netconsole deadlock with virtnet
Subsystems (3)
Hi,
Approximately two weeks ago, our regression team started to experience those netconsole splats. The tested code is Linus's master (-rc4) + netdev net-next
+ netdev net-rc.
Such splats are random and we can't bisect because there is no stable reproducer.
Any idea, what is the root cause?
[ 21.141117] netpoll: netconsole: local port 6666
[ 21.141947] netpoll: netconsole: local IPv4 address 10.236.251.7
[ 21.143138] netpoll: netconsole: interface 'eth1'
[ 21.143988] netpoll: netconsole: remote port 62517
[ 21.145127] netpoll: netconsole: remote IPv4 address 10.236.251.1
[ 21.146206] netpoll: netconsole: remote ethernet address 68:05:ca:aa:99:49
[ 21.149464] =====================================================
[ 21.149466] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected
[ 21.149467] 5.10.0-rc4_for_upstream_min_debug_2020_11_16_13_03 #1 Not tainted
[ 21.149468] -----------------------------------------------------
[ 21.149469] modprobe/596 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire:
[ 21.149471] ffff000194b4e298 (_xmit_ETHER#2){+.-.}-{2:2}, at: virtnet_poll_tx+0x84/0x120
[ 21.149478] and this task is already holding:
[ 21.149478] ffff8000093d4018 (target_list_lock){....}-{2:2}, at: write_msg+0x6c/0x120 [netconsole]
[ 21.149483] which would create a new lock dependency:
[ 21.149484] (target_list_lock){....}-{2:2} -> (_xmit_ETHER#2){+.-.}-{2:2}
[ 21.149491] but this new dependency connects a HARDIRQ-irq-safe lock:
[ 21.149492] (console_owner){-.-.}-{0:0}
[ 21.149496] ... which became HARDIRQ-irq-safe at:
[ 21.149496] __lock_acquire+0xa78/0x1a94
[ 21.149498] lock_acquire.part.0+0x170/0x360
[ 21.149498] lock_acquire+0x68/0x8c
[ 21.149500] console_unlock+0x1e8/0x6a4
[ 21.149500] vprintk_emit+0x1c4/0x3c4
[ 21.149501] vprintk_default+0x40/0x4c
[ 21.149502] vprintk_func+0x10c/0x220
[ 21.149503] printk+0x68/0x90
[ 21.149504] crng_fast_load+0x1bc/0x1c0
[ 21.149505] add_interrupt_randomness+0x280/0x290
[ 21.149506] handle_irq_event+0x80/0x120
[ 21.149507] handle_fasteoi_irq+0xac/0x200
[ 21.149508] __handle_domain_irq+0x84/0xf0
[ 21.149509] gic_handle_irq+0xd4/0x320
[ 21.149510] el1_irq+0xd0/0x180
[ 21.149511] arch_cpu_idle+0x24/0x44
[ 21.149512] default_idle_call+0x48/0xa0
[ 21.149513] do_idle+0x260/0x300
[ 21.149514] cpu_startup_entry+0x30/0x6c
[ 21.149515] rest_init+0x1b4/0x288
[ 21.149515] arch_call_rest_init+0x18/0x24
[ 21.149516] start_kernel+0x5cc/0x608
[ 21.149518] to a HARDIRQ-irq-unsafe lock:
[ 21.149519] (_xmit_ETHER#2){+.-.}-{2:2}
[ 21.149523] ... which became HARDIRQ-irq-unsafe at:
[ 21.149524] ... __lock_acquire+0x8bc/0x1a94
[ 21.149525] lock_acquire.part.0+0x170/0x360
[ 21.149526] lock_acquire+0x68/0x8c
[ 21.149527] _raw_spin_trylock+0x80/0xd0
[ 21.149527] virtnet_poll+0xac/0x360
[ 21.149528] net_rx_action+0x1b0/0x4e0
[ 21.149529] __do_softirq+0x1f4/0x638
[ 21.149530] do_softirq+0xb8/0xcc
[ 21.149531] __local_bh_enable_ip+0x18c/0x200
[ 21.149532] virtnet_napi_enable+0xc0/0xd4
[ 21.149533] virtnet_open+0x98/0x1c0
[ 21.149534] __dev_open+0x12c/0x200
[ 21.149535] __dev_change_flags+0x1a0/0x220
[ 21.149536] dev_change_flags+0x2c/0x70
[ 21.149536] do_setlink+0x214/0xe20
[ 21.149537] __rtnl_newlink+0x514/0x820
[ 21.149538] rtnl_newlink+0x58/0x84
[ 21.149539] rtnetlink_rcv_msg+0x184/0x4b4
[ 21.149540] netlink_rcv_skb+0x60/0x124
[ 21.149541] rtnetlink_rcv+0x20/0x30
[ 21.149542] netlink_unicast+0x1b4/0x270
[ 21.149543] netlink_sendmsg+0x1f0/0x400
[ 21.149544] sock_sendmsg+0x5c/0x70
[ 21.149545] ____sys_sendmsg+0x24c/0x280
[ 21.149545] ___sys_sendmsg+0x88/0xd0
[ 21.149546] __sys_sendmsg+0x70/0xd0
[ 21.149547] __arm64_sys_sendmsg+0x2c/0x40
[ 21.149548] el0_svc_common.constprop.0+0x84/0x200
[ 21.149549] do_el0_svc+0x2c/0x90
[ 21.149550] el0_svc+0x18/0x50
[ 21.149551] el0_sync_handler+0xe0/0x350
[ 21.149552] el0_sync+0x158/0x180
[ 21.149553] other info that might help us debug this:
[ 21.149555] Chain exists of:
[ 21.149556] console_owner --> target_list_lock --> _xmit_ETHER#2
[ 21.149563] Possible interrupt unsafe locking scenario:
[ 21.149564] CPU0 CPU1
[ 21.149565] ---- ----
[ 21.149566] lock(_xmit_ETHER#2);
[ 21.149569] local_irq_disable();
[ 21.149570] lock(console_owner);
[ 21.149572] lock(target_list_lock);
[ 21.149575] <Interrupt>
[ 21.149576] lock(console_owner);
[ 21.149579] *** DEADLOCK ***
[ 21.149580] 3 locks held by modprobe/596:
[ 21.149581] #0: ffff800011a1efe0 (console_lock){+.+.}-{0:0}, at: register_console+0x144/0x2f4
[ 21.149586] #1: ffff800011a1f108 (console_owner){-.-.}-{0:0}, at: console_unlock+0x17c/0x6a4
[ 21.149590] #2: ffff8000093d4018 (target_list_lock){....}-{2:2}, at: write_msg+0x6c/0x120 [netconsole]
[ 21.149596] the dependencies between HARDIRQ-irq-safe lock and the holding lock:
[ 21.149596] -> (console_owner){-.-.}-{0:0} ops: 1191 {
[ 21.149601] IN-HARDIRQ-W at:
[ 21.149602] __lock_acquire+0xa78/0x1a94
[ 21.149603] lock_acquire.part.0+0x170/0x360
[ 21.149604] lock_acquire+0x68/0x8c
[ 21.149605] console_unlock+0x1e8/0x6a4
[ 21.149606] vprintk_emit+0x1c4/0x3c4
[ 21.149607] vprintk_default+0x40/0x4c
[ 21.149608] vprintk_func+0x10c/0x220
[ 21.149610] printk+0x68/0x90
[ 21.149611] crng_fast_load+0x1bc/0x1c0
[ 21.149612] add_interrupt_randomness+0x280/0x290
[ 21.149613] handle_irq_event+0x80/0x120
[ 21.149614] handle_fasteoi_irq+0xac/0x200
[ 21.149615] __handle_domain_irq+0x84/0xf0
[ 21.149616] gic_handle_irq+0xd4/0x320
[ 21.149617] el1_irq+0xd0/0x180
[ 21.149618] arch_cpu_idle+0x24/0x44
[ 21.149619] default_idle_call+0x48/0xa0
[ 21.149620] do_idle+0x260/0x300
[ 21.149621] cpu_startup_entry+0x30/0x6c
[ 21.149622] rest_init+0x1b4/0x288
[ 21.149624] arch_call_rest_init+0x18/0x24
[ 21.149625] start_kernel+0x5cc/0x608
[ 21.149625] IN-SOFTIRQ-W at:
[ 21.149627] __lock_acquire+0x894/0x1a94
[ 21.149628] lock_acquire.part.0+0x170/0x360
[ 21.149629] lock_acquire+0x68/0x8c
[ 21.149630] console_unlock+0x1e8/0x6a4
[ 21.149631] vprintk_emit+0x1c4/0x3c4
[ 21.149632] vprintk_default+0x40/0x4c
[ 21.149633] vprintk_func+0x10c/0x220
[ 21.149634] printk+0x68/0x90
[ 21.149635] hrtimer_interrupt+0x290/0x294
[ 21.149636] arch_timer_handler_virt+0x3c/0x50
[ 21.149637] handle_percpu_devid_irq+0x94/0x164
[ 21.149673] __handle_domain_irq+0x84/0xf0
[ 21.149674] gic_handle_irq+0xd4/0x320
[ 21.149675] el1_irq+0xd0/0x180
[ 21.149676] __do_softirq+0x108/0x638
[ 21.149677] __irq_exit_rcu+0x17c/0x1b0
[ 21.149678] irq_exit+0x18/0x44
[ 21.149679] __handle_domain_irq+0x88/0xf0
[ 21.149680] gic_handle_irq+0xd4/0x320
[ 21.149681] el1_irq+0xd0/0x180
[ 21.149682] smp_call_function_many_cond+0x3cc/0x3f0
[ 21.149683] kick_all_cpus_sync+0x4c/0x80
[ 21.149684] load_module+0x1eec/0x2734
[ 21.149685] __do_sys_finit_module+0xbc/0x12c
[ 21.149686] __arm64_sys_finit_module+0x28/0x34
[ 21.149687] el0_svc_common.constprop.0+0x84/0x200
[ 21.149688] do_el0_svc+0x2c/0x90
[ 21.149689] el0_svc+0x18/0x50
[ 21.149690] el0_sync_handler+0xe0/0x350
[ 21.149691] el0_sync+0x158/0x180
[ 21.149692] INITIAL USE at:
[ 21.149694] __lock_acquire+0x2e0/0x1a94
[ 21.149695] lock_acquire.part.0+0x170/0x360
[ 21.149696] lock_acquire+0x68/0x8c
[ 21.149697] console_unlock+0x1e8/0x6a4
[ 21.149698] vprintk_emit+0x1c4/0x3c4
[ 21.149699] vprintk_default+0x40/0x4c
[ 21.149700] vprintk_func+0x10c/0x220
[ 21.149701] printk+0x68/0x90
[ 21.149702] start_kernel+0x8c/0x608
[ 21.149703] }
[ 21.149704] ... key at: [<ffff800011a1f108>] console_owner_dep_map+0x0/0x28
[ 21.149705] ... acquired at:
[ 21.149706] lock_acquire.part.0+0x170/0x360
[ 21.149707] lock_acquire+0x68/0x8c
[ 21.149708] _raw_spin_lock_irqsave+0x88/0x15c
[ 21.149709] write_msg+0x6c/0x120 [netconsole]
[ 21.149710] console_unlock+0x3ec/0x6a4
[ 21.149711] register_console+0x17c/0x2f4
[ 21.149712] init_netconsole+0x20c/0x1000 [netconsole]
[ 21.149713] do_one_initcall+0x8c/0x480
[ 21.149714] do_init_module+0x60/0x270
[ 21.149715] load_module+0x21f8/0x2734
[ 21.149716] __do_sys_finit_module+0xbc/0x12c
[ 21.149717] __arm64_sys_finit_module+0x28/0x34
[ 21.149718] el0_svc_common.constprop.0+0x84/0x200
[ 21.149719] do_el0_svc+0x2c/0x90
[ 21.149720] el0_svc+0x18/0x50
[ 21.149722] el0_sync_handler+0xe0/0x350
[ 21.149723] el0_sync+0x158/0x180
[ 21.149724] -> (target_list_lock){....}-{2:2} ops: 4 {
[ 21.149729] INITIAL USE at:
[ 21.149730] __lock_acquire+0x2e0/0x1a94
[ 21.149731] lock_acquire.part.0+0x170/0x360
[ 21.149732] lock_acquire+0x68/0x8c
[ 21.149734] _raw_spin_lock_irqsave+0x88/0x15c
[ 21.149735] init_netconsole+0x148/0x1000 [netconsole]
[ 21.149736] do_one_initcall+0x8c/0x480
[ 21.149737] do_init_module+0x60/0x270
[ 21.149738] load_module+0x21f8/0x2734
[ 21.149739] __do_sys_finit_module+0xbc/0x12c
[ 21.149740] __arm64_sys_finit_module+0x28/0x34
[ 21.149741] el0_svc_common.constprop.0+0x84/0x200
[ 21.149742] do_el0_svc+0x2c/0x90
[ 21.149743] el0_svc+0x18/0x50
[ 21.149744] el0_sync_handler+0xe0/0x350
[ 21.149745] el0_sync+0x158/0x180
[ 21.149746] }
[ 21.149747] ... key at: [<ffff8000093d4018>] target_list_lock+0x18/0xfffffffffffff000 [netconsole]
[ 21.149748] ..
