expand scope of da refcount to protect I/Os in flight

[chs]

Summary

When a disk disappears and the periph is invalidated, any I/Os that are pending with the controller can cause a crash when they complete. This diff holds the softc reference count taken in dastart() until the I/O is complete rather than only until xpt_action() returns. (This approach was suggested by Ken Merry.)

Note that the ada and nda drivers likely need similar changes, this addresses only the da driver since that's all I can test so far.

Test Plan

This patch has been tested with script that repeated uses lsiutil to repeated cause a drive to disappear while a bunch of I/O in progress. With the patch the crashes in the CAM code are gone (though panics in other subsystems are also triggered, I'll submit patches for those as well).

[imp]

Looks good to me. One tiny question that occurred to me after I looked at it this morning. I think we're fine delaying the close for up to a tick so we don't have to try to send a bunch of wake ups during normal operation. Also did the obligatory lock check, and we're good there as well: refcount is only changed while holding the periph lock, and cam_periph_sleep uses the same lock to sleep on.

FreeBSD/sys/cam/scsi/scsi_da.c

4409 ↗

(On Diff #42548)

in daclose we have:

while (softc->refcount != 0)
        cam_periph_sleep(periph, &softc->refcount, PRIBIO, "daclose", 1);

do we need to wake that up, or just rely on the timeout?

[ken]

Looks good. I'm glad this seems to have fixed the problem in the da(4) driver at least. I agree, ada(4) and nda(4) should probably have the same change since they'll have the same issue.

FreeBSD/sys/cam/scsi/scsi_da.c

4409 ↗

(On Diff #42548)

I guess you could do either. With a 1/hz timeout, this turns into something of a busy wait.