| Message ID | 20200403200757.886443-4-lyude@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | drm/dp_mst: drm_dp_check_act_status() fixes | expand |
On Fri, Apr 3, 2020 at 4:08 PM Lyude Paul <lyude@redhat.com> wrote: > > Currently we only poll for an ACT up to 30 times, with a busy-wait delay > of 100µs between each attempt - giving us a timeout of 2900µs. While > this might seem sensible, it would appear that in certain scenarios it > can take dramatically longer then that for us to receive an ACT. On one > of the EVGA MST hubs that I have available, I observed said hub > sometimes taking longer then a second before signalling the ACT. These > delays mostly seem to occur when previous sideband messages we've sent > are NAKd by the hub, however it wouldn't be particularly surprising if > it's possible to reproduce times like this simply by introducing branch > devices with large LCTs since payload allocations have to take effect on > every downstream device up to the payload's target. > > So, instead of just retrying 30 times we poll for the ACT for up to 3ms, > and additionally use usleep_range() to avoid a very long and rude > busy-wait. Note that the previous retry count of 30 appears to have been > arbitrarily chosen, as I can't find any mention of a recommended timeout > or retry count for ACTs in the DisplayPort 2.0 specification. This also > goes for the range we were previously using for udelay(), although I > suspect that was just copied from the recommended delay for link > training on SST devices. > > Signed-off-by: Lyude Paul <lyude@redhat.com> > Fixes: ad7f8a1f9ced ("drm/helper: add Displayport multi-stream helper (v0.6)") > Cc: Sean Paul <sean@poorly.run> > Cc: <stable@vger.kernel.org> # v3.17+ > --- > drivers/gpu/drm/drm_dp_mst_topology.c | 26 +++++++++++++++++++------- > 1 file changed, 19 insertions(+), 7 deletions(-) > > diff --git a/drivers/gpu/drm/drm_dp_mst_topology.c b/drivers/gpu/drm/drm_dp_mst_topology.c > index 7aaf184a2e5f..f313407374ed 100644 > --- a/drivers/gpu/drm/drm_dp_mst_topology.c > +++ b/drivers/gpu/drm/drm_dp_mst_topology.c > @@ -4466,17 +4466,30 @@ static int drm_dp_dpcd_write_payload(struct drm_dp_mst_topology_mgr *mgr, > * @mgr: manager to use > * > * Tries waiting for the MST hub to finish updating it's payload table by > - * polling for the ACT handled bit. > + * polling for the ACT handled bit for up to 3 seconds (yes-some hubs really > + * take that long). > * > * Returns: > * 0 if the ACT was handled in time, negative error code on failure. > */ > int drm_dp_check_act_status(struct drm_dp_mst_topology_mgr *mgr) > { > - int count = 0, ret; > + /* > + * There doesn't seem to be any recommended retry count or timeout in > + * the MST specification. Since some hubs have been observed to take > + * over 1 second to update their payload allocations under certain > + * conditions, we use a rather large timeout value. > + */ > + const int timeout_ms = 3000; > + unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms); > + int ret; > + bool retrying = false; > u8 status; > > do { > + if (retrying) > + usleep_range(100, 1000); > + > ret = drm_dp_dpcd_readb(mgr->aux, > DP_PAYLOAD_TABLE_UPDATE_STATUS, > &status); > @@ -4488,13 +4501,12 @@ int drm_dp_check_act_status(struct drm_dp_mst_topology_mgr *mgr) > > if (status & DP_PAYLOAD_ACT_HANDLED) > break; > - count++; > - udelay(100); > - } while (count < 30); > + retrying = true; > + } while (jiffies < timeout); Somewhat academic, but I think there's an overflow possibility here if timeout is near ulong_max and jiffies overflows during the usleep. In that case we'll be retrying for a very loong time. I wish we had i915's wait_for() macro available to all drm... Sean > > if (!(status & DP_PAYLOAD_ACT_HANDLED)) { > - DRM_DEBUG_KMS("failed to get ACT bit %d after %d retries\n", > - status, count); > + DRM_DEBUG_KMS("failed to get ACT bit %d after %dms\n", > + status, timeout_ms); > return -EINVAL; > } > return 0; > -- > 2.25.1 >