[ 21.149750] Lost 190 message(s)!_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 81 lines · On Tue, 17 Nov 2020 12:23:41 +0200 Leon Romanovsky <[email protected]> wrote:
Hi,
Approximately two weeks ago, our regression team started to experience those netconsole splats. The tested code is Linus's master (-rc4) + netdev net-next
+ netdev net-rc.Such splats are random and we can't bisect because there is no stable reproducer.
Any idea, what is the root cause?
[ 21.141117] netpoll: netconsole: local port 6666 [ 21.141947] netpoll: netconsole: local IPv4 address 10.236.251.7 [ 21.143138] netpoll: netconsole: interface 'eth1' [ 21.143988] netpoll: netconsole: remote port 62517 [ 21.145127] netpoll: netconsole: remote IPv4 address 10.236.251.1 [ 21.146206] netpoll: netconsole: remote ethernet address 68:05:ca:aa:99:49 [ 21.149464] ===================================================== [ 21.149466] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected [ 21.149467] 5.10.0-rc4_for_upstream_min_debug_2020_11_16_13_03 #1 Not tainted [ 21.149468] ----------------------------------------------------- [ 21.149469] modprobe/596 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire: [ 21.149471] ffff000194b4e298 (_xmit_ETHER#2){+.-.}-{2:2}, at: virtnet_poll_tx+0x84/0x120 [ 21.149478] and this task is already holding: [ 21.149478] ffff8000093d4018 (target_list_lock){....}-{2:2}, at: write_msg+0x6c/0x120 [netconsole] [ 21.149483] which would create a new lock dependency: [ 21.149484] (target_list_lock){....}-{2:2} -> (_xmit_ETHER#2){+.-.}-{2:2} [ 21.149491] but this new dependency connects a HARDIRQ-irq-safe lock: [ 21.149492] (console_owner){-.-.}-{0:0} [ 21.149496] ... which became HARDIRQ-irq-safe at: [ 21.149496] __lock_acquire+0xa78/0x1a94 [ 21.149498] lock_acquire.part.0+0x170/0x360 [ 21.149498] lock_acquire+0x68/0x8c [ 21.149500] console_unlock+0x1e8/0x6a4 [ 21.149500] vprintk_emit+0x1c4/0x3c4 [ 21.149501] vprintk_default+0x40/0x4c [ 21.149502] vprintk_func+0x10c/0x220 [ 21.149503] printk+0x68/0x90 [ 21.149504] crng_fast_load+0x1bc/0x1c0 [ 21.149505] add_interrupt_randomness+0x280/0x290 [ 21.149506] handle_irq_event+0x80/0x120 [ 21.149507] handle_fasteoi_irq+0xac/0x200 [ 21.149508] __handle_domain_irq+0x84/0xf0 [ 21.149509] gic_handle_irq+0xd4/0x320 [ 21.149510] el1_irq+0xd0/0x180 [ 21.149511] arch_cpu_idle+0x24/0x44 [ 21.149512] default_idle_call+0x48/0xa0 [ 21.149513] do_idle+0x260/0x300 [ 21.149514] cpu_startup_entry+0x30/0x6c [ 21.149515] rest_init+0x1b4/0x288 [ 21.149515] arch_call_rest_init+0x18/0x24 [ 21.149516] start_kernel+0x5cc/0x608 [ 21.149518] to a HARDIRQ-irq-unsafe lock: [ 21.149519] (_xmit_ETHER#2){+.-.}-{2:2} [ 21.149523] ... which became HARDIRQ-irq-unsafe at: [ 21.149524] ... __lock_acquire+0x8bc/0x1a94 [ 21.149525] lock_acquire.part.0+0x170/0x360 [ 21.149526] lock_acquire+0x68/0x8c [ 21.149527] _raw_spin_trylock+0x80/0xd0 [ 21.149527] virtnet_poll+0xac/0x360 [ 21.149528] net_rx_action+0x1b0/0x4e0 [ 21.149529] __do_softirq+0x1f4/0x638 [ 21.149530] do_softirq+0xb8/0xcc [ 21.149531] __local_bh_enable_ip+0x18c/0x200 [ 21.149532] virtnet_napi_enable+0xc0/0xd4 [ 21.149533] virtnet_open+0x98/0x1c0 [ 21.149534] __dev_open+0x12c/0x200 [ 21.149535] __dev_change_flags+0x1a0/0x220 [ 21.149536] dev_change_flags+0x2c/0x70 [ 21.149536] do_setlink+0x214/0xe20 [ 21.149537] __rtnl_newlink+0x514/0x820 [ 21.149538] rtnl_newlink+0x58/0x84 [ 21.149539] rtnetlink_rcv_msg+0x184/0x4b4 [ 21.149540] netlink_rcv_skb+0x60/0x124 [ 21.149541] rtnetlink_rcv+0x20/0x30 [ 21.149542] netlink_unicast+0x1b4/0x270 [ 21.149543] netlink_sendmsg+0x1f0/0x400 [ 21.149544] sock_sendmsg+0x5c/0x70 [ 21.149545] ____sys_sendmsg+0x24c/0x280 [ 21.149545] ___sys_sendmsg+0x88/0xd0 [ 21.149546] __sys_sendmsg+0x70/0xd0 [ 21.149547] __arm64_sys_sendmsg+0x2c/0x40
Quoted 8 lines
[ 21.149548] el0_svc_common.constprop.0+0x84/0x200 [ 21.149549] do_el0_svc+0x2c/0x90 [ 21.149550] el0_svc+0x18/0x50 [ 21.149551] el0_sync_handler+0xe0/0x350 [ 21.149552] el0_sync+0x158/0x180 [ 21.149553] other info that might help us debug this: [ 21.149555] Chain exists of: [ 21.149556] console_owner --> target_list_lock --> _xmit_ETHER#2
So somewhere we have target_list_lock taken while holding the console_owner lock (which can happen in interrupt). And what this is saying, target_list_lock is held somewhere where _xmit_ETHER#2 is taken (which I'm guessing from the back traces is the xmit_lock taken in virtnet_poll_tx().
Thus, you can have a deadlock with three CPUs (I need to update the lockdep output to make it proper one of these days).
CPU0 CPU1 CPU2
---- ---- ----
lock(xmit_ETHER#2)
local_irq_disable()
lock(target_list_lock)
lock(xmit_ETHER#2)
[blocked]
local_irq_disable()
lock(console_owner)
lock(target_list_lock)
[blocked]
<interrupt>
lock(console_owner)
[blocked]*** DEADLOCK! ***
Quoted 113 lines
[ 21.149563] Possible interrupt unsafe locking scenario:
[ 21.149564] CPU0 CPU1
[ 21.149565] ---- ----
[ 21.149566] lock(_xmit_ETHER#2);
[ 21.149569] local_irq_disable();
[ 21.149570] lock(console_owner);
[ 21.149572] lock(target_list_lock);
[ 21.149575] <Interrupt>
[ 21.149576] lock(console_owner);
[ 21.149579] *** DEADLOCK ***
[ 21.149580] 3 locks held by modprobe/596:
[ 21.149581] #0: ffff800011a1efe0 (console_lock){+.+.}-{0:0}, at: register_console+0x144/0x2f4
[ 21.149586] #1: ffff800011a1f108 (console_owner){-.-.}-{0:0}, at: console_unlock+0x17c/0x6a4
[ 21.149590] #2: ffff8000093d4018 (target_list_lock){....}-{2:2}, at: write_msg+0x6c/0x120 [netconsole]
[ 21.149596] the dependencies between HARDIRQ-irq-safe lock and the holding lock:
[ 21.149596] -> (console_owner){-.-.}-{0:0} ops: 1191 {
[ 21.149601] IN-HARDIRQ-W at:
[ 21.149602] __lock_acquire+0xa78/0x1a94
[ 21.149603] lock_acquire.part.0+0x170/0x360
[ 21.149604] lock_acquire+0x68/0x8c
[ 21.149605] console_unlock+0x1e8/0x6a4
[ 21.149606] vprintk_emit+0x1c4/0x3c4
[ 21.149607] vprintk_default+0x40/0x4c
[ 21.149608] vprintk_func+0x10c/0x220
[ 21.149610] printk+0x68/0x90
[ 21.149611] crng_fast_load+0x1bc/0x1c0
[ 21.149612] add_interrupt_randomness+0x280/0x290
[ 21.149613] handle_irq_event+0x80/0x120
[ 21.149614] handle_fasteoi_irq+0xac/0x200
[ 21.149615] __handle_domain_irq+0x84/0xf0
[ 21.149616] gic_handle_irq+0xd4/0x320
[ 21.149617] el1_irq+0xd0/0x180
[ 21.149618] arch_cpu_idle+0x24/0x44
[ 21.149619] default_idle_call+0x48/0xa0
[ 21.149620] do_idle+0x260/0x300
[ 21.149621] cpu_startup_entry+0x30/0x6c
[ 21.149622] rest_init+0x1b4/0x288
[ 21.149624] arch_call_rest_init+0x18/0x24
[ 21.149625] start_kernel+0x5cc/0x608
[ 21.149625] IN-SOFTIRQ-W at:
[ 21.149627] __lock_acquire+0x894/0x1a94
[ 21.149628] lock_acquire.part.0+0x170/0x360
[ 21.149629] lock_acquire+0x68/0x8c
[ 21.149630] console_unlock+0x1e8/0x6a4
[ 21.149631] vprintk_emit+0x1c4/0x3c4
[ 21.149632] vprintk_default+0x40/0x4c
[ 21.149633] vprintk_func+0x10c/0x220
[ 21.149634] printk+0x68/0x90
[ 21.149635] hrtimer_interrupt+0x290/0x294
[ 21.149636] arch_timer_handler_virt+0x3c/0x50
[ 21.149637] handle_percpu_devid_irq+0x94/0x164
[ 21.149673] __handle_domain_irq+0x84/0xf0
[ 21.149674] gic_handle_irq+0xd4/0x320
[ 21.149675] el1_irq+0xd0/0x180
[ 21.149676] __do_softirq+0x108/0x638
[ 21.149677] __irq_exit_rcu+0x17c/0x1b0
[ 21.149678] irq_exit+0x18/0x44
[ 21.149679] __handle_domain_irq+0x88/0xf0
[ 21.149680] gic_handle_irq+0xd4/0x320
[ 21.149681] el1_irq+0xd0/0x180
[ 21.149682] smp_call_function_many_cond+0x3cc/0x3f0
[ 21.149683] kick_all_cpus_sync+0x4c/0x80
[ 21.149684] load_module+0x1eec/0x2734
[ 21.149685] __do_sys_finit_module+0xbc/0x12c
[ 21.149686] __arm64_sys_finit_module+0x28/0x34
[ 21.149687] el0_svc_common.constprop.0+0x84/0x200
[ 21.149688] do_el0_svc+0x2c/0x90
[ 21.149689] el0_svc+0x18/0x50
[ 21.149690] el0_sync_handler+0xe0/0x350
[ 21.149691] el0_sync+0x158/0x180
[ 21.149692] INITIAL USE at:
[ 21.149694] __lock_acquire+0x2e0/0x1a94
[ 21.149695] lock_acquire.part.0+0x170/0x360
[ 21.149696] lock_acquire+0x68/0x8c
[ 21.149697] console_unlock+0x1e8/0x6a4
[ 21.149698] vprintk_emit+0x1c4/0x3c4
[ 21.149699] vprintk_default+0x40/0x4c
[ 21.149700] vprintk_func+0x10c/0x220
[ 21.149701] printk+0x68/0x90
[ 21.149702] start_kernel+0x8c/0x608
[ 21.149703] }
[ 21.149704] ... key at: [<ffff800011a1f108>] console_owner_dep_map+0x0/0x28
[ 21.149705] ... acquired at:
[ 21.149706] lock_acquire.part.0+0x170/0x360
[ 21.149707] lock_acquire+0x68/0x8c
[ 21.149708] _raw_spin_lock_irqsave+0x88/0x15c
[ 21.149709] write_msg+0x6c/0x120 [netconsole]
[ 21.149710] console_unlock+0x3ec/0x6a4
[ 21.149711] register_console+0x17c/0x2f4
[ 21.149712] init_netconsole+0x20c/0x1000 [netconsole]
[ 21.149713] do_one_initcall+0x8c/0x480
[ 21.149714] do_init_module+0x60/0x270
[ 21.149715] load_module+0x21f8/0x2734
[ 21.149716] __do_sys_finit_module+0xbc/0x12c
[ 21.149717] __arm64_sys_finit_module+0x28/0x34
[ 21.149718] el0_svc_common.constprop.0+0x84/0x200
[ 21.149719] do_el0_svc+0x2c/0x90
[ 21.149720] el0_svc+0x18/0x50
[ 21.149722] el0_sync_handler+0xe0/0x350
[ 21.149723] el0_sync+0x158/0x180
[ 21.149724] -> (target_list_lock){....}-{2:2} ops: 4 {
[ 21.149729] INITIAL USE at:
[ 21.149730] __lock_acquire+0x2e0/0x1a94
[ 21.149731] lock_acquire.part.0+0x170/0x360
[ 21.149732] lock_acquire+0x68/0x8c
[ 21.149734] _raw_spin_lock_irqsave+0x88/0x15c
[ 21.149735] init_netconsole+0x148/0x1000 [netconsole]
[ 21.149736] do_one_initcall+0x8c/0x480
[ 21.149737] do_init_module+0x60/0x270
[ 21.149738] load_module+0x21f8/0x2734
[ 21.149739] __do_sys_finit_module+0xbc/0x12c
[ 21.149740] __arm64_sys_finit_module+0x28/0x34
[ 21.149741] el0_svc_common.constprop.0+0x84/0x200Quoted 8 lines
[ 21.149742] do_el0_svc+0x2c/0x90 [ 21.149743] el0_svc+0x18/0x50 [ 21.149744] el0_sync_handler+0xe0/0x350 [ 21.149745] el0_sync+0x158/0x180 [ 21.149746] } [ 21.149747] ... key at: [<ffff8000093d4018>] target_list_lock+0x18/0xfffffffffffff000 [netconsole] [ 21.149748] .. [ 21.149750] Lost 190 message(s)!
