On Tue, Mar 31, 2020 at 10:19:36PM +0000, Bryan Stenson wrote:
> I started "doas ktrace -p <pid_of_pflogd>" after boot.
>
> However, your clue about "...quotaon(8) could miss the trace file
> vnode..." seems relevant here. I did use "doas quotaoff /home" and
> "doas quotaon -u /home" while proving to myself how the quota system
> works.
>
> The kernel didn't immediately panic after turning quotas on/off, so I
> didn't think it was related.
>
> I'd love to be able to help test/prove this patch, but I haven't been
> able to reproduce it. I've been running a script to
> create/delete/modify files by the user with quota limits. At the same
> time, I've run another script to enable/disable quotas, hoping to
> trigger the bug...but I haven't been able to reproduce at all.
You should be able to reproduce the panic by starting a process with
ktrace, followed by enabling of quotas on the file system where the
trace file is being collected.
# start top(1) to generate system calls
ktrace -f $FS_WITH_QUOTAS/ktrace.out top
# on another console
doas quotaoff -av
doas quotaon -av
Below is an updated diff that fixes a bug in the previous version.
Index: kern/kern_ktrace.c
===================================================================
RCS file: src/sys/kern/kern_ktrace.c,v
retrieving revision 1.102
diff -u -p -r1.102 kern_ktrace.c
--- kern/kern_ktrace.c 23 Mar 2020 15:45:39 -0000 1.102
+++ kern/kern_ktrace.c 1 Apr 2020 15:32:55 -0000
@@ -83,6 +83,7 @@ ktrcleartrace(struct process *pr)
pr->ps_tracevp = NULL;
pr->ps_tracecred = NULL;
+ vp->v_writecount--;
vrele(vp);
crfree(cred);
}
@@ -109,6 +110,7 @@ ktrsettrace(struct process *pr, int facs
vref(newvp);
crhold(newcred);
+ newvp->v_writecount++;
oldvp = pr->ps_tracevp;
oldcred = pr->ps_tracecred;
@@ -117,6 +119,7 @@ ktrsettrace(struct process *pr, int facs
pr->ps_tracecred = newcred;
if (oldvp != NULL) {
+ oldvp->v_writecount--;
vrele(oldvp);
crfree(oldcred);
}
