| Message ID | 1413551588-14877-1-git-send-email-galak@codeaurora.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 10/17/2014 06:13 AM, Kumar Gala wrote: > With CONFIG_DEBUG_SPINLOCK enabled we see: > > BUG: spinlock bad magic on CPU#0, swapper/0 > lock: early_console_dev+0x8/0x160, .magic: 00000000, .owner: <none>/-1, .owner_cpu: 0 > CPU: 0 PID: 0 Comm: swapper Not tainted 3.17.0-09733-gb446b33-dirty #18 [...] > diff --git a/drivers/tty/serial/msm_serial.c b/drivers/tty/serial/msm_serial.c > index 4b6c783..8ab27c9 100644 > --- a/drivers/tty/serial/msm_serial.c > +++ b/drivers/tty/serial/msm_serial.c > @@ -973,6 +973,8 @@ msm_serial_early_console_setup(struct earlycon_device *device, const char *opt) > if (!device->port.membase) > return -ENODEV; > > + spin_lock_init(&device->port.lock); > + > device->con->write = msm_serial_early_write; > return 0; > } > @@ -995,6 +997,8 @@ msm_serial_early_console_setup_dm(struct earlycon_device *device, > if (!device->port.membase) > return -ENODEV; > > + spin_lock_init(&device->port.lock); > + > device->con->write = msm_serial_early_write_dm; > return 0; > } Could this be done in the tty core instead?
On 10/20/2014 02:27 PM, Stephen Boyd wrote: > On 10/17/2014 06:13 AM, Kumar Gala wrote: >> With CONFIG_DEBUG_SPINLOCK enabled we see: >> >> BUG: spinlock bad magic on CPU#0, swapper/0 >> lock: early_console_dev+0x8/0x160, .magic: 00000000, .owner: <none>/-1, .owner_cpu: 0 >> CPU: 0 PID: 0 Comm: swapper Not tainted 3.17.0-09733-gb446b33-dirty #18 > > [...] > >> diff --git a/drivers/tty/serial/msm_serial.c b/drivers/tty/serial/msm_serial.c >> index 4b6c783..8ab27c9 100644 >> --- a/drivers/tty/serial/msm_serial.c >> +++ b/drivers/tty/serial/msm_serial.c >> @@ -973,6 +973,8 @@ msm_serial_early_console_setup(struct earlycon_device *device, const char *opt) >> if (!device->port.membase) >> return -ENODEV; >> + spin_lock_init(&device->port.lock); >> + >> device->con->write = msm_serial_early_write; >> return 0; >> } >> @@ -995,6 +997,8 @@ msm_serial_early_console_setup_dm(struct earlycon_device *device, >> if (!device->port.membase) >> return -ENODEV; >> + spin_lock_init(&device->port.lock); >> + >> device->con->write = msm_serial_early_write_dm; >> return 0; >> } > > Could this be done in the tty core instead? Or even better: refactor __msm_console_write() so no spin_lock() is required for the earlycon write. FWIW, __msm_console_write() is a deadlock-waiting-to_happen since it does not disable interrupts with spin_lock_irqsave(). Regards, Peter Hurley
On 10/20, Peter Hurley wrote: > On 10/20/2014 02:27 PM, Stephen Boyd wrote: > > On 10/17/2014 06:13 AM, Kumar Gala wrote: > >> With CONFIG_DEBUG_SPINLOCK enabled we see: > >> > >> BUG: spinlock bad magic on CPU#0, swapper/0 > >> lock: early_console_dev+0x8/0x160, .magic: 00000000, .owner: <none>/-1, .owner_cpu: 0 > >> CPU: 0 PID: 0 Comm: swapper Not tainted 3.17.0-09733-gb446b33-dirty #18 > > > > [...] > > > >> diff --git a/drivers/tty/serial/msm_serial.c b/drivers/tty/serial/msm_serial.c > >> index 4b6c783..8ab27c9 100644 > >> --- a/drivers/tty/serial/msm_serial.c > >> +++ b/drivers/tty/serial/msm_serial.c > >> @@ -973,6 +973,8 @@ msm_serial_early_console_setup(struct earlycon_device *device, const char *opt) > >> if (!device->port.membase) > >> return -ENODEV; > >> + spin_lock_init(&device->port.lock); > >> + > >> device->con->write = msm_serial_early_write; > >> return 0; > >> } > >> @@ -995,6 +997,8 @@ msm_serial_early_console_setup_dm(struct earlycon_device *device, > >> if (!device->port.membase) > >> return -ENODEV; > >> + spin_lock_init(&device->port.lock); > >> + > >> device->con->write = msm_serial_early_write_dm; > >> return 0; > >> } > > > > Could this be done in the tty core instead? > > Or even better: refactor __msm_console_write() so no spin_lock() is required > for the earlycon write. Can't we keep the early console forever with keep_bootcon? In which case we may want to prevent things from trampling on each other? I don't see how the lock will help us here though because the port isn't even the same between the earlycon and the ttyMSM console. Removing the lock from the earlycon path is probably easier than caring about this debugging scenario. > > FWIW, __msm_console_write() is a deadlock-waiting-to_happen since it does > not disable interrupts with spin_lock_irqsave(). > Yeah we should fix that. As far as I can tell we've never used the irqsave variant of the lock in the console path. Did this become necessary at some point since this driver was introduced? I'm wondering how many stable trees need the fix.