It really sucks that we lose 190 messages that would help to decipher this more. :-p
Because I'm not sure where the xmit_lock is taken while holding the target_list_lock. But the above does show that printk() calls write_msg() while holding the console_lock, and write_msg() takes the target_list_lock.
Thus, the fix would ether require disabling interrupts every time the xmit_lock is taken, or to get it from being taken while holding the target_list_lock.
-- Steve
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 246 lines · Steven Rostedt
Quoted 81 lines · On Tue, 17 Nov 2020 12:23:41 +0200 Leon Romanovsky <[email protected]> wrote:
Hi,
Approximately two weeks ago, our regression team started to experience those netconsole splats. The tested code is Linus's master (-rc4) + netdev net-next
+ netdev net-rc.Such splats are random and we can't bisect because there is no stable reproducer.
Any idea, what is the root cause?
[ 21.141117] netpoll: netconsole: local port 6666 [ 21.141947] netpoll: netconsole: local IPv4 address 10.236.251.7 [ 21.143138] netpoll: netconsole: interface 'eth1' [ 21.143988] netpoll: netconsole: remote port 62517 [ 21.145127] netpoll: netconsole: remote IPv4 address 10.236.251.1 [ 21.146206] netpoll: netconsole: remote ethernet address 68:05:ca:aa:99:49 [ 21.149464] ===================================================== [ 21.149466] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected [ 21.149467] 5.10.0-rc4_for_upstream_min_debug_2020_11_16_13_03 #1 Not tainted [ 21.149468] ----------------------------------------------------- [ 21.149469] modprobe/596 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire: [ 21.149471] ffff000194b4e298 (_xmit_ETHER#2){+.-.}-{2:2}, at: virtnet_poll_tx+0x84/0x120 [ 21.149478] and this task is already holding: [ 21.149478] ffff8000093d4018 (target_list_lock){....}-{2:2}, at: write_msg+0x6c/0x120 [netconsole] [ 21.149483] which would create a new lock dependency: [ 21.149484] (target_list_lock){....}-{2:2} -> (_xmit_ETHER#2){+.-.}-{2:2} [ 21.149491] but this new dependency connects a HARDIRQ-irq-safe lock: [ 21.149492] (console_owner){-.-.}-{0:0} [ 21.149496] ... which became HARDIRQ-irq-safe at: [ 21.149496] __lock_acquire+0xa78/0x1a94 [ 21.149498] lock_acquire.part.0+0x170/0x360 [ 21.149498] lock_acquire+0x68/0x8c [ 21.149500] console_unlock+0x1e8/0x6a4 [ 21.149500] vprintk_emit+0x1c4/0x3c4 [ 21.149501] vprintk_default+0x40/0x4c [ 21.149502] vprintk_func+0x10c/0x220 [ 21.149503] printk+0x68/0x90 [ 21.149504] crng_fast_load+0x1bc/0x1c0 [ 21.149505] add_interrupt_randomness+0x280/0x290 [ 21.149506] handle_irq_event+0x80/0x120 [ 21.149507] handle_fasteoi_irq+0xac/0x200 [ 21.149508] __handle_domain_irq+0x84/0xf0 [ 21.149509] gic_handle_irq+0xd4/0x320 [ 21.149510] el1_irq+0xd0/0x180 [ 21.149511] arch_cpu_idle+0x24/0x44 [ 21.149512] default_idle_call+0x48/0xa0 [ 21.149513] do_idle+0x260/0x300 [ 21.149514] cpu_startup_entry+0x30/0x6c [ 21.149515] rest_init+0x1b4/0x288 [ 21.149515] arch_call_rest_init+0x18/0x24 [ 21.149516] start_kernel+0x5cc/0x608 [ 21.149518] to a HARDIRQ-irq-unsafe lock: [ 21.149519] (_xmit_ETHER#2){+.-.}-{2:2} [ 21.149523] ... which became HARDIRQ-irq-unsafe at: [ 21.149524] ... __lock_acquire+0x8bc/0x1a94 [ 21.149525] lock_acquire.part.0+0x170/0x360 [ 21.149526] lock_acquire+0x68/0x8c [ 21.149527] _raw_spin_trylock+0x80/0xd0 [ 21.149527] virtnet_poll+0xac/0x360 [ 21.149528] net_rx_action+0x1b0/0x4e0 [ 21.149529] __do_softirq+0x1f4/0x638 [ 21.149530] do_softirq+0xb8/0xcc [ 21.149531] __local_bh_enable_ip+0x18c/0x200 [ 21.149532] virtnet_napi_enable+0xc0/0xd4 [ 21.149533] virtnet_open+0x98/0x1c0 [ 21.149534] __dev_open+0x12c/0x200 [ 21.149535] __dev_change_flags+0x1a0/0x220 [ 21.149536] dev_change_flags+0x2c/0x70 [ 21.149536] do_setlink+0x214/0xe20 [ 21.149537] __rtnl_newlink+0x514/0x820 [ 21.149538] rtnl_newlink+0x58/0x84 [ 21.149539] rtnetlink_rcv_msg+0x184/0x4b4 [ 21.149540] netlink_rcv_skb+0x60/0x124 [ 21.149541] rtnetlink_rcv+0x20/0x30 [ 21.149542] netlink_unicast+0x1b4/0x270 [ 21.149543] netlink_sendmsg+0x1f0/0x400 [ 21.149544] sock_sendmsg+0x5c/0x70 [ 21.149545] ____sys_sendmsg+0x24c/0x280 [ 21.149545] ___sys_sendmsg+0x88/0xd0 [ 21.149546] __sys_sendmsg+0x70/0xd0 [ 21.149547] __arm64_sys_sendmsg+0x2c/0x40Quoted 8 lines
[ 21.149548] el0_svc_common.constprop.0+0x84/0x200 [ 21.149549] do_el0_svc+0x2c/0x90 [ 21.149550] el0_svc+0x18/0x50 [ 21.149551] el0_sync_handler+0xe0/0x350 [ 21.149552] el0_sync+0x158/0x180 [ 21.149553] other info that might help us debug this: [ 21.149555] Chain exists of: [ 21.149556] console_owner --> target_list_lock --> _xmit_ETHER#2So somewhere we have
target_list_locktaken while holding theconsole_ownerlock (which can happen in interrupt). And what this is saying,target_list_lockis held somewhere where _xmit_ETHER#2 is taken (which I'm guessing from the back traces is thexmit_locktaken invirtnet_poll_tx().Thus, you can have a deadlock with three CPUs (I need to update the lockdep output to make it proper one of these days).
CPU0 CPU1 CPU2 ---- ---- ---- lock(xmit_ETHER#2) local_irq_disable() lock(target_list_lock) lock(xmit_ETHER#2) [blocked] local_irq_disable() lock(console_owner) lock(target_list_lock) [blocked] <interrupt> lock(console_owner) [blocked]*** DEADLOCK! ***
Quoted 113 lines
[ 21.149563] Possible interrupt unsafe locking scenario: [ 21.149564] CPU0 CPU1 [ 21.149565] ---- ---- [ 21.149566] lock(_xmit_ETHER#2); [ 21.149569] local_irq_disable(); [ 21.149570] lock(console_owner); [ 21.149572] lock(target_list_lock); [ 21.149575] <Interrupt> [ 21.149576] lock(console_owner); [ 21.149579] *** DEADLOCK *** [ 21.149580] 3 locks held by modprobe/596: [ 21.149581] #0: ffff800011a1efe0 (console_lock){+.+.}-{0:0}, at: register_console+0x144/0x2f4 [ 21.149586] #1: ffff800011a1f108 (console_owner){-.-.}-{0:0}, at: console_unlock+0x17c/0x6a4 [ 21.149590] #2: ffff8000093d4018 (target_list_lock){....}-{2:2}, at: write_msg+0x6c/0x120 [netconsole] [ 21.149596] the dependencies between HARDIRQ-irq-safe lock and the holding lock: [ 21.149596] -> (console_owner){-.-.}-{0:0} ops: 1191 { [ 21.149601] IN-HARDIRQ-W at: [ 21.149602] __lock_acquire+0xa78/0x1a94 [ 21.149603] lock_acquire.part.0+0x170/0x360 [ 21.149604] lock_acquire+0x68/0x8c [ 21.149605] console_unlock+0x1e8/0x6a4 [ 21.149606] vprintk_emit+0x1c4/0x3c4 [ 21.149607] vprintk_default+0x40/0x4c [ 21.149608] vprintk_func+0x10c/0x220 [ 21.149610] printk+0x68/0x90 [ 21.149611] crng_fast_load+0x1bc/0x1c0 [ 21.149612] add_interrupt_randomness+0x280/0x290 [ 21.149613] handle_irq_event+0x80/0x120 [ 21.149614] handle_fasteoi_irq+0xac/0x200 [ 21.149615] __handle_domain_irq+0x84/0xf0 [ 21.149616] gic_handle_irq+0xd4/0x320 [ 21.149617] el1_irq+0xd0/0x180 [ 21.149618] arch_cpu_idle+0x24/0x44 [ 21.149619] default_idle_call+0x48/0xa0 [ 21.149620] do_idle+0x260/0x300 [ 21.149621] cpu_startup_entry+0x30/0x6c [ 21.149622] rest_init+0x1b4/0x288 [ 21.149624] arch_call_rest_init+0x18/0x24 [ 21.149625] start_kernel+0x5cc/0x608 [ 21.149625] IN-SOFTIRQ-W at: [ 21.149627] __lock_acquire+0x894/0x1a94 [ 21.149628] lock_acquire.part.0+0x170/0x360 [ 21.149629] lock_acquire+0x68/0x8c [ 21.149630] console_unlock+0x1e8/0x6a4 [ 21.149631] vprintk_emit+0x1c4/0x3c4 [ 21.149632] vprintk_default+0x40/0x4c [ 21.149633] vprintk_func+0x10c/0x220 [ 21.149634] printk+0x68/0x90 [ 21.149635] hrtimer_interrupt+0x290/0x294 [ 21.149636] arch_timer_handler_virt+0x3c/0x50 [ 21.149637] handle_percpu_devid_irq+0x94/0x164 [ 21.149673] __handle_domain_irq+0x84/0xf0 [ 21.149674] gic_handle_irq+0xd4/0x320 [ 21.149675] el1_irq+0xd0/0x180 [ 21.149676] __do_softirq+0x108/0x638 [ 21.149677] __irq_exit_rcu+0x17c/0x1b0 [ 21.149678] irq_exit+0x18/0x44 [ 21.149679] __handle_domain_irq+0x88/0xf0 [ 21.149680] gic_handle_irq+0xd4/0x320 [ 21.149681] el1_irq+0xd0/0x180 [ 21.149682] smp_call_function_many_cond+0x3cc/0x3f0 [ 21.149683] kick_all_cpus_sync+0x4c/0x80 [ 21.149684] load_module+0x1eec/0x2734 [ 21.149685] __do_sys_finit_module+0xbc/0x12c [ 21.149686] __arm64_sys_finit_module+0x28/0x34 [ 21.149687] el0_svc_common.constprop.0+0x84/0x200 [ 21.149688] do_el0_svc+0x2c/0x90 [ 21.149689] el0_svc+0x18/0x50 [ 21.149690] el0_sync_handler+0xe0/0x350 [ 21.149691] el0_sync+0x158/0x180 [ 21.149692] INITIAL USE at: [ 21.149694] __lock_acquire+0x2e0/0x1a94 [ 21.149695] lock_acquire.part.0+0x170/0x360 [ 21.149696] lock_acquire+0x68/0x8c [ 21.149697] console_unlock+0x1e8/0x6a4 [ 21.149698] vprintk_emit+0x1c4/0x3c4 [ 21.149699] vprintk_default+0x40/0x4c [ 21.149700] vprintk_func+0x10c/0x220 [ 21.149701] printk+0x68/0x90 [ 21.149702] start_kernel+0x8c/0x608 [ 21.149703] } [ 21.149704] ... key at: [<ffff800011a1f108>] console_owner_dep_map+0x0/0x28 [ 21.149705] ... acquired at: [ 21.149706] lock_acquire.part.0+0x170/0x360 [ 21.149707] lock_acquire+0x68/0x8c [ 21.149708] _raw_spin_lock_irqsave+0x88/0x15c [ 21.149709] write_msg+0x6c/0x120 [netconsole] [ 21.149710] console_unlock+0x3ec/0x6a4 [ 21.149711] register_console+0x17c/0x2f4 [ 21.149712] init_netconsole+0x20c/0x1000 [netconsole] [ 21.149713] do_one_initcall+0x8c/0x480 [ 21.149714] do_init_module+0x60/0x270 [ 21.149715] load_module+0x21f8/0x2734 [ 21.149716] __do_sys_finit_module+0xbc/0x12c [ 21.149717] __arm64_sys_finit_module+0x28/0x34 [ 21.149718] el0_svc_common.constprop.0+0x84/0x200 [ 21.149719] do_el0_svc+0x2c/0x90 [ 21.149720] el0_svc+0x18/0x50 [ 21.149722] el0_sync_handler+0xe0/0x350 [ 21.149723] el0_sync+0x158/0x180 [ 21.149724] -> (target_list_lock){....}-{2:2} ops: 4 { [ 21.149729] INITIAL USE at: [ 21.149730] __lock_acquire+0x2e0/0x1a94 [ 21.149731] lock_acquire.part.0+0x170/0x360 [ 21.149732] lock_acquire+0x68/0x8c [ 21.149734] _raw_spin_lock_irqsave+0x88/0x15c [ 21.149735] init_netconsole+0x148/0x1000 [netconsole] [ 21.149736] do_one_initcall+0x8c/0x480 [ 21.149737] do_init_module+0x60/0x270 [ 21.149738] load_module+0x21f8/0x2734 [ 21.149739] __do_sys_finit_module+0xbc/0x12c [ 21.149740] __arm64_sys_finit_module+0x28/0x34 [ 21.149741] el0_svc_common.constprop.0+0x84/0x200Quoted 8 lines
[ 21.149742] do_el0_svc+0x2c/0x90 [ 21.149743] el0_svc+0x18/0x50 [ 21.149744] el0_sync_handler+0xe0/0x350 [ 21.149745] el0_sync+0x158/0x180 [ 21.149746] } [ 21.149747] ... key at: [<ffff8000093d4018>] target_list_lock+0x18/0xfffffffffffff000 [netconsole] [ 21.149748] .. [ 21.149750] Lost 190 message(s)!It really sucks that we lose 190 messages that would help to decipher this more. :-p
Quoted 7 lines
Because I'm not sure where the
xmit_lockis taken while holding thetarget_list_lock. But the above does show thatprintk()callswrite_msg()while holding theconsole_lock, andwrite_msg()takes thetarget_list_lock.Thus, the fix would ether require disabling interrupts every time the
xmit_lockis taken, or to get it from being taken while holding thetarget_list_lock.
