usb: reducing an usb-port auto-resume latency.
Subsystems (2)
PATCH 1/1 usb: reducing an usb-port auto-resume latency.
From: Anshuman Gupta <[email protected]>
This patch will improve the variable auto-resume latency of an usb-port.
When xhci gets a port status change event interrupt due to PORT_PLC (port link state transition), linux Host controller driver drives the resume signalling on the bus for the amount of time defined by USB_REUME_TIMEOUT(40ms) macro.
This 40ms delay for resume signalling is in acceptable limit, but it get worse when xhci goes for polling mode in order to detect other events on its ports and modify rh_timer timer with a variable time out of 1ms to (HZ/4)ms.
drivers/usb/core/hcd.c line 799mod_timer (&hcd->rh_timer, (jiffies/(HZ/4) + 1) * (HZ/4)).
Due to above variable timeout usb auto-resume latency varies from 40ms to ~300ms.
Log Snippet:
~128ms latency
[ 53.112049] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 53.229200] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0004 [ 53.240177] usb 1-2: usb wakeup-resume [ 53.240195] usb 1-2: finish resume [ 53.240357] usb usb1-port2: resume, status 0 ----------------------------------------------------------------- ~300ms latency [ 59.946620] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 59.979341] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 60.229342] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0004 [ 60.251321] usb 1-2: usb wakeup-resume [ 60.251335] usb 1-2: finish resume [ 60.251539] usb usb1-port2: resume, status 0
This variable resume latency can be optimized, as in case of PORT_PLC change event rh_timer has already been modified with USB_RESUME_TIMEOUT (40ms) delay,leaving the rest to GetPortStatus and started polling for root hub status (invoking usb_hcd_poll_rh_status).
We can avoid polling as we have already modified rh_timer with delay of 40ms.
This patch set the HCD_FLAG_POLL_RH to hcd->flags after modification of rh_timer, and avoids polling of root hub status. so rh_timer can fire after 40ms and usb device auto-resuem latency will be around 40ms.
--- drivers/usb/host/xhci-ring.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/usb/host/xhci-ring.c +6 -1
@@ -1677,9 +1677,14 @@ static void handle_port_status(struct xhci_hcd *xhci, bus_state->resume_done[faked_port_index] = jiffies +msecs_to_jiffies (USB_RESUME_TIMEOUT );set_bit (faked_port_index, &bus_state->resuming_ports); +/* Do the rest in GetPortStatus after resume time delay. +* Avoid polling roothub status before that so that a +* usb device auto-resume latency around ~40ms. +*/ +set_bit (HCD_FLAG_POLL_RH , &hcd->flags);mod_timer (&hcd->rh_timer, bus_state->resume_done[faked_port_index]); -/* Do the rest in GetPortStatus */ + bogus_port_status = true; } }
Signature · 1 line
2.7.4
Quoted 43 lines · On 23.08.2017 09:18, [email protected] wrote:
From: Anshuman Gupta <[email protected]>
This patch will improve the variable auto-resume latency of an usb-port.
When xhci gets a port status change event interrupt due to
PORT_PLC(port link state transition), linux Host controller driver drives the resume signalling on the bus for the amount of time defined byUSB_REUME_TIMEOUT(40ms)macro.This 40ms delay for resume signalling is in acceptable limit, but it get worse when xhci goes for polling mode in order to detect other events on its ports and modify
rh_timertimer with a variable time out of 1ms to (HZ/4)ms.
drivers/usb/core/hcd.cline 799mod_timer(&hcd->rh_timer, (jiffies/(HZ/4) + 1) * (HZ/4)).Due to above variable timeout usb auto-resume latency varies from 40ms to ~300ms.
Log Snippet:
~128ms latency[ 53.112049] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 53.229200] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0004 [ 53.240177] usb 1-2: usb wakeup-resume [ 53.240195] usb 1-2: finish resume [ 53.240357] usb usb1-port2: resume, status 0 ----------------------------------------------------------------- ~300ms latency [ 59.946620] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 59.979341] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 60.229342] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0004 [ 60.251321] usb 1-2: usb wakeup-resume [ 60.251335] usb 1-2: finish resume [ 60.251539] usb usb1-port2: resume, status 0This variable resume latency can be optimized, as in case of
PORT_PLCchange eventrh_timerhas already been modified withUSB_RESUME_TIMEOUT(40ms) delay,leaving the rest to GetPortStatus and started polling for root hub status (invokingusb_hcd_poll_rh_status).
We can avoid polling as we have already modifiedrh_timerwith delay of 40ms.
Quoted 6 lines
This patch set the
HCD_FLAG_POLL_RHtohcd->flagsafter modification ofrh_timer, and avoids polling of root hub status. sorh_timercan fire after 40ms and usb device auto-resuem latency will be around 40ms.Signed-off-by Anshuman Gupta <[email protected]>---
Thanks,
adding and sending forward after 4.14-rc1
-Mathias
Quoted 48 lines · Mathias Nyman
Quoted 47 lines · On 23.08.2017 09:18, [email protected] wrote:
From: Anshuman Gupta <[email protected]>
This patch will improve the variable auto-resume latency of an usb-port.