On 10/20, Stephen Boyd wrote: > On 10/20, Peter Hurley wrote: > > > > FWIW, __msm_console_write() is a deadlock-waiting-to_happen since it does > > not disable interrupts with spin_lock_irqsave(). > > > > Yeah we should fix that. As far as I can tell we've never used > the irqsave variant of the lock in the console path. Did this > become necessary at some point since this driver was introduced? > I'm wondering how many stable trees need the fix. Actually I don't see this case. It looks like the console is always called in call_console_drivers() with irqs saved and disabled via the logbuf_lock in console_cont_flush() or console_unlock().
On 10/20/2014 06:20 PM, Stephen Boyd wrote: > On 10/20, Stephen Boyd wrote: >> On 10/20, Peter Hurley wrote: >>> >>> FWIW, __msm_console_write() is a deadlock-waiting-to_happen since it does >>> not disable interrupts with spin_lock_irqsave(). >>> >> >> Yeah we should fix that. As far as I can tell we've never used >> the irqsave variant of the lock in the console path. Did this >> become necessary at some point since this driver was introduced? >> I'm wondering how many stable trees need the fix. > > Actually I don't see this case. It looks like the console is > always called in call_console_drivers() with irqs saved and > disabled via the logbuf_lock in console_cont_flush() or > console_unlock(). Thanks for bringing that to my attention. That looks like a vestige of the initial transition away from BKL. I might look into just disabling preemption; even that might not really be necessary. Regards, Peter Hurley
diff --git a/drivers/tty/serial/msm_serial.c b/drivers/tty/serial/msm_serial.c index 4b6c783..8ab27c9 100644 --- a/drivers/tty/serial/msm_serial.c +++ b/drivers/tty/serial/msm_serial.c @@ -973,6 +973,8 @@ msm_serial_early_console_setup(struct earlycon_device *device, const char *opt) if (!device->port.membase) return -ENODEV; + spin_lock_init(&device->port.lock); + device->con->write = msm_serial_early_write; return 0; } @@ -995,6 +997,8 @@ msm_serial_early_console_setup_dm(struct earlycon_device *device, if (!device->port.membase) return -ENODEV; + spin_lock_init(&device->port.lock); + device->con->write = msm_serial_early_write_dm; return 0; }
With CONFIG_DEBUG_SPINLOCK enabled we see: BUG: spinlock bad magic on CPU#0, swapper/0 lock: early_console_dev+0x8/0x160, .magic: 00000000, .owner: <none>/-1, .owner_cpu: 0 CPU: 0 PID: 0 Comm: swapper Not tainted 3.17.0-09733-gb446b33-dirty #18 Call trace: [<c0087ba0>] dump_backtrace+0x0/0x140 [<c0087cf0>] show_stack+0x10/0x1c [<c0598be0>] dump_stack+0x74/0xb8 [<c00cb720>] spin_dump+0x78/0x88 [<c00cb754>] spin_bug+0x24/0x34 [<c00cb7e4>] do_raw_spin_lock+0x30/0x158 [<c059ea2c>] _raw_spin_lock+0x1c/0x2c [<c02e0e30>] __msm_console_write+0x68/0x154 [<c02e0f50>] msm_serial_early_write_dm+0x14/0x20 [<c00cf040>] call_console_drivers.constprop.23+0xa4/0xc4 [<c00d00c4>] console_unlock+0x2a8/0x330 [<c00d121c>] register_console+0x26c/0x338 [<c0805af4>] setup_earlycon+0x2c0/0x300 [<c0805c48>] msm_serial_dm_setup_earlycon+0x18/0x24 [<c07ed158>] do_early_param+0x84/0xd0 [<c00aec18>] parse_args+0x2b4/0x3ec [<c07ed588>] parse_early_options+0x2c/0x38 [<c07ed5c8>] parse_early_param+0x34/0x50 [<c07f03c0>] setup_arch+0x290/0x604 [<c07ed67c>] start_kernel+0x94/0x38c As the earlycon write function for msm requires a lock, we need to init it as part of the setup call. Signed-off-by: Kumar Gala <galak@codeaurora.org> --- drivers/tty/serial/msm_serial.c | 4 ++++ 1 file changed, 4 insertions(+)