| Message ID | 1438668598-6678-1-git-send-email-drinkcat@chromium.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Tue, Aug 04, 2015 at 02:09:56PM +0800, Nicolas Boichat wrote: > Enabling CONFIG_DEBUG_ATOMIC_SLEEP in kernel configuration, we get > this warning in spi_gpio_setup: > [ 1.177747] BUG: sleeping function called from invalid context at drivers/gpio/gpiolib.c:1431 > [ 1.190182] in_atomic(): 1, irqs_disabled(): 128, pid: 1, name: swapper/0 > [ 1.196922] 3 locks held by swapper/0/1: Please don't include entire stack traces in commit logs, they're enormous and overwhelm the actual content (like here where the trace is much bigger than the actual commit message). If you feel the information from the trace adds something please present *edited* highlights instead. > Actually, I'm not sure if I understand the existing code: why are we not > waiting for busy to go down to 0, then call chipselect, instead of not calling > it at all if the bus happens to be busy when we setup the device? With the > current approach, it would be easy to just use an unconditional mutex_lock. We shouldn't be blocking waiting for the bus to become free, that's not how the interface works. > Also, is it harmful to deactivate the newly setup device in spi_bitbang_setup, > even if the bus is busy with another device? chipselect should be independent > for each device (or is it not?). So I'm not clear why we need any locking at > all... If you assert chip select on one device while another device is still being interacted with then the new device will see the traffic for the old device and become confused. > Anyway, this patch series does not change the existing behaviour, applies on > top of broonie-sound/for-next, and, along with the 2 follow-up patches, was > compile-tested on x86-64/arm (allyesconfig) and ppc44x (defconfig+SPI driver), > and runtime-tested on an arm platform. I'm not seeing enough analysis in the commit log of what's being locked and why - I don't fully understand what the busy stuff is for either (not that I've looked at the code in detail just now) but I think that's going to be the key thing here.
On Tue, Aug 4, 2015 at 6:59 PM, Mark Brown <broonie@kernel.org> wrote: > On Tue, Aug 04, 2015 at 02:09:56PM +0800, Nicolas Boichat wrote: [snip] >> Actually, I'm not sure if I understand the existing code: why are we not >> waiting for busy to go down to 0, then call chipselect, instead of not calling >> it at all if the bus happens to be busy when we setup the device? With the >> current approach, it would be easy to just use an unconditional mutex_lock. > > We shouldn't be blocking waiting for the bus to become free, that's not > how the interface works. Noted. >> Also, is it harmful to deactivate the newly setup device in spi_bitbang_setup, >> even if the bus is busy with another device? chipselect should be independent >> for each device (or is it not?). So I'm not clear why we need any locking at >> all... > > If you assert chip select on one device while another device is still > being interacted with then the new device will see the traffic for the > old device and become confused. Here in spi_bitbang_setup, we do: bitbang->chipselect(spi, BITBANG_CS_INACTIVE); That is, we make sure that the current device is _not_ selected. Looking a bit more into this, the issue here is that some drivers use chipselect to chip select individual devices separately, while, on others, the above command unselects all devices (which is a bad idea if the bus is currently transferring data...). So, in short, it's probably better not to touch this code. >> Anyway, this patch series does not change the existing behaviour, applies on >> top of broonie-sound/for-next, and, along with the 2 follow-up patches, was >> compile-tested on x86-64/arm (allyesconfig) and ppc44x (defconfig+SPI driver), >> and runtime-tested on an arm platform. > > I'm not seeing enough analysis in the commit log of what's being locked > and why - I don't fully understand what the busy stuff is for either > (not that I've looked at the code in detail just now) but I think that's > going to be the key thing here. Regarding deleting prepare/unprepare and moving the lock into transfer_one. spi.c:__spi_pump_messages, which calls these functions, does the following: - If no more messages need to be sent, call unprepare_transfer_hardware - If messages need to be sent: + call prepare_transfer_hardware + call prepare_message (NULL for bitbang) + call transfer_one_message I don't think it makes a big difference if we set busy (or hold a mutex) in prepare/unprepare, or we just do it in transfer_one_message, since chipselect is only set in transfer_one_message, and is deselected at the end of the function anyway (in most cases, and if it's not, it will be selected again at the next transfer anyway). Anyway, the "safer" way to fix this would be to keep the prepare/unprepare functions, busy variable, and just protect it with a mutex instead of a spinlock... Thanks. Best, -- To unsubscribe from this list: send the line "unsubscribe linux-spi" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Aug 05, 2015 at 06:24:05PM +0800, Nicolas Boichat wrote: > Anyway, the "safer" way to fix this would be to keep the > prepare/unprepare functions, busy variable, and just protect it with a > mutex instead of a spinlock... OK, that seems reasonable.
