Fix FFS panic on ENXIO.

[trasz]

Summary

Fix FFS panic that happens after the mounted device went away with
softupdates enabled. Backtrace would look like this:

#11 0xffffffff80a50063 in panic (fmt=0xffffffff81c6e7f0 <cnputs_mtx+24> "\004") at /usr/home/trasz/svn-ssh/android/sys/kern/kern_shutdown.c:710
#12 0xffffffff80d05dfe in softdep_deallocate_dependencies (bp=0xfffffe01ea78cc40) at /usr/home/trasz/svn-ssh/android/sys/ufs/ffs/ffs_softdep.c:14257
#13 0xffffffff80af662c in buf_deallocate (bp=0xfffffe01ea78cc40) at /usr/home/trasz/svn-ssh/android/sys/sys/buf.h:429
#14 brelse (bp=0xfffffe01ea78cc40) at /usr/home/trasz/svn-ssh/android/sys/kern/vfs_bio.c:2318
#15 0xffffffff80b14b57 in flushbuflist (bufv=<optimized out>, flags=<optimized out>, bo=<optimized out>, slpflag=<optimized out>, slptimeo=0) at /usr/home/trasz/svn-ssh/android/sys/kern/vfs_subr.c:1779
#16 0xffffffff80b147af in bufobj_invalbuf (bo=0xfffff8007e334658, flags=0, slpflag=0, slptimeo=0) at /usr/home/trasz/svn-ssh/android/sys/kern/vfs_subr.c:1661
#17 0xffffffff80b17dde in vinvalbuf (vp=<optimized out>, flags=0, slpflag=0, slptimeo=0) at /usr/home/trasz/svn-ssh/android/sys/kern/vfs_subr.c:1721
#18 vgonel (vp=0xfffff8007e334588) at /usr/home/trasz/svn-ssh/android/sys/kern/vfs_subr.c:3353
#19 0xffffffff80b1838f in vgone (vp=0xfffff8007e334588) at /usr/home/trasz/svn-ssh/android/sys/kern/vfs_subr.c:3229
#20 0xffffffff8091c654 in devfs_delete (dm=0xfffff8000d71ad80, de=0xfffff80004565100, flags=<optimized out>) at /usr/home/trasz/svn-ssh/android/sys/fs/devfs/devfs_devs.c:393

[kib]

Can you explain how this patch fixes things (instead of papering over and allowing corruption or panic later) ?

From my understanding of the situation where your patch makes a difference, the buffer' dependencies are still not handled, and kernel will get its data corrupted when trying to write buffers which depend on the buffer you hacked on: the depended on buffer is destroyed and will not be found.

Also, why do you allow B_INVAL but not B_NOCACHE ?

[trasz]

The patch fixes an instapanic. Granted, the buffer won't get correctly handled later on - but that's a separate issue, and can't really be addressed until this one is fixed. That said, I don't quite follow: which buffer gets destroyed?

No idea about B_NOCACHE. Should I just add that, ie s/B_INVAL/B_INVAL | B_NOCACHE/?

[kib]

In D10653#220964, @trasz wrote:

The patch fixes an instapanic.

No, it does not. It incorrectly papers over ot.

Granted, the buffer won't get correctly handled later on - but that's a separate issue, and can't really be addressed until this one is fixed. That said, I don't quite follow: which buffer gets destroyed?

The buffer for which the call to softdep_deallocate_dependencies() is performed.

No idea about B_NOCACHE. Should I just add that, ie s/B_INVAL/B_INVAL | B_NOCACHE/?

Read the else if () condition in the brelse() code path which results in buf_deallocate() call.

[trasz]

Sigh.

Okay. Do you have any suggestions on what should happen in this case? How the buf should be handled?

[kib]

In D10653#220977, @trasz wrote:

Sigh.

Okay. Do you have any suggestions on what should happen in this case? How the buf should be handled?

We already discussed this in detail in D9674, see https://reviews.freebsd.org/D9674#201637 and below for the summary.