[LTP] [PATCH v3 3/3] tst_find_backing_dev(): add support for overlayfs

[Cyril Hrubis]

Hi!
> diff --git a/lib/tst_device.c b/lib/tst_device.c
> index 76c3a3e1e..29dc6974f 100644
> --- a/lib/tst_device.c
> +++ b/lib/tst_device.c
> @@ -3,6 +3,7 @@
>   * Copyright (C) 2014 Cyril Hrubis chrubis@suse.cz
>   */
>  
> +#define _GNU_SOURCE
>  #include <sys/types.h>
>  #include <sys/stat.h>
>  #include <sys/ioctl.h>
> @@ -572,6 +573,136 @@ static void btrfs_get_uevent_path(char *tmp_path, char *uevent_path)
>  	SAFE_CLOSEDIR(NULL, dir);
>  }
>  
> +static char *overlay_mount_from_dev(dev_t dev)
> +{
> +	unsigned dev_major, dev_minor, mnt_major, mnt_minor;
> +	FILE *fp;
> +	char line[PATH_MAX];

PATH_MAX does not really make any sense here. It's as good as any other
number so I would just hardcode 4096 here.

> +	char *mountpoint;
> +	int ret;
> +
> +	dev_major = major(dev);
> +	dev_minor = minor(dev);
> +
> +	fp = SAFE_FOPEN(NULL, "/proc/self/mountinfo", "r");
> +	while (fgets(line, sizeof(line), fp) != NULL) {
> +		ret = sscanf(line, "%*d %*d %u:%u %*s %ms",
> +			     &mnt_major, &mnt_minor, &mountpoint);
> +		if (ret != 3)
> +			tst_brkm(TBROK, NULL,
> +				 "failed to parse mountinfo line: \"%s\"",
                                                                   ^
                We usually use ' instead of " inside of strings in LTP.


> +				 line);
> +		if (mnt_major == dev_major && mnt_minor == dev_minor)
> +			break;
> +		free(mountpoint);
> +		mountpoint = NULL;
> +	}
> +	SAFE_FCLOSE(NULL, fp);
> +	if (!mountpoint)
> +		tst_brkm(TBROK, NULL,
> +			 "Unable to find mount entry for device %u:%u\n",
                                                                      ^
							No newlines in
							tst_*()
							messages please.
> +			 dev_major, dev_minor);
> +
> +	return mountpoint;
> +}
> +
> +static char *overlay_get_upperdir(char *mountpoint)
> +{
> +	FILE *mntf;
> +	struct mntent *mnt;
> +	char *optstr, *optstart, *optend;
> +	char *upperdir = NULL;
> +
> +	mntf = setmntent("/proc/self/mounts", "r");
> +	if (!mntf)
> +		tst_brkm(TBROK | TERRNO, NULL, "Can't open /proc/self/mounts");
> +
> +	while ((mnt = getmntent(mntf)) != NULL) {
> +		if (strncmp(mnt->mnt_dir, mountpoint, strlen(mountpoint)))
> +			continue;

Why strncmp() here? Isn't this possibly generating false positives in
the case that we there is more mounts that have the same prefix that
matches mountpoint?

> +		if (strncmp(mnt->mnt_type, "overlay", strlen("overlay")))
> +			tst_brkm(TBROK, NULL,
> +				 "expected overlayfs on mount point \"%s\", but it is of type %s.",
> +				 mountpoint, mnt->mnt_type);

Here as well, I suppose that the probability of false positive here is
close to zero, but I do not see the reason for strncmp() here either.

> +		optstr = hasmntopt(mnt, "upperdir");
> +		if (optstr) {
> +			optstart = strchr(optstr, '=');
> +			optstart++;
> +			optend = strchrnul(optstr, ',');
> +			upperdir = calloc(optend - optstart + 1, 1);
> +			memcpy(upperdir, optstart, optend - optstart);

Isn't this just a complicated way how to re-implement strndup()?

