linux/drivers/base/power/runtime.c

1813 lines
50 KiB

PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
locking/atomics: COCCINELLE/treewide: Convert trivial ACCESS_ONCE() patterns to READ_ONCE()/WRITE_ONCE() Please do not apply this to mainline directly, instead please re-run the coccinelle script shown below and apply its output. For several reasons, it is desirable to use {READ,WRITE}_ONCE() in preference to ACCESS_ONCE(), and new code is expected to use one of the former. So far, there's been no reason to change most existing uses of ACCESS_ONCE(), as these aren't harmful, and changing them results in churn. However, for some features, the read/write distinction is critical to correct operation. To distinguish these cases, separate read/write accessors must be used. This patch migrates (most) remaining ACCESS_ONCE() instances to {READ,WRITE}_ONCE(), using the following coccinelle script: ---- // Convert trivial ACCESS_ONCE() uses to equivalent READ_ONCE() and // WRITE_ONCE() // $ make coccicheck COCCI=/home/mark/once.cocci SPFLAGS="--include-headers" MODE=patch virtual patch @ depends on patch @ expression E1, E2; @@ - ACCESS_ONCE(E1) = E2 + WRITE_ONCE(E1, E2) @ depends on patch @ expression E; @@ - ACCESS_ONCE(E) + READ_ONCE(E) ---- Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: davem@davemloft.net Cc: linux-arch@vger.kernel.org Cc: mpe@ellerman.id.au Cc: shuah@kernel.org Cc: snitzer@redhat.com Cc: thor.thayer@linux.intel.com Cc: tj@kernel.org Cc: viro@zeniv.linux.org.uk Cc: will.deacon@arm.com Link: http://lkml.kernel.org/r/1508792849-3115-19-git-send-email-paulmck@linux.vnet.ibm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-23 14:07:29 -07:00
driver core: Fix handling of runtime PM flags in device_link_add() After commit ead18c23c263 ("driver core: Introduce device links reference counting"), if there is a link between the given supplier and the given consumer already, device_link_add() will refcount it and return it unconditionally without updating its flags. It is possible, however, that the second (or any subsequent) caller of device_link_add() for the same consumer-supplier pair will pass DL_FLAG_PM_RUNTIME, possibly along with DL_FLAG_RPM_ACTIVE, in flags to it and the existing link may not behave as expected then. First, if DL_FLAG_PM_RUNTIME is not set in the existing link's flags at all, it needs to be set like during the original initialization of the link. Second, if DL_FLAG_RPM_ACTIVE is passed to device_link_add() in flags (in addition to DL_FLAG_PM_RUNTIME), the existing link should to be updated to reflect the "active" runtime PM configuration of the consumer-supplier pair and extra care must be taken here to avoid possible destructive races with runtime PM of the consumer. To that end, redefine the rpm_active field in struct device_link as a refcount, initialize it to 1 and make rpm_resume() (for the consumer) and device_link_add() increment it whenever they acquire a runtime PM reference on the supplier device. Accordingly, make rpm_suspend() (for the consumer) and pm_runtime_clean_up_links() decrement it and drop runtime PM references to the supplier device in a loop until rpm_active becones 1 again. Fixes: ead18c23c263 ("driver core: Introduce device links reference counting") Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-01 01:49:14 +01:00
driver core: Fix handling of runtime PM flags in device_link_add() After commit ead18c23c263 ("driver core: Introduce device links reference counting"), if there is a link between the given supplier and the given consumer already, device_link_add() will refcount it and return it unconditionally without updating its flags. It is possible, however, that the second (or any subsequent) caller of device_link_add() for the same consumer-supplier pair will pass DL_FLAG_PM_RUNTIME, possibly along with DL_FLAG_RPM_ACTIVE, in flags to it and the existing link may not behave as expected then. First, if DL_FLAG_PM_RUNTIME is not set in the existing link's flags at all, it needs to be set like during the original initialization of the link. Second, if DL_FLAG_RPM_ACTIVE is passed to device_link_add() in flags (in addition to DL_FLAG_PM_RUNTIME), the existing link should to be updated to reflect the "active" runtime PM configuration of the consumer-supplier pair and extra care must be taken here to avoid possible destructive races with runtime PM of the consumer. To that end, redefine the rpm_active field in struct device_link as a refcount, initialize it to 1 and make rpm_resume() (for the consumer) and device_link_add() increment it whenever they acquire a runtime PM reference on the supplier device. Accordingly, make rpm_suspend() (for the consumer) and pm_runtime_clean_up_links() decrement it and drop runtime PM references to the supplier device in a loop until rpm_active becones 1 again. Fixes: ead18c23c263 ("driver core: Introduce device links reference counting") Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-01 01:49:14 +01:00
driver core: Fix handling of runtime PM flags in device_link_add() After commit ead18c23c263 ("driver core: Introduce device links reference counting"), if there is a link between the given supplier and the given consumer already, device_link_add() will refcount it and return it unconditionally without updating its flags. It is possible, however, that the second (or any subsequent) caller of device_link_add() for the same consumer-supplier pair will pass DL_FLAG_PM_RUNTIME, possibly along with DL_FLAG_RPM_ACTIVE, in flags to it and the existing link may not behave as expected then. First, if DL_FLAG_PM_RUNTIME is not set in the existing link's flags at all, it needs to be set like during the original initialization of the link. Second, if DL_FLAG_RPM_ACTIVE is passed to device_link_add() in flags (in addition to DL_FLAG_PM_RUNTIME), the existing link should to be updated to reflect the "active" runtime PM configuration of the consumer-supplier pair and extra care must be taken here to avoid possible destructive races with runtime PM of the consumer. To that end, redefine the rpm_active field in struct device_link as a refcount, initialize it to 1 and make rpm_resume() (for the consumer) and device_link_add() increment it whenever they acquire a runtime PM reference on the supplier device. Accordingly, make rpm_suspend() (for the consumer) and pm_runtime_clean_up_links() decrement it and drop runtime PM references to the supplier device in a loop until rpm_active becones 1 again. Fixes: ead18c23c263 ("driver core: Introduce device links reference counting") Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-01 01:49:14 +01:00
driver core: Fix handling of runtime PM flags in device_link_add() After commit ead18c23c263 ("driver core: Introduce device links reference counting"), if there is a link between the given supplier and the given consumer already, device_link_add() will refcount it and return it unconditionally without updating its flags. It is possible, however, that the second (or any subsequent) caller of device_link_add() for the same consumer-supplier pair will pass DL_FLAG_PM_RUNTIME, possibly along with DL_FLAG_RPM_ACTIVE, in flags to it and the existing link may not behave as expected then. First, if DL_FLAG_PM_RUNTIME is not set in the existing link's flags at all, it needs to be set like during the original initialization of the link. Second, if DL_FLAG_RPM_ACTIVE is passed to device_link_add() in flags (in addition to DL_FLAG_PM_RUNTIME), the existing link should to be updated to reflect the "active" runtime PM configuration of the consumer-supplier pair and extra care must be taken here to avoid possible destructive races with runtime PM of the consumer. To that end, redefine the rpm_active field in struct device_link as a refcount, initialize it to 1 and make rpm_resume() (for the consumer) and device_link_add() increment it whenever they acquire a runtime PM reference on the supplier device. Accordingly, make rpm_suspend() (for the consumer) and pm_runtime_clean_up_links() decrement it and drop runtime PM references to the supplier device in a loop until rpm_active becones 1 again. Fixes: ead18c23c263 ("driver core: Introduce device links reference counting") Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-01 01:49:14 +01:00
