Re: [PATCH 3/4] cxl: add a firmware update mechanism using the sysfs firmware loader

[Dan Williams]

Verma, Vishal L wrote:
> On Mon, 2023-05-22 at 20:21 -0700, Dan Williams wrote:
> > Vishal Verma wrote:
> 
> <snip>
> Everything else not addressed here sounds good and I've made those
> changes.
> 
> > > 
> > > +       remaining = size - cur_size;
> > > +       size_in = cur_size + sizeof(*transfer);
> > > +
> > > +       mutex_lock(&cxlds->fw.fw_mutex);
> > 
> > What is this lock protecting? I.e. will the fw_loader really try to send
> > multiple overlapping firmware update attempts?
> 
> The lock is just to provide predictable points at which a cancel
> request may be intercepted. The loader won't try overlapping firmware
> transfer requests, but the ->cancel request comes from user space, and
> could happen while there is a transfer in progress. With the lock, the
> cancel will only be 'processed' after the current chunk's transfer is
> done. 

So right now cancel is only considered at certain points within either
the ->write() or ->poll_complete() callbacks. The firmware upload core
is guaranteeing that ->prepare(), ->write() and ->poll_complete() never
overlap for a given session, and that if any of those return an error
the upload session is terminated.

While the lock does flush in flight ->write() and ->prepare() it does
nothing to enforce when the cancellation is processed. It will still be
the case that the next invocation of ->write() or ->poll_complete() will
consider the cancel state before doing the next step.

I am failing to see what the lock is protecting. The other usage is for
checking that ->prepare() has completed before ->write() is invoked, but
again that is enforced by the firmware uploader workqueue.

I think the lock and the clear_to_send bit can be eliminated. Clear to
send is implied by ->prepare() succeeding. Convert cancel to an atomic
flag where cxl_fw_cancel() does:

set_bit(CXL_FW_CANCEL, &cxlds->fw.state);

...and cxl_fw_write() and cxl_fw_poll_complete() can just do:

if (test_and_clear_bit(CXL_FW_CANCEL, &cxlds->fw.state))
	do_cancel();