Re: NFSv4 crash of CURRENT

On Mon, Jan 15, 2024 at 2:53 AM Peter Blok <pblok@bsd4all.org> wrote:
>
> Hi,
>
> Forgot to mention I’m on 13-stable. The fix that is causing the crash with automounted NFS is:
>
> commit cc5cda1dbaa907ce52074f47264cc45b5a7d6c8b
> Author: Konstantin Belousov <kib@FreeBSD.org>
> Date:   Tue Jan 2 00:22:44 2024 +0200
>
>     nfsclient: limit situations when we do unlocked read-ahead by nfsiod
>
>     (cherry picked from commit 70dc6b2ce314a0f32755005ad02802fca7ed186e)
>
> When I remove the fix, the problem is gone. Add it back and the crash happens.
Kostik has already come up with a probable fix. If you want it right
away, here it is,
but he'll probably commit it soon anyhow:
diff --git a/sys/fs/nfsclient/nfs_clbio.c b/sys/fs/nfsclient/nfs_clbio.c
index c027d7d7c3fd..1cf45bb0c924 100644
--- a/sys/fs/nfsclient/nfs_clbio.c
+++ b/sys/fs/nfsclient/nfs_clbio.c
@@ -414,6 +414,18 @@ nfs_bioread_check_cons(struct vnode *vp, struct
thread *td, struct ucred *cred)
        return (error);
 }

+static bool
+ncl_bioread_dora(struct vnode *vp)
+{
+       vm_object_t obj;
+
+       obj = vp->v_object;
+       if (obj == NULL)
+               return (true);
+       return (!vm_object_mightbedirty(vp->v_object) &&
+           vp->v_object->un_pager.vnp.writemappings == 0);
+}
+
 /*
  * Vnode op for read using bio
  */
@@ -486,9 +498,7 @@ ncl_bioread(struct vnode *vp, struct uio *uio, int
ioflag, struct ucred *cred)
                 * unlocked read by nfsiod could obliterate changes
                 * done by userspace.
                 */
-               if (nmp->nm_readahead > 0 &&
-                   !vm_object_mightbedirty(vp->v_object) &&
-                   vp->v_object->un_pager.vnp.writemappings == 0) {
+               if (nmp->nm_readahead > 0 && ncl_bioread_dora(vp)) {
                    for (nra = 0; nra < nmp->nm_readahead && nra < seqcount &&
                        (off_t)(lbn + 1 + nra) * biosize < nsize; nra++) {
                        rabn = lbn + 1 + nra;
@@ -675,9 +685,7 @@ ncl_bioread(struct vnode *vp, struct uio *uio, int
ioflag, struct ucred *cred)
                 *  directory offset cookie of the next block.)
                 */
                NFSLOCKNODE(np);
-               if (nmp->nm_readahead > 0 &&
-                   !vm_object_mightbedirty(vp->v_object) &&
-                   vp->v_object->un_pager.vnp.writemappings == 0 &&
+               if (nmp->nm_readahead > 0 && ncl_bioread_dora(vp) &&
                    (bp->b_flags & B_INVAL) == 0 &&
                    (np->n_direofoffset == 0 ||
                    (lbn + 1) * NFS_DIRBLKSIZ < np->n_direofoffset) &&

rick
ps: It appears that autofs causes the directory to be read before it
is open'd for
      some reason. I've never looked at autofs.