PM / runtime: Add _rcuidle suffix to allow rpm_idle() use from idle This commit appends a few _rcuidle suffixes to fix the following RCU-used-from-idle bug: > =============================== > [ INFO: suspicious RCU usage. ] > 4.6.0-rc5-next-20160426+ #1116 Not tainted > ------------------------------- > include/trace/events/rpm.h:95 suspicious rcu_dereference_check() usage! > > other info that might help us debug this: > > > RCU used illegally from idle CPU! > rcu_scheduler_active = 1, debug_locks = 0 > RCU used illegally from extended quiescent state! > 1 lock held by swapper/0/0: > #0: (&(&dev->power.lock)->rlock){-.-...}, at: [<c052cc2c>] __rpm_callback+0x58/0x60 > > stack backtrace: > CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.6.0-rc5-next-20160426+ #1116 > Hardware name: Generic OMAP36xx (Flattened Device Tree) > [<c0110290>] (unwind_backtrace) from [<c010c3a8>] (show_stack+0x10/0x14) > [<c010c3a8>] (show_stack) from [<c047fd68>] (dump_stack+0xb0/0xe4) > [<c047fd68>] (dump_stack) from [<c052d5d0>] (rpm_suspend+0x580/0x768) > [<c052d5d0>] (rpm_suspend) from [<c052ec58>] (__pm_runtime_suspend+0x64/0x84) > [<c052ec58>] (__pm_runtime_suspend) from [<c04bf25c>] (omap2_gpio_prepare_for_idle+0x5c/0x70) > [<c04bf25c>] (omap2_gpio_prepare_for_idle) from [<c0125568>] (omap_sram_idle+0x140/0x244) > [<c0125568>] (omap_sram_idle) from [<c01269dc>] (omap3_enter_idle_bm+0xfc/0x1ec) > [<c01269dc>] (omap3_enter_idle_bm) from [<c0601bdc>] (cpuidle_enter_state+0x80/0x3d4) > [<c0601bdc>] (cpuidle_enter_state) from [<c0183b08>] (cpu_startup_entry+0x198/0x3a0) > [<c0183b08>] (cpu_startup_entry) from [<c0b00c0c>] (start_kernel+0x354/0x3c8) > [<c0b00c0c>] (start_kernel) from [<8000807c>] (0x8000807c) In the immortal words of Steven Rostedt, "*Whack* *Whack* *Whack*!!!" Reported-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Tony Lindgren <tony@atomide.com> Tested-by: Guenter Roeck <linux@roeck-us.net> WhACKED-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-04-26 13:03:51 -07:00
PM / runtime: Add _rcuidle suffix to allow rpm_idle() use from idle This commit appends a few _rcuidle suffixes to fix the following RCU-used-from-idle bug: > =============================== > [ INFO: suspicious RCU usage. ] > 4.6.0-rc5-next-20160426+ #1116 Not tainted > ------------------------------- > include/trace/events/rpm.h:95 suspicious rcu_dereference_check() usage! > > other info that might help us debug this: > > > RCU used illegally from idle CPU! > rcu_scheduler_active = 1, debug_locks = 0 > RCU used illegally from extended quiescent state! > 1 lock held by swapper/0/0: > #0: (&(&dev->power.lock)->rlock){-.-...}, at: [<c052cc2c>] __rpm_callback+0x58/0x60 > > stack backtrace: > CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.6.0-rc5-next-20160426+ #1116 > Hardware name: Generic OMAP36xx (Flattened Device Tree) > [<c0110290>] (unwind_backtrace) from [<c010c3a8>] (show_stack+0x10/0x14) > [<c010c3a8>] (show_stack) from [<c047fd68>] (dump_stack+0xb0/0xe4) > [<c047fd68>] (dump_stack) from [<c052d5d0>] (rpm_suspend+0x580/0x768) > [<c052d5d0>] (rpm_suspend) from [<c052ec58>] (__pm_runtime_suspend+0x64/0x84) > [<c052ec58>] (__pm_runtime_suspend) from [<c04bf25c>] (omap2_gpio_prepare_for_idle+0x5c/0x70) > [<c04bf25c>] (omap2_gpio_prepare_for_idle) from [<c0125568>] (omap_sram_idle+0x140/0x244) > [<c0125568>] (omap_sram_idle) from [<c01269dc>] (omap3_enter_idle_bm+0xfc/0x1ec) > [<c01269dc>] (omap3_enter_idle_bm) from [<c0601bdc>] (cpuidle_enter_state+0x80/0x3d4) > [<c0601bdc>] (cpuidle_enter_state) from [<c0183b08>] (cpu_startup_entry+0x198/0x3a0) > [<c0183b08>] (cpu_startup_entry) from [<c0b00c0c>] (start_kernel+0x354/0x3c8) > [<c0b00c0c>] (start_kernel) from [<8000807c>] (0x8000807c) In the immortal words of Steven Rostedt, "*Whack* *Whack* *Whack*!!!" Reported-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Tony Lindgren <tony@atomide.com> Tested-by: Guenter Roeck <linux@roeck-us.net> WhACKED-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-04-26 13:03:51 -07:00
PM / runtime: Add _rcuidle suffix to allow rpm_idle() use from idle This commit appends a few _rcuidle suffixes to fix the following RCU-used-from-idle bug: > =============================== > [ INFO: suspicious RCU usage. ] > 4.6.0-rc5-next-20160426+ #1116 Not tainted > ------------------------------- > include/trace/events/rpm.h:95 suspicious rcu_dereference_check() usage! > > other info that might help us debug this: > > > RCU used illegally from idle CPU! > rcu_scheduler_active = 1, debug_locks = 0 > RCU used illegally from extended quiescent state! > 1 lock held by swapper/0/0: > #0: (&(&dev->power.lock)->rlock){-.-...}, at: [<c052cc2c>] __rpm_callback+0x58/0x60 > > stack backtrace: > CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.6.0-rc5-next-20160426+ #1116 > Hardware name: Generic OMAP36xx (Flattened Device Tree) > [<c0110290>] (unwind_backtrace) from [<c010c3a8>] (show_stack+0x10/0x14) > [<c010c3a8>] (show_stack) from [<c047fd68>] (dump_stack+0xb0/0xe4) > [<c047fd68>] (dump_stack) from [<c052d5d0>] (rpm_suspend+0x580/0x768) > [<c052d5d0>] (rpm_suspend) from [<c052ec58>] (__pm_runtime_suspend+0x64/0x84) > [<c052ec58>] (__pm_runtime_suspend) from [<c04bf25c>] (omap2_gpio_prepare_for_idle+0x5c/0x70) > [<c04bf25c>] (omap2_gpio_prepare_for_idle) from [<c0125568>] (omap_sram_idle+0x140/0x244) > [<c0125568>] (omap_sram_idle) from [<c01269dc>] (omap3_enter_idle_bm+0xfc/0x1ec) > [<c01269dc>] (omap3_enter_idle_bm) from [<c0601bdc>] (cpuidle_enter_state+0x80/0x3d4) > [<c0601bdc>] (cpuidle_enter_state) from [<c0183b08>] (cpu_startup_entry+0x198/0x3a0) > [<c0183b08>] (cpu_startup_entry) from [<c0b00c0c>] (start_kernel+0x354/0x3c8) > [<c0b00c0c>] (start_kernel) from [<8000807c>] (0x8000807c) In the immortal words of Steven Rostedt, "*Whack* *Whack* *Whack*!!!" Reported-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Tony Lindgren <tony@atomide.com> Tested-by: Guenter Roeck <linux@roeck-us.net> WhACKED-by: Steven Rostedt <rostedt@goodmis.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-04-26 13:03:51 -07:00
PM / runtime: Use _rcuidle for runtime suspend tracepoints Further testing with false negatives suppressed by commit 293e2421fe25 ("rcu: Remove superfluous versions of rcu_read_lock_sched_held()") identified a few more unprotected uses of RCU from the idle loop. Because RCU actively ignores idle-loop code (for energy-efficiency reasons, among other things), using RCU from the idle loop can result in too-short grace periods, in turn resulting in arbitrary misbehavior. The affected function is rpm_suspend(). The resulting lockdep-RCU splat is as follows: ------------------------------------------------------------------------ Warning from omap3 =============================== [ INFO: suspicious RCU usage. ] 4.6.0-rc5-next-20160426+ #1112 Not tainted ------------------------------- include/trace/events/rpm.h:63 suspicious rcu_dereference_check() usage! other info that might help us debug this: RCU used illegally from idle CPU! rcu_scheduler_active = 1, debug_locks = 0 RCU used illegally from extended quiescent state! 1 lock held by swapper/0/0: #0: (&(&dev->power.lock)->rlock){-.-...}, at: [<c052ee24>] __pm_runtime_suspend+0x54/0x84 stack backtrace: CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.6.0-rc5-next-20160426+ #1112 Hardware name: Generic OMAP36xx (Flattened Device Tree) [<c0110308>] (unwind_backtrace) from [<c010c3a8>] (show_stack+0x10/0x14) [<c010c3a8>] (show_stack) from [<c047fec8>] (dump_stack+0xb0/0xe4) [<c047fec8>] (dump_stack) from [<c052d7b4>] (rpm_suspend+0x604/0x7e4) [<c052d7b4>] (rpm_suspend) from [<c052ee34>] (__pm_runtime_suspend+0x64/0x84) [<c052ee34>] (__pm_runtime_suspend) from [<c04bf3bc>] (omap2_gpio_prepare_for_idle+0x5c/0x70) [<c04bf3bc>] (omap2_gpio_prepare_for_idle) from [<c01255e8>] (omap_sram_idle+0x140/0x244) [<c01255e8>] (omap_sram_idle) from [<c0126b48>] (omap3_enter_idle_bm+0xfc/0x1ec) [<c0126b48>] (omap3_enter_idle_bm) from [<c0601db8>] (cpuidle_enter_state+0x80/0x3d4) [<c0601db8>] (cpuidle_enter_state) from [<c0183c74>] (cpu_startup_entry+0x198/0x3a0) [<c0183c74>] (cpu_startup_entry) from [<c0b00c0c>] (start_kernel+0x354/0x3c8) [<c0b00c0c>] (start_kernel) from [<8000807c>] (0x8000807c) ------------------------------------------------------------------------ Reported-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Tony Lindgren <tony@atomide.com> Tested-by: Guenter Roeck <linux@roeck-us.net> [ rjw: Subject ] Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-04-26 10:42:25 -07:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM / wakeirq: Fix dedicated wakeirq for drivers not using autosuspend I noticed some wakeirq flakeyness with consumer drivers not using autosuspend. For drivers not using autosuspend, the wakeirq may never get unmasked in rpm_suspend() because of irq desc->depth. We are configuring dedicated wakeirqs to start with IRQ_NOAUTOEN as we naturally don't want them running until rpm_suspend() is called. However, when a consumer driver initially calls pm_runtime_get(), we now wrongly start with disable_irq_nosync() call on the dedicated wakeirq that is disabled to start with. This causes desc->depth to toggle between 1 and 2 instead of the usual 0 and 1. This can prevent enable_irq() from unmasking the wakeirq as that only happens at desc->depth 1. This does not necessarily show up with drivers using autosuspend as there is time for disable_irq_nosync() before rpm_suspend() gets called after the autosuspend timeout. Let's fix the issue by adding wirq->status that lazily gets set on the first rpm_suspend(). We also need PM runtime core private functions for dev_pm_enable_wake_irq_check() and dev_pm_disable_wake_irq_check() so we can enable the dedicated wakeirq on the first rpm_suspend(). While at it, let's also fix the comments for dev_pm_enable_wake_irq() and dev_pm_disable_wake_irq(). Those can still be used by the consumer drivers as needed because the IRQ core manages the interrupt usecount for us. Fixes: 4990d4fe327b (PM / Wakeirq: Add automated device wake IRQ handling) Signed-off-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-12-05 16:38:16 -08:00
