Re: [RFC] hw/scsi: avoid deadlock upon TMF request canceling with VirtIO

[Fiona Ebner <f.ebner@proxmox.com>]

Hi Stefan,

Am 16.10.25 um 8:25 PM schrieb Stefan Hajnoczi:
> On Wed, Oct 15, 2025 at 03:43:50PM +0200, Fiona Ebner wrote:
>> RFC, because it's a naive approach, maybe somebody has a better idea?
> 
> Another approach is the one taken in scsi_device_for_each_req_async_bh()
> where requests from this AioContext are collected into a GList and then
> processed after releasing s->requests_lock. It's safe because the
> function runs in the request's AioContext and we know nothing else can
> modify the request while we are running. The same constraint applies in
> this case too.
> 
> That solution is more localized because various function prototypes
> don't need to be extended with holds_requests_lock. Either it can be
> open coded inside virtio_scsi_do_tmf_aio_context() or you could extract
> a helper function from scsi_device_for_each_req_async_bh() in scsi-bus.c
> and call the new helper from virtio_scsi_do_tmf_aio_context().
> 
> I slightly prefer a localized fix so that other parts of the codebase
> don't need to worry about whether or not requests_lock is held. Do you
> want to try implementing that?
thank you for the suggestion and pointers! I also feel like a localized
solution would be nicer. I'll try this approach.

Kind Regards,
Fiona