diff --git a/drivers/spi/spi-bitbang.c b/drivers/spi/spi-bitbang.c index 840a498..931c37e 100644 --- a/drivers/spi/spi-bitbang.c +++ b/drivers/spi/spi-bitbang.c @@ -180,7 +180,6 @@ int spi_bitbang_setup(struct spi_device *spi) { struct spi_bitbang_cs *cs = spi->controller_state; struct spi_bitbang *bitbang; - unsigned long flags; bitbang = spi_master_get_devdata(spi->master); @@ -210,12 +209,11 @@ int spi_bitbang_setup(struct spi_device *spi) */ /* deselect chip (low or high) */ - spin_lock_irqsave(&bitbang->lock, flags); - if (!bitbang->busy) { + if (mutex_trylock(&bitbang->lock)) { bitbang->chipselect(spi, BITBANG_CS_INACTIVE); ndelay(cs->nsecs); + mutex_unlock(&bitbang->lock); } - spin_unlock_irqrestore(&bitbang->lock, flags); return 0; } @@ -252,20 +250,6 @@ static int spi_bitbang_bufs(struct spi_device *spi, struct spi_transfer *t) * transfer-at-a-time ones to leverage dma or fifo hardware. */ -static int spi_bitbang_prepare_hardware(struct spi_master *spi) -{ - struct spi_bitbang *bitbang; - unsigned long flags; - - bitbang = spi_master_get_devdata(spi); - - spin_lock_irqsave(&bitbang->lock, flags); - bitbang->busy = 1; - spin_unlock_irqrestore(&bitbang->lock, flags); - - return 0; -} - static int spi_bitbang_transfer_one(struct spi_master *master, struct spi_message *m) { @@ -279,6 +263,8 @@ static int spi_bitbang_transfer_one(struct spi_master *master, bitbang = spi_master_get_devdata(master); + mutex_lock(&bitbang->lock); + /* FIXME this is made-up ... the correct value is known to * word-at-a-time bitbang code, and presumably chipselect() * should enforce these requirements too? @@ -372,21 +358,9 @@ static int spi_bitbang_transfer_one(struct spi_master *master, spi_finalize_current_message(master); - return status; -} - -static int spi_bitbang_unprepare_hardware(struct spi_master *spi) -{ - struct spi_bitbang *bitbang; - unsigned long flags; + mutex_unlock(&bitbang->lock); - bitbang = spi_master_get_devdata(spi); - - spin_lock_irqsave(&bitbang->lock, flags); - bitbang->busy = 0; - spin_unlock_irqrestore(&bitbang->lock, flags); - - return 0; + return status; } /*----------------------------------------------------------------------*/ @@ -427,7 +401,7 @@ int spi_bitbang_start(struct spi_bitbang *bitbang) if (!master || !bitbang->chipselect) return -EINVAL; - spin_lock_init(&bitbang->lock); + mutex_init(&bitbang->lock); if (!master->mode_bits) master->mode_bits = SPI_CPOL | SPI_CPHA | bitbang->flags; @@ -435,8 +409,6 @@ int spi_bitbang_start(struct spi_bitbang *bitbang) if (master->transfer || master->transfer_one_message) return -EINVAL; - master->prepare_transfer_hardware = spi_bitbang_prepare_hardware; - master->unprepare_transfer_hardware = spi_bitbang_unprepare_hardware; master->transfer_one_message = spi_bitbang_transfer_one; if (!bitbang->txrx_bufs) { diff --git a/include/linux/spi/spi_bitbang.h b/include/linux/spi/spi_bitbang.h index 85578d4..1329053 100644 --- a/include/linux/spi/spi_bitbang.h +++ b/include/linux/spi/spi_bitbang.h @@ -4,8 +4,7 @@ #include <linux/workqueue.h> struct spi_bitbang { - spinlock_t lock; - u8 busy; + struct mutex lock; u8 use_dma; u8 flags; /* extra spi->mode support */
