Hey Andreas,
I've been reviewing some of your patches that you've submitted,
and I came across e2fsprogs-lustre_ismounted.patch. It violated a
number of basic libext2fs rules, such as never calling printf/fprintf
from inside a libext2fs routine, and using a non-standard return code
convention. So I started reworking it so it would be acceptable, and
then I realized it defined two new static functions, neither of which
were actually used in the patch.
I hooked the patch into the appropriate place, and was going to
send it to you for review and testing, since I don't have a Lustre
system to play with, when I came to my second realization --- perhaps
the reason why it was never linked is because it isn't needed any more!
At least in 2.6 kernels, hopefully devices which are in use by Lustre
should be marked as busy, so that when they are opened O_EXCL, the
kernel should return EBUSY --- and that should be enough for e2fsprogs
without needing to use all sorts of complicated Lustre-specific code,
right?
So could you verify for me that without applying the patch
below, that if you build tst_ismounted, and run it on a device which is
currently in use by Lustre, that it is reporting the device to be busy
(EXT2_MF_BUSY). And if not, would you agree that is a bug that should
be fixed in the Lustre code?
If for some reason there is a good reason why Lustre isn't
marking the device as busy, then we can look at the attached patch, but
hopefully it isn't required at all.
Regards,
- Ted
--------------
Add checks to see if the partition is in use by Lustre
Signed-off-by: Andreas Dilger <[email protected]>
Signed-off-by: "Theodore Ts'o" <[email protected]>
diff -r 32c7d2357ee2 lib/ext2fs/ismounted.c
--- a/lib/ext2fs/ismounted.c Wed Oct 04 09:12:35 2006 -0400
+++ b/lib/ext2fs/ismounted.c Sun Oct 08 15:03:22 2006 -0400
@@ -30,6 +30,8 @@
#endif /* HAVE_GETMNTINFO */
#include <string.h>
#include <sys/stat.h>
+#include <sys/types.h>
+#include <dirent.h>
#include "ext2_fs.h"
#include "ext2fs.h"
@@ -276,6 +278,109 @@ static int is_swap_device(const char *fi
return ret;
}
+/*
+ * Check to see if filesystem is in use by Lustre
+ */
+static errcode_t check_if_lustre_busy(const char *devname, int *mount_flags)
+{
+ errcode_t retval = 0;
+ struct dirent *direntp;
+ const char *procname;
+ char *ptr, *mnt_device = 0, *proc_val = 0;
+ char *real_devname = 0, *real_mnt_device = 0;
+ DIR *dirp;
+ int fd, numr;
+
+ procname = "/proc/fs/lustre/obdfilter";
+ dirp = opendir(procname);
+ if (!dirp) {
+ procname = "/proc/fs/lustre/mds";
+ dirp = opendir(procname);
+ }
+ if (!dirp)
+ return 0;
+
+ retval = ENOMEM;
+ if ((real_devname = malloc(PATH_MAX)) == NULL)
+ goto out;
+ if ((mnt_device = malloc(PATH_MAX)) == NULL)
+ goto out;
+ if ((proc_val = malloc(PATH_MAX)) == NULL)
+ goto out;
+ if ((real_mnt_device = malloc(PATH_MAX)) == NULL)
+ goto out;
+
+ if (realpath(devname, real_devname) == NULL) {
+ retval = errno;
+ goto out;
+ }
+
+ while ((direntp = readdir(dirp)) != NULL) {
+ if ((strncmp(direntp->d_name, ".", 1) == 0) ||
+ (strncmp(direntp->d_name, "..", 2) == 0)) {
+ continue;
+ }
+
+ memset(proc_val, 0, PATH_MAX);
+ snprintf(proc_val, PATH_MAX, "%s/%s/mntdev", procname,
+ direntp->d_name);
+ fd = open(proc_val, O_RDONLY);
+ if (fd < 0) {
+ if (errno == ENOENT || errno == ENOTDIR)
+ continue;
+ retval = errno;
+ goto out;
+ }
+
+ memset(mnt_device, 0, PATH_MAX);
+ numr = 0;
+ ptr = mnt_device;
+ do {
+ numr = read(fd, ptr, PATH_MAX);
+ if (numr < 0) {
+ retval = errno;
+ close(fd);
+ goto out;
+ }
+ ptr += numr;
+ } while (numr != 0);
+ close(fd);
+
+ ptr = strchr(mnt_device, '\n');
+ if (ptr)
+ *ptr = '\0';
+
+ memset(real_mnt_device, 0, PATH_MAX);
+ if (realpath(mnt_device, real_mnt_device) == NULL) {
+ retval = errno;
+ goto out;
+ }
+
+ if (strcmp(real_devname, real_mnt_device) == 0) {
+#ifdef DEBUG
+ fprintf(stderr,
+ "device %s mounted by lustre per %s\n",
+ real_devname, proc_val);
+#endif
+ *mount_flags = EXT2_MF_BUSY;
+ retval = 0;
+ goto out;
+ }
+ }
+
+out:
+ if (proc_val)
+ free(proc_val);
+ if (mnt_device)
+ free(mnt_device);
+ if (real_mnt_device)
+ free(real_mnt_device);
+ if (dirp)
+ closedir(dirp);
+
+ return retval;
+}
+
/*
* ext2fs_check_mount_point() fills determines if the device is
@@ -327,7 +432,10 @@ errcode_t ext2fs_check_mount_point(const
close(fd);
#endif
- return 0;
+ if (*mount_flags == 0)
+ retval = check_if_lustre_busy(device, mount_flags);
+
+ return retval;
}
/*
On Oct 08, 2006 15:50 -0400, Theodore Ts'o wrote:
> So I started reworking it so it would be acceptable, and
> then I realized it defined two new static functions, neither of which
> were actually used in the patch.
Correct, my initial testing was on 2.6, which as you say works OK, but
we still have some customers (sadly) running 2.4 kernels where O_EXCL
doesn't work.
> If for some reason there is a good reason why Lustre isn't
> marking the device as busy, then we can look at the attached patch, but
> hopefully it isn't required at all.
>
> +static errcode_t check_if_lustre_busy(const char *devname, int *mount_flags)
> +{
> + procname = "/proc/fs/lustre/obdfilter";
> + dirp = opendir(procname);
> + if (!dirp) {
> + procname = "/proc/fs/lustre/mds";
> + dirp = opendir(procname);
> + }
This isn't quite correct, compared to the earlier patch that calls
check_lustre_proc_vals() twice. Lustre uses filesystems as either a
metadata target or storage target, and more than one of each can exist
on the same node (though usually they don't for large systems). That
means this function needs to always check both the .../obdfilter tree
and the .../mds tree on a given node. The above code will skip the
.../mds tree if the .../obdfilter tree exists.
> + if (strcmp(real_devname, real_mnt_device) == 0) {
> +#ifdef DEBUG
> + fprintf(stderr,
> + "device %s mounted by lustre per %s\n",
> + real_devname, proc_val);
> +#endif
The reason this was printed is that Lustre mounts (done internally
by the kernel) do not show up in /proc/mounts, or this whole exercise
wouldn't be needed.
Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.