| Message ID | 1493983066-28202-1-git-send-email-sunil.kovvuri@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, May 5, 2017 at 4:47 PM, <sunil.kovvuri@gmail.com> wrote: > From: Sunil Goutham <sgoutham@cavium.com> > > Processing queue full of TLB invalidation commands might > take more time on some platforms than current timeout > of 100us. So increased drain timeout value. > > Also now udelay time is increased exponentially for each poll. > > Signed-off-by: Sunil Goutham <sgoutham@cavium.com> > --- > > v2 > - Addressed Will's and Robin's comments i.e > increased poll timeout only for drain case and removed spin loop > whose logic anyway is screwed up. > - Also changed commit subject and message as this patch no longer > exactly reflects what is done in SMMU driver i.e > 8513c8930069 iommu/arm-smmu: Poll for TLB sync completion more effectively > > drivers/iommu/arm-smmu-v3.c | 15 +++++++++++++-- > 1 file changed, 13 insertions(+), 2 deletions(-) > > diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c > index d412bdd..cbc8309 100644 > --- a/drivers/iommu/arm-smmu-v3.c > +++ b/drivers/iommu/arm-smmu-v3.c > @@ -379,6 +379,8 @@ > #define CMDQ_SYNC_0_CS_NONE (0UL << CMDQ_SYNC_0_CS_SHIFT) > #define CMDQ_SYNC_0_CS_SEV (2UL << CMDQ_SYNC_0_CS_SHIFT) > > +#define CMDQ_DRAIN_TIMEOUT_US 1000 > + > /* Event queue */ > #define EVTQ_ENT_DWORDS 4 > #define EVTQ_MAX_SZ_SHIFT 7 > @@ -737,7 +739,12 @@ static void queue_inc_prod(struct arm_smmu_queue *q) > */ > static int queue_poll_cons(struct arm_smmu_queue *q, bool drain, bool wfe) > { > - ktime_t timeout = ktime_add_us(ktime_get(), ARM_SMMU_POLL_TIMEOUT_US); > + ktime_t timeout; > + unsigned int delay = 1; > + > + /* Wait longer if it's queue drain */ > + timeout = ktime_add_us(ktime_get(), drain ? CMDQ_DRAIN_TIMEOUT_US : > + ARM_SMMU_POLL_TIMEOUT_US); > > while (queue_sync_cons(q), (drain ? !queue_empty(q) : queue_full(q))) { > if (ktime_compare(ktime_get(), timeout) > 0) > @@ -747,7 +754,11 @@ static int queue_poll_cons(struct arm_smmu_queue *q, bool drain, bool wfe) > wfe(); > } else { > cpu_relax(); > - udelay(1); > + udelay(delay); > + /* No point in sleeping for fixed time, > + * if cons isn't moving fast. > + */ > + delay *= 2; > } > } > > -- > 2.7.4 > Gentle reminder for patch review. Thanks, Sunil.
Hi Sunil, On Mon, May 29, 2017 at 02:41:59PM +0530, Sunil Kovvuri wrote: > On Fri, May 5, 2017 at 4:47 PM, <sunil.kovvuri@gmail.com> wrote: > > From: Sunil Goutham <sgoutham@cavium.com> > > > > Processing queue full of TLB invalidation commands might > > take more time on some platforms than current timeout > > of 100us. So increased drain timeout value. > > > > Also now udelay time is increased exponentially for each poll. > > > > Signed-off-by: Sunil Goutham <sgoutham@cavium.com> [...] > Gentle reminder for patch review. Sorry, I forgot to push this out. Should be on my iommu/devel branch now. Will
diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c index d412bdd..cbc8309 100644 --- a/drivers/iommu/arm-smmu-v3.c +++ b/drivers/iommu/arm-smmu-v3.c @@ -379,6 +379,8 @@ #define CMDQ_SYNC_0_CS_NONE (0UL << CMDQ_SYNC_0_CS_SHIFT) #define CMDQ_SYNC_0_CS_SEV (2UL << CMDQ_SYNC_0_CS_SHIFT) +#define CMDQ_DRAIN_TIMEOUT_US 1000 + /* Event queue */ #define EVTQ_ENT_DWORDS 4 #define EVTQ_MAX_SZ_SHIFT 7 @@ -737,7 +739,12 @@ static void queue_inc_prod(struct arm_smmu_queue *q) */ static int queue_poll_cons(struct arm_smmu_queue *q, bool drain, bool wfe) { - ktime_t timeout = ktime_add_us(ktime_get(), ARM_SMMU_POLL_TIMEOUT_US); + ktime_t timeout; + unsigned int delay = 1; + + /* Wait longer if it's queue drain */ + timeout = ktime_add_us(ktime_get(), drain ? CMDQ_DRAIN_TIMEOUT_US : + ARM_SMMU_POLL_TIMEOUT_US); while (queue_sync_cons(q), (drain ? !queue_empty(q) : queue_full(q))) { if (ktime_compare(ktime_get(), timeout) > 0) @@ -747,7 +754,11 @@ static int queue_poll_cons(struct arm_smmu_queue *q, bool drain, bool wfe) wfe(); } else { cpu_relax(); - udelay(1); + udelay(delay); + /* No point in sleeping for fixed time, + * if cons isn't moving fast. + */ + delay *= 2; } }