I can add any debug code to our night runs that you think may help.
Thanks
Quoted 2 lines
-- Steve
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 54 lines · Steven Rostedt
Quoted 54 lines
[ 21.149601] IN-HARDIRQ-W at: [ 21.149602] __lock_acquire+0xa78/0x1a94 [ 21.149603] lock_acquire.part.0+0x170/0x360 [ 21.149604] lock_acquire+0x68/0x8c [ 21.149605] console_unlock+0x1e8/0x6a4 [ 21.149606] vprintk_emit+0x1c4/0x3c4 [ 21.149607] vprintk_default+0x40/0x4c [ 21.149608] vprintk_func+0x10c/0x220 [ 21.149610] printk+0x68/0x90 [ 21.149611] crng_fast_load+0x1bc/0x1c0 [ 21.149612] add_interrupt_randomness+0x280/0x290 [ 21.149613] handle_irq_event+0x80/0x120 [ 21.149614] handle_fasteoi_irq+0xac/0x200 [ 21.149615] __handle_domain_irq+0x84/0xf0 [ 21.149616] gic_handle_irq+0xd4/0x320 [ 21.149617] el1_irq+0xd0/0x180 [ 21.149618] arch_cpu_idle+0x24/0x44 [ 21.149619] default_idle_call+0x48/0xa0 [ 21.149620] do_idle+0x260/0x300 [ 21.149621] cpu_startup_entry+0x30/0x6c [ 21.149622] rest_init+0x1b4/0x288 [ 21.149624] arch_call_rest_init+0x18/0x24 [ 21.149625] start_kernel+0x5cc/0x608 [ 21.149625] IN-SOFTIRQ-W at: [ 21.149627] __lock_acquire+0x894/0x1a94 [ 21.149628] lock_acquire.part.0+0x170/0x360 [ 21.149629] lock_acquire+0x68/0x8c [ 21.149630] console_unlock+0x1e8/0x6a4 [ 21.149631] vprintk_emit+0x1c4/0x3c4 [ 21.149632] vprintk_default+0x40/0x4c [ 21.149633] vprintk_func+0x10c/0x220 [ 21.149634] printk+0x68/0x90 [ 21.149635] hrtimer_interrupt+0x290/0x294 [ 21.149636] arch_timer_handler_virt+0x3c/0x50 [ 21.149637] handle_percpu_devid_irq+0x94/0x164 [ 21.149673] __handle_domain_irq+0x84/0xf0 [ 21.149674] gic_handle_irq+0xd4/0x320 [ 21.149675] el1_irq+0xd0/0x180 [ 21.149676] __do_softirq+0x108/0x638 [ 21.149677] __irq_exit_rcu+0x17c/0x1b0 [ 21.149678] irq_exit+0x18/0x44 [ 21.149679] __handle_domain_irq+0x88/0xf0 [ 21.149680] gic_handle_irq+0xd4/0x320 [ 21.149681] el1_irq+0xd0/0x180 [ 21.149682] smp_call_function_many_cond+0x3cc/0x3f0 [ 21.149683] kick_all_cpus_sync+0x4c/0x80 [ 21.149684] load_module+0x1eec/0x2734 [ 21.149685] __do_sys_finit_module+0xbc/0x12c [ 21.149686] __arm64_sys_finit_module+0x28/0x34 [ 21.149687] el0_svc_common.constprop.0+0x84/0x200 [ 21.149688] do_el0_svc+0x2c/0x90 [ 21.149689] el0_svc+0x18/0x50 [ 21.149690] el0_sync_handler+0xe0/0x350 [ 21.149691] el0_sync+0x158/0x180
[..]
Quoted 2 lines
It really sucks that we lose 190 messages that would help to decipher this more. :-p
Indeed.
Quoted 2 lines
Because I'm not sure where the
xmit_lockis taken while holding thetarget_list_lock.
I don't see where does this happen. It seems to me that the report is not about broken locking order, but more about:
- soft-irq can be preempted (while holding _xmit_lock) by a hardware interrupt, that will attempt to acquire the same _xmit_lock lock.
CPU0
<<soft IRQ>>
virtnet_poll_tx()
__netif_tx_lock()
spin_lock(_xmit_lock)
<<hard IRQ>>
add_interrupt_randomness()
crng_fast_load()
printk()
call_console_drivers()
spin_lock_irqsave(&target_list_lock)
spin_lock(_xmit_lock);Does this make sense?
-ss
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
On (20/11/18 11:46), Sergey Senozhatsky wrote:
[..]
Quoted 21 lines
Quoted 2 lines
Because I'm not sure where the
xmit_lockis taken while holding thetarget_list_lock.I don't see where does this happen. It seems to me that the report is not about broken locking order, but more about:
- soft-irq can be preempted (while holding_xmit_lock) by a hardware interrupt, that will attempt to acquire the same_xmit_locklock.CPU0 <<soft IRQ>> virtnet_poll_tx() __netif_tx_lock() spin_lock(_xmit_lock) <<hard IRQ>> add_interrupt_randomness() crng_fast_load() printk() call_console_drivers() spin_lock_irqsave(&target_list_lock) spin_lock(_xmit_lock);
Quoted 1 line
Does this make sense?
Hmm, lockdep says something similar, but there are 2 printk() happening - both on local and remote CPUs.
[ 21.149564] CPU0 CPU1 [ 21.149565] ---- ---- [ 21.149566] lock(_xmit_ETHER#2); [ 21.149569] local_irq_disable(); [ 21.149570] lock(console_owner); [ 21.149572] lock(target_list_lock); [ 21.149575] <Interrupt> [ 21.149576] lock(console_owner);
This CPU0 lock(_xmit_ETHER#2) -> hard IRQ -> lock(console_owner) is basically
soft IRQ -> lock(_xmit_ETHER#2) -> hard IRQ -> printk()
Then CPU1 spins on xmit, which is owned by CPU0, CPU0 spins on console_owner, which is owned by CPU1?
A quick-and-dirty idea (it doesn't fix the lockdep report) - can we add some sort of max_loops variable to console_trylock_spinning(), so that it will not spin forever in while (READ_ONCE(console_waiter)) waiting for a console_owner to pass the lock?
-ss
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 35 lines · Sergey Senozhatsky
On (20/11/18 11:46), Sergey Senozhatsky wrote:
[..]Quoted 20 lines
Quoted 2 lines
Because I'm not sure where the
xmit_lockis taken while holding thetarget_list_lock.I don't see where does this happen. It seems to me that the report is not about broken locking order, but more about:
- soft-irq can be preempted (while holding_xmit_lock) by a hardware interrupt, that will attempt to acquire the same_xmit_locklock.CPU0 <<soft IRQ>> virtnet_poll_tx() __netif_tx_lock() spin_lock(_xmit_lock) <<hard IRQ>> add_interrupt_randomness() crng_fast_load() printk() call_console_drivers() spin_lock_irqsave(&target_list_lock) spin_lock(_xmit_lock);Quoted 1 line
Does this make sense?
Hmm, lockdep says something similar, but there are 2
printk()happening - both on local and remote CPUs.[ 21.149564] CPU0 CPU1 [ 21.149565] ---- ---- [ 21.149566] lock(_xmit_ETHER#2); [ 21.149569] local_irq_disable(); [ 21.149570] lock(console_owner); [ 21.149572] lock(target_list_lock); [ 21.149575] <Interrupt> [ 21.149576] lock(console_owner);
Quoted 6 lines
This CPU0
lock(_xmit_ETHER#2)-> hard IRQ ->lock(console_owner)is basicallysoft IRQ -> lock(_xmit_ETHER#2) -> hard IRQ -> printk()Then CPU1 spins on xmit, which is owned by CPU0, CPU0 spins on
console_owner, which is owned by CPU1?
If this is true, it looks not a virtio-net specific issue but somewhere else.
I think all network driver will synchronize through bh instead of hardirq.
Thanks
Quoted 8 lines
A quick-and-dirty idea (it doesn't fix the lockdep report) - can we add some sort of
max_loopsvariable toconsole_trylock_spinning(), so that it will not spin forever inwhile (READ_ONCE(console_waiter))waiting for aconsole_ownerto pass the lock?-ss
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
[ Adding netdev as perhaps someone there knows ]
Quoted 6 lines · On Wed, 18 Nov 2020 12:09:59 +0800 Jason Wang <[email protected]> wrote:
Quoted 6 lines
This CPU0
lock(_xmit_ETHER#2)-> hard IRQ ->lock(console_owner)is basicallysoft IRQ -> lock(_xmit_ETHER#2) -> hard IRQ -> printk()Then CPU1 spins on xmit, which is owned by CPU0, CPU0 spins on
console_owner, which is owned by CPU1?
It still looks to me that the target_list_lock is taken in IRQ, (which can be the case because printk calls write_msg() which takes that lock). And someplace there's a:
lock(target_list_lock)
lock(xmit_lock)
which means you can remove the console lock from this scenario completely, and you still have a possible deadlock between target_list_lock and xmit_lock.
Quoted 6 lines
If this is true, it looks not a virtio-net specific issue but somewhere else.
I think all network driver will synchronize through bh instead of hardirq.
I think the issue is where target_list_lock is held when we take xmit_lock. Is there anywhere in netconsole.c that can end up taking xmit_lock while holding the target_list_lock? If so, that's the problem. As target_list_lock is something that can be taken in IRQ context, which means *any* other lock that is taking while holding the target_list_lock must also protect against interrupts from happening while it they are held.
-- Steve
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 25 lines · Steven Rostedt
Quoted 10 lines · On Tue, 17 Nov 2020 12:23:41 +0200 Leon Romanovsky <[email protected]> wrote:
Hi,
Approximately two weeks ago, our regression team started to experience those netconsole splats. The tested code is Linus's master (-rc4) + netdev net-next
+ netdev net-rc.Such splats are random and we can't bisect because there is no stable reproducer.
Any idea, what is the root cause?
Quoted 11 lines
[ 21.149739] __do_sys_finit_module+0xbc/0x12c [ 21.149740] __arm64_sys_finit_module+0x28/0x34 [ 21.149741] el0_svc_common.constprop.0+0x84/0x200 [ 21.149742] do_el0_svc+0x2c/0x90 [ 21.149743] el0_svc+0x18/0x50 [ 21.149744] el0_sync_handler+0xe0/0x350 [ 21.149745] el0_sync+0x158/0x180 [ 21.149746] } [ 21.149747] ... key at: [<ffff8000093d4018>] target_list_lock+0x18/0xfffffffffffff000 [netconsole] [ 21.149748] .. [ 21.149750] Lost 190 message(s)!
Quoted 2 lines
It really sucks that we lose 190 messages that would help to decipher this more. :-p
The message commes from the printk_safe code. The size can be increased by CONFIG_PRINTK_SAFE_LOG_BUF_SHIFT.