On Mon, 2020-04-06 at 15:41 -0400, Sean Paul wrote: > On Fri, Apr 3, 2020 at 4:08 PM Lyude Paul <lyude@redhat.com> wrote: > > Currently we only poll for an ACT up to 30 times, with a busy-wait delay > > of 100µs between each attempt - giving us a timeout of 2900µs. While > > this might seem sensible, it would appear that in certain scenarios it > > can take dramatically longer then that for us to receive an ACT. On one > > of the EVGA MST hubs that I have available, I observed said hub > > sometimes taking longer then a second before signalling the ACT. These > > delays mostly seem to occur when previous sideband messages we've sent > > are NAKd by the hub, however it wouldn't be particularly surprising if > > it's possible to reproduce times like this simply by introducing branch > > devices with large LCTs since payload allocations have to take effect on > > every downstream device up to the payload's target. > > > > So, instead of just retrying 30 times we poll for the ACT for up to 3ms, > > and additionally use usleep_range() to avoid a very long and rude > > busy-wait. Note that the previous retry count of 30 appears to have been > > arbitrarily chosen, as I can't find any mention of a recommended timeout > > or retry count for ACTs in the DisplayPort 2.0 specification. This also > > goes for the range we were previously using for udelay(), although I > > suspect that was just copied from the recommended delay for link > > training on SST devices. > > > > Signed-off-by: Lyude Paul <lyude@redhat.com> > > Fixes: ad7f8a1f9ced ("drm/helper: add Displayport multi-stream helper > > (v0.6)") > > Cc: Sean Paul <sean@poorly.run> > > Cc: <stable@vger.kernel.org> # v3.17+ > > --- > > drivers/gpu/drm/drm_dp_mst_topology.c | 26 +++++++++++++++++++------- > > 1 file changed, 19 insertions(+), 7 deletions(-) > > > > diff --git a/drivers/gpu/drm/drm_dp_mst_topology.c > > b/drivers/gpu/drm/drm_dp_mst_topology.c > > index 7aaf184a2e5f..f313407374ed 100644 > > --- a/drivers/gpu/drm/drm_dp_mst_topology.c > > +++ b/drivers/gpu/drm/drm_dp_mst_topology.c > > @@ -4466,17 +4466,30 @@ static int drm_dp_dpcd_write_payload(struct > > drm_dp_mst_topology_mgr *mgr, > > * @mgr: manager to use > > * > > * Tries waiting for the MST hub to finish updating it's payload table by > > - * polling for the ACT handled bit. > > + * polling for the ACT handled bit for up to 3 seconds (yes-some hubs > > really > > + * take that long). > > * > > * Returns: > > * 0 if the ACT was handled in time, negative error code on failure. > > */ > > int drm_dp_check_act_status(struct drm_dp_mst_topology_mgr *mgr) > > { > > - int count = 0, ret; > > + /* > > + * There doesn't seem to be any recommended retry count or timeout > > in > > + * the MST specification. Since some hubs have been observed to > > take > > + * over 1 second to update their payload allocations under certain > > + * conditions, we use a rather large timeout value. > > + */ > > + const int timeout_ms = 3000; > > + unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms); > > + int ret; > > + bool retrying = false; > > u8 status; > > > > do { > > + if (retrying) > > + usleep_range(100, 1000); > > + > > ret = drm_dp_dpcd_readb(mgr->aux, > > DP_PAYLOAD_TABLE_UPDATE_STATUS, > > &status); > > @@ -4488,13 +4501,12 @@ int drm_dp_check_act_status(struct > > drm_dp_mst_topology_mgr *mgr) > > > > if (status & DP_PAYLOAD_ACT_HANDLED) > > break; > > - count++; > > - udelay(100); > > - } while (count < 30); > > + retrying = true; > > + } while (jiffies < timeout); > > Somewhat academic, but I think there's an overflow possibility here if > timeout is near ulong_max and jiffies overflows during the usleep. In > that case we'll be retrying for a very loong time. > > I wish we had i915's wait_for() macro available to all drm... Maybe we could add it to the kernel library somewhere? I don't see why we'd need to stop at DRM > > Sean > > > if (!(status & DP_PAYLOAD_ACT_HANDLED)) { > > - DRM_DEBUG_KMS("failed to get ACT bit %d after %d > > retries\n", > > - status, count); > > + DRM_DEBUG_KMS("failed to get ACT bit %d after %dms\n", > > + status, timeout_ms); > > return -EINVAL; > > } > > return 0; > > -- > > 2.25.1 > >