Enabling CONFIG_DEBUG_ATOMIC_SLEEP in kernel configuration, we get this warning in spi_gpio_setup: [ 1.177747] BUG: sleeping function called from invalid context at drivers/gpio/gpiolib.c:1431 [ 1.190182] in_atomic(): 1, irqs_disabled(): 128, pid: 1, name: swapper/0 [ 1.196922] 3 locks held by swapper/0/1: [ 1.200812] #0: (&dev->mutex){......}, at: [<ffffffc00051db78>] __driver_attach+0x58/0x98 [ 1.209147] #1: (spi_add_lock){+.+.+.}, at: [<ffffffc000578164>] spi_add_device+0x80/0x14c [ 1.217564] #2: (&(&bitbang->lock)->rlock){......}, at: [<ffffffc00057b054>] spi_bitbang_setup+0x84/0xc4 [ 1.227185] irq event stamp: 279856 [ 1.230645] hardirqs last enabled at (279855): [<ffffffc0007cce70>] __mutex_unlock_slowpath+0x158/0x16c [ 1.240070] hardirqs last disabled at (279856): [<ffffffc0007cead0>] _raw_spin_lock_irqsave+0x20/0x6c [ 1.249233] softirqs last enabled at (262072): [<ffffffc0002f0954>] bdi_register+0x124/0x1d0 [ 1.257707] softirqs last disabled at (262070): [<ffffffc0002f0930>] bdi_register+0x100/0x1d0 [ 1.266185] CPU: 3 PID: 1 Comm: swapper/0 Not tainted 3.18.0 #608 [ 1.277419] Call trace: [ 1.279848] [<ffffffc000208754>] dump_backtrace+0x0/0x12c [ 1.285209] [<ffffffc000208890>] show_stack+0x10/0x1c [ 1.290223] [<ffffffc0007c897c>] dump_stack+0x80/0xb4 [ 1.295238] [<ffffffc000244248>] __might_sleep+0x110/0x11c [ 1.300687] [<ffffffc0004722d4>] gpiod_set_raw_value_cansleep+0x24/0x4c [ 1.307255] [<ffffffc00057b464>] spi_gpio_chipselect+0x74/0x88 [ 1.313045] [<ffffffc00057b068>] spi_bitbang_setup+0x98/0xc4 [ 1.318664] [<ffffffc00057bb44>] spi_gpio_setup+0x50/0xc8 [ 1.324022] [<ffffffc0005780d0>] spi_setup+0xe4/0xf8 [ 1.328950] [<ffffffc0005781b4>] spi_add_device+0xd0/0x14c [ 1.334396] [<ffffffc000579ba8>] spi_register_master+0x6a8/0x718 [ 1.340359] [<ffffffc00057b358>] spi_bitbang_start+0xe8/0x108 [ 1.346064] [<ffffffc00057bf70>] spi_gpio_probe+0x3b4/0x448 [ 1.351595] [<ffffffc00051f65c>] platform_drv_probe+0x4c/0x9c [ 1.357301] [<ffffffc00051d930>] driver_probe_device+0xd4/0x23c [ 1.363180] [<ffffffc00051db88>] __driver_attach+0x68/0x98 [ 1.368627] [<ffffffc00051ca30>] bus_for_each_dev+0x7c/0xb0 [ 1.374160] [<ffffffc00051d4bc>] driver_attach+0x1c/0x28 [ 1.379434] [<ffffffc00051d07c>] bus_add_driver+0xd8/0x1e0 [ 1.384881] [<ffffffc00051e654>] driver_register+0xbc/0x10c [ 1.390412] [<ffffffc00051f58c>] __platform_driver_register+0x5c/0x68 [ 1.396808] [<ffffffc000c28584>] spi_gpio_driver_init+0x14/0x20 [ 1.402685] [<ffffffc000200a04>] do_one_initcall+0x18c/0x1ac [ 1.408306] [<ffffffc000c00c64>] kernel_init_freeable+0x228/0x2e0 [ 1.414356] [<ffffffc0007c5f60>] kernel_init+0x10/0xd8 chipselect (in this case, spi_gpio_chipselect, which calls gpiod_set_raw_value_cansleep), can sleep, so we should not hold a spinlock while calling it. This issue was introduced by this commit, which converted spi-gpio to cansleep variants: d9dda5a191 "spi: spi-gpio: Use 'cansleep' variants to access GPIO" Replace spinlock + busy variable by a mutex, and get rid of spi_bitbang_prepare_hardware and spi_bitbang_unprepare_hardware, which are not useful anymore. Signed-off-by: Nicolas Boichat <drinkcat@chromium.org> --- Actually, I'm not sure if I understand the existing code: why are we not waiting for busy to go down to 0, then call chipselect, instead of not calling it at all if the bus happens to be busy when we setup the device? With the current approach, it would be easy to just use an unconditional mutex_lock. Also, is it harmful to deactivate the newly setup device in spi_bitbang_setup, even if the bus is busy with another device? chipselect should be independent for each device (or is it not?). So I'm not clear why we need any locking at all... Hopefully someone can shine some light on this... Anyway, this patch series does not change the existing behaviour, applies on top of broonie-sound/for-next, and, along with the 2 follow-up patches, was compile-tested on x86-64/arm (allyesconfig) and ppc44x (defconfig+SPI driver), and runtime-tested on an arm platform. drivers/spi/spi-bitbang.c | 42 +++++++---------------------------------- include/linux/spi/spi_bitbang.h | 3 +-- 2 files changed, 8 insertions(+), 37 deletions(-)