Quoted 7 lines
Because I'm not sure where the
xmit_lockis taken while holding thetarget_list_lock. But the above does show thatprintk()callswrite_msg()while holding theconsole_lock, andwrite_msg()takes thetarget_list_lock.Thus, the fix would ether require disabling interrupts every time the
xmit_lockis taken, or to get it from being taken while holding thetarget_list_lock.
It seems that the missing messages might help to find the root of the problem.
Best Regards,
Petr
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 36 lines · Petr Mladek
Quoted 25 lines · Steven Rostedt
Quoted 10 lines · On Tue, 17 Nov 2020 12:23:41 +0200 Leon Romanovsky <[email protected]> wrote:
Hi,
Approximately two weeks ago, our regression team started to experience those netconsole splats. The tested code is Linus's master (-rc4) + netdev net-next
+ netdev net-rc.Such splats are random and we can't bisect because there is no stable reproducer.
Any idea, what is the root cause?
Quoted 11 lines
[ 21.149739] __do_sys_finit_module+0xbc/0x12c [ 21.149740] __arm64_sys_finit_module+0x28/0x34 [ 21.149741] el0_svc_common.constprop.0+0x84/0x200 [ 21.149742] do_el0_svc+0x2c/0x90 [ 21.149743] el0_svc+0x18/0x50 [ 21.149744] el0_sync_handler+0xe0/0x350 [ 21.149745] el0_sync+0x158/0x180 [ 21.149746] } [ 21.149747] ... key at: [<ffff8000093d4018>] target_list_lock+0x18/0xfffffffffffff000 [netconsole] [ 21.149748] .. [ 21.149750] Lost 190 message(s)!Quoted 2 lines
It really sucks that we lose 190 messages that would help to decipher this more. :-p
The message commes from the
printk_safecode. The size can be increased byCONFIG_PRINTK_SAFE_LOG_BUF_SHIFT.Quoted 4 lines
Because I'm not sure where the
xmit_lockis taken while holding thetarget_list_lock. But the above does show thatprintk()callswrite_msg()while holding theconsole_lock, andwrite_msg()takes thetarget_list_lock.
Quoted 6 lines
Quoted 3 lines
Thus, the fix would ether require disabling interrupts every time the
xmit_lockis taken, or to get it from being taken while holding thetarget_list_lock.It seems that the missing messages might help to find the root of the problem.
Sorry for not being very responsive, I was in internet-free zone :).
I'll increase CONFIG_PRINTK_SAFE_LOG_BUF_SHIFT from 13 to be 26, let's see what night run will give us.
Thanks
Quoted 3 lines
Best Regards,
Petr
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 29 lines · Steven Rostedt
[ Adding netdev as perhaps someone there knows ]
Quoted 6 lines · On Wed, 18 Nov 2020 12:09:59 +0800 Jason Wang <[email protected]> wrote:
Quoted 6 lines
This CPU0
lock(_xmit_ETHER#2)-> hard IRQ ->lock(console_owner)is basicallysoft IRQ -> lock(_xmit_ETHER#2) -> hard IRQ -> printk()Then CPU1 spins on xmit, which is owned by CPU0, CPU0 spins on
console_owner, which is owned by CPU1?It still looks to me that the
target_list_lockis taken in IRQ, (which can be the case because printk callswrite_msg()which takes that lock). And someplace there's a:
lock(target_list_lock)
lock(xmit_lock)which means you can remove the console lock from this scenario completely, and you still have a possible deadlock between
target_list_lockandxmit_lock.Quoted 5 lines
If this is true, it looks not a virtio-net specific issue but somewhere else.
Quoted 8 lines
Quoted 1 line
I think all network driver will synchronize through bh instead of hardirq.
I think the issue is where
target_list_lockis held when we takexmit_lock. Is there anywhere innetconsole.cthat can end up takingxmit_lockwhile holding thetarget_list_lock? If so, that's the problem. Astarget_list_lockis something that can be taken in IRQ context, which means *any* other lock that is taking while holding thetarget_list_lockmust also protect against interrupts from happening while it they are held.
I increased printk buffer like Petr suggested and the splat is below. It doesn't happening on x86, but on ARM65 and ppc64.
[ 10.027975] =====================================================
[ 10.027976] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected
[ 10.027976] 5.10.0-rc4_for_upstream_min_debug_2020_11_22_19_37 #1 Not tainted
[ 10.027977] -----------------------------------------------------
[ 10.027978] modprobe/638 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire:
[ 10.027979] ffff0000c9f63c98 (_xmit_ETHER#2){+.-.}-{2:2}, at: virtnet_poll_tx+0x84/0x120
[ 10.027982]
[ 10.027982] and this task is already holding:
[ 10.027983] ffff800009007018 (target_list_lock){....}-{2:2}, at: write_msg+0x6c/0x120 [netconsole]
[ 10.027985] which would create a new lock dependency:
[ 10.027985] (target_list_lock){....}-{2:2} -> (_xmit_ETHER#2){+.-.}-{2:2}
[ 10.027989]
[ 10.027989] but this new dependency connects a HARDIRQ-irq-safe lock:
[ 10.027990] (console_owner){-...}-{0:0}
[ 10.027991]
[ 10.027992] ... which became HARDIRQ-irq-safe at:
[ 10.027992] __lock_acquire+0xa78/0x1a94
[ 10.027993] lock_acquire.part.0+0x170/0x360
[ 10.027993] lock_acquire+0x68/0x8c
[ 10.027994] console_unlock+0x1e8/0x6a4
[ 10.027994] vprintk_emit+0x1c4/0x3c4
[ 10.027995] vprintk_default+0x40/0x4c
[ 10.027995] vprintk_func+0x10c/0x220
[ 10.027995] printk+0x68/0x90
[ 10.027996] crng_fast_load+0x1bc/0x1c0
[ 10.027997] add_interrupt_randomness+0x280/0x290
[ 10.027997] handle_irq_event+0x80/0x120
[ 10.027997] handle_fasteoi_irq+0xac/0x200
[ 10.027998] __handle_domain_irq+0x84/0xf0
[ 10.027999] gic_handle_irq+0xd4/0x320
[ 10.027999] el1_irq+0xd0/0x180
[ 10.028000] arch_cpu_idle+0x24/0x44
[ 10.028000] default_idle_call+0x48/0xa0
[ 10.028001] do_idle+0x260/0x300
[ 10.028001] cpu_startup_entry+0x30/0x6c
[ 10.028001] rest_init+0x1b4/0x288
[ 10.028002] arch_call_rest_init+0x18/0x24
[ 10.028002] start_kernel+0x5cc/0x608
[ 10.028003]
[ 10.028003] to a HARDIRQ-irq-unsafe lock:
[ 10.028004] (_xmit_ETHER#2){+.-.}-{2:2}
[ 10.028005]
[ 10.028006] ... which became HARDIRQ-irq-unsafe at:
[ 10.028006] ... __lock_acquire+0x8bc/0x1a94
[ 10.028007] lock_acquire.part.0+0x170/0x360
[ 10.028007] lock_acquire+0x68/0x8c
[ 10.028008] _raw_spin_trylock+0x80/0xd0
[ 10.028008] virtnet_poll+0xac/0x360
[ 10.028009] net_rx_action+0x1b0/0x4e0
[ 10.028010] __do_softirq+0x1f4/0x638
[ 10.028010] do_softirq+0xb8/0xcc
[ 10.028010] __local_bh_enable_ip+0x18c/0x200
[ 10.028011] virtnet_napi_enable+0xc0/0xd4
[ 10.028011] virtnet_open+0x98/0x1c0
[ 10.028012] __dev_open+0x12c/0x200
[ 10.028013] __dev_change_flags+0x1a0/0x220
[ 10.028013] dev_change_flags+0x2c/0x70
[ 10.028014] do_setlink+0x214/0xe20
[ 10.028014] __rtnl_newlink+0x514/0x820
[ 10.028015] rtnl_newlink+0x58/0x84
[ 10.028015] rtnetlink_rcv_msg+0x184/0x4b4
[ 10.028016] netlink_rcv_skb+0x60/0x124
[ 10.028016] rtnetlink_rcv+0x20/0x30
[ 10.028017] netlink_unicast+0x1b4/0x270
[ 10.028017] netlink_sendmsg+0x1f0/0x400
[ 10.028018] sock_sendmsg+0x5c/0x70
[ 10.028018] ____sys_sendmsg+0x24c/0x280
[ 10.028019] ___sys_sendmsg+0x88/0xd0
[ 10.028019] __sys_sendmsg+0x70/0xd0
[ 10.028020] __arm64_sys_sendmsg+0x2c/0x40
[ 10.028021] el0_svc_common.constprop.0+0x84/0x200
[ 10.028021] do_el0_svc+0x2c/0x90
[ 10.028021] el0_svc+0x18/0x50
[ 10.028022] el0_sync_handler+0xe0/0x350
[ 10.028023] el0_sync+0x158/0x180
[ 10.028023]
[ 10.028023] other info that might help us debug this:
[ 10.028024]
[ 10.028024] Chain exists of:
[ 10.028025] console_owner --> target_list_lock --> _xmit_ETHER#2
[ 10.028028]
[ 10.028028] Possible interrupt unsafe locking scenario:
[ 10.028029]
[ 10.028029] CPU0 CPU1
[ 10.028030] ---- ----
[ 10.028030] lock(_xmit_ETHER#2);
[ 10.028032] local_irq_disable();
[ 10.028032] lock(console_owner);
[ 10.028034] lock(target_list_lock);
[ 10.028035] <Interrupt>
[ 10.028035] lock(console_owner);
[ 10.028036]
[ 10.028037] *** DEADLOCK ***
[ 10.028037]
[ 10.028038] 3 locks held by modprobe/638:
[ 10.028038] #0: ffff800011e1efe0 (console_lock){+.+.}-{0:0}, at: register_console+0x144/0x2f4
[ 10.028040] #1: ffff800011e1f108 (console_owner){-...}-{0:0}, at: console_unlock+0x17c/0x6a4
[ 10.028043] #2: ffff800009007018 (target_list_lock){....}-{2:2}, at: write_msg+0x6c/0x120 [netconsole]
[ 10.028045]
[ 10.028046] the dependencies between HARDIRQ-irq-safe lock and the holding lock:
[ 10.028046] -> (console_owner){-...}-{0:0} ops: 1574 {
[ 10.028049] IN-HARDIRQ-W at:
[ 10.028050] __lock_acquire+0xa78/0x1a94
[ 10.028050] lock_acquire.part.0+0x170/0x360
[ 10.028051] lock_acquire+0x68/0x8c
[ 10.028051] console_unlock+0x1e8/0x6a4
[ 10.028052] vprintk_emit+0x1c4/0x3c4
[ 10.028052] vprintk_default+0x40/0x4c
[ 10.028053] vprintk_func+0x10c/0x220
[ 10.028054] printk+0x68/0x90
[ 10.028054] crng_fast_load+0x1bc/0x1c0
[ 10.028055] add_interrupt_randomness+0x280/0x290
[ 10.028056] handle_irq_event+0x80/0x120
[ 10.028056] handle_fasteoi_irq+0xac/0x200
[ 10.028057] __handle_domain_irq+0x84/0xf0
[ 10.028057] gic_handle_irq+0xd4/0x320
[ 10.028058] el1_irq+0xd0/0x180
[ 10.028058] arch_cpu_idle+0x24/0x44
[ 10.028059] default_idle_call+0x48/0xa0
[ 10.028060] do_idle+0x260/0x300
[ 10.028061] cpu_startup_entry+0x30/0x6c
[ 10.028061] rest_init+0x1b4/0x288
[ 10.028062] arch_call_rest_init+0x18/0x24
[ 10.028062] start_kernel+0x5cc/0x608
[ 10.028063] INITIAL USE at:
[ 10.028064] __lock_acquire+0x2e0/0x1a94
[ 10.028064] lock_acquire.part.0+0x170/0x360
[ 10.028065] lock_acquire+0x68/0x8c
[ 10.028066] console_unlock+0x1e8/0x6a4
[ 10.028067] vprintk_emit+0x1c4/0x3c4
[ 10.028067] vprintk_default+0x40/0x4c
[ 10.028068] vprintk_func+0x10c/0x220
[ 10.028068] printk+0x68/0x90
[ 10.028069] start_kernel+0x8c/0x608
[ 10.028069] }
[ 10.028070] ... key at: [<ffff800011e1f108>] console_owner_dep_map+0x0/0x28
