From: Daniel Wagner <hidden> Date: 2016-08-18 13:22:38
From: Daniel Wagner <redacted>
Hi,
Using complete_all() is not wrong per se but it suggest that there
might be more than one reader. For -rt I am reviewing all
complete_all() users and would like to leave only the real ones in the
tree. The main problem for -rt about complete_all() is that it can be
uses inside IRQ context and that can lead to unbounded amount work
inside the interrupt handler. That is a no no for -rt.
The patches grouped per subsystem and in small batches to allow
reviewing.
This series ignores all complete_all() usages in the firmware loading
path. They will be hopefully address by Luis' sysdata patches [0].
That leaves a couple of complete_all() calls.
The first patch fixes a real glitch for the carl9170 driver. I was
able to test it because I have the hardware. For the second one I
haven't found any dongle with that chip in my drawers.
This series against net-next of today.
cheers,
daniel
[0] https://lkml.kernel.org/r/1466117661-22075-1-git-send-email-mcgrof@kernel.org
Daniel Wagner (2):
carl9170: Fix wrong completion usage
ath10k: use complete() instead complete_all()
drivers/net/wireless/ath/ath10k/core.c | 16 ++++++++--------
drivers/net/wireless/ath/ath10k/mac.c | 2 +-
drivers/net/wireless/ath/carl9170/usb.c | 6 ++----
3 files changed, 11 insertions(+), 13 deletions(-)
--
2.7.4
From: Daniel Wagner <hidden> Date: 2016-08-18 13:22:39
From: Daniel Wagner <redacted>
carl9170_usb_stop() is used from several places to flush and cleanup any
pending work. The normal pattern is to send a request and wait for the
irq handler to call complete(). The completion is not reinitialized
during normal operation and as the old comment indicates it is important
to keep calls to wait_for_completion_timeout() and complete() balanced.
Calling complete_all() brings this equilibirum out of balance and needs
to be fixed by a reinit_completion(). But that opens a small race
window. It is possible that the sequence of complete_all(),
reinit_completion() is faster than the wait_for_completion_timeout() can
do its work. The wake up is not lost but the done counter test is after
reinit_completion() has been executed. The only reason we don't see
carl9170_exec_cmd() hang forever is we use the timeout version of
wait_for_copletion().
Let's fix this by reinitializing the completion (that is just setting
done counter to 0) just before we send out an request. Now,
carl9170_usb_stop() can be sure a complete() call is enough to make
progess since there is only one waiter at max. This is a common pattern
also seen in various drivers which use completion.
Signed-off-by: Daniel Wagner <redacted>
---
drivers/net/wireless/ath/carl9170/usb.c | 6 ++----
1 file changed, 2 insertions(+), 4 deletions(-)
@@ -778,10 +779,7 @@ void carl9170_usb_stop(struct ar9170 *ar)spin_lock_bh(&ar->cmd_lock);ar->readlen=0;spin_unlock_bh(&ar->cmd_lock);-complete_all(&ar->cmd_wait);--/* This is required to prevent an early completion on _start */-reinit_completion(&ar->cmd_wait);+complete(&ar->cmd_wait);/**Note:
From: Daniel Wagner <hidden> Date: 2016-08-18 13:23:16
From: Daniel Wagner <redacted>
There is only one waiter for the completion, therefore there
is no need to use complete_all(). Let's make that clear by
using complete() instead of complete_all().
The usage pattern of the completion is:
waiter context waker context
scan.started
------------
ath10k_start_scan()
lockdep_assert_held(conf_mutex)
auth10k_wmi_start_scan()
wait_for_completion_timeout(scan.started)
ath10k_wmi_event_scan_start_failed()
complete(scan.started)
ath10k_wmi_event_scan_started()
complete(scan.started)
scan.completed
--------------
ath10k_scan_stop()
lockdep_assert_held(conf_mutex)
ath10k_wmi_stop_scan()
wait_for_completion_timeout(scan.completed)
__ath10k_scan_finish()
complete(scan.completed)
scan.on_channel
---------------
ath10k_remain_on_channel()
mutex_lock(conf_mutex)
ath10k_start_scan()
wait_for_completion_timeout(scan.on_channel)
ath10k_wmi_event_scan_foreign_chan()
complete(scan.on_channel)
offchan_tx_completed
--------------------
ath10k_offchan_tx_work()
mutex_lock(conf_mutex)
reinit_completion(offchan_tx_completed)
wait_for_completion_timeout(offchan_tx_completed)
ath10k_report_offchain_tx()
complete(offchan_tx_completed)
install_key_done
----------------
ath10k_install_key()
lockep_assert_held(conf_mutex)
reinit_completion(install_key_done)
wait_for_completion_timeout(install_key_done)
ath10k_htt_t2h_msg_handler()
complete(install_key_done)
vdev_setup_done
---------------
ath10k_monitor_vdev_start()
lockdep_assert_held(conf_mutex)
reinit_completion(vdev_setup_done)
ath10k_vdev_setup_sync()
wait_for_completion_timeout(vdev_setup_done)
ath10k_wmi_event_vdev_start_resp()
complete(vdev_setup_done)
ath10k_monitor_vdev_stop()
lockdep_assert_held(conf_mutex)
reinit_completion(vdev_setup_done()
ath10k_vdev_setup_sync()
wait_for_completion_timeout(vdev_setup_done)
ath10k_wmi_event_vdev_stopped()
complete(vdev_setup_done)
thermal.wmi_sync
----------------
ath10k_thermal_show_temp()
mutex_lock(conf_mutex)
reinit_completion(thermal.wmi_sync)
wait_for_completion_timeout(thermal.wmi_sync)
ath10k_thermal_event_temperature()
complete(thermal.wmi_sync)
bss_survey_done
---------------
ath10k_mac_update_bss_chan_survey
lockdep_assert_held(conf_mutex)
reinit_completion(bss_survey_done)
wait_for_completion_timeout(bss_survey_done)
ath10k_wmi_event_pdev_bss_chan_info()
complete(bss_survey_done)
All complete() calls happen while the conf_mutex is taken. That means
at max one waiter is possible.
