On Mon, Jan 28, 2019 at 08:05:13AM -0800, Doug Anderson wrote:
Hi, On Mon, Jan 28, 2019 at 7:44 AM Sasha Levin <sashal@xxxxxxxxxx> wrote:From: Douglas Anderson <dianders@xxxxxxxxxxxx> [ Upstream commit d6e1935819db0c91ce4a5af82466f3ab50d17346 ] Right now serial drivers process sysrq keys deep in their character receiving code. This means that they've already grabbed their port->lock spinlock. This can end up getting in the way if we've go to do serial stuff (especially kgdb) in response to the sysrq. Serial drivers have various hacks in them to handle this. Looking at '8250_port.c' you can see that the console_write() skips locking if we're in the sysrq handler. Looking at 'msm_serial.c' you can see that the port lock is dropped around uart_handle_sysrq_char(). It turns out that these hacks aren't exactly perfect. If you have lockdep turned on and use something like the 8250_port hack you'll get a splat that looks like: WARNING: possible circular locking dependency detected [...] is trying to acquire lock: ... (console_owner){-.-.}, at: console_unlock+0x2e0/0x5e4 but task is already holding lock: ... (&port_lock_key){-.-.}, at: serial8250_handle_irq+0x30/0xe4 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #1 (&port_lock_key){-.-.}: _raw_spin_lock_irqsave+0x58/0x70 serial8250_console_write+0xa8/0x250 univ8250_console_write+0x40/0x4c console_unlock+0x528/0x5e4 register_console+0x2c4/0x3b0 uart_add_one_port+0x350/0x478 serial8250_register_8250_port+0x350/0x3a8 dw8250_probe+0x67c/0x754 platform_drv_probe+0x58/0xa4 really_probe+0x150/0x294 driver_probe_device+0xac/0xe8 __driver_attach+0x98/0xd0 bus_for_each_dev+0x84/0xc8 driver_attach+0x2c/0x34 bus_add_driver+0xf0/0x1ec driver_register+0xb4/0x100 __platform_driver_register+0x60/0x6c dw8250_platform_driver_init+0x20/0x28 ... -> #0 (console_owner){-.-.}: lock_acquire+0x1e8/0x214 console_unlock+0x35c/0x5e4 vprintk_emit+0x230/0x274 vprintk_default+0x7c/0x84 vprintk_func+0x190/0x1bc printk+0x80/0xa0 __handle_sysrq+0x104/0x21c handle_sysrq+0x30/0x3c serial8250_read_char+0x15c/0x18c serial8250_rx_chars+0x34/0x74 serial8250_handle_irq+0x9c/0xe4 dw8250_handle_irq+0x98/0xcc serial8250_interrupt+0x50/0xe8 ... other info that might help us debug this: Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&port_lock_key); lock(console_owner); lock(&port_lock_key); lock(console_owner); *** DEADLOCK *** The hack used in 'msm_serial.c' doesn't cause the above splats but it seems a bit ugly to unlock / lock our spinlock deep in our irq handler. It seems like we could defer processing the sysrq until the end of the interrupt handler right after we've unlocked the port. With this scheme if a whole batch of sysrq characters comes in one irq then we won't handle them all, but that seems like it should be a fine compromise. Signed-off-by: Douglas Anderson <dianders@xxxxxxxxxxxx> Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx> --- include/linux/serial_core.h | 37 ++++++++++++++++++++++++++++++++++++- 1 file changed, 36 insertions(+), 1 deletion(-)FWIW this patch shouldn't hurt to be backported (I haven't heard any problems report with it), but it is effectively a no-op unless you also pick a patch that uses the new API. For instance commit 596f63da42b9 ("serial: 8250: Process sysrq at port unlock time"). ...and if you want that patch I think you also need commit 3e6f88068314 ("serial: core: Include console.h from serial_core.h"). In theory you could think about adding the "qcom_geni_serial" patches related to sysrq processing too--dunno if anyone really cares about those on 4.20 stable...
Since no one actually tagged it for stable, probably not... I'll drop it, thanks! -- Thanks, Sasha