[ 10.028071] ... acquired at:
[ 10.028071] lock_acquire.part.0+0x170/0x360
[ 10.028072] lock_acquire+0x68/0x8c
[ 10.028072] _raw_spin_lock_irqsave+0x88/0x15c
[ 10.028073] write_msg+0x6c/0x120 [netconsole]
[ 10.028073] console_unlock+0x3ec/0x6a4
[ 10.028074] register_console+0x17c/0x2f4
[ 10.028075] init_netconsole+0x20c/0x1000 [netconsole]
[ 10.028075] do_one_initcall+0x8c/0x480
[ 10.028076] do_init_module+0x60/0x270
[ 10.028076] load_module+0x21f8/0x2734
[ 10.028077] __do_sys_finit_module+0xbc/0x12c
[ 10.028077] __arm64_sys_finit_module+0x28/0x34
[ 10.028078] el0_svc_common.constprop.0+0x84/0x200
[ 10.028078] do_el0_svc+0x2c/0x90
[ 10.028079] el0_svc+0x18/0x50
[ 10.028079] el0_sync_handler+0xe0/0x350
[ 10.028080] el0_sync+0x158/0x180
[ 10.028080]
[ 10.028081] -> (target_list_lock){....}-{2:2} ops: 34 {
[ 10.028083] INITIAL USE at:
[ 10.028084] __lock_acquire+0x2e0/0x1a94
[ 10.028084] lock_acquire.part.0+0x170/0x360
[ 10.028085] lock_acquire+0x68/0x8c
[ 10.028085] _raw_spin_lock_irqsave+0x88/0x15c
[ 10.028086] init_netconsole+0x148/0x1000 [netconsole]
[ 10.028087] do_one_initcall+0x8c/0x480
[ 10.028087] do_init_module+0x60/0x270
[ 10.028088] load_module+0x21f8/0x2734
[ 10.028088] __do_sys_finit_module+0xbc/0x12c
[ 10.028089] __arm64_sys_finit_module+0x28/0x34
[ 10.028090] el0_svc_common.constprop.0+0x84/0x200
[ 10.028090] do_el0_svc+0x2c/0x90
[ 10.028091] el0_svc+0x18/0x50
[ 10.028092] el0_sync_handler+0xe0/0x350
[ 10.028092] el0_sync+0x158/0x180
[ 10.028093] }
[ 10.028093] ... key at: [<ffff800009007018>] target_list_lock+0x18/0xfffffffffffff000 [netconsole]
[ 10.028094] ... acquired at:
[ 10.028094] __lock_acquire+0x134c/0x1a94
[ 10.028095] lock_acquire.part.0+0x170/0x360
[ 10.028095] lock_acquire+0x68/0x8c
[ 10.028096] _raw_spin_lock+0x64/0x90
[ 10.028096] virtnet_poll_tx+0x84/0x120
[ 10.028097] netpoll_poll_dev+0x12c/0x350
[ 10.028097] netpoll_send_skb+0x39c/0x400
[ 10.028098] netpoll_send_udp+0x2b8/0x440
[ 10.028098] write_msg+0xfc/0x120 [netconsole]
[ 10.028099] console_unlock+0x3ec/0x6a4
[ 10.028100] register_console+0x17c/0x2f4
[ 10.028100] init_netconsole+0x20c/0x1000 [netconsole]
[ 10.028101] do_one_initcall+0x8c/0x480
[ 10.028101] do_init_module+0x60/0x270
[ 10.028102] load_module+0x21f8/0x2734
[ 10.028102] __do_sys_finit_module+0xbc/0x12c
[ 10.028103] __arm64_sys_finit_module+0x28/0x34
[ 10.028103] el0_svc_common.constprop.0+0x84/0x200
[ 10.028104] do_el0_svc+0x2c/0x90
[ 10.028104] el0_svc+0x18/0x50
[ 10.028105] el0_sync_handler+0xe0/0x350
[ 10.028105] el0_sync+0x158/0x180
[ 10.028106]
[ 10.028106]
[ 10.028107] the dependencies between the lock to be acquired
[ 10.028107] and HARDIRQ-irq-unsafe lock:
[ 10.028108] -> (_xmit_ETHER#2){+.-.}-{2:2} ops: 217 {
[ 10.028110] HARDIRQ-ON-W at:
[ 10.028111] __lock_acquire+0x8bc/0x1a94
[ 10.028111] lock_acquire.part.0+0x170/0x360
[ 10.028112] lock_acquire+0x68/0x8c
[ 10.028113] _raw_spin_trylock+0x80/0xd0
[ 10.028113] virtnet_poll+0xac/0x360
[ 10.028114] net_rx_action+0x1b0/0x4e0
[ 10.028115] __do_softirq+0x1f4/0x638
[ 10.028115] do_softirq+0xb8/0xcc
[ 10.028116] __local_bh_enable_ip+0x18c/0x200
[ 10.028116] virtnet_napi_enable+0xc0/0xd4
[ 10.028117] virtnet_open+0x98/0x1c0
[ 10.028118] __dev_open+0x12c/0x200
[ 10.028118] __dev_change_flags+0x1a0/0x220
[ 10.028119] dev_change_flags+0x2c/0x70
[ 10.028119] do_setlink+0x214/0xe20
[ 10.028120] __rtnl_newlink+0x514/0x820
[ 10.028120] rtnl_newlink+0x58/0x84
[ 10.028121] rtnetlink_rcv_msg+0x184/0x4b4
[ 10.028122] netlink_rcv_skb+0x60/0x124
[ 10.028122] rtnetlink_rcv+0x20/0x30
[ 10.028123] netlink_unicast+0x1b4/0x270
[ 10.028124] netlink_sendmsg+0x1f0/0x400
[ 10.028124] sock_sendmsg+0x5c/0x70
[ 10.028125] ____sys_sendmsg+0x24c/0x280
[ 10.028125] ___sys_sendmsg+0x88/0xd0
[ 10.028126] __sys_sendmsg+0x70/0xd0
[ 10.028127] __arm64_sys_sendmsg+0x2c/0x40
[ 10.028128] el0_svc_common.constprop.0+0x84/0x200
[ 10.028128] do_el0_svc+0x2c/0x90
[ 10.028129] el0_svc+0x18/0x50
[ 10.028129] el0_sync_handler+0xe0/0x350
[ 10.028130] el0_sync+0x158/0x180
[ 10.028130] IN-SOFTIRQ-W at:
[ 10.028131] __lock_acquire+0x894/0x1a94
[ 10.028132] lock_acquire.part.0+0x170/0x360
[ 10.028132] lock_acquire+0x68/0x8c
[ 10.028133] _raw_spin_lock+0x64/0x90
[ 10.028134] virtnet_poll_tx+0x84/0x120
[ 10.028134] net_rx_action+0x1b0/0x4e0
[ 10.028135] __do_softirq+0x1f4/0x638
[ 10.028135] do_softirq+0xb8/0xcc
[ 10.028136] __local_bh_enable_ip+0x18c/0x200
[ 10.028137] virtnet_napi_enable+0xc0/0xd4
[ 10.028137] virtnet_open+0x14c/0x1c0
[ 10.028138] __dev_open+0x12c/0x200
[ 10.028138] __dev_change_flags+0x1a0/0x220
[ 10.028139] dev_change_flags+0x2c/0x70
[ 10.028140] do_setlink+0x214/0xe20
[ 10.028140] __rtnl_newlink+0x514/0x820
[ 10.028141] rtnl_newlink+0x58/0x84
[ 10.028141] rtnetlink_rcv_msg+0x184/0x4b4
[ 10.028142] netlink_rcv_skb+0x60/0x124
[ 10.028142] rtnetlink_rcv+0x20/0x30
[ 10.028143] netlink_unicast+0x1b4/0x270
[ 10.028144] netlink_sendmsg+0x1f0/0x400
[ 10.028144] sock_sendmsg+0x5c/0x70
[ 10.028145] ____sys_sendmsg+0x24c/0x280
[ 10.028146] ___sys_sendmsg+0x88/0xd0
[ 10.028146] __sys_sendmsg+0x70/0xd0
[ 10.028147] __arm64_sys_sendmsg+0x2c/0x40
[ 10.028148] el0_svc_common.constprop.0+0x84/0x200
[ 10.028148] do_el0_svc+0x2c/0x90
[ 10.028149] el0_svc+0x18/0x50
[ 10.028149] el0_sync_handler+0xe0/0x350
[ 10.028150] el0_sync+0x158/0x180
[ 10.028150] INITIAL USE at:
[ 10.028151] __lock_acquire+0x2e0/0x1a94
[ 10.028152] lock_acquire.part.0+0x170/0x360
[ 10.028153] lock_acquire+0x68/0x8c
[ 10.028153] _raw_spin_trylock+0x80/0xd0
[ 10.028154] virtnet_poll+0xac/0x360
[ 10.028154] net_rx_action+0x1b0/0x4e0
[ 10.028155] __do_softirq+0x1f4/0x638
[ 10.028155] do_softirq+0xb8/0xcc
[ 10.028156] __local_bh_enable_ip+0x18c/0x200
[ 10.028157] virtnet_napi_enable+0xc0/0xd4
[ 10.028157] virtnet_open+0x98/0x1c0
[ 10.028158] __dev_open+0x12c/0x200
[ 10.028158] __dev_change_flags+0x1a0/0x220
[ 10.028159] dev_change_flags+0x2c/0x70
[ 10.028159] do_setlink+0x214/0xe20
[ 10.028160] __rtnl_newlink+0x514/0x820
[ 10.028161] rtnl_newlink+0x58/0x84
[ 10.028161] rtnetlink_rcv_msg+0x184/0x4b4
[ 10.028162] netlink_rcv_skb+0x60/0x124
[ 10.028162] rtnetlink_rcv+0x20/0x30
[ 10.028163] netlink_unicast+0x1b4/0x270
[ 10.028163] netlink_sendmsg+0x1f0/0x400
[ 10.028164] sock_sendmsg+0x5c/0x70
[ 10.028165] ____sys_sendmsg+0x24c/0x280
[ 10.028165] ___sys_sendmsg+0x88/0xd0
[ 10.028166] __sys_sendmsg+0x70/0xd0
[ 10.028166] __arm64_sys_sendmsg+0x2c/0x40
[ 10.028167] el0_svc_common.constprop.0+0x84/0x200
[ 10.028168] do_el0_svc+0x2c/0x90
[ 10.028168] el0_svc+0x18/0x50
[ 10.028169] el0_sync_handler+0xe0/0x350
[ 10.028169] el0_sync+0x158/0x180
[ 10.028170] }
[ 10.028171] ... key at: [<ffff80001312aef8>] netdev_xmit_lock_key+0x10/0x390
[ 10.028171] ... acquired at:
[ 10.028172] __lock_acquire+0x134c/0x1a94
[ 10.028172] lock_acquire.part.0+0x170/0x360
[ 10.028173] lock_acquire+0x68/0x8c
[ 10.028173] _raw_spin_lock+0x64/0x90
[ 10.028174] virtnet_poll_tx+0x84/0x120
[ 10.028174] netpoll_poll_dev+0x12c/0x350
[ 10.028175] netpoll_send_skb+0x39c/0x400
[ 10.028175] netpoll_send_udp+0x2b8/0x440
[ 10.028176] write_msg+0xfc/0x120 [netconsole]
[ 10.028176] console_unlock+0x3ec/0x6a4
[ 10.028177] register_console+0x17c/0x2f4
[ 10.028178] init_netconsole+0x20c/0x1000 [netconsole]
[ 10.028178] do_one_initcall+0x8c/0x480
[ 10.028179] do_init_module+0x60/0x270
[ 10.028179] load_module+0x21f8/0x2734
[ 10.028180] __do_sys_finit_module+0xbc/0x12c
[ 10.028180] __arm64_sys_finit_module+0x28/0x34
[ 10.028181] el0_svc_common.constprop.0+0x84/0x200
[ 10.028181] do_el0_svc+0x2c/0x90
[ 10.028182] el0_svc+0x18/0x50
[ 10.028182] el0_sync_handler+0xe0/0x350
[ 10.028183] el0_sync+0x158/0x180
[ 10.028183]
[ 10.028183]
[ 10.028184] stack backtrace:
[ 10.028185] CPU: 14 PID: 638 Comm: modprobe Not tainted 5.10.0-rc4_for_upstream_min_debug_2020_11_22_19_37 #1
[ 10.028186] Hardware name: QEMU KVM Virtual Machine, BIOS 0.0.0 02/06/2015
[ 10.028186] Call trace:
[ 10.028186] dump_backtrace+0x0/0x1d0
[ 10.028187] show_stack+0x20/0x3c
[ 10.028187] dump_stack+0xec/0x138
[ 10.028188] check_irq_usage+0x6b8/0x6cc
[ 10.028188] __lock_acquire+0x134c/0x1a94
[ 10.028189] lock_acquire.part.0+0x170/0x360
[ 10.028189] lock_acquire+0x68/0x8c
[ 10.028190] _raw_spin_lock+0x64/0x90
[ 10.028191] virtnet_poll_tx+0x84/0x120
[ 10.028191] netpoll_poll_dev+0x12c/0x350
[ 10.028192] netpoll_send_skb+0x39c/0x400
[ 10.028192] netpoll_send_udp+0x2b8/0x440
[ 10.028193] write_msg+0xfc/0x120 [netconsole]
[ 10.028193] console_unlock+0x3ec/0x6a4
[ 10.028194] register_console+0x17c/0x2f4
[ 10.028194] init_netconsole+0x20c/0x1000 [netconsole]
[ 10.028195] do_one_initcall+0x8c/0x480
[ 10.028195] do_init_module+0x60/0x270
[ 10.028196] load_module+0x21f8/0x2734
[ 10.028197] __do_sys_finit_module+0xbc/0x12c
[ 10.028197] __arm64_sys_finit_module+0x28/0x34
[ 10.028198] el0_svc_common.constprop.0+0x84/0x200
[ 10.028198] do_el0_svc+0x2c/0x90
[ 10.028199] el0_svc+0x18/0x50
[ 10.028199] el0_sync_handler+0xe0/0x350
[ 10.028200] el0_sync+0x158/0x180
[ 10.073569] random: crng init done
[ 10.073964] printk: console [netcon0] enabled
[ 10.074704] random: 7 urandom warning(s) missed due to ratelimiting
[ 10.075340] netconsole: network logging startedQuoted 2 lines
-- Steve
Quoted 2 lines · On Mon, 23 Nov 2020 13:08:55 +0200 Leon Romanovsky <[email protected]> wrote:
[ 10.028024] Chain exists of: [ 10.028025] console_owner --> target_list_lock --> _xmit_ETHER#2
Note, the problem is that we have a location that grabs the xmit_lock while holding target_list_lock (and possibly console_owner).
