Re: xen/evtchn: Interrupt for port 34, but apparently not enabled; per-user 00000000a86a4c1b on 5.10

["Jürgen Groß" <jgross@suse.com>]

[-- Attachment #1.1.1: Type: text/plain, Size: 4772 bytes --]

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.


Juergen

[-- Attachment #1.1.2: OpenPGP_0xB0DE9DD628BF132F.asc --]
[-- Type: application/pgp-keys, Size: 3135 bytes --]

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 495 bytes --]