> +			break;
> +		} else {
> +			tst_brkm(TBROK, NULL,
> +				 "mount point %s does not contain an upperdir",
> +				 mountpoint);
> +		}
> +	}
> +	endmntent(mntf);
> +
> +	if (!upperdir)
> +		tst_brkm(TBROK, NULL,
> +			 "Unable to find mount point \"%s\" in mount table",
> +			 mountpoint);
> +
> +	return upperdir;
> +}
> +
> +/*
> + * To get from a file or directory on an overlayfs to a device
> + * for an underlying mountpoint requires the following steps:
> + *
> + * 1. stat() the pathname and pick out st_dev.
> + * 2. use the st_dev to look up the mount point of the file
> + *    system in /proc/self/mountinfo
> + *
> + * Because 'mountinfo' is a much more complicated file format than
> + * 'mounts', we switch to looking up the mount point in /proc/mounts,
> + * and use the mntent.h helpers to parse the entries.
> + *
> + * 3. Using getmntent(), find the entry for the mount point identified
> + *    in step 2.
> + * 4. Call hasmntopt() to find the upperdir option, and parse that
> + *    option to get to the path name for the upper directory.
> + * 5. Call stat on the upper directory.  This should now contain
> + *    the major and minor number for the underlying device (assuming
> + *    that there aren't nested overlay file systems).
> + * 6. Populate the uevent path.
> + *
> + * Example /proc/self/mountinfo line for overlayfs:
> + *    471 69 0:48 / /tmp rw,relatime shared:242 - overlay overlay rw,seclabel,lowerdir=/tmp,upperdir=/mnt/upper/upper,workdir=/mnt/upper/work,uuid=null
> + *
> + * See section 3.5 of the kernel's Documentation/filesystems/proc.rst file
> + * for a detailed explanation of the mountinfo format.
> + */
> +static void overlay_get_uevent_path(char *tmp_path, char *uevent_path)
> +{
> +	int ret;
> +	struct stat st;
> +	char *mountpoint, *upperdir;
> +
> +	tst_resm(TINFO, "Use OVERLAYFS specific strategy");
> +
> +	ret = stat(tmp_path, &st);
> +	if (ret)
> +		tst_brkm(TBROK | TERRNO, NULL, "stat failed");
> +
> +	mountpoint = overlay_mount_from_dev(st.st_dev);
> +	upperdir = overlay_get_upperdir(mountpoint);
> +	free(mountpoint);

Since the mntpoint is only intermediate result, why can't we pass the
st.dev to the overlay_get_upperdir() and call overlay_mount_from_dev()
from there?

> +	ret = stat(upperdir, &st);
> +	free(upperdir);
> +	if (ret)
> +		tst_brkm(TBROK | TERRNO, NULL, "stat failed");
> +
> +	tst_resm(TINFO, "Warning: used first of multiple backing devices.");
> +	sprintf(uevent_path, "/sys/dev/block/%d:%d/uevent",
> +		major(st.st_dev), minor(st.st_dev));
> +}
> +
>  __attribute__((nonnull))
>  void tst_find_backing_dev(const char *path, char *dev, size_t dev_size)
>  {
> @@ -599,6 +730,8 @@ void tst_find_backing_dev(const char *path, char *dev, size_t dev_size)
>  
>  	if (fsbuf.f_type == TST_BTRFS_MAGIC) {
>  		btrfs_get_uevent_path(tmp_path, uevent_path);
> +	} else if (fsbuf.f_type == TST_OVERLAYFS_MAGIC) {
> +		overlay_get_uevent_path(tmp_path, uevent_path);
>  	} else if (dev_major == 0) {
>  		tst_brkm(TBROK, NULL, "%s resides on an unsupported pseudo-file system.", path);
>  	} else {
> -- 
> 2.43.5
> 
> 
> -- 
> Mailing list info: https://lists.linux.it/listinfo/ltp

-- 
Cyril Hrubis
chrubis@suse.cz