[dropped fugang.duan@nxp.com, mail bounces with 550 5.4.1 Recipient
address rejected: Access denied]

Am 2021-05-12 12:07, schrieb Johan Hovold:

On Wed, May 12, 2021 at 11:46:28AM +0200, Michael Walle wrote:

quoted

Am 2021-05-12 11:30, schrieb Johan Hovold:

quoted

On Tue, May 11, 2021 at 10:01:44PM +0200, Michael Walle wrote:

quoted

Although there is already (broken) sysrq characters handling, a break
condition was never detected. There is also a possible deadlock
because
we might call handle_sysrq() while still holding the port lock.

Where's the possible deadlock?

[   17.866874] ======================================================
[   17.866876] WARNING: possible circular locking dependency detected
[   17.866878] 5.13.0-rc1-next-20210511+ #555 Not tainted
[   17.866880] ------------------------------------------------------
[   17.866882] sl28-variant.sh/1934 is trying to acquire lock:
[   17.866884] ffff800011d16a00 (console_owner){-.-.}-{0:0}, at:
console_unlock+0x1c0/0x660
[   17.866892]
[   17.866893] but task is already holding lock:
[   17.866895] ffff0020026ea098 (&port_lock_key){-.-.}-{2:2}, at:
lpuart32_int+0x1b0/0x7c8
[   17.866902]
[   17.866904] which lock already depends on the new lock.
[   17.866906]
[   17.866907]
[   17.866909] the existing dependency chain (in reverse order) is:
[   17.866910]
[   17.866912] -> #1 (&port_lock_key){-.-.}-{2:2}:
[   17.866918]        _raw_spin_lock_irqsave+0x80/0xd0
[   17.866920]        lpuart32_console_write+0x214/0x2b8
[   17.866922]        console_unlock+0x404/0x660
[   17.866924]        register_console+0x170/0x2a8
[   17.866925]        uart_add_one_port+0x464/0x478
[   17.866927]        lpuart_probe+0x218/0x3a8
[   17.866928]        platform_probe+0x70/0xe0
[   17.866930]        really_probe+0xec/0x3c0
[   17.866931]        driver_probe_device+0x6c/0xd0
[   17.866933]        device_driver_attach+0x7c/0x88
[   17.866935]        __driver_attach+0x6c/0xf8
[   17.866936]        bus_for_each_dev+0x7c/0xd0
[   17.866938]        driver_attach+0x2c/0x38
[   17.866939]        bus_add_driver+0x194/0x1f8
[   17.866941]        driver_register+0x6c/0x128
[   17.866943]        __platform_driver_register+0x30/0x40
[   17.866944]        lpuart_serial_init+0x44/0x6c
[   17.866946]        do_one_initcall+0x90/0x470
[   17.866948]        kernel_init_freeable+0x2d4/0x344
[   17.866949]        kernel_init+0x1c/0x120
[   17.866951]        ret_from_fork+0x10/0x30
[   17.866952]
[   17.866953] -> #0 (console_owner){-.-.}-{0:0}:
[   17.866959]        __lock_acquire+0xf60/0x17e8
[   17.866961]        lock_acquire+0x138/0x4c0
[   17.866963]        console_unlock+0x224/0x660
[   17.866964]        vprintk_emit+0x11c/0x338
[   17.866966]        vprintk_default+0x40/0x50
[   17.866967]        vprintk+0xfc/0x320
[   17.866969]        printk+0x6c/0x90
[   17.866970]        __handle_sysrq+0x16c/0x1d8
[   17.866972]        handle_sysrq+0x2c/0x48
[   17.866973]        lpuart32_int+0x70c/0x7c8
[   17.866975]        __handle_irq_event_percpu+0xcc/0x430
[   17.866977]        handle_irq_event_percpu+0x40/0x98
[   17.866978]        handle_irq_event+0x50/0x100
[   17.866980]        handle_fasteoi_irq+0xc0/0x178
[   17.866981]        generic_handle_irq+0x38/0x50
[   17.866983]        __handle_domain_irq+0x6c/0xc8
