svn commit: r237367 - head/sys/fs/nfsclient
Konstantin Belousov
kostikbel at gmail.com
Sun Jul 1 21:43:12 UTC 2012
On Sun, Jul 01, 2012 at 08:33:09PM +0200, Andreas Tobler wrote:
> On 01.07.12 19:05, Konstantin Belousov wrote:
> >On Sun, Jul 01, 2012 at 03:56:52PM +0200, Andreas Tobler wrote:
> >>On 01.07.12 15:41, Konstantin Belousov wrote:
> >>>On Sun, Jul 01, 2012 at 03:37:18PM +0200, Andreas Tobler wrote:
> >>>>On 01.07.12 14:04, Konstantin Belousov wrote:
> >>>>>On Sun, Jul 01, 2012 at 01:23:02PM +0200, Andreas Tobler wrote:
> >>>>>>On 21.06.12 11:26, Konstantin Belousov wrote:
> >>>>>>>Author: kib
> >>>>>>>Date: Thu Jun 21 09:26:06 2012
> >>>>>>>New Revision: 237367
> >>>>>>>URL: http://svn.freebsd.org/changeset/base/237367
> >>>>>>>
> >>>>>>>Log:
> >>>>>>> Enable deadlock avoidance code for NFS client.
> >>>>>>
> >>>>>>
> >>>>>>Hm, since this commit I fail with my nfs installworld/kernel.
> >>>>>>
> >>>>>>I have a builder which installs world/kernel to a nfs mounted
> >>>>>>directory.
> >>>>>>Namely used for cross builds.
> >>>>>>
> >>>>>>Now since this commit I get the following when I install kernel to the
> >>>>>>nfs directory:
> >>>>>>
> >>>>>>..
> >>>>>>install -o root -g wheel -m 555 zfs.ko.symbols
> >>>>>>/netboot/sparc64/boot/kernel
> >>>>>>install: /netboot/sparc64/boot/kernel/zfs.ko.symbols: No such file or
> >>>>>>directory
> >>>>>>*** [_kmodinstall] Error code 71
> >>>>>>..
> >>>>>>
> >>>>>>The file is there, a local install of the tree works without problems.
> >>>>>>Reverting to r237366 also makes it work again.
> >>>>>>
> >>>>>>The server is a -CURRENT, r237880, The client, -CURRENT too.
> >>>>>>
> >>>>>>How can I help to track down the real issue?
> >>>>>
> >>>>>Is it always the same file in the install procedure which causes the
> >>>>>failure ? Even more, is the failure pattern always the same ?
> >>>>
> >>>>I'd say so yes. When installing a kernel onto a nfs mounted fs then
> >>>>always (in my cases) the zfs.ko.symbols was the failing pattern.
> >>>>I tried ppc64 and sparc64 as target. With both it was the above file.
> >>>>
> >>>>When doing a installworld, it was, also in both cases, ppc64/sparc64,
> >>>>the cc1 in libexec which failed.
> >>>>
> >>>>>Might be, start with ktrace-ing the whole make invocation, including
> >>>>>the children processes.
> >>>>
> >>>>Some recipes how to start?
> >>>ktrace -o <file on local fs> -i make installkernel
> >>>Then kdump and cut the lines around relevant failure.
> >>
> >>ktrace -f, right?
> >Right, but without -i it is useless.
>
> Ah, yes, seems clear now after reading the man page.
>
> >>I placed the whole kdump here:
> >>
> >>http://people.freebsd.org/~andreast/dumped_installkernel.log
> >>
> >>It is not clear to me where the failure starts :)
> >Because logs do not contain tracepoints from the children.
> >See above about -i.
> >
> >I asked about excerpt because I expect the proper log to have an order
> >of magnitude bigger size.
>
> Ok. The dump is around 100MB, I hope I extracted as much as needed:
>
> http://people.freebsd.org/~andreast/dumped_installkernel-7.log
>
> >>>>>I used buildworld on the NFS-mounted obj/ as the test for the changes.
> >>>>
> >>>>Here the obj is local, only the src and the destination is on the
> >>>>nfs/netboot server.
> >>>
> >>>I just finished build on NFS obj/ and did several rounds of installs
> >>>for world and kernel into nfs-mounted destdir. It seems I cannot
> >>>reproduce
> >>>this locally.