On Mon, Apr 6, 2020 at 3:43 PM Lyude Paul <lyude@redhat.com> wrote: > > On Mon, 2020-04-06 at 15:41 -0400, Sean Paul wrote: > > On Fri, Apr 3, 2020 at 4:08 PM Lyude Paul <lyude@redhat.com> wrote: > > > Currently we only poll for an ACT up to 30 times, with a busy-wait delay > > > of 100µs between each attempt - giving us a timeout of 2900µs. While > > > this might seem sensible, it would appear that in certain scenarios it > > > can take dramatically longer then that for us to receive an ACT. On one > > > of the EVGA MST hubs that I have available, I observed said hub > > > sometimes taking longer then a second before signalling the ACT. These > > > delays mostly seem to occur when previous sideband messages we've sent > > > are NAKd by the hub, however it wouldn't be particularly surprising if > > > it's possible to reproduce times like this simply by introducing branch > > > devices with large LCTs since payload allocations have to take effect on > > > every downstream device up to the payload's target. > > > > > > So, instead of just retrying 30 times we poll for the ACT for up to 3ms, > > > and additionally use usleep_range() to avoid a very long and rude > > > busy-wait. Note that the previous retry count of 30 appears to have been > > > arbitrarily chosen, as I can't find any mention of a recommended timeout > > > or retry count for ACTs in the DisplayPort 2.0 specification. This also > > > goes for the range we were previously using for udelay(), although I > > > suspect that was just copied from the recommended delay for link > > > training on SST devices. > > > > > > Signed-off-by: Lyude Paul <lyude@redhat.com> > > > Fixes: ad7f8a1f9ced ("drm/helper: add Displayport multi-stream helper > > > (v0.6)") > > > Cc: Sean Paul <sean@poorly.run> > > > Cc: <stable@vger.kernel.org> # v3.17+ > > > --- > > > drivers/gpu/drm/drm_dp_mst_topology.c | 26 +++++++++++++++++++------- > > > 1 file changed, 19 insertions(+), 7 deletions(-) > > > > > > diff --git a/drivers/gpu/drm/drm_dp_mst_topology.c > > > b/drivers/gpu/drm/drm_dp_mst_topology.c > > > index 7aaf184a2e5f..f313407374ed 100644 > > > --- a/drivers/gpu/drm/drm_dp_mst_topology.c > > > +++ b/drivers/gpu/drm/drm_dp_mst_topology.c > > > @@ -4466,17 +4466,30 @@ static int drm_dp_dpcd_write_payload(struct > > > drm_dp_mst_topology_mgr *mgr, > > > * @mgr: manager to use > > > * > > > * Tries waiting for the MST hub to finish updating it's payload table by > > > - * polling for the ACT handled bit. > > > + * polling for the ACT handled bit for up to 3 seconds (yes-some hubs > > > really > > > + * take that long). > > > * > > > * Returns: > > > * 0 if the ACT was handled in time, negative error code on failure. > > > */ > > > int drm_dp_check_act_status(struct drm_dp_mst_topology_mgr *mgr) > > > { > > > - int count = 0, ret; > > > + /* > > > + * There doesn't seem to be any recommended retry count or timeout > > > in > > > + * the MST specification. Since some hubs have been observed to > > > take > > > + * over 1 second to update their payload allocations under certain > > > + * conditions, we use a rather large timeout value. > > > + */ > > > + const int timeout_ms = 3000; > > > + unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms); > > > + int ret; > > > + bool retrying = false; > > > u8 status; > > > > > > do { > > > + if (retrying) > > > + usleep_range(100, 1000); > > > + > > > ret = drm_dp_dpcd_readb(mgr->aux, > > > DP_PAYLOAD_TABLE_UPDATE_STATUS, > > > &status); > > > @@ -4488,13 +4501,12 @@ int drm_dp_check_act_status(struct > > > drm_dp_mst_topology_mgr *mgr) > > > > > > if (status & DP_PAYLOAD_ACT_HANDLED) > > > break; > > > - count++; > > > - udelay(100); > > > - } while (count < 30); > > > + retrying = true; > > > + } while (jiffies < timeout); > > > > Somewhat academic, but I think there's an overflow possibility here if > > timeout is near ulong_max and jiffies overflows during the usleep. In > > that case we'll be retrying for a very loong time. > > > > I wish we had i915's wait_for() macro available to all drm... > > Maybe we could add it to the kernel library somewhere? I don't see why we'd > need to stop at DRM So You Want To Build A Bikeshed... Seriously though, I'd be very happy with that. Alternatively you could shoehorn this into readx_poll_timeout as well. Sean > > > > > Sean > > > > > if (!