PM / Runtime: Use device PM QoS constraints (v2) Make the runtime PM core use device PM QoS constraints to check if it is allowed to suspend a given device, so that an error code is returned if the device's own PM QoS constraint is negative or one of its children has already been suspended for too long. If this is not the case, the maximum estimated time the device is allowed to be suspended, computed as the minimum of the device's PM QoS constraint and the PM QoS constraints of its children (reduced by the difference between the current time and their suspend times) is stored in a new device's PM field power.max_time_suspended_ns that can be used by the device's subsystem or PM domain to decide whether or not to put the device into lower-power (and presumably higher-latency) states later (if the constraint is 0, which means "no constraint", the power.max_time_suspended_ns is set to -1). Additionally, the time of execution of the subsystem-level .runtime_suspend() callback for the device is recorded in the new power.suspend_time field for later use by the device's subsystem or PM domain along with power.max_time_suspended_ns (it also is used by the core code when the device's parent is suspended). Introduce a new helper function, pm_runtime_update_max_time_suspended(), allowing subsystems and PM domains (or device drivers) to update the power.max_time_suspended_ns field, for example after changing the power state of a suspended device. Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
2011-12-01 00:01:31 +01:00
PM / runtime: Use _rcuidle for runtime suspend tracepoints Further testing with false negatives suppressed by commit 293e2421fe25 ("rcu: Remove superfluous versions of rcu_read_lock_sched_held()") identified a few more unprotected uses of RCU from the idle loop. Because RCU actively ignores idle-loop code (for energy-efficiency reasons, among other things), using RCU from the idle loop can result in too-short grace periods, in turn resulting in arbitrary misbehavior. The affected function is rpm_suspend(). The resulting lockdep-RCU splat is as follows: ------------------------------------------------------------------------ Warning from omap3 =============================== [ INFO: suspicious RCU usage. ] 4.6.0-rc5-next-20160426+ #1112 Not tainted ------------------------------- include/trace/events/rpm.h:63 suspicious rcu_dereference_check() usage! other info that might help us debug this: RCU used illegally from idle CPU! rcu_scheduler_active = 1, debug_locks = 0 RCU used illegally from extended quiescent state! 1 lock held by swapper/0/0: #0: (&(&dev->power.lock)->rlock){-.-...}, at: [<c052ee24>] __pm_runtime_suspend+0x54/0x84 stack backtrace: CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.6.0-rc5-next-20160426+ #1112 Hardware name: Generic OMAP36xx (Flattened Device Tree) [<c0110308>] (unwind_backtrace) from [<c010c3a8>] (show_stack+0x10/0x14) [<c010c3a8>] (show_stack) from [<c047fec8>] (dump_stack+0xb0/0xe4) [<c047fec8>] (dump_stack) from [<c052d7b4>] (rpm_suspend+0x604/0x7e4) [<c052d7b4>] (rpm_suspend) from [<c052ee34>] (__pm_runtime_suspend+0x64/0x84) [<c052ee34>] (__pm_runtime_suspend) from [<c04bf3bc>] (omap2_gpio_prepare_for_idle+0x5c/0x70) [<c04bf3bc>] (omap2_gpio_prepare_for_idle) from [<c01255e8>] (omap_sram_idle+0x140/0x244) [<c01255e8>] (omap_sram_idle) from [<c0126b48>] (omap3_enter_idle_bm+0xfc/0x1ec) [<c0126b48>] (omap3_enter_idle_bm) from [<c0601db8>] (cpuidle_enter_state+0x80/0x3d4) [<c0601db8>] (cpuidle_enter_state) from [<c0183c74>] (cpu_startup_entry+0x198/0x3a0) [<c0183c74>] (cpu_startup_entry) from [<c0b00c0c>] (start_kernel+0x354/0x3c8) [<c0b00c0c>] (start_kernel) from [<8000807c>] (0x8000807c) ------------------------------------------------------------------------ Reported-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Tony Lindgren <tony@atomide.com> Tested-by: Guenter Roeck <linux@roeck-us.net> [ rjw: Subject ] Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-04-26 10:42:25 -07:00
PM / Runtime: Use device PM QoS constraints (v2) Make the runtime PM core use device PM QoS constraints to check if it is allowed to suspend a given device, so that an error code is returned if the device's own PM QoS constraint is negative or one of its children has already been suspended for too long. If this is not the case, the maximum estimated time the device is allowed to be suspended, computed as the minimum of the device's PM QoS constraint and the PM QoS constraints of its children (reduced by the difference between the current time and their suspend times) is stored in a new device's PM field power.max_time_suspended_ns that can be used by the device's subsystem or PM domain to decide whether or not to put the device into lower-power (and presumably higher-latency) states later (if the constraint is 0, which means "no constraint", the power.max_time_suspended_ns is set to -1). Additionally, the time of execution of the subsystem-level .runtime_suspend() callback for the device is recorded in the new power.suspend_time field for later use by the device's subsystem or PM domain along with power.max_time_suspended_ns (it also is used by the core code when the device's parent is suspended). Introduce a new helper function, pm_runtime_update_max_time_suspended(), allowing subsystems and PM domains (or device drivers) to update the power.max_time_suspended_ns field, for example after changing the power state of a suspended device. Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
2011-12-01 00:01:31 +01:00
PM / wakeirq: Fix dedicated wakeirq for drivers not using autosuspend I noticed some wakeirq flakeyness with consumer drivers not using autosuspend. For drivers not using autosuspend, the wakeirq may never get unmasked in rpm_suspend() because of irq desc->depth. We are configuring dedicated wakeirqs to start with IRQ_NOAUTOEN as we naturally don't want them running until rpm_suspend() is called. However, when a consumer driver initially calls pm_runtime_get(), we now wrongly start with disable_irq_nosync() call on the dedicated wakeirq that is disabled to start with. This causes desc->depth to toggle between 1 and 2 instead of the usual 0 and 1. This can prevent enable_irq() from unmasking the wakeirq as that only happens at desc->depth 1. This does not necessarily show up with drivers using autosuspend as there is time for disable_irq_nosync() before rpm_suspend() gets called after the autosuspend timeout. Let's fix the issue by adding wirq->status that lazily gets set on the first rpm_suspend(). We also need PM runtime core private functions for dev_pm_enable_wake_irq_check() and dev_pm_disable_wake_irq_check() so we can enable the dedicated wakeirq on the first rpm_suspend(). While at it, let's also fix the comments for dev_pm_enable_wake_irq() and dev_pm_disable_wake_irq(). Those can still be used by the consumer drivers as needed because the IRQ core manages the interrupt usecount for us. Fixes: 4990d4fe327b (PM / Wakeirq: Add automated device wake IRQ handling) Signed-off-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-12-05 16:38:16 -08:00
PM / Runtime: Use device PM QoS constraints (v2) Make the runtime PM core use device PM QoS constraints to check if it is allowed to suspend a given device, so that an error code is returned if the device's own PM QoS constraint is negative or one of its children has already been suspended for too long. If this is not the case, the maximum estimated time the device is allowed to be suspended, computed as the minimum of the device's PM QoS constraint and the PM QoS constraints of its children (reduced by the difference between the current time and their suspend times) is stored in a new device's PM field power.max_time_suspended_ns that can be used by the device's subsystem or PM domain to decide whether or not to put the device into lower-power (and presumably higher-latency) states later (if the constraint is 0, which means "no constraint", the power.max_time_suspended_ns is set to -1). Additionally, the time of execution of the subsystem-level .runtime_suspend() callback for the device is recorded in the new power.suspend_time field for later use by the device's subsystem or PM domain along with power.max_time_suspended_ns (it also is used by the core code when the device's parent is suspended). Introduce a new helper function, pm_runtime_update_max_time_suspended(), allowing subsystems and PM domains (or device drivers) to update the power.max_time_suspended_ns field, for example after changing the power state of a suspended device. Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
2011-12-01 00:01:31 +01:00