> >>
> >>Ok. I try with an nfs obj too.
>
> So, I was not able to reproduce the failure with an nfs mounted obj dir.
>
> But I was able to reproduce the failure with three different machines
> which all have the obj local and the destination mounted via nfs.
>
> Are you able to try with a local obj too?
Below are two patches. Please follow my instructions literally to get
most of your bug report.
First, please apply the usr.bin/xinstall patch only, and retry installkernel
(no need to use ktrace). It should show the proper error, short write, with
zero-sized result, instead of garbage ENOENT from errno.
Next, please apply the sys/fs/nfsclient patch, which should fix the core
cause.
diff --git a/sys/fs/nfsclient/nfs_clbio.c b/sys/fs/nfsclient/nfs_clbio.c
index 71286e3..f7af6fb 100644
--- a/sys/fs/nfsclient/nfs_clbio.c
+++ b/sys/fs/nfsclient/nfs_clbio.c
@@ -897,7 +897,7 @@ ncl_write(struct vop_write_args *ap)
struct nfsmount *nmp = VFSTONFS(vp->v_mount);
daddr_t lbn;
int bcount;
- int bp_cached, n, on, error = 0;
+ int bp_cached, n, on, error = 0, error1;
size_t orig_resid, local_resid;
off_t orig_size, tmp_off;
@@ -1259,9 +1259,12 @@ again:
if ((ioflag & IO_SYNC)) {
if (ioflag & IO_INVAL)
bp->b_flags |= B_NOCACHE;
- error = bwrite(bp);
- if (error)
+ error1 = bwrite(bp);
+ if (error1 != 0) {
+ if (error == 0)
+ error = error1;
break;
+ }
} else if ((n + on) == biosize) {
bp->b_flags |= B_ASYNC;
(void) ncl_writebp(bp, 0, NULL);
diff --git a/usr.bin/xinstall/xinstall.c b/usr.bin/xinstall/xinstall.c
index a920f85..3eba4f7 100644
--- a/usr.bin/xinstall/xinstall.c
+++ b/usr.bin/xinstall/xinstall.c
@@ -53,6 +53,7 @@ __FBSDID("$FreeBSD$");
#include <errno.h>
#include <fcntl.h>
#include <grp.h>
+#include <inttypes.h>
#include <paths.h>
#include <pwd.h>
#include <stdio.h>
@@ -671,11 +672,18 @@ copy(int from_fd, const char *from_name, int to_fd, const char *to_name,
if (size <= 8 * 1048576 && trymmap(from_fd) &&
(p = mmap(NULL, (size_t)size, PROT_READ, MAP_SHARED,
from_fd, (off_t)0)) != (char *)MAP_FAILED) {
- if ((nw = write(to_fd, p, size)) != size) {
+ nw = write(to_fd, p, size);
+ if (nw != size) {
serrno = errno;
(void)unlink(to_name);
- errno = nw > 0 ? EIO : serrno;
- err(EX_OSERR, "%s", to_name);
+ if (nw >= 0) {
+ errx(EX_OSERR,
+ "short write to %s: %jd bytes written, %jd bytes asked to write",
+ to_name, (uintmax_t)nw, (uintmax_t)size);
+ } else {
+ errno = serrno;
+ err(EX_OSERR, "%s", to_name);
+ }
}
done_copy = 1;
}
@@ -684,8 +692,15 @@ copy(int from_fd, const char *from_name, int to_fd, const char *to_name,
if ((nw = write(to_fd, buf, nr)) != nr) {
serrno = errno;
(void)unlink(to_name);
- errno = nw > 0 ? EIO : serrno;
- err(EX_OSERR, "%s", to_name);
+ if (nw >= 0) {
+ errx(EX_OSERR,
+ "short write to %s: %jd bytes written, %jd bytes asked to write",
+ to_name, (uintmax_t)nw,
+ (uintmax_t)size);
+ } else {
+ errno = serrno;
+ err(EX_OSERR, "%s", to_name);
+ }
}
if (nr != 0) {
serrno = errno;
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 196 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/svn-src-all/attachments/20120701/76a6b2a9/attachment.pgp
More information about the svn-src-all
mailing list