Re: Urgent Bug report: PPPoE ioctl(PPPIOCCONNECT): Transport endpoint is not connected
From: Martin Zaharinov <hidden>
Date: 2021-09-16 20:00:41
Small Updates
After switch from frr to bird bgp reduce load from frr
but still when have disconnect 5k+ users have slow pppoe_flush_dev
PerfTop: 15606 irqs/sec kernel:77.7% exact: 100.0% lost: 0/0 drop: 0/0 [4000Hz cycles], (all, 12 CPUs)
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
8.24% [kernel] [k] osq_lock
7.55% [kernel] [k] mutex_spin_on_owner
7.04% [pppoe] [k] pppoe_flush_dev
2.77% libteam.so.5.6.1 [.] 0x0000000000009e7d
2.67% libteam.so.5.6.1 [.] 0x0000000000006470
1.90% [kernel] [k] fib_table_flush
1.73% [kernel] [k] queued_read_lock_slowpath
1.68% [kernel] [k] next_uptodate_page
1.36% ip [.] 0x0000000000011b74
1.23% ip [.] 0x00000000000121b0
1.09% [kernel] [k] zap_pte_range
0.99% libteam.so.5.6.1 [.] 0x0000000000006501
0.88% dtvbras [.] 0x0000000000014be8
0.87% [kernel] [k] inet_dump_ifaddr
0.74% [kernel] [k] filemap_map_pages
0.72% [kernel] [k] neigh_flush_dev.isra.0
0.66% [kernel] [k] snmp_get_cpu_field
0.65% [kernel] [k] fib_table_insert
0.63% [kernel] [k] native_irq_return_iret
0.63% libteam.so.5.6.1 [.] 0x0000000000005c78
0.60% [kernel] [k] copy_page
0.52% libteam.so.5.6.1 [.] 0x000000000000647f
0.50% [kernel] [k] _raw_spin_lock
0.48% libc.so.6 [.] 0x00000000000965a2
0.45% [kernel] [k] _raw_read_lock_bh
0.44% [kernel] [k] release_pages
0.42% [kernel] [k] clear_page_erms
0.42% [kernel] [k] page_remove_rmap
0.41% [kernel] [k] queued_spin_lock_slowpath
0.38% [kernel] [k] kmem_cache_alloc
0.36% [kernel] [k] vma_interval_tree_insert
0.36% libteam.so.5.6.1 [.] 0x0000000000009e6f
0.36% [kernel] [k] do_set_pte
sessions:
starting: 296
active: 3868
finishing: 6748
On 14 Sep 2021, at 14:00, Florian Westphal [off-list ref] wrote: Martin Zaharinov [off-list ref] wrote: [ Trimming CC list ]quoted
Florian: If you make patch send to test please.Attached. No idea if it helps, but 'ip' should stay responsive even when masquerade processes netdevice events. <defer_masq_work.diff>