PM / Runtime: Use device PM QoS constraints (v2) Make the runtime PM core use device PM QoS constraints to check if it is allowed to suspend a given device, so that an error code is returned if the device's own PM QoS constraint is negative or one of its children has already been suspended for too long. If this is not the case, the maximum estimated time the device is allowed to be suspended, computed as the minimum of the device's PM QoS constraint and the PM QoS constraints of its children (reduced by the difference between the current time and their suspend times) is stored in a new device's PM field power.max_time_suspended_ns that can be used by the device's subsystem or PM domain to decide whether or not to put the device into lower-power (and presumably higher-latency) states later (if the constraint is 0, which means "no constraint", the power.max_time_suspended_ns is set to -1). Additionally, the time of execution of the subsystem-level .runtime_suspend() callback for the device is recorded in the new power.suspend_time field for later use by the device's subsystem or PM domain along with power.max_time_suspended_ns (it also is used by the core code when the device's parent is suspended). Introduce a new helper function, pm_runtime_update_max_time_suspended(), allowing subsystems and PM domains (or device drivers) to update the power.max_time_suspended_ns field, for example after changing the power state of a suspended device. Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
2011-12-01 00:01:31 +01:00
PM / runtime: Add _rcuidle suffix to allow rpm_resume() to be called from idle This commit applies another _rcuidle suffix to fix an RCU use from idle. > =============================== > [ INFO: suspicious RCU usage. ] > 4.6.0-rc5-next-20160426+ #1122 Not tainted > ------------------------------- > include/trace/events/rpm.h:69 suspicious rcu_dereference_check() usage! > > other info that might help us debug this: > > > RCU used illegally from idle CPU! > rcu_scheduler_active = 1, debug_locks = 0 > RCU used illegally from extended quiescent state! > 1 lock held by swapper/0/0: > #0: (&(&dev->power.lock)->rlock){-.-...}, at: [<c052e3dc>] __pm_runtime_resume+0x3c/0x64 > > stack backtrace: > CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.6.0-rc5-next-20160426+ #1122 > Hardware name: Generic OMAP36xx (Flattened Device Tree) > [<c0110290>] (unwind_backtrace) from [<c010c3a8>] (show_stack+0x10/0x14) > [<c010c3a8>] (show_stack) from [<c047fd68>] (dump_stack+0xb0/0xe4) > [<c047fd68>] (dump_stack) from [<c052e178>] (rpm_resume+0x5cc/0x7f4) > [<c052e178>] (rpm_resume) from [<c052e3ec>] (__pm_runtime_resume+0x4c/0x64) > [<c052e3ec>] (__pm_runtime_resume) from [<c04bf2c4>] (omap2_gpio_resume_after_idle+0x54/0x68) > [<c04bf2c4>] (omap2_gpio_resume_after_idle) from [<c01269dc>] (omap3_enter_idle_bm+0xfc/0x1ec) > [<c01269dc>] (omap3_enter_idle_bm) from [<c060198c>] (cpuidle_enter_state+0x80/0x3d4) > [<c060198c>] (cpuidle_enter_state) from [<c0183b08>] (cpu_startup_entry+0x198/0x3a0) > [<c0183b08>] (cpu_startup_entry) from [<c0b00c0c>] (start_kernel+0x354/0x3c8) > [<c0b00c0c>] (start_kernel) from [<8000807c>] (0x8000807c) Reported-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Tony Lindgren <tony@atomide.com> Tested-by: Guenter Roeck <linux@roeck-us.net> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-04-26 13:38:55 -07:00
PM / Runtime: let rpm_resume() succeed if RPM_ACTIVE, even when disabled, v2 There are several drivers where the return value of pm_runtime_get_sync() is used to decide whether or not it is safe to access hardware and that don't provide .suspend() callbacks for system suspend (but may use late/noirq callbacks.) If such a driver happens to call pm_runtime_get_sync() during system suspend, after the core has disabled runtime PM, it will get the error code and will decide that the hardware should not be accessed, although this may be a wrong conclusion, depending on the state of the device when runtime PM was disabled. Drivers might work around this problem by using a test like: ret = pm_runtime_get_sync(dev); if (!ret || (ret == -EACCES && driver_private_data(dev)->suspended)) { /* access hardware */ } where driver_private_data(dev)->suspended is a flag set by the driver's .suspend() method (that would have to be added for this purpose). However, that potentially would need to be done by multiple drivers which means quite a lot of duplicated code and bloat. To avoid that we can use the observation that the core sets dev->power.is_suspended before disabling runtime PM and use that instead of the driver's private flag. Still, potentially many drivers would need to repeat that same check in quite a few places, so it's better to let the core do it. Then we can be a bit smarter and check whether or not runtime PM was disabled by the core only (disable_depth == 1) or by someone else in addition to the core (disable_depth > 1). In the former case rpm_resume() can return 1 if the runtime PM status is RPM_ACTIVE, because it means the device was active when the core disabled runtime PM. In the latter case it should still return -EACCES, because it isn't clear why runtime PM has been disabled. Tested on AM3730/Beagle-xM where a wakeup IRQ firing during the late suspend phase triggers runtime PM activity in the I2C driver since the wakeup IRQ is on an I2C-connected PMIC. [rjw: Modified whitespace to follow the file's convention.] Signed-off-by: Kevin Hilman <khilman@ti.com> Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
2012-09-21 22:47:34 +00:00
PM / wakeirq: Fix dedicated wakeirq for drivers not using autosuspend I noticed some wakeirq flakeyness with consumer drivers not using autosuspend. For drivers not using autosuspend, the wakeirq may never get unmasked in rpm_suspend() because of irq desc->depth. We are configuring dedicated wakeirqs to start with IRQ_NOAUTOEN as we naturally don't want them running until rpm_suspend() is called. However, when a consumer driver initially calls pm_runtime_get(), we now wrongly start with disable_irq_nosync() call on the dedicated wakeirq that is disabled to start with. This causes desc->depth to toggle between 1 and 2 instead of the usual 0 and 1. This can prevent enable_irq() from unmasking the wakeirq as that only happens at desc->depth 1. This does not necessarily show up with drivers using autosuspend as there is time for disable_irq_nosync() before rpm_suspend() gets called after the autosuspend timeout. Let's fix the issue by adding wirq->status that lazily gets set on the first rpm_suspend(). We also need PM runtime core private functions for dev_pm_enable_wake_irq_check() and dev_pm_disable_wake_irq_check() so we can enable the dedicated wakeirq on the first rpm_suspend(). While at it, let's also fix the comments for dev_pm_enable_wake_irq() and dev_pm_disable_wake_irq(). Those can still be used by the consumer drivers as needed because the IRQ core manages the interrupt usecount for us. Fixes: 4990d4fe327b (PM / Wakeirq: Add automated device wake IRQ handling) Signed-off-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-12-05 16:38:16 -08:00
PM / wakeirq: Fix dedicated wakeirq for drivers not using autosuspend I noticed some wakeirq flakeyness with consumer drivers not using autosuspend. For drivers not using autosuspend, the wakeirq may never get unmasked in rpm_suspend() because of irq desc->depth. We are configuring dedicated wakeirqs to start with IRQ_NOAUTOEN as we naturally don't want them running until rpm_suspend() is called. However, when a consumer driver initially calls pm_runtime_get(), we now wrongly start with disable_irq_nosync() call on the dedicated wakeirq that is disabled to start with. This causes desc->depth to toggle between 1 and 2 instead of the usual 0 and 1. This can prevent enable_irq() from unmasking the wakeirq as that only happens at desc->depth 1. This does not necessarily show up with drivers using autosuspend as there is time for disable_irq_nosync() before rpm_suspend() gets called after the autosuspend timeout. Let's fix the issue by adding wirq->status that lazily gets set on the first rpm_suspend(). We also need PM runtime core private functions for dev_pm_enable_wake_irq_check() and dev_pm_disable_wake_irq_check() so we can enable the dedicated wakeirq on the first rpm_suspend(). While at it, let's also fix the comments for dev_pm_enable_wake_irq() and dev_pm_disable_wake_irq(). Those can still be used by the consumer drivers as needed because the IRQ core manages the interrupt usecount for us. Fixes: 4990d4fe327b (PM / Wakeirq: Add automated device wake IRQ handling) Signed-off-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-12-05 16:38:16 -08:00