Quoted 17 lines
[ 10.028028] [ 10.028028] Possible interrupt unsafe locking scenario: [ 10.028029] [ 10.028029] CPU0 CPU1 [ 10.028030] ---- ---- [ 10.028030] lock(_xmit_ETHER#2); [ 10.028032] local_irq_disable(); [ 10.028032] lock(console_owner); [ 10.028034] lock(target_list_lock); [ 10.028035] <Interrupt> [ 10.028035] lock(console_owner); [ 10.028036] [ 10.028037] *** DEADLOCK *** [ 10.028037]
Quoted 9 lines
[ 10.028107] the dependencies between the lock to be acquired
[ 10.028107] and HARDIRQ-irq-unsafe lock:
[ 10.028108] -> (_xmit_ETHER#2){+.-.}-{2:2} ops: 217 {
[ 10.028110] HARDIRQ-ON-W at:
[ 10.028111] __lock_acquire+0x8bc/0x1a94
[ 10.028111] lock_acquire.part.0+0x170/0x360
[ 10.028112] lock_acquire+0x68/0x8c
[ 10.028113] _raw_spin_trylock+0x80/0xd0
[ 10.028113] virtnet_poll+0xac/0x360xmit_lock is taken in virtnet_poll() (via virtnet_poll_cleantx()).
This is called from the softirq, and interrupts are not disabled.
Quoted 27 lines
[ 10.028114] net_rx_action+0x1b0/0x4e0 [ 10.028115] __do_softirq+0x1f4/0x638 [ 10.028115] do_softirq+0xb8/0xcc [ 10.028116] __local_bh_enable_ip+0x18c/0x200 [ 10.028116] virtnet_napi_enable+0xc0/0xd4 [ 10.028117] virtnet_open+0x98/0x1c0 [ 10.028118] __dev_open+0x12c/0x200 [ 10.028118] __dev_change_flags+0x1a0/0x220 [ 10.028119] dev_change_flags+0x2c/0x70 [ 10.028119] do_setlink+0x214/0xe20 [ 10.028120] __rtnl_newlink+0x514/0x820 [ 10.028120] rtnl_newlink+0x58/0x84 [ 10.028121] rtnetlink_rcv_msg+0x184/0x4b4 [ 10.028122] netlink_rcv_skb+0x60/0x124 [ 10.028122] rtnetlink_rcv+0x20/0x30 [ 10.028123] netlink_unicast+0x1b4/0x270 [ 10.028124] netlink_sendmsg+0x1f0/0x400 [ 10.028124] sock_sendmsg+0x5c/0x70 [ 10.028125] ____sys_sendmsg+0x24c/0x280 [ 10.028125] ___sys_sendmsg+0x88/0xd0 [ 10.028126] __sys_sendmsg+0x70/0xd0 [ 10.028127] __arm64_sys_sendmsg+0x2c/0x40 [ 10.028128] el0_svc_common.constprop.0+0x84/0x200 [ 10.028128] do_el0_svc+0x2c/0x90 [ 10.028129] el0_svc+0x18/0x50 [ 10.028129] el0_sync_handler+0xe0/0x350 [ 10.028130] el0_sync+0x158/0x180
[..]
Quoted 4 lines
[ 10.028171] ... key at: [<ffff80001312aef8>] netdev_xmit_lock_key+0x10/0x390 [ 10.028171] ... acquired at: [ 10.028172] __lock_acquire+0x134c/0x1a94 [ 10.028172] lock_acquire.part.0+0x170/0x360
Quoted 8 lines
[ 10.028173] lock_acquire+0x68/0x8c [ 10.028173] _raw_spin_lock+0x64/0x90 [ 10.028174] virtnet_poll_tx+0x84/0x120 [ 10.028174] netpoll_poll_dev+0x12c/0x350 [ 10.028175] netpoll_send_skb+0x39c/0x400 [ 10.028175] netpoll_send_udp+0x2b8/0x440 [ 10.028176] write_msg+0xfc/0x120 [netconsole] [ 10.028176] console_unlock+0x3ec/0x6a4
The above shows the problem. We have:
console_unlock() (which holds the console_owner lock) write_msg() (which holds the target_list_lock)
Then we write_msg() calls:
netpoll_send_udp() {
netpoll_send_skb() {
netpoll_poll_dev() {
virtnet_poll_tx() (which takes the xmit_lock!)DEADLOCK!
In netpoll_send_skb() I see this:
/* tickle device maybe there is some cleanup */ netpoll_poll_dev(np->dev);
Which looks to me that it will call some code that should only be used in softirq context. It's called with locks held that are taken in interrupt context, and any locks that are taken in netpoll_poll_dev() must always be taken with interrupts disabled. That is, if xmit_lock is taken within netpoll_poll_dev(), then it must always be taken with interrupts disabled. Otherwise you can have the deadlock that lockdep reported.
-- Steve
Quoted 40 lines
[ 10.028177] register_console+0x17c/0x2f4 [ 10.028178] init_netconsole+0x20c/0x1000 [netconsole] [ 10.028178] do_one_initcall+0x8c/0x480 [ 10.028179] do_init_module+0x60/0x270 [ 10.028179] load_module+0x21f8/0x2734 [ 10.028180] __do_sys_finit_module+0xbc/0x12c [ 10.028180] __arm64_sys_finit_module+0x28/0x34 [ 10.028181] el0_svc_common.constprop.0+0x84/0x200 [ 10.028181] do_el0_svc+0x2c/0x90 [ 10.028182] el0_svc+0x18/0x50 [ 10.028182] el0_sync_handler+0xe0/0x350 [ 10.028183] el0_sync+0x158/0x180 [ 10.028183] [ 10.028183] [ 10.028184] stack backtrace: [ 10.028185] CPU: 14 PID: 638 Comm: modprobe Not tainted 5.10.0-rc4_for_upstream_min_debug_2020_11_22_19_37 #1 [ 10.028186] Hardware name: QEMU KVM Virtual Machine, BIOS 0.0.0 02/06/2015 [ 10.028186] Call trace: [ 10.028186] dump_backtrace+0x0/0x1d0 [ 10.028187] show_stack+0x20/0x3c [ 10.028187] dump_stack+0xec/0x138 [ 10.028188] check_irq_usage+0x6b8/0x6cc [ 10.028188] __lock_acquire+0x134c/0x1a94 [ 10.028189] lock_acquire.part.0+0x170/0x360 [ 10.028189] lock_acquire+0x68/0x8c [ 10.028190] _raw_spin_lock+0x64/0x90 [ 10.028191] virtnet_poll_tx+0x84/0x120 [ 10.028191] netpoll_poll_dev+0x12c/0x350 [ 10.028192] netpoll_send_skb+0x39c/0x400 [ 10.028192] netpoll_send_udp+0x2b8/0x440 [ 10.028193] write_msg+0xfc/0x120 [netconsole] [ 10.028193] console_unlock+0x3ec/0x6a4 [ 10.028194] register_console+0x17c/0x2f4 [ 10.028194] init_netconsole+0x20c/0x1000 [netconsole] [ 10.028195] do_one_initcall+0x8c/0x480 [ 10.028195] do_init_module+0x60/0x270 [ 10.028196] load_module+0x21f8/0x2734 [ 10.028197] __do_sys_finit_module+0xbc/0x12c [ 10.028197] __arm64_sys_finit_module+0x28/0x34 [ 10.028198] el0_svc_common.constprop.0+0x84/0x200
Quoted 9 lines
[ 10.028198] do_el0_svc+0x2c/0x90 [ 10.028199] el0_svc+0x18/0x50 [ 10.028199] el0_sync_handler+0xe0/0x350 [ 10.028200] el0_sync+0x158/0x180 [ 10.073569] random: crng init done [ 10.073964] printk: console [netcon0] enabled [ 10.074704] random: 7 urandom warning(s) missed due to ratelimiting [ 10.075340] netconsole: network logging started
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Jakub Kicinski
maintainerQuoted 9 lines · On Mon, 23 Nov 2020 09:31:28 -0500 Steven Rostedt wrote:
Quoted 2 lines · On Mon, 23 Nov 2020 13:08:55 +0200 Leon Romanovsky <[email protected]> wrote:
[ 10.028024] Chain exists of: [ 10.028025] console_owner --> target_list_lock --> _xmit_ETHER#2Note, the problem is that we have a location that grabs the
xmit_lockwhile holdingtarget_list_lock(and possiblyconsole_owner).
Well, it try_locks the xmit_lock. Does lockdep understand try-locks?
(not that I condone the shenanigans that are going on here)
Quoted 8 lines · On Mon, 23 Nov 2020 10:52:52 -0800 Jakub Kicinski <[email protected]> wrote:
Quoted 7 lines · On Mon, 23 Nov 2020 09:31:28 -0500 Steven Rostedt wrote:
Quoted 2 lines · On Mon, 23 Nov 2020 13:08:55 +0200 Leon Romanovsky <[email protected]> wrote:
[ 10.028024] Chain exists of: [ 10.028025] console_owner --> target_list_lock --> _xmit_ETHER#2
Quoted 6 lines
Quoted 2 lines
Note, the problem is that we have a location that grabs the
xmit_lockwhile holdingtarget_list_lock(and possiblyconsole_owner).Well, it
try_locksthexmit_lock. Does lockdep understand try-locks?(not that I condone the shenanigans that are going on here)
Does it?
virtnet_poll_tx() {
__netif_tx_lock() {
spin_lock(&txq->_xmit_lock);That looks like we can have:
CPU0 CPU1 ---- ----
lock(xmit_lock)
lock(console) lock(target_list_lock) __netif_tx_lock() lock(xmit_lock); [BLOCKED]
<interrupt>
lock(console)
[BLOCKED]
DEADLOCK.
So where is the trylock here?
Perhaps you need the trylock in virtnet_poll_tx()?
-- Steve
Jakub Kicinski
maintainerQuoted 16 lines · On Mon, 23 Nov 2020 14:09:34 -0500 Steven Rostedt wrote:
Quoted 13 lines · On Mon, 23 Nov 2020 10:52:52 -0800 Jakub Kicinski <[email protected]> wrote:
Quoted 9 lines · On Mon, 23 Nov 2020 09:31:28 -0500 Steven Rostedt wrote:
Quoted 2 lines · On Mon, 23 Nov 2020 13:08:55 +0200 Leon Romanovsky <[email protected]> wrote:
[ 10.028024] Chain exists of: [ 10.028025] console_owner --> target_list_lock --> _xmit_ETHER#2Note, the problem is that we have a location that grabs the
xmit_lockwhile holdingtarget_list_lock(and possiblyconsole_owner).Well, it
try_locksthexmit_lock. Does lockdep understand try-locks?
Quoted 7 lines
Quoted 1 line
(not that I condone the shenanigans that are going on here)
Does it?
virtnet_poll_tx() { __netif_tx_lock() { spin_lock(&txq->_xmit_lock);
Umpf. Right. I was looking at virtnet_poll_cleantx()
Quoted 21 lines
That looks like we can have:
CPU0 CPU1 ---- ----
lock(xmit_lock)lock(console) lock(target_list_lock) __netif_tx_lock() lock(xmit_lock); [BLOCKED]<interrupt>
lock(console)[BLOCKED]
Quoted 6 lines
DEADLOCK.
So where is the trylock here?
Perhaps you need the trylock in
virtnet_poll_tx()?
That could work. Best if we used normal lock if !!budget, and trylock when budget is 0. But maybe that's too hairy.
I'm assuming all this trickiness comes from virtqueue_get_buf() needing locking vs the TX path? It's pretty unusual for the completion path to need locking vs xmit path.
Quoted 44 lines · Jakub Kicinski
Quoted 12 lines · On Mon, 23 Nov 2020 14:09:34 -0500 Steven Rostedt wrote:
Quoted 9 lines · On Mon, 23 Nov 2020 10:52:52 -0800 Jakub Kicinski <[email protected]> wrote:
Quoted 8 lines · On Mon, 23 Nov 2020 09:31:28 -0500 Steven Rostedt wrote:
Quoted 2 lines · On Mon, 23 Nov 2020 13:08:55 +0200 Leon Romanovsky <[email protected]> wrote:
[ 10.028024] Chain exists of: [ 10.028025] console_owner --> target_list_lock --> _xmit_ETHER#2Note, the problem is that we have a location that grabs the
xmit_lockwhile holdingtarget_list_lock(and possiblyconsole_owner).Quoted 8 lines
Quoted 3 lines
Well, it
try_locksthexmit_lock. Does lockdep understand try-locks?(not that I condone the shenanigans that are going on here)
Does it?
virtnet_poll_tx() { __netif_tx_lock() { spin_lock(&txq->_xmit_lock);Umpf. Right. I was looking at
virtnet_poll_cleantx()Quoted 21 lines
That looks like we can have:
CPU0 CPU1 ---- ---- lock(xmit_lock) lock(console) lock(target_list_lock) __netif_tx_lock() lock(xmit_lock); [BLOCKED] <interrupt> lock(console) [BLOCKED]
Quoted 8 lines
Quoted 6 lines
DEADLOCK.
