On 15.12.20 08:27, Jürgen Groß wrote:
> On 14.12.20 22:25, Julien Grall wrote:
>> Hi Juergen,
>>
>> When testing Linux 5.10 dom0, I could reliably hit the following 
>> warning with using event 2L ABI:
>>
>> [  589.591737] Interrupt for port 34, but apparently not enabled; 
>> per-user 00000000a86a4c1b
>> [  589.593259] WARNING: CPU: 0 PID: 1111 at 
>> /home/ANT.AMAZON.COM/jgrall/works/oss/linux/drivers/xen/evtchn.c:170 
>> evtchn_interrupt+0xeb/0x100
>> [  589.595514] Modules linked in:
>> [  589.596145] CPU: 0 PID: 1111 Comm: qemu-system-i38 Tainted: G 
>> W         5.10.0+ #180
>> [  589.597708] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), 
>> BIOS rel-1.12.0-59-gc9ba5276e321-prebuilt.qemu.org 04/01/2014
>> [  589.599782] RIP: e030:evtchn_interrupt+0xeb/0x100
>> [  589.600698] Code: 48 8d bb d8 01 00 00 ba 01 00 00 00 be 1d 00 00 
>> 00 e8 d9 10 ca ff eb b2 8b 75 20 48 89 da 48 c7 c7 a8 31 3d 82 e8 65 
>> 29 a0 ff <0f> 0b e9 42 ff ff ff 0f 1f 40 00 66 2e 0f 1f 84 00 00 00 00 
>> 00 0f
>> [  589.604087] RSP: e02b:ffffc90040003e70 EFLAGS: 00010086
>> [  589.605102] RAX: 0000000000000000 RBX: ffff888102091800 RCX: 
>> 0000000000000027
>> [  589.606445] RDX: 0000000000000000 RSI: ffff88817fe19150 RDI: 
>> ffff88817fe19158
>> [  589.607790] RBP: ffff88810f5ab980 R08: 0000000000000001 R09: 
>> 0000000000328980
>> [  589.609134] R10: 0000000000000000 R11: ffffc90040003c70 R12: 
>> ffff888107fd3c00
>> [  589.610484] R13: ffffc90040003ed4 R14: 0000000000000000 R15: 
>> ffff88810f5ffd80
>> [  589.611828] FS:  00007f960c4b8ac0(0000) GS:ffff88817fe00000(0000) 
>> knlGS:0000000000000000
>> [  589.613348] CS:  10000e030 DS: 0000 ES: 0000 CR0: 0000000080050033
>> [  589.614525] CR2: 00007f17ee72e000 CR3: 000000010f5b6000 CR4: 
>> 0000000000050660
>> [  589.615874] Call Trace:
>> [  589.616402]  <IRQ>
>> [  589.616855]  __handle_irq_event_percpu+0x4e/0x2c0
>> [  589.617784]  handle_irq_event_percpu+0x30/0x80
>> [  589.618660]  handle_irq_event+0x3a/0x60
>> [  589.619428]  handle_edge_irq+0x9b/0x1f0
>> [  589.620209]  generic_handle_irq+0x4f/0x60
>> [  589.621008]  evtchn_2l_handle_events+0x160/0x280
>> [  589.621913]  __xen_evtchn_do_upcall+0x66/0xb0
>> [  589.622767]  __xen_pv_evtchn_do_upcall+0x11/0x20
>> [  589.623665]  asm_call_irq_on_stack+0x12/0x20
>> [  589.624511]  </IRQ>
>> [  589.624978]  xen_pv_evtchn_do_upcall+0x77/0xf0
>> [  589.625848]  exc_xen_hypervisor_callback+0x8/0x10
>>
>> This can be reproduced when creating/destroying guest in a loop. 
>> Although, I have struggled to reproduce it on a vanilla Xen.
>>
>> After several hours of debugging, I think I have found the root cause.
>>
>> While we only expect the unmask to happen when the event channel is 
>> EOIed, there is an unmask happening as part of handle_edge_irq() 
>> because the interrupt was seen as pending by another vCPU 
>> (IRQS_PENDING is set).
>>
>> It turns out that the event channel is set for multiple vCPU is in 
>> cpu_evtchn_mask. This is happening because the affinity is not cleared 
>> when freeing an event channel.
>>
>> The implementation of evtchn_2l_handle_events() will look for all the 
>> active interrupts for the current vCPU and later on clear the pending 
>> bit (via the ack() callback). IOW, I believe, this is not an atomic 
>> operation.
>>
>> Even if Xen will notify the event to a single vCPU, evtchn_pending_sel 
>> may still be set on the other vCPU (thanks to a different event 
>> channel). Therefore, there is a chance that two vCPUs will try to 
>> handle the same interrupt.
>>
>> The IRQ handler handle_edge_irq() is able to deal with that and will 
>> mask/unmask the interrupt. This will mess us with the lateeoi logic 
>> (although, I managed to reproduce it once without XSA-332).
> 
> Thanks for the analysis!
> 
>> My initial idea to fix the problem was to switch the affinity from CPU 
>> X to CPU0 when the event channel is freed.
>>
>> However, I am not sure this is enough because I haven't found anything 
>> yet preventing a race between evtchn_2l_handle_events9) and 
>> evtchn_2l_bind_vcpu().
>>
>> So maybe we want to introduce a refcounting (if there is nothing 
>> provided by the IRQ framework) and only unmask when the counter drop 
>> to 0.
>>
>> Any opinions?
> 
> I think we don't need a refcount, but just the internal states "masked"
> and "eoi_pending" and unmask only if both are false. "masked" will be
> set when the event is being masked. When delivering a lateeoi irq
> "eoi_pending" will be set and "masked "reset. "masked" will be reset
> when a normal unmask is happening. And "eoi_pending" will be reset
> when a lateeoi is signaled. Any reset of "masked" and "eoi_pending"
> will check the other flag and do an unmask if both are false.
> 
> I'll write a patch.

Julien, could you please test the attached (only build tested) patch?


Juergen