On Thu, 3 Dec 2020 18:53:20 +0100 Oleksij Rempel wrote: > On Thu, Dec 03, 2020 at 08:35:17AM -0800, Jakub Kicinski wrote: > > On Thu, 3 Dec 2020 09:50:11 +0100 Oleksij Rempel wrote: > > > @Jakub, > > > > > > > You can't take sleeping locks from .ndo_get_stats64. > > > > > > > > Also regmap may sleep? > > > > > > > > + ret = regmap_read(priv->regmap, reg, &val); > > > > > > Yes. And underling layer is mdio bus which is by default sleeping as > > > well. > > > > > > > Am I missing something? > > > > > > In this log, the ar9331_get_stats64() was never called from atomic or > > > irq context. Why it should not be sleeping? > > > > You missed some long discussions about this within last week on netdev. > > Also Documentation/networking/statistics.rst. > > > > To answer your direct question - try: > > > > # cat /proc/net/dev > > > > procfs iterates over devices while holding only an RCU read lock. > > Now i can reproduce it :) > > [33683.199864] BUG: sleeping function called from invalid context at kernel/locking/mutex.c:935 > [33683.210737] in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid: 593, name: cat > [33683.216796] INFO: lockdep is turned off. > [33683.222972] CPU: 0 PID: 593 Comm: cat Not tainted 5.10.0-rc3-ar9331-00733-gff7090915bb7-dirty #28 > [33683.231743] Stack : 808f0000 80885ffc 820eba5c 00000000 00000000 d4a19200 80980000 819a93c8 > [33683.240093] 80980ca7 80d43358 804ee1f4 80980000 00000002 800afe08 820eba08 d4a19200 > [33683.247181] 00000000 00000000 8089ffb0 00000000 820ebfe0 00000000 00000000 00000000 > [33683.257767] 820ebab4 77bbfdc0 00fae587 77e859a0 80980000 80000000 00000000 80990000 > [33683.266107] 804ee1f4 80980000 00000002 8200f750 8097ca9c d4a19200 000859df 00000001 > [33683.274529] ... > [33683.275626] Call Trace: > [33683.280156] [<80069ce0>] show_stack+0x9c/0x140 > [33683.283200] [<800afe08>] ___might_sleep+0x220/0x244 > [33683.290441] [<8073c030>] __mutex_lock+0x70/0x374 > [33683.293651] [<8073c360>] mutex_lock_nested+0x2c/0x38 > [33683.300793] [<804ee1f4>] ar9331_read_stats+0x34/0x834 > [33683.304441] [<804eea34>] ar9331_get_stats64+0x40/0x394 > [33683.311797] [<80526584>] dev_get_stats+0x58/0xfc > [33683.315013] [<805657bc>] dev_seq_printf_stats+0x44/0x228 > [33683.322476] [<805659e8>] dev_seq_show+0x48/0x50 > [33683.325601] [<8021dd28>] seq_read_iter+0x3d8/0x4d0 > [33683.332585] [<8021df60>] seq_read+0x140/0x198 > [33683.335532] [<8026f950>] proc_reg_read+0xe4/0xf8 > [33683.342397] [<801f0840>] vfs_read+0xc8/0x1a8 > [33683.345260] [<801f0b7c>] ksys_read+0x9c/0xfc > [33683.352056] [<80071aa4>] syscall_common+0x34/0x58 > > Hm.. There is no way i can guarantee that underlying mdio system is > not using mutexes. So, i can't read stats directly from HW within > ar9331_get_stats64(), only driver internal storage can be used. It is possible > to poll it more frequently, but it make no reals sense on this low power > devices. > > What kind of options do we have? Vladimir has been looking at solving this, I'll let him answer with his latest thoughts.