Signed-off-by: Daniel Wagner <redacted>
---
drivers/net/wireless/ath/ath10k/core.c | 16 ++++++++--------
drivers/net/wireless/ath/ath10k/mac.c | 2 +-
2 files changed, 9 insertions(+), 9 deletions(-)
From: "Luis R. Rodriguez" <mcgrof@kernel.org> Date: 2016-08-19 02:02:04
On Thu, Aug 18, 2016 at 03:12:04PM +0200, Daniel Wagner wrote:
This series ignores all complete_all() usages in the firmware loading
path. They will be hopefully address by Luis' sysdata patches [0].
That leaves a couple of complete_all() calls.
I had not considered this as a gain, but glad to know the sysdata series
could help with RT as well, thanks for the clarification.
Luis
From: Kalle Valo <hidden> Date: 2016-09-02 15:55:36
Daniel Wagner [off-list ref] wrote:
From: Daniel Wagner <redacted>
There is only one waiter for the completion, therefore there
is no need to use complete_all(). Let's make that clear by
using complete() instead of complete_all().
The usage pattern of the completion is:
waiter context waker context
scan.started
------------
ath10k_start_scan()
lockdep_assert_held(conf_mutex)
auth10k_wmi_start_scan()
wait_for_completion_timeout(scan.started)
ath10k_wmi_event_scan_start_failed()
complete(scan.started)
ath10k_wmi_event_scan_started()
complete(scan.started)
scan.completed
--------------
ath10k_scan_stop()
lockdep_assert_held(conf_mutex)
ath10k_wmi_stop_scan()
wait_for_completion_timeout(scan.completed)
__ath10k_scan_finish()
complete(scan.completed)
scan.on_channel
---------------
ath10k_remain_on_channel()
mutex_lock(conf_mutex)
ath10k_start_scan()
wait_for_completion_timeout(scan.on_channel)
ath10k_wmi_event_scan_foreign_chan()
complete(scan.on_channel)
offchan_tx_completed
--------------------
ath10k_offchan_tx_work()
mutex_lock(conf_mutex)
reinit_completion(offchan_tx_completed)
wait_for_completion_timeout(offchan_tx_completed)
ath10k_report_offchain_tx()
complete(offchan_tx_completed)
install_key_done
----------------
ath10k_install_key()
lockep_assert_held(conf_mutex)
reinit_completion(install_key_done)
wait_for_completion_timeout(install_key_done)
ath10k_htt_t2h_msg_handler()
complete(install_key_done)
vdev_setup_done
---------------
ath10k_monitor_vdev_start()
lockdep_assert_held(conf_mutex)
reinit_completion(vdev_setup_done)
ath10k_vdev_setup_sync()
wait_for_completion_timeout(vdev_setup_done)
ath10k_wmi_event_vdev_start_resp()
complete(vdev_setup_done)
ath10k_monitor_vdev_stop()
lockdep_assert_held(conf_mutex)
reinit_completion(vdev_setup_done()
ath10k_vdev_setup_sync()
wait_for_completion_timeout(vdev_setup_done)
ath10k_wmi_event_vdev_stopped()
complete(vdev_setup_done)
thermal.wmi_sync
----------------
ath10k_thermal_show_temp()
mutex_lock(conf_mutex)
reinit_completion(thermal.wmi_sync)
wait_for_completion_timeout(thermal.wmi_sync)
ath10k_thermal_event_temperature()
complete(thermal.wmi_sync)
bss_survey_done
---------------
ath10k_mac_update_bss_chan_survey
lockdep_assert_held(conf_mutex)
reinit_completion(bss_survey_done)
wait_for_completion_timeout(bss_survey_done)
ath10k_wmi_event_pdev_bss_chan_info()
complete(bss_survey_done)
All complete() calls happen while the conf_mutex is taken. That means
at max one waiter is possible.
Signed-off-by: Daniel Wagner <redacted>
Thanks, 1 patch applied to ath-next branch of ath.git:
881ed54ecc13 ath10k: use complete() instead complete_all()
--
Sent by pwcli
https://patchwork.kernel.org/patch/9287731/
From: Kalle Valo <hidden> Date: 2016-09-09 12:16:28
Daniel Wagner [off-list ref] wrote:
From: Daniel Wagner <redacted>
carl9170_usb_stop() is used from several places to flush and cleanup any
pending work. The normal pattern is to send a request and wait for the
irq handler to call complete(). The completion is not reinitialized
during normal operation and as the old comment indicates it is important
to keep calls to wait_for_completion_timeout() and complete() balanced.
Calling complete_all() brings this equilibirum out of balance and needs
to be fixed by a reinit_completion(). But that opens a small race
window. It is possible that the sequence of complete_all(),
reinit_completion() is faster than the wait_for_completion_timeout() can
do its work. The wake up is not lost but the done counter test is after
reinit_completion() has been executed. The only reason we don't see
carl9170_exec_cmd() hang forever is we use the timeout version of
wait_for_copletion().
Let's fix this by reinitializing the completion (that is just setting
done counter to 0) just before we send out an request. Now,
carl9170_usb_stop() can be sure a complete() call is enough to make
progess since there is only one waiter at max. This is a common pattern
also seen in various drivers which use completion.
Signed-off-by: Daniel Wagner <redacted>