On Fri, Oct 02, 2020 at 12:35:17PM +0200, Sebastian Andrzej Siewior wrote:
> I was trying to get rid of the in in_softirq() in ezusb_req_ctx_wait()
> within the orinoco usb driver,
> drivers/net/wireless/intersil/orinoco/orinoco_usb.c. A small snippet:
>
> | static void ezusb_req_ctx_wait(struct ezusb_priv *upriv,
> | struct request_context *ctx)
> …
> | if (in_softirq()) {
> | /* If we get called from a timer, timeout timers don't
> | * get the chance to run themselves. So we make sure
> | * that we don't sleep for ever */
> | int msecs = DEF_TIMEOUT * (1000 / HZ);
> |
> | while (!try_wait_for_completion(&ctx->done) && msecs--)
> | udelay(1000);
> | } else {
> | wait_for_completion(&ctx->done);
> …
> | }
>
> This is broken. The EHCI and XHCI HCD will complete the URB in
> BH/tasklet. Should we ever get here in_softirq() then we will spin
> here/wait here until the timeout passes because the tasklet won't be
> able to run. OHCI/UHCI HCDs still complete in hard-IRQ so it would work
> here.
>
> Is it possible to end up here in softirq context or is this a relic?
I think it's a relic of where USB host controllers completed their urbs
in hard-irq mode. The BH/tasklet change is a pretty recent change.
> Well I have no hardware but I see this:
>
> orinoco_set_monitor_channel() [I assume that this is fully preemtible]
> -> orinoco_lock() [this should point to ezusb_lock_irqsave() which
> does spin_lock_bh(lock), so from here on
> in_softirq() returns true]
> -> hw->ops->cmd_wait() [-> ezusb_docmd_wait()]
> -> ezusb_alloc_ctx() [ sets ctx->in_rid to EZUSB_RID_ACK/0x0710 ]
> -> ezusb_access_ltv()
> -> if (ctx->in_rid)
> -> ezusb_req_ctx_wait(upriv, ctx);
> -> ctx->state should be EZUSB_CTX_REQ_COMPLETE so we end up in
> the while loop above. So we udelay() 3 * 1000 * 1ms = 3sec.
> -> Then ezusb_access_ltv() should return with an error due to
> timeout.
>
> This isn't limited to exotic features like monitor mode. orinoco_open()
> does orinoco_lock() followed by orinoco_hw_program_rids() which in the
> end invokes ezusb_write_ltv(,, EZUSB_RID_ACK) which is non-zero and also
> would block (ezusb_xmit() would use 0 as the last argument so it won't
> block).
>
> I don't see how this driver can work on EHCI/XHCI HCD as of today.
> The driver is an orphan since commit
> 3a59babbee409 ("orinoco: update status in MAINTAINERS")
>
> which is ten years ago. If I replace in_softirq() with a `may_sleep'
> argument then it is still broken.
> Should it be removed?
We can move it out to drivers/staging/ and then drop it to see if anyone
complains that they have the device and is willing to test any changes.
thanks,
greg k-h