So where is the trylock here?
Perhaps you need the trylock in
virtnet_poll_tx()?That could work. Best if we used normal lock if !!budget, and trylock when budget is 0. But maybe that's too hairy.
If we use trylock, we probably lose(or delay) tx notification that may have side effects to the stack.
Quoted 4 lines
I'm assuming all this trickiness comes from
virtqueue_get_buf()needing locking vs the TX path? It's pretty unusual for the completion path to need locking vs xmit path.
Two reasons for doing this:
1) For some historical reason, we try to free transmitted tx packets in xmit (see free_old_xmit_skbs() in start_xmit()), we can probably remove this if we remove the non tx interrupt mode.
2) virtio core requires virtqueue_get_buf() to be synchronized with virtqueue_add(), we probably can solve this but it requires some non trivial refactoring in the virtio core
Btw, have a quick search, there are several other drivers that uses tx lock in the tx NAPI.
Thanks
Quoted 1 line
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 65 lines · Jason Wang
Quoted 43 lines · Jakub Kicinski
Quoted 11 lines · On Mon, 23 Nov 2020 14:09:34 -0500 Steven Rostedt wrote:
Quoted 8 lines · On Mon, 23 Nov 2020 10:52:52 -0800 Jakub Kicinski <[email protected]> wrote:
Quoted 7 lines · On Mon, 23 Nov 2020 09:31:28 -0500 Steven Rostedt wrote:
Quoted 2 lines · On Mon, 23 Nov 2020 13:08:55 +0200 Leon Romanovsky <[email protected]> wrote:
[ 10.028024] Chain exists of: [ 10.028025] console_owner --> target_list_lock --> _xmit_ETHER#2Note, the problem is that we have a location that grabs the
xmit_lockwhile holdingtarget_list_lock(and possiblyconsole_owner).Quoted 8 lines
Quoted 3 lines
Well, it
try_locksthexmit_lock. Does lockdep understand try-locks?(not that I condone the shenanigans that are going on here)
Does it?
virtnet_poll_tx() { __netif_tx_lock() { spin_lock(&txq->_xmit_lock);Umpf. Right. I was looking at
virtnet_poll_cleantx()Quoted 21 lines
That looks like we can have:
CPU0 CPU1 ---- ---- lock(xmit_lock) lock(console) lock(target_list_lock) __netif_tx_lock() lock(xmit_lock); [BLOCKED] <interrupt> lock(console) [BLOCKED]Quoted 8 lines
Quoted 6 lines
DEADLOCK.
So where is the trylock here?
Perhaps you need the trylock in
virtnet_poll_tx()?That could work. Best if we used normal lock if !!budget, and trylock when budget is 0. But maybe that's too hairy.
If we use trylock, we probably lose(or delay) tx notification that may have side effects to the stack.
Quoted 4 lines
I'm assuming all this trickiness comes from
virtqueue_get_buf()needing locking vs the TX path? It's pretty unusual for the completion path to need locking vs xmit path.
Quoted 8 lines
Two reasons for doing this:
1) For some historical reason, we try to free transmitted tx packets in xmit (see
free_old_xmit_skbs()instart_xmit()), we can probably remove this if we remove the non tx interrupt mode.
2) virtio core requiresvirtqueue_get_buf()to be synchronized withvirtqueue_add(), we probably can solve this but it requires some non trivial refactoring in the virtio core
So how will we solve our lockdep issues?
Thanks
Quoted 8 lines
Btw, have a quick search, there are several other drivers that uses tx lock in the tx NAPI.
Thanks
Quoted 1 line
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 70 lines · Leon Romanovsky
Quoted 61 lines · Jason Wang
Quoted 43 lines · Jakub Kicinski
Quoted 11 lines · On Mon, 23 Nov 2020 14:09:34 -0500 Steven Rostedt wrote:
Quoted 8 lines · On Mon, 23 Nov 2020 10:52:52 -0800 Jakub Kicinski <[email protected]> wrote:
Quoted 7 lines · On Mon, 23 Nov 2020 09:31:28 -0500 Steven Rostedt wrote:
Quoted 2 lines · On Mon, 23 Nov 2020 13:08:55 +0200 Leon Romanovsky <[email protected]> wrote:
[ 10.028024] Chain exists of: [ 10.028025] console_owner --> target_list_lock --> _xmit_ETHER#2Note, the problem is that we have a location that grabs the
xmit_lockwhile holdingtarget_list_lock(and possiblyconsole_owner).Quoted 8 lines
Quoted 3 lines
Well, it
try_locksthexmit_lock. Does lockdep understand try-locks?(not that I condone the shenanigans that are going on here)
Does it?
virtnet_poll_tx() { __netif_tx_lock() { spin_lock(&txq->_xmit_lock);Umpf. Right. I was looking at
virtnet_poll_cleantx()Quoted 21 lines
That looks like we can have:
CPU0 CPU1 ---- ---- lock(xmit_lock) lock(console) lock(target_list_lock) __netif_tx_lock() lock(xmit_lock); [BLOCKED] <interrupt> lock(console) [BLOCKED]Quoted 8 lines
Quoted 6 lines
DEADLOCK.
So where is the trylock here?
Perhaps you need the trylock in
virtnet_poll_tx()?That could work. Best if we used normal lock if !!budget, and trylock when budget is 0. But maybe that's too hairy.
If we use trylock, we probably lose(or delay) tx notification that may have side effects to the stack.
Quoted 3 lines
I'm assuming all this trickiness comes from
virtqueue_get_buf()needing locking vs the TX path? It's pretty unusual for the completion path to need locking vs xmit path.Quoted 8 lines
Two reasons for doing this:
1) For some historical reason, we try to free transmitted tx packets in xmit (see
free_old_xmit_skbs()instart_xmit()), we can probably remove this if we remove the non tx interrupt mode.
2) virtio core requiresvirtqueue_get_buf()to be synchronized withvirtqueue_add(), we probably can solve this but it requires some non trivial refactoring in the virtio core
Quoted 3 lines
So how will we solve our lockdep issues?
Thanks
It's not clear to me that whether it's a virtio-net specific issue. E.g the above deadlock looks like a generic issue so workaround it via virtio-net may not help for other drivers.
Thanks
Quoted 6 lines
Quoted 6 lines
Btw, have a quick search, there are several other drivers that uses tx lock in the tx NAPI.
Thanks
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 72 lines · Jason Wang
Quoted 70 lines · Leon Romanovsky
Quoted 69 lines · Jason Wang
Quoted 43 lines · Jakub Kicinski
Quoted 11 lines · On Mon, 23 Nov 2020 14:09:34 -0500 Steven Rostedt wrote:
Quoted 8 lines · On Mon, 23 Nov 2020 10:52:52 -0800 Jakub Kicinski <[email protected]> wrote:
Quoted 7 lines · On Mon, 23 Nov 2020 09:31:28 -0500 Steven Rostedt wrote:
Quoted 2 lines · On Mon, 23 Nov 2020 13:08:55 +0200 Leon Romanovsky <[email protected]> wrote:
[ 10.028024] Chain exists of: [ 10.028025] console_owner --> target_list_lock --> _xmit_ETHER#2Note, the problem is that we have a location that grabs the
xmit_lockwhile holdingtarget_list_lock(and possiblyconsole_owner).Quoted 8 lines
Quoted 3 lines
Well, it
try_locksthexmit_lock. Does lockdep understand try-locks?(not that I condone the shenanigans that are going on here)
Does it?
virtnet_poll_tx() { __netif_tx_lock() { spin_lock(&txq->_xmit_lock);Umpf. Right. I was looking at
virtnet_poll_cleantx()Quoted 21 lines
That looks like we can have:
CPU0 CPU1 ---- ---- lock(xmit_lock) lock(console) lock(target_list_lock) __netif_tx_lock() lock(xmit_lock); [BLOCKED] <interrupt> lock(console) [BLOCKED]Quoted 8 lines
Quoted 6 lines
DEADLOCK.
So where is the trylock here?
Perhaps you need the trylock in
virtnet_poll_tx()?That could work. Best if we used normal lock if !!budget, and trylock when budget is 0. But maybe that's too hairy.
If we use trylock, we probably lose(or delay) tx notification that may have side effects to the stack.
Quoted 3 lines
I'm assuming all this trickiness comes from
virtqueue_get_buf()needing locking vs the TX path? It's pretty unusual for the completion path to need locking vs xmit path.Two reasons for doing this:
1) For some historical reason, we try to free transmitted tx packets in xmit (see
free_old_xmit_skbs()instart_xmit()), we can probably remove this if we remove the non tx interrupt mode.
2) virtio core requiresvirtqueue_get_buf()to be synchronized withvirtqueue_add(), we probably can solve this but it requires some non trivial refactoring in the virtio core
Quoted 8 lines
Quoted 3 lines
So how will we solve our lockdep issues?
Thanks
It's not clear to me that whether it's a virtio-net specific issue. E.g the above deadlock looks like a generic issue so workaround it via virtio-net may not help for other drivers.
It is hard to say, no one else complained except me who is using virtio :).
Thanks
Quoted 11 lines
Thanks
Quoted 6 lines
Quoted 6 lines
Btw, have a quick search, there are several other drivers that uses tx lock in the tx NAPI.
Thanks
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Quoted 2 lines · On Tue, 24 Nov 2020 11:22:03 +0800 Jason Wang <[email protected]> wrote:
Btw, have a quick search, there are several other drivers that uses tx lock in the tx NAPI.
tx NAPI is not the issue. The issue is that write_msg() (in netconsole.c) calls this polling logic with the target_list_lock held.
Are those other drivers called by netconsole? If not, then this is unique to virtio-net.
-- Steve
Jakub Kicinski
maintainerQuoted 6 lines · On Tue, 24 Nov 2020 11:22:03 +0800 Jason Wang wrote:
Quoted 3 lines
Quoted 1 line
Perhaps you need the trylock in
virtnet_poll_tx()?That could work. Best if we used normal lock if !!budget, and trylock when budget is 0. But maybe that's too hairy.
If we use trylock, we probably lose(or delay) tx notification that may have side effects to the stack.
That's why I said only trylock with budget == 0. Only netpoll calls with budget == 0, AFAIK.
Quoted 13 lines
Quoted 3 lines
I'm assuming all this trickiness comes from
virtqueue_get_buf()needing locking vs the TX path? It's pretty unusual for the completion path to need locking vs xmit path.Two reasons for doing this:
1) For some historical reason, we try to free transmitted tx packets in xmit (see
free_old_xmit_skbs()instart_xmit()), we can probably remove this if we remove the non tx interrupt mode.
2) virtio core requiresvirtqueue_get_buf()to be synchronized withvirtqueue_add(), we probably can solve this but it requires some non trivial refactoring in the virtio core
Quoted 2 lines
Btw, have a quick search, there are several other drivers that uses tx lock in the tx NAPI.
Unless they do:
netdev->priv_flags |= IFF_DISABLE_NETPOLL;
they are all broken.
Quoted 7 lines · Steven Rostedt
Quoted 2 lines · On Tue, 24 Nov 2020 11:22:03 +0800 Jason Wang <[email protected]> wrote:
Btw, have a quick search, there are several other drivers that uses tx lock in the tx NAPI.
tx NAPI is not the issue. The issue is that
write_msg()(innetconsole.c) calls this polling logic with thetarget_list_lockheld.
But in the tx NAPI poll it tries to lock TX instead of using trylock.
Quoted 3 lines
Are those other drivers called by netconsole? If not, then this is unique to virtio-net.
I think the answer is yes, since net console is not disabled in the codes.
Thanks
Quoted 3 lines
-- Steve
Quoted 8 lines · Jakub Kicinski
Quoted 5 lines · On Tue, 24 Nov 2020 11:22:03 +0800 Jason Wang wrote:
Quoted 3 lines
Quoted 1 line
Perhaps you need the trylock in
virtnet_poll_tx()?That could work. Best if we used normal lock if !!budget, and trylock when budget is 0. But maybe that's too hairy.
If we use trylock, we probably lose(or delay) tx notification that may have side effects to the stack.
That's why I said only trylock with budget == 0. Only netpoll calls with budget == 0, AFAIK.
Oh right.
So I think maybe we can switch to use trylock when budget is zero and try to schedule another TX NAPI if we trylock fail.
Quoted 13 lines
Quoted 13 lines
Quoted 4 lines
I'm assuming all this trickiness comes from
virtqueue_get_buf()needing locking vs the TX path? It's pretty unusual for the completion path to need locking vs xmit path.Two reasons for doing this:
1) For some historical reason, we try to free transmitted tx packets in xmit (see
free_old_xmit_skbs()instart_xmit()), we can probably remove this if we remove the non tx interrupt mode.
2) virtio core requiresvirtqueue_get_buf()to be synchronized withvirtqueue_add(), we probably can solve this but it requires some non trivial refactoring in the virtio core
Quoted 7 lines
Quoted 2 lines
Btw, have a quick search, there are several other drivers that uses tx lock in the tx NAPI.
Unless they do:
netdev->priv_flags |= IFF_DISABLE_NETPOLL;they are all broken.
Yes.
Thanks