[   17.866985]        gic_handle_irq+0xdc/0x340
[   17.866986]        el1_irq+0xb8/0x150
[   17.866988]        arch_local_irq_restore+0x8/0x20
[   17.866989]        page_add_file_rmap+0x24/0x1f8
[   17.866991]        do_set_pte+0xd4/0x1a0
[   17.866992]        filemap_map_pages+0x358/0x590
[   17.866994]        __handle_mm_fault+0xbc0/0xdd0
[   17.866995]        handle_mm_fault+0x170/0x3e0
[   17.866997]        do_page_fault+0x1e8/0x448
[   17.866998]        do_translation_fault+0x60/0x70
[   17.867000]        do_mem_abort+0x48/0xb8
[   17.867001]        el0_da+0x44/0x80
[   17.867002]        el0_sync_handler+0x68/0xb8
[   17.867004]        el0_sync+0x178/0x180
[   17.867005]
[   17.867007] other info that might help us debug this:
[   17.867008]
[   17.867009]  Possible unsafe locking scenario:
[   17.867011]
[   17.867012]        CPU0                    CPU1
[   17.867013]        ----                    ----
[   17.867015]   lock(&port_lock_key);
[   17.867019]                                lock(console_owner);
[   17.867023]                                lock(&port_lock_key);
[   17.867027]   lock(console_owner);
[   17.867030]
[   17.867031]  *** DEADLOCK ***
[   17.867033]
[   17.867034] 7 locks held by sl28-variant.sh/1934:
[   17.867035]  #0: ffff002003a37b08 (&mm->mmap_lock){++++}-{3:3}, at:
do_page_fault+0x180/0x448
[   17.867043]  #1: ffff800011d87660 (rcu_read_lock){....}-{1:2}, at:
filemap_map_pages+0x8/0x590
[   17.867051]  #2: ffff0020048d8318 (ptlock_ptr(page)){+.+.}-{2:2}, 
at:
filemap_map_pages+0x27c/0x590
[   17.867059]  #3: ffff800011d87660 (rcu_read_lock){....}-{1:2}, at:
lock_page_memcg+0x8/0x1d8
[   17.867067]  #4: ffff0020026ea098 (&port_lock_key){-.-.}-{2:2}, at:
lpuart32_int+0x1b0/0x7c8
[   17.867074]  #5: ffff800011d87660 (rcu_read_lock){....}-{1:2}, at:
__handle_sysrq+0x8/0x1d8
[   17.867082]  #6: ffff800011d168a0 (console_lock){+.+.}-{0:0}, at:
vprintk_emit+0x114/0x338

Note that it says "possible" deadlock; the lockdep validator probably
isn't smart enough to understand the trylock hack in the console write
callback.

quoted

quoted

First, as you point out above the driver currently doesn't detect
breaks
so the sysrq handler is never called and there's no risk for deadlocks
in the console code.

But this commit introduces it? Therefore, I don't get your point.

My point is that your commit message makes it sound like an actual
deadlock in the current code. Something which, for example, can cause
commits to get backported to stable when it is not needed.

I see. I'll rephrase the commit message. FWIW I intentionally didn't
put a Fixes: tag here, because of that.

quoted

quoted

Second, the driver's console implementation explicitly handles being
called recursively so would not deadlock after you start detecting
breaks either.

See above. Or there is something wrong with the lock debugging.

Seems to work as intended.

quoted

quoted

quoted

Add support for break detection and use the proper
uart_unlock_and_check_sysrq() to defer calling handle_sysrq().

But you should get rid of the sysrq trylock hack when switching to
uart_unlock_and_check_sysrq().

Ok. But only for the sport->port.sysrq part right? We'll still
need it for oops_in_progress.

Thanks for reviewing,
-michael