PM / runtime: Add _rcuidle suffix to allow rpm_resume() to be called from idle This commit applies another _rcuidle suffix to fix an RCU use from idle. > =============================== > [ INFO: suspicious RCU usage. ] > 4.6.0-rc5-next-20160426+ #1122 Not tainted > ------------------------------- > include/trace/events/rpm.h:69 suspicious rcu_dereference_check() usage! > > other info that might help us debug this: > > > RCU used illegally from idle CPU! > rcu_scheduler_active = 1, debug_locks = 0 > RCU used illegally from extended quiescent state! > 1 lock held by swapper/0/0: > #0: (&(&dev->power.lock)->rlock){-.-...}, at: [<c052e3dc>] __pm_runtime_resume+0x3c/0x64 > > stack backtrace: > CPU: 0 PID: 0 Comm: swapper/0 Not tainted 4.6.0-rc5-next-20160426+ #1122 > Hardware name: Generic OMAP36xx (Flattened Device Tree) > [<c0110290>] (unwind_backtrace) from [<c010c3a8>] (show_stack+0x10/0x14) > [<c010c3a8>] (show_stack) from [<c047fd68>] (dump_stack+0xb0/0xe4) > [<c047fd68>] (dump_stack) from [<c052e178>] (rpm_resume+0x5cc/0x7f4) > [<c052e178>] (rpm_resume) from [<c052e3ec>] (__pm_runtime_resume+0x4c/0x64) > [<c052e3ec>] (__pm_runtime_resume) from [<c04bf2c4>] (omap2_gpio_resume_after_idle+0x54/0x68) > [<c04bf2c4>] (omap2_gpio_resume_after_idle) from [<c01269dc>] (omap3_enter_idle_bm+0xfc/0x1ec) > [<c01269dc>] (omap3_enter_idle_bm) from [<c060198c>] (cpuidle_enter_state+0x80/0x3d4) > [<c060198c>] (cpuidle_enter_state) from [<c0183b08>] (cpu_startup_entry+0x198/0x3a0) > [<c0183b08>] (cpu_startup_entry) from [<c0b00c0c>] (start_kernel+0x354/0x3c8) > [<c0b00c0c>] (start_kernel) from [<8000807c>] (0x8000807c) Reported-by: Tony Lindgren <tony@atomide.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Tested-by: Tony Lindgren <tony@atomide.com> Tested-by: Guenter Roeck <linux@roeck-us.net> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2016-04-26 13:38:55 -07:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM / runtime: Avoid false-positive warnings from might_sleep_if() The might_sleep_if() assertions in __pm_runtime_idle(), __pm_runtime_suspend() and __pm_runtime_resume() may generate false-positive warnings in some situations. For example, that happens if a nested pm_runtime_get_sync()/pm_runtime_put() pair is executed with disabled interrupts within an outer pm_runtime_get_sync()/pm_runtime_put() section for the same device. [Generally, pm_runtime_get_sync() may sleep, so it should not be called with disabled interrupts, but in this particular case the previous pm_runtime_get_sync() guarantees that the device will not be suspended, so the inner pm_runtime_get_sync() will return immediately after incrementing the device's usage counter.] That started to happen in the i915 driver in 4.10-rc, leading to the following splat: BUG: sleeping function called from invalid context at drivers/base/power/runtime.c:1032 in_atomic(): 1, irqs_disabled(): 0, pid: 1500, name: Xorg 1 lock held by Xorg/1500: #0: (&dev->struct_mutex){+.+.+.}, at: [<ffffffffa0680c13>] i915_mutex_lock_interruptible+0x43/0x140 [i915] CPU: 0 PID: 1500 Comm: Xorg Not tainted Call Trace: dump_stack+0x85/0xc2 ___might_sleep+0x196/0x260 __might_sleep+0x53/0xb0 __pm_runtime_resume+0x7a/0x90 intel_runtime_pm_get+0x25/0x90 [i915] aliasing_gtt_bind_vma+0xaa/0xf0 [i915] i915_vma_bind+0xaf/0x1e0 [i915] i915_gem_execbuffer_relocate_entry+0x513/0x6f0 [i915] i915_gem_execbuffer_relocate_vma.isra.34+0x188/0x250 [i915] ? trace_hardirqs_on+0xd/0x10 ? i915_gem_execbuffer_reserve_vma.isra.31+0x152/0x1f0 [i915] ? i915_gem_execbuffer_reserve.isra.32+0x372/0x3a0 [i915] i915_gem_do_execbuffer.isra.38+0xa70/0x1a40 [i915] ? __might_fault+0x4e/0xb0 i915_gem_execbuffer2+0xc5/0x260 [i915] ? __might_fault+0x4e/0xb0 drm_ioctl+0x206/0x450 [drm] ? i915_gem_execbuffer+0x340/0x340 [i915] ? __fget+0x5/0x200 do_vfs_ioctl+0x91/0x6f0 ? __fget+0x111/0x200 ? __fget+0x5/0x200 SyS_ioctl+0x79/0x90 entry_SYSCALL_64_fastpath+0x23/0xc6 even though the code triggering it is correct. Unfortunately, the might_sleep_if() assertions in question are too coarse-grained to cover such cases correctly, so make them a bit less sensitive in order to avoid the false-positives. Reported-and-tested-by: Sedat Dilek <sedat.dilek@gmail.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2017-02-04 00:44:36 +01:00
PM / runtime: Avoid false-positive warnings from might_sleep_if() The might_sleep_if() assertions in __pm_runtime_idle(), __pm_runtime_suspend() and __pm_runtime_resume() may generate false-positive warnings in some situations. For example, that happens if a nested pm_runtime_get_sync()/pm_runtime_put() pair is executed with disabled interrupts within an outer pm_runtime_get_sync()/pm_runtime_put() section for the same device. [Generally, pm_runtime_get_sync() may sleep, so it should not be called with disabled interrupts, but in this particular case the previous pm_runtime_get_sync() guarantees that the device will not be suspended, so the inner pm_runtime_get_sync() will return immediately after incrementing the device's usage counter.] That started to happen in the i915 driver in 4.10-rc, leading to the following splat: BUG: sleeping function called from invalid context at drivers/base/power/runtime.c:1032 in_atomic(): 1, irqs_disabled(): 0, pid: 1500, name: Xorg 1 lock held by Xorg/1500: #0: (&dev->struct_mutex){+.+.+.}, at: [<ffffffffa0680c13>] i915_mutex_lock_interruptible+0x43/0x140 [i915] CPU: 0 PID: 1500 Comm: Xorg Not tainted Call Trace: dump_stack+0x85/0xc2 ___might_sleep+0x196/0x260 __might_sleep+0x53/0xb0 __pm_runtime_resume+0x7a/0x90 intel_runtime_pm_get+0x25/0x90 [i915] aliasing_gtt_bind_vma+0xaa/0xf0 [i915] i915_vma_bind+0xaf/0x1e0 [i915] i915_gem_execbuffer_relocate_entry+0x513/0x6f0 [i915] i915_gem_execbuffer_relocate_vma.isra.34+0x188/0x250 [i915] ? trace_hardirqs_on+0xd/0x10 ? i915_gem_execbuffer_reserve_vma.isra.31+0x152/0x1f0 [i915] ? i915_gem_execbuffer_reserve.isra.32+0x372/0x3a0 [i915] i915_gem_do_execbuffer.isra.38+0xa70/0x1a40 [i915] ? __might_fault+0x4e/0xb0 i915_gem_execbuffer2+0xc5/0x260 [i915] ? __might_fault+0x4e/0xb0 drm_ioctl+0x206/0x450 [drm] ? i915_gem_execbuffer+0x340/0x340 [i915] ? __fget+0x5/0x200 do_vfs_ioctl+0x91/0x6f0 ? __fget+0x111/0x200 ? __fget+0x5/0x200 SyS_ioctl+0x79/0x90 entry_SYSCALL_64_fastpath+0x23/0xc6 even though the code triggering it is correct. Unfortunately, the might_sleep_if() assertions in question are too coarse-grained to cover such cases correctly, so make them a bit less sensitive in order to avoid the false-positives. Reported-and-tested-by: Sedat Dilek <sedat.dilek@gmail.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2017-02-04 00:44:36 +01:00
PM / runtime: Avoid false-positive warnings from might_sleep_if() The might_sleep_if() assertions in __pm_runtime_idle(), __pm_runtime_suspend() and __pm_runtime_resume() may generate false-positive warnings in some situations. For example, that happens if a nested pm_runtime_get_sync()/pm_runtime_put() pair is executed with disabled interrupts within an outer pm_runtime_get_sync()/pm_runtime_put() section for the same device. [Generally, pm_runtime_get_sync() may sleep, so it should not be called with disabled interrupts, but in this particular case the previous pm_runtime_get_sync() guarantees that the device will not be suspended, so the inner pm_runtime_get_sync() will return immediately after incrementing the device's usage counter.] That started to happen in the i915 driver in 4.10-rc, leading to the following splat: BUG: sleeping function called from invalid context at drivers/base/power/runtime.c:1032 in_atomic(): 1, irqs_disabled(): 0, pid: 1500, name: Xorg 1 lock held by Xorg/1500: #0: (&dev->struct_mutex){+.+.+.}, at: [<ffffffffa0680c13>] i915_mutex_lock_interruptible+0x43/0x140 [i915] CPU: 0 PID: 1500 Comm: Xorg Not tainted Call Trace: dump_stack+0x85/0xc2 ___might_sleep+0x196/0x260 __might_sleep+0x53/0xb0 __pm_runtime_resume+0x7a/0x90 intel_runtime_pm_get+0x25/0x90 [i915] aliasing_gtt_bind_vma+0xaa/0xf0 [i915] i915_vma_bind+0xaf/0x1e0 [i915] i915_gem_execbuffer_relocate_entry+0x513/0x6f0 [i915] i915_gem_execbuffer_relocate_vma.isra.34+0x188/0x250 [i915] ? trace_hardirqs_on+0xd/0x10 ? i915_gem_execbuffer_reserve_vma.isra.31+0x152/0x1f0 [i915] ? i915_gem_execbuffer_reserve.isra.32+0x372/0x3a0 [i915] i915_gem_do_execbuffer.isra.38+0xa70/0x1a40 [i915] ? __might_fault+0x4e/0xb0 i915_gem_execbuffer2+0xc5/0x260 [i915] ? __might_fault+0x4e/0xb0 drm_ioctl+0x206/0x450 [drm] ? i915_gem_execbuffer+0x340/0x340 [i915] ? __fget+0x5/0x200 do_vfs_ioctl+0x91/0x6f0 ? __fget+0x111/0x200 ? __fget+0x5/0x200 SyS_ioctl+0x79/0x90 entry_SYSCALL_64_fastpath+0x23/0xc6 even though the code triggering it is correct. Unfortunately, the might_sleep_if() assertions in question are too coarse-grained to cover such cases correctly, so make them a bit less sensitive in order to avoid the false-positives. Reported-and-tested-by: Sedat Dilek <sedat.dilek@gmail.com> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2017-02-04 00:44:36 +01:00