When xhci gets a port status change event interrupt due to
PORT_PLC(port link state transition), linux Host controller driver drives the resume signalling on the bus for the amount of time defined byUSB_REUME_TIMEOUT(40ms)macro.This 40ms delay for resume signalling is in acceptable limit, but it get worse when xhci goes for polling mode in order to detect other events on its ports and modify
rh_timertimer with a variable time out of 1ms to (HZ/4)ms.
drivers/usb/core/hcd.cline 799mod_timer(&hcd->rh_timer, (jiffies/(HZ/4) + 1) * (HZ/4)).Due to above variable timeout usb auto-resume latency varies from 40ms to ~300ms.
Log Snippet:
~128ms latency[ 53.112049] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 53.229200] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0004 [ 53.240177] usb 1-2: usb wakeup-resume [ 53.240195] usb 1-2: finish resume [ 53.240357] usb usb1-port2: resume, status 0 ----------------------------------------------------------------- ~300ms latency [ 59.946620] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 59.979341] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 60.229342] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0004 [ 60.251321] usb 1-2: usb wakeup-resume [ 60.251335] usb 1-2: finish resume [ 60.251539] usb usb1-port2: resume, status 0This variable resume latency can be optimized, as in case of
PORT_PLCchange eventrh_timerhas already been modified withUSB_RESUME_TIMEOUT(40ms) delay,leaving the rest to GetPortStatus and started polling for root hub status (invokingusb_hcd_poll_rh_status).
We can avoid polling as we have already modifiedrh_timerwith delay of 40ms.This patch set the
HCD_FLAG_POLL_RHtohcd->flagsafter modification ofrh_timer, and avoids polling of root hub status. sorh_timercan fire after 40ms and usb device auto-resuem latency will be around 40ms.
Quoted 5 lines
Quoted 2 lines
Signed-off-by Anshuman Gupta <[email protected]>---
Thanks,
adding and sending forward after 4.14-rc1
This patch description doesn't make sense. If you want to avoid resetting the root-hub timer by 250 ms, you should _clear_ the HCD_FLAG_POLL_RH flag, not _set_ it! See this (slightly tricky) code in usb_hcd_poll_rh_status():
if (hcd->uses_new_polling ? HCD_POLL_RH(hcd) : (length == 0 && hcd->status_urb != NULL)) mod_timer (&hcd->rh_timer, (jiffies/(HZ/4) + 1) * (HZ/4));
Since xhci-hcd does set the uses_new_polling flag, the "if" condition just tests the HCD_FLAG_POLL_RH bit.
Alan Stern
Mathias Nyman
maintainerQuoted 61 lines · Alan Stern
Quoted 48 lines · Mathias Nyman
Quoted 47 lines · On 23.08.2017 09:18, [email protected] wrote:
From: Anshuman Gupta <[email protected]>
This patch will improve the variable auto-resume latency of an usb-port.
When xhci gets a port status change event interrupt due to
PORT_PLC(port link state transition), linux Host controller driver drives the resume signalling on the bus for the amount of time defined byUSB_REUME_TIMEOUT(40ms)macro.This 40ms delay for resume signalling is in acceptable limit, but it get worse when xhci goes for polling mode in order to detect other events on its ports and modify
rh_timertimer with a variable time out of 1ms to (HZ/4)ms.
drivers/usb/core/hcd.cline 799mod_timer(&hcd->rh_timer, (jiffies/(HZ/4) + 1) * (HZ/4)).Due to above variable timeout usb auto-resume latency varies from 40ms to ~300ms.
Log Snippet:
~128ms latency[ 53.112049] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 53.229200] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0004 [ 53.240177] usb 1-2: usb wakeup-resume [ 53.240195] usb 1-2: finish resume [ 53.240357] usb usb1-port2: resume, status 0 ----------------------------------------------------------------- ~300ms latency [ 59.946620] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 59.979341] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 60.229342] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0004 [ 60.251321] usb 1-2: usb wakeup-resume [ 60.251335] usb 1-2: finish resume [ 60.251539] usb usb1-port2: resume, status 0This variable resume latency can be optimized, as in case of
PORT_PLCchange eventrh_timerhas already been modified withUSB_RESUME_TIMEOUT(40ms) delay,leaving the rest to GetPortStatus and started polling for root hub status (invokingusb_hcd_poll_rh_status).
We can avoid polling as we have already modifiedrh_timerwith delay of 40ms.This patch set the
HCD_FLAG_POLL_RHtohcd->flagsafter modification ofrh_timer, and avoids polling of root hub status. sorh_timercan fire after 40ms and usb device auto-resuem latency will be around 40ms.Quoted 5 lines
Quoted 2 lines
Signed-off-by Anshuman Gupta <[email protected]>---
Thanks,
adding and sending forward after 4.14-rc1This patch description doesn't make sense. If you want to avoid resetting the root-hub timer by 250 ms, you should _clear_ the
HCD_FLAG_POLL_RHflag, not _set_ it! See this (slightly tricky) code inusb_hcd_poll_rh_status():
Quoted 6 lines
if (hcd->uses_new_polling ? HCD_POLL_RH(hcd) : (length == 0 && hcd->status_urb != NULL)) mod_timer (&hcd->rh_timer, (jiffies/(HZ/4) + 1) * (HZ/4));Since xhci-hcd does set the
uses_new_pollingflag, the "if" condition just tests theHCD_FLAG_POLL_RHbit.