>
> Peter
>
> On 15 Jan 2024, at 09:31, Peter Blok <pblok@bsd4all.org> wrote:
>
> Hi,
>
> I do have a crash on a NFS client with stable of today (4c4633fdffbe8e4b6d328c2bc9bb3edacc9ab50a). It is also autofs related. Maybe it is the same problem.
>
> I have ports automounted on /am/ports. When I do cd /am/ports/sys and type tab to autocomplete it crashes with the below stack trace. If I plainly mount ports on /usr/ports and do the same everything works. I am using NFSv3
>
> Peter
>
>
>
>
> Fatal trap 12: page fault while in kernel mode
> cpuid = 2; apic id = 04
> fault virtual address = 0x89
> fault code = supervisor read data, page not present
> instruction pointer = 0x20:0xffffffff809645d4
> stack pointer        = 0x28:0xfffffe00acadb830
> frame pointer        = 0x28:0xfffffe00acadb830
> code segment = base 0x0, limit 0xfffff, type 0x1b
> = DPL 0, pres 1, long 1, def32 0, gran 1
> processor eflags = interrupt enabled, resume, IOPL = 0
> current process = 6869 (csh)
> trap number = 12
> panic: page fault
> cpuid = 2
> time = 1705306940
> KDB: stack backtrace:
> #0 0xffffffff806232f5 at kdb_backtrace+0x65
> #1 0xffffffff805d7a02 at vpanic+0x152
> #2 0xffffffff805d78a3 at panic+0x43
> #3 0xffffffff809d58ad at trap_fatal+0x38d
> #4 0xffffffff809d58ff at trap_pfault+0x4f
> #5 0xffffffff809af048 at calltrap+0x8
> #6 0xffffffff804c7a7e at ncl_bioread+0xb7e
> #7 0xffffffff804b9d90 at nfs_readdir+0x1f0
> #8 0xffffffff8069c61a at vop_sigdefer+0x2a
> #9 0xffffffff809f8ae0 at VOP_READDIR_APV+0x20
> #10 0xffffffff81ce75de at autofs_readdir+0x2ce
> #11 0xffffffff809f8ae0 at VOP_READDIR_APV+0x20
> #12 0xffffffff806c3002 at kern_getdirentries+0x222
> #13 0xffffffff806c33a9 at sys_getdirentries+0x29
> #14 0xffffffff809d6180 at amd64_syscall+0x110
> #15 0xffffffff809af95b at fast_syscall_common+0xf8
>
>
>
> On 15 Jan 2024, at 06:46, FreeBSD User <freebsd@walstatt-de.de> wrote:
>
> Am Sun, 14 Jan 2024 20:34:12 -0800
> Cy Schubert <Cy.Schubert@cschubert.com> schrieb:
>
> In message <CAM5tNy5aat8vUn2fsX9jV=D9yGZdnO20Q0Ea7qtszx+zSES2bw@mail.gmail.c
> om>
> , Rick Macklem writes:
>
> On Sat, Jan 13, 2024 at 12:39=E2=80=AFPM Ronald Klop <ronald-lists@klop.ws>=
> wrote:
>
>
>
> Van: FreeBSD User <freebsd@walstatt-de.de>
> Datum: 13 januari 2024 19:34
> Aan: FreeBSD CURRENT <freebsd-current@freebsd.org>
> Onderwerp: NFSv4 crash of CURRENT
>
> Hello,
>
> running CURRENT client (FreeBSD 15.0-CURRENT #4 main-n267556-69748e62e82a=
>
> : Sat Jan 13 18:08:32
>
> CET 2024 amd64). One NFSv4 server is same OS revision as the mentioned cl=
>
> ient, other is FreeBSD
>
> 13.2-RELEASE-p8. Both offer NFSv4 filesystems, non-kerberized.
>
> I can crash the client reproducable by accessing the one or other NFSv4 F=
>
> S (a simple ls -la).
>
> The NFSv4 FS is backed by ZFS (if this matters). I do not have physicla a=
>
> ccess to the client
>
> host, luckily the box recovers.
>
> Did you rebuild both the nfscommon and nfscl modules from the same sources?
> I did a commit to main that changes the interface between these two
> modules and did bump the
> __FreeBSD_version to 1500010, which should cause both to be rebuilt.
> (If you have "options NFSCL" in your kernel config, both should have
> been rebuilt as a part of
> the kernel build.)
>
>
> Is anyone by chance seeing autofs in the backtrace too?
>
>
>
> Hello Cy Shubert,
>
> I forgot to mention that those crashes occur with autofs mounted filesystems. Good question,
> by the way, I will check whether crashes also happen when mounting the tradidional way.
>
> Kind regards,
>
> oh
>
> --
> O. Hartmann
>
>
>