PM-runtime: Switch autosuspend over to using hrtimers PM-runtime uses the timer infrastructure for autosuspend. This implies that the minimum time before autosuspending a device is in the range of 1 tick included to 2 ticks excluded -On arm64 this means between 4ms and 8ms with default jiffies configuration -And on arm, it is between 10ms and 20ms These values are quite high for embedded systems which sometimes want the duration to be in the range of 1 ms. It is possible to switch autosuspend over to using hrtimers to get finer granularity for short durations and take advantage of slack to retain some margins and get long timeouts with minimum wakeups. On an arm64 platform that uses 1ms for autosuspending timeout of its GPU, idle power is reduced by 10% with hrtimer. The latency impact on arm64 hikey octo cores is: - mark_last_busy: from 1.11 us to 1.25 us - rpm_suspend: from 15.54 us to 15.38 us [Only the code path of rpm_suspend() that starts hrtimer has been measured.] arm64 image (arm64 default defconfig) decreases by around 3KB with following details: $ size vmlinux-timer text data bss dec hex filename 12034646 6869268 386840 19290754 1265a82 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 12030550 6870164 387032 19287746 1264ec2 vmlinux The latency impact on arm 32bits snowball dual cores is : - mark_last_busy: from 0.31 us usec to 0.77 us - rpm_suspend: from 6.83 us to 6.67 usec The increase of the image for snowball platform that I used for testing performance impact, is neglictable (244B). $ size vmlinux-timer text data bss dec hex filename 7157961 2119580 264120 9541661 91981d build-ux500/vmlinux size vmlinux-hrtimer text data bss dec hex filename 7157773 2119884 264248 9541905 919911 vmlinux-hrtimer And arm 32bits image (multi_v7_defconfig) increases by around 1.7KB with following details: $ size vmlinux-timer text data bss dec hex filename 13304443 6803420 402768 20510631 138f7a7 vmlinux $ size vmlinux-hrtimer text data bss dec hex filename 13304299 6805276 402768 20512343 138fe57 vmlinux Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2018-12-14 15:22:25 +01:00
PM / runtime: Re-init runtime PM states at probe error and driver unbind There are two common expectations among several subsystems/drivers that deploys runtime PM support, but which isn't met by the driver core. Expectation 1) At ->probe() the subsystem/driver expects the runtime PM status of the device to be RPM_SUSPENDED, which is the initial status being assigned at device registration. This expectation is especially common among some of those subsystems/ drivers that manages devices with an attached PM domain, as those requires the ->runtime_resume() callback at the PM domain level to be invoked during ->probe(). Moreover these subsystems/drivers entirely relies on runtime PM resources being managed at the PM domain level, thus don't implement their own set of runtime PM callbacks. These are two scenarios that suffers from this unmet expectation. i) A failed ->probe() sequence requests probe deferral: ->probe() ... pm_runtime_enable() pm_runtime_get_sync() ... err: pm_runtime_put() pm_runtime_disable() ... As there are no guarantees that such sequence turns the runtime PM status of the device into RPM_SUSPENDED, the re-trying ->probe() may start with the status in RPM_ACTIVE. In such case the runtime PM core won't invoke the ->runtime_resume() callback because of a pm_runtime_get_sync(), as it considers the device to be already runtime resumed. ii) A driver re-bind sequence: At driver unbind, the subsystem/driver's >remove() callback invokes a sequence of runtime PM APIs, to undo actions during ->probe() and to put the device into low power state. ->remove() ... pm_runtime_put() pm_runtime_disable() ... Similar as in the failing ->probe() case, this sequence don't guarantee the runtime PM status of the device to turn into RPM_SUSPENDED. Trying to re-bind the driver thus causes the same issue as when re-trying ->probe(), in the probe deferral scenario. Expectation 2) Drivers that invokes the pm_runtime_irq_safe() API during ->probe(), triggers the runtime PM core to increase the usage count for the device's parent and permanently make it runtime resumed. The usage count is only dropped at device removal, which also allows it to be runtime suspended again. A re-trying ->probe() repeats the call to pm_runtime_irq_safe() and thus once more triggers the usage count of the device's parent to be increased. This leads to not only an imbalance issue of the usage count of the device's parent, but also to keep it runtime resumed permanently even if ->probe() fails. To address these issues, let's change the policy of the driver core to meet these expectations. More precisely, at ->probe() failures and driver unbind, restore the initial states of runtime PM. Although to still allow subsystem's to control PM for devices that doesn't ->probe() successfully, don't restore the initial states unless runtime PM is disabled. Signed-off-by: Ulf Hansson <ulf.hansson@linaro.org> Reviewed-by: Kevin Hilman <khilman@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2015-11-18 11:48:39 +01:00
PM / runtime: Re-init runtime PM states at probe error and driver unbind There are two common expectations among several subsystems/drivers that deploys runtime PM support, but which isn't met by the driver core. Expectation 1) At ->probe() the subsystem/driver expects the runtime PM status of the device to be RPM_SUSPENDED, which is the initial status being assigned at device registration. This expectation is especially common among some of those subsystems/ drivers that manages devices with an attached PM domain, as those requires the ->runtime_resume() callback at the PM domain level to be invoked during ->probe(). Moreover these subsystems/drivers entirely relies on runtime PM resources being managed at the PM domain level, thus don't implement their own set of runtime PM callbacks. These are two scenarios that suffers from this unmet expectation. i) A failed ->probe() sequence requests probe deferral: ->probe() ... pm_runtime_enable() pm_runtime_get_sync() ... err: pm_runtime_put() pm_runtime_disable() ... As there are no guarantees that such sequence turns the runtime PM status of the device into RPM_SUSPENDED, the re-trying ->probe() may start with the status in RPM_ACTIVE. In such case the runtime PM core won't invoke the ->runtime_resume() callback because of a pm_runtime_get_sync(), as it considers the device to be already runtime resumed. ii) A driver re-bind sequence: At driver unbind, the subsystem/driver's >remove() callback invokes a sequence of runtime PM APIs, to undo actions during ->probe() and to put the device into low power state. ->remove() ... pm_runtime_put() pm_runtime_disable() ... Similar as in the failing ->probe() case, this sequence don't guarantee the runtime PM status of the device to turn into RPM_SUSPENDED. Trying to re-bind the driver thus causes the same issue as when re-trying ->probe(), in the probe deferral scenario. Expectation 2) Drivers that invokes the pm_runtime_irq_safe() API during ->probe(), triggers the runtime PM core to increase the usage count for the device's parent and permanently make it runtime resumed. The usage count is only dropped at device removal, which also allows it to be runtime suspended again. A re-trying ->probe() repeats the call to pm_runtime_irq_safe() and thus once more triggers the usage count of the device's parent to be increased. This leads to not only an imbalance issue of the usage count of the device's parent, but also to keep it runtime resumed permanently even if ->probe() fails. To address these issues, let's change the policy of the driver core to meet these expectations. More precisely, at ->probe() failures and driver unbind, restore the initial states of runtime PM. Although to still allow subsystem's to control PM for devices that doesn't ->probe() successfully, don't restore the initial states unless runtime PM is disabled. Signed-off-by: Ulf Hansson <ulf.hansson@linaro.org> Reviewed-by: Kevin Hilman <khilman@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
2015-11-18 11:48:39 +01:00
driver core: Fix handling of runtime PM flags in device_link_add() After commit ead18c23c263 ("driver core: Introduce device links reference counting"), if there is a link between the given supplier and the given consumer already, device_link_add() will refcount it and return it unconditionally without updating its flags. It is possible, however, that the second (or any subsequent) caller of device_link_add() for the same consumer-supplier pair will pass DL_FLAG_PM_RUNTIME, possibly along with DL_FLAG_RPM_ACTIVE, in flags to it and the existing link may not behave as expected then. First, if DL_FLAG_PM_RUNTIME is not set in the existing link's flags at all, it needs to be set like during the original initialization of the link. Second, if DL_FLAG_RPM_ACTIVE is passed to device_link_add() in flags (in addition to DL_FLAG_PM_RUNTIME), the existing link should to be updated to reflect the "active" runtime PM configuration of the consumer-supplier pair and extra care must be taken here to avoid possible destructive races with runtime PM of the consumer. To that end, redefine the rpm_active field in struct device_link as a refcount, initialize it to 1 and make rpm_resume() (for the consumer) and device_link_add() increment it whenever they acquire a runtime PM reference on the supplier device. Accordingly, make rpm_suspend() (for the consumer) and pm_runtime_clean_up_links() decrement it and drop runtime PM references to the supplier device in a loop until rpm_active becones 1 again. Fixes: ead18c23c263 ("driver core: Introduce device links reference counting") Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-01 01:49:14 +01:00