The description could be a bit clearer, I agree, but the issue it fixes seems correct.
We want start the rh polling when a HS port is found in resume state,
and we want to time it so that the it runs just after resume is complete (40ms) This is why we have the mod_timer(&hcd->rh_timer, bus_state->resume_done[faked_port_index]);
This would all be fine, but the xhci event handler has a bug and it calls usb_hcd_poll_rh_status(hcd) manually right after modifying rh_timer.
This causes hub thread to read port status before resume is ready, and we need to wait for another 250ms before the rh thread can handle the port.
What this patch in the end does is just removing the manual usb_hcd_poll_rh_status(hcd) in this case.
Thanks
-Mathias
PATCH v2 1/1 USB: xhci: reducing HS port auto-resume latency.
From: Anshuman Gupta <[email protected]>
This patch will improve the variable auto-resume latency of an usb-port.
When xhci gets a HS port status change event interrupt due to PORT_PLC (port link state transition), linux Host controller driver drives the resume signalling on the bus for the amount of time defined by USB_RESUME_TIMEOUT(40ms) macro.
This 40ms delay for resume signalling is in acceptable limit, but it get worse when xhci goes for polling mode in order to detect other events on its ports and modify rh_timer timer with a variable time out of 1ms to (HZ/4)ms.
drivers/usb/core/hcd.c line 799mod_timer (&hcd->rh_timer, (jiffies/(HZ/4) + 1) * (HZ/4)).
Due to above variable timeout usb auto-resume latency varies from 40ms to ~300ms(when HZ=1000).
Log Snippet:
~128ms latency
[ 53.112049] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 53.229200] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0004 [ 53.240177] usb 1-2: usb wakeup-resume [ 53.240195] usb 1-2: finish resume [ 53.240357] usb usb1-port2: resume, status 0 ----------------------------------------------------------------- ~300ms latency [ 59.946620] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 59.979341] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0000 [ 60.229342] hub 1-0:1.0: state 7 ports 12 chg 0000 evt 0004 [ 60.251321] usb 1-2: usb wakeup-resume [ 60.251335] usb 1-2: finish resume [ 60.251539] usb usb1-port2: resume, status 0
xhci bug behind the higher latency:
When HS port is resuming, xhci passes the resume event to hub event thread after USB_RESUME_TIMEOUT delay. it modify rh_timer with USB_RESUME_TIMEOUT delay, and set the polling mode flag and invoke usb_hcd_poll_rh_status() function manually.
This eventually makes hub event thread to read the HS port status before it resumes, since polling mode flag is set, host controller driver modify rh_timer again with (HZ/4)ms delay, This makes hub event thread to wait for (HZ/4)ms. which causes higher auto-resume latency.
This variable resume latency can be optimized, as in case of HS port resume event rh_timer has already been modified with USB_RESUME_TIMEOUT (40ms) delay, leaving the rest to GetPortStatus.
We can avoid calling usb_hcd_poll_rh_status() function manually in case of HS port resume event.
This patch set the HCD_FLAG_POLL_RH to hcd->flags after modification of rh_timer, and avoid calling usb_hcd_poll_rh_status(), This makes to call usb_hcd_poll_rh_status() by rh_timer and passing the HS port resume event to hub event thread after USB_RESUME_TIMEOUT delay.
---
Changes in v2
- Make the commit message more clearer.
- Changing the comment.
drivers/usb/host/xhci-ring.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/usb/host/xhci-ring.c +6 -1
@@ -1677,9 +1677,14 @@ static void handle_port_status(struct xhci_hcd *xhci, bus_state->resume_done[faked_port_index] = jiffies +msecs_to_jiffies (USB_RESUME_TIMEOUT );set_bit (faked_port_index, &bus_state->resuming_ports); +/* Do the rest in GetPortStatus after resume time delay. +* Avoid calling usb_hcd_poll_rh_status before that,so +* an usb port can auto-resume after resume time delay. +*/ +set_bit (HCD_FLAG_POLL_RH , &hcd->flags);mod_timer (&hcd->rh_timer, bus_state->resume_done[faked_port_index]); -/* Do the rest in GetPortStatus */ + bogus_port_status = true; } }
Signature · 1 line
2.7.4
A message that is not in the archive (sent off-list or not yet received): [email protected]@intel.com
Mathias Nyman
maintainerRe: [PATCH v2] USB: xhci: reducing HS port auto-resume latency.
Hi
Quoted 4 lines · anshuman gupta
I have addressed the feedback on the patch description with reference to thread https://patchwork.kernel.org/patch/9916641/ with [PATCH V2] please provide your review comments.
I'm taking the patch and sending it forward to Greg after the merge window closes (when 4.14-rc1 is out)
Thanks
-Mathias