(status & DP_PAYLOAD_ACT_HANDLED)) { > > > - DRM_DEBUG_KMS("failed to get ACT bit %d after %d > > > retries\n", > > > - status, count); > > > + DRM_DEBUG_KMS("failed to get ACT bit %d after %dms\n", > > > + status, timeout_ms); > > > return -EINVAL; > > > } > > > return 0; > > > -- > > > 2.25.1 > > > > -- > Cheers, > Lyude Paul (she/her) > Associate Software Engineer at Red Hat >
diff --git a/drivers/gpu/drm/drm_dp_mst_topology.c b/drivers/gpu/drm/drm_dp_mst_topology.c index 7aaf184a2e5f..f313407374ed 100644 --- a/drivers/gpu/drm/drm_dp_mst_topology.c +++ b/drivers/gpu/drm/drm_dp_mst_topology.c @@ -4466,17 +4466,30 @@ static int drm_dp_dpcd_write_payload(struct drm_dp_mst_topology_mgr *mgr, * @mgr: manager to use * * Tries waiting for the MST hub to finish updating it's payload table by - * polling for the ACT handled bit. + * polling for the ACT handled bit for up to 3 seconds (yes-some hubs really + * take that long). * * Returns: * 0 if the ACT was handled in time, negative error code on failure. */ int drm_dp_check_act_status(struct drm_dp_mst_topology_mgr *mgr) { - int count = 0, ret; + /* + * There doesn't seem to be any recommended retry count or timeout in + * the MST specification. Since some hubs have been observed to take + * over 1 second to update their payload allocations under certain + * conditions, we use a rather large timeout value. + */ + const int timeout_ms = 3000; + unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms); + int ret; + bool retrying = false; u8 status; do { + if (retrying) + usleep_range(100, 1000); + ret = drm_dp_dpcd_readb(mgr->aux, DP_PAYLOAD_TABLE_UPDATE_STATUS, &status); @@ -4488,13 +4501,12 @@ int drm_dp_check_act_status(struct drm_dp_mst_topology_mgr *mgr) if (status & DP_PAYLOAD_ACT_HANDLED) break; - count++; - udelay(100); - } while (count < 30); + retrying = true; + } while (jiffies < timeout); if (!(status & DP_PAYLOAD_ACT_HANDLED)) { - DRM_DEBUG_KMS("failed to get ACT bit %d after %d retries\n", - status, count); + DRM_DEBUG_KMS("failed to get ACT bit %d after %dms\n", + status, timeout_ms); return -EINVAL; } return 0;
Currently we only poll for an ACT up to 30 times, with a busy-wait delay of 100µs between each attempt - giving us a timeout of 2900µs. While this might seem sensible, it would appear that in certain scenarios it can take dramatically longer then that for us to receive an ACT. On one of the EVGA MST hubs that I have available, I observed said hub sometimes taking longer then a second before signalling the ACT. These delays mostly seem to occur when previous sideband messages we've sent are NAKd by the hub, however it wouldn't be particularly surprising if it's possible to reproduce times like this simply by introducing branch devices with large LCTs since payload allocations have to take effect on every downstream device up to the payload's target. So, instead of just retrying 30 times we poll for the ACT for up to 3ms, and additionally use usleep_range() to avoid a very long and rude busy-wait. Note that the previous retry count of 30 appears to have been arbitrarily chosen, as I can't find any mention of a recommended timeout or retry count for ACTs in the DisplayPort 2.0 specification. This also goes for the range we were previously using for udelay(), although I suspect that was just copied from the recommended delay for link training on SST devices. Signed-off-by: Lyude Paul <lyude@redhat.com> Fixes: ad7f8a1f9ced ("drm/helper: add Displayport multi-stream helper (v0.6)") Cc: Sean Paul <sean@poorly.run> Cc: <stable@vger.kernel.org> # v3.17+ --- drivers/gpu/drm/drm_dp_mst_topology.c | 26 +++++++++++++++++++------- 1 file changed, 19 insertions(+), 7 deletions(-)