driver core: Remove device link creation limitation If device_link_add() is called for a consumer/supplier pair with an existing device link between them and the existing link's type is not in agreement with the flags passed to that function by its caller, NULL will be returned. That is seriously inconvenient, because it forces the callers of device_link_add() to worry about what others may or may not do even if that is not relevant to them for any other reasons. It turns out, however, that this limitation can be made go away relatively easily. The underlying observation is that if DL_FLAG_STATELESS has been passed to device_link_add() in flags for the given consumer/supplier pair at least once, calling either device_link_del() or device_link_remove() to release the link returned by it should work, but there are no other requirements associated with that flag. In turn, if at least one of the callers of device_link_add() for the given consumer/supplier pair has not passed DL_FLAG_STATELESS to it in flags, the driver core should track the status of the link and act on it as appropriate (ie. the link should be treated as "managed"). This means that DL_FLAG_STATELESS needs to be set for managed device links and it should be valid to call device_link_del() or device_link_remove() to drop references to them in certain sutiations. To allow that to happen, introduce a new (internal) device link flag called DL_FLAG_MANAGED and make device_link_add() set it automatically whenever DL_FLAG_STATELESS is not passed to it. Also make it take additional references to existing device links that were previously stateless (that is, with DL_FLAG_STATELESS set and DL_FLAG_MANAGED unset) and will need to be managed going forward and initialize their status (which has been DL_STATE_NONE so far). Accordingly, when a managed device link is dropped automatically by the driver core, make it clear DL_FLAG_MANAGED, reset the link's status back to DL_STATE_NONE and drop the reference to it associated with DL_FLAG_MANAGED instead of just deleting it right away (to allow it to stay around in case it still needs to be released explicitly by someone). With that, since setting DL_FLAG_STATELESS doesn't mean that the device link in question is not managed any more, replace all of the status-tracking checks against DL_FLAG_STATELESS with analogous checks against DL_FLAG_MANAGED and update the documentation to reflect these changes. While at it, make device_link_add() reject flags that it does not recognize, including DL_FLAG_MANAGED. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Saravana Kannan <saravanak@google.com> Tested-by: Marek Szyprowski <m.szyprowski@samsung.com> Review-by: Saravana Kannan <saravanak@google.com> Link: https://lore.kernel.org/r/2305283.AStDPdUUnE@kreacher Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-07-16 17:21:06 +02:00
driver core: Remove device link creation limitation If device_link_add() is called for a consumer/supplier pair with an existing device link between them and the existing link's type is not in agreement with the flags passed to that function by its caller, NULL will be returned. That is seriously inconvenient, because it forces the callers of device_link_add() to worry about what others may or may not do even if that is not relevant to them for any other reasons. It turns out, however, that this limitation can be made go away relatively easily. The underlying observation is that if DL_FLAG_STATELESS has been passed to device_link_add() in flags for the given consumer/supplier pair at least once, calling either device_link_del() or device_link_remove() to release the link returned by it should work, but there are no other requirements associated with that flag. In turn, if at least one of the callers of device_link_add() for the given consumer/supplier pair has not passed DL_FLAG_STATELESS to it in flags, the driver core should track the status of the link and act on it as appropriate (ie. the link should be treated as "managed"). This means that DL_FLAG_STATELESS needs to be set for managed device links and it should be valid to call device_link_del() or device_link_remove() to drop references to them in certain sutiations. To allow that to happen, introduce a new (internal) device link flag called DL_FLAG_MANAGED and make device_link_add() set it automatically whenever DL_FLAG_STATELESS is not passed to it. Also make it take additional references to existing device links that were previously stateless (that is, with DL_FLAG_STATELESS set and DL_FLAG_MANAGED unset) and will need to be managed going forward and initialize their status (which has been DL_STATE_NONE so far). Accordingly, when a managed device link is dropped automatically by the driver core, make it clear DL_FLAG_MANAGED, reset the link's status back to DL_STATE_NONE and drop the reference to it associated with DL_FLAG_MANAGED instead of just deleting it right away (to allow it to stay around in case it still needs to be released explicitly by someone). With that, since setting DL_FLAG_STATELESS doesn't mean that the device link in question is not managed any more, replace all of the status-tracking checks against DL_FLAG_STATELESS with analogous checks against DL_FLAG_MANAGED and update the documentation to reflect these changes. While at it, make device_link_add() reject flags that it does not recognize, including DL_FLAG_MANAGED. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Saravana Kannan <saravanak@google.com> Tested-by: Marek Szyprowski <m.szyprowski@samsung.com> Review-by: Saravana Kannan <saravanak@google.com> Link: https://lore.kernel.org/r/2305283.AStDPdUUnE@kreacher Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-07-16 17:21:06 +02:00
driver core: Fix handling of runtime PM flags in device_link_add() After commit ead18c23c263 ("driver core: Introduce device links reference counting"), if there is a link between the given supplier and the given consumer already, device_link_add() will refcount it and return it unconditionally without updating its flags. It is possible, however, that the second (or any subsequent) caller of device_link_add() for the same consumer-supplier pair will pass DL_FLAG_PM_RUNTIME, possibly along with DL_FLAG_RPM_ACTIVE, in flags to it and the existing link may not behave as expected then. First, if DL_FLAG_PM_RUNTIME is not set in the existing link's flags at all, it needs to be set like during the original initialization of the link. Second, if DL_FLAG_RPM_ACTIVE is passed to device_link_add() in flags (in addition to DL_FLAG_PM_RUNTIME), the existing link should to be updated to reflect the "active" runtime PM configuration of the consumer-supplier pair and extra care must be taken here to avoid possible destructive races with runtime PM of the consumer. To that end, redefine the rpm_active field in struct device_link as a refcount, initialize it to 1 and make rpm_resume() (for the consumer) and device_link_add() increment it whenever they acquire a runtime PM reference on the supplier device. Accordingly, make rpm_suspend() (for the consumer) and pm_runtime_clean_up_links() decrement it and drop runtime PM references to the supplier device in a loop until rpm_active becones 1 again. Fixes: ead18c23c263 ("driver core: Introduce device links reference counting") Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-01 01:49:14 +01:00
driver core: Fix possible supplier PM-usage counter imbalance If a stateless device link to a certain supplier with DL_FLAG_PM_RUNTIME set in the flags is added and then removed by the consumer driver's probe callback, the supplier's PM-runtime usage counter will be nonzero after that which effectively causes the supplier to remain "always on" going forward. Namely, device_link_add() called to add the link invokes device_link_rpm_prepare() which notices that the consumer driver is probing, so it increments the supplier's PM-runtime usage counter with the assumption that the link will stay around until pm_runtime_put_suppliers() is called by driver_probe_device(), but if the link goes away before that point, the supplier's PM-runtime usage counter will remain nonzero. To prevent that from happening, first rework pm_runtime_get_suppliers() and pm_runtime_put_suppliers() to use the rpm_active refounts of device links and make the latter only drop rpm_active and the supplier's PM-runtime usage counter for each link by one, unless rpm_active is one already for it. Next, modify device_link_add() to bump up the new link's rpm_active refcount and the suppliers PM-runtime usage counter by two, to prevent pm_runtime_put_suppliers(), if it is called subsequently, from suspending the supplier prematurely (in case its PM-runtime usage counter goes down to 0 in there). Due to the way rpm_put_suppliers() works, this change does not affect runtime suspend of the consumer ends of new device links (or, generally, device links for which DL_FLAG_PM_RUNTIME has just been set). Fixes: e2f3cd831a28 ("driver core: Fix handling of runtime PM flags in device_link_add()") Reported-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Tested-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-12 13:08:10 +01:00
