrary large numbers in tv_nsec), while keeping this specific corner case working.
I've added a patch on top of my 2038 branch, please have a look at that.
Arnd