driver core: Fix PM-runtime for links added during consumer probe Commit 4c06c4e6cf63 ("driver core: Fix possible supplier PM-usage counter imbalance") introduced a regression that causes suppliers to be suspended prematurely for device links added during consumer driver probe if the initial PM-runtime status of the consumer is "suspended" and the consumer is resumed after adding the link and before pm_runtime_put_suppliers() is called. In that case, pm_runtime_put_suppliers() will drop the rpm_active refcount for the link by one and (since rpm_active is equal to two after the preceding consumer resume) the supplier's PM-runtime usage counter will be decremented, which may cause the supplier to suspend even though the consumer's PM-runtime status is "active". For this reason, partially revert commit 4c06c4e6cf63 as the problem it tried to fix needs to be addressed somewhat differently, and change pm_runtime_get_suppliers() and pm_runtime_put_suppliers() so that the latter only drops rpm_active references acquired by the former. [This requires adding a new field to struct device_link, but I coulnd't find a cleaner way to address the issue that would work in all cases.] This causes pm_runtime_put_suppliers() to effectively ignore device links added during consumer probe, so device_link_add() doesn't need to worry about ensuring that suppliers will remain active after pm_runtime_put_suppliers() for links created with DL_FLAG_RPM_ACTIVE set and it only needs to bump up rpm_active by one for those links, so pm_runtime_active_link() is not necessary any more. Fixes: 4c06c4e6cf63 ("driver core: Fix possible supplier PM-usage counter imbalance") Reported-by: Jon Hunter <jonathanh@nvidia.com> Tested-by: Jon Hunter <jonathanh@nvidia.com> Tested-by: Ulf Hansson <ulf.hansson@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Tested-by: Thierry Reding <treding@nvidia.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-19 17:53:26 +01:00
driver core: Fix possible supplier PM-usage counter imbalance If a stateless device link to a certain supplier with DL_FLAG_PM_RUNTIME set in the flags is added and then removed by the consumer driver's probe callback, the supplier's PM-runtime usage counter will be nonzero after that which effectively causes the supplier to remain "always on" going forward. Namely, device_link_add() called to add the link invokes device_link_rpm_prepare() which notices that the consumer driver is probing, so it increments the supplier's PM-runtime usage counter with the assumption that the link will stay around until pm_runtime_put_suppliers() is called by driver_probe_device(), but if the link goes away before that point, the supplier's PM-runtime usage counter will remain nonzero. To prevent that from happening, first rework pm_runtime_get_suppliers() and pm_runtime_put_suppliers() to use the rpm_active refounts of device links and make the latter only drop rpm_active and the supplier's PM-runtime usage counter for each link by one, unless rpm_active is one already for it. Next, modify device_link_add() to bump up the new link's rpm_active refcount and the suppliers PM-runtime usage counter by two, to prevent pm_runtime_put_suppliers(), if it is called subsequently, from suspending the supplier prematurely (in case its PM-runtime usage counter goes down to 0 in there). Due to the way rpm_put_suppliers() works, this change does not affect runtime suspend of the consumer ends of new device links (or, generally, device links for which DL_FLAG_PM_RUNTIME has just been set). Fixes: e2f3cd831a28 ("driver core: Fix handling of runtime PM flags in device_link_add()") Reported-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Tested-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-12 13:08:10 +01:00
driver core: Fix possible supplier PM-usage counter imbalance If a stateless device link to a certain supplier with DL_FLAG_PM_RUNTIME set in the flags is added and then removed by the consumer driver's probe callback, the supplier's PM-runtime usage counter will be nonzero after that which effectively causes the supplier to remain "always on" going forward. Namely, device_link_add() called to add the link invokes device_link_rpm_prepare() which notices that the consumer driver is probing, so it increments the supplier's PM-runtime usage counter with the assumption that the link will stay around until pm_runtime_put_suppliers() is called by driver_probe_device(), but if the link goes away before that point, the supplier's PM-runtime usage counter will remain nonzero. To prevent that from happening, first rework pm_runtime_get_suppliers() and pm_runtime_put_suppliers() to use the rpm_active refounts of device links and make the latter only drop rpm_active and the supplier's PM-runtime usage counter for each link by one, unless rpm_active is one already for it. Next, modify device_link_add() to bump up the new link's rpm_active refcount and the suppliers PM-runtime usage counter by two, to prevent pm_runtime_put_suppliers(), if it is called subsequently, from suspending the supplier prematurely (in case its PM-runtime usage counter goes down to 0 in there). Due to the way rpm_put_suppliers() works, this change does not affect runtime suspend of the consumer ends of new device links (or, generally, device links for which DL_FLAG_PM_RUNTIME has just been set). Fixes: e2f3cd831a28 ("driver core: Fix handling of runtime PM flags in device_link_add()") Reported-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Tested-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-12 13:08:10 +01:00
driver core: Fix PM-runtime for links added during consumer probe Commit 4c06c4e6cf63 ("driver core: Fix possible supplier PM-usage counter imbalance") introduced a regression that causes suppliers to be suspended prematurely for device links added during consumer driver probe if the initial PM-runtime status of the consumer is "suspended" and the consumer is resumed after adding the link and before pm_runtime_put_suppliers() is called. In that case, pm_runtime_put_suppliers() will drop the rpm_active refcount for the link by one and (since rpm_active is equal to two after the preceding consumer resume) the supplier's PM-runtime usage counter will be decremented, which may cause the supplier to suspend even though the consumer's PM-runtime status is "active". For this reason, partially revert commit 4c06c4e6cf63 as the problem it tried to fix needs to be addressed somewhat differently, and change pm_runtime_get_suppliers() and pm_runtime_put_suppliers() so that the latter only drops rpm_active references acquired by the former. [This requires adding a new field to struct device_link, but I coulnd't find a cleaner way to address the issue that would work in all cases.] This causes pm_runtime_put_suppliers() to effectively ignore device links added during consumer probe, so device_link_add() doesn't need to worry about ensuring that suppliers will remain active after pm_runtime_put_suppliers() for links created with DL_FLAG_RPM_ACTIVE set and it only needs to bump up rpm_active by one for those links, so pm_runtime_active_link() is not necessary any more. Fixes: 4c06c4e6cf63 ("driver core: Fix possible supplier PM-usage counter imbalance") Reported-by: Jon Hunter <jonathanh@nvidia.com> Tested-by: Jon Hunter <jonathanh@nvidia.com> Tested-by: Ulf Hansson <ulf.hansson@linaro.org> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Tested-by: Thierry Reding <treding@nvidia.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2019-02-19 17:53:26 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00
PM / runtime: Rework pm_runtime_force_suspend/resume() One of the limitations of pm_runtime_force_suspend/resume() is that if a parent driver wants to use these functions, all of its child drivers generally have to do that too because of the parent usage counter manipulations necessary to get the correct state of the parent during system-wide transitions to the working state (system resume). However, that limitation turns out to be artificial, so remove it. Namely, pm_runtime_force_suspend() only needs to update the children counter of its parent (if there's is a parent) when the device can stay in suspend after the subsequent system resume transition, as that counter is correct already otherwise. Now, if the parent's children counter is not updated, it is not necessary to increment the parent's usage counter in that case any more, as long as the children counters of devices are checked along with their usage counters in order to decide whether or not the devices may be left in suspend after the subsequent system resume transition. Accordingly, modify pm_runtime_force_suspend() to only call pm_runtime_set_suspended() for devices whose usage and children counters are at the "no references" level (the runtime PM status of the device needs to be updated to "suspended" anyway in case this function is called once again for the same device during the transition under way), drop the parent usage counter incrementation from it and update pm_runtime_force_resume() to compensate for these changes. Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>
2018-01-12 14:12:05 +01:00