Convert slashes in the UNC parameter to backslashes

This patch ensures that slashes used as separators in the UNC are properly converted to backslashes. The existing implementation did not perform that conversion and that lead to a "invalid argument" error when specifying an UNC explicitly which contained slashes in it.

---
  fs/cifs/connect.c |   12 ++++++------
  1 files changed, 6 insertions(+), 6 deletions(-)

fs/cifs/connect.c +1 -0

index 5c670b9..8f4c76f 100644
@@ -1108,6 +1108,7 @@ cifs_parse_mount_options(const char *mountdata, 
 const char *devname,
  	char *string = NULL;
  	char *tmp_end, *value;
  	char delim;
+	char *p, *q;
 
  	separator[0] = ',';

  	separator[1] = 0;

fs/cifs/connect.c +5 -6

@@ -1573,12 +1574,11 @@ cifs_parse_mount_options(const char *mountdata, 
 const char *devname,
  				printk(KERN_WARNING "CIFS: no memory for UNC\n");
  				goto cifs_parse_mount_err;
  			}
-			strcpy(vol->UNC, string);
-
-			if (strncmp(string, "//", 2) == 0) {
-				vol->UNC[0] = '\\';
-				vol->UNC[1] = '\\';
-			} else if (strncmp(string, "\\\\", 2) != 0) {
+			for (p = string, q = vol->UNC; *p; ++p, ++q) {
+				*q = *p == '/'? '\\' : *p;
+			}
+			*q = '\0';
+			if (strncmp(vol->UNC, "\\\\", 2) != 0) { 
  				printk(KERN_WARNING "CIFS: UNC Path does not "
  						    "begin with // or \\\\\n");

goto cifs_parse_mount_err;

Quoted 41 lines · On Tue, 06 Nov 2012 13:29:47 +0100 Federico Sauter…

This patch ensures that slashes used as separators in the UNC are properly converted to backslashes. The existing implementation did not perform that conversion and that lead to a "invalid argument" error when specifying an UNC explicitly which contained slashes in it.

---
   fs/cifs/connect.c |   12 ++++++------
   1 files changed, 6 insertions(+), 6 deletions(-)

fs/cifs/connect.c +1 -0

index 5c670b9..8f4c76f 100644
@@ -1108,6 +1108,7 @@ cifs_parse_mount_options(const char *mountdata,
 const char *devname,
   	char *string = NULL;
   	char *tmp_end, *value;
   	char delim;
+	char *p, *q;
 
   	separator[0] = ',';

   	separator[1] = 0;

fs/cifs/connect.c +5 -6

@@ -1573,12 +1574,11 @@ cifs_parse_mount_options(const char *mountdata,
 const char *devname,
   				printk(KERN_WARNING "CIFS: no memory for UNC\n");
   				goto cifs_parse_mount_err;
   			}
-			strcpy(vol->UNC, string);
-
-			if (strncmp(string, "//", 2) == 0) {
-				vol->UNC[0] = '\\';
-				vol->UNC[1] = '\\';
-			} else if (strncmp(string, "\\\\", 2) != 0) {
+			for (p = string, q = vol->UNC; *p; ++p, ++q) {
+				*q = *p == '/'? '\\' : *p;
+			}
+			*q = '\0';
+			if (strncmp(vol->UNC, "\\\\", 2) != 0) {
   				printk(KERN_WARNING "CIFS: UNC Path does not "
   						    "begin with // or \\\\\n");

Quoted 1 line

   				goto cifs_parse_mount_err;

Looks harmless enough, but it would be clearer to use the convert_delimiter() helper function here, rather than open-coding that functionality.

Long term, it would be good to get rid of the unc= and prefixpath= options altogether and simply have the kernel extract this information from the device name at mount time instead.

Here the corrected (and tested) version.

Jeff: I agree with you regarding removing the UNC parameter altogether. Is there any scenario in which an UNC parameter could differ from the specified source parameter?

Regards,
Fred

fs/cifs/connect.c +2 -4

index 5c670b9..c52faf8 100644
@@ -1574,11 +1574,9 @@ cifs_parse_mount_options(const char *mountdata, 
 const char *devname,
                                 goto cifs_parse_mount_err;
                         }
                         strcpy(vol->UNC, string);
+                       convert_delimiter(vol->UNC, '\\');
 
-                       if (strncmp(string, "//", 2) == 0) {
-                               vol->UNC[0] = '\\';
-                               vol->UNC[1] = '\\';
-                       } else if (strncmp(string, "\\\\", 2) != 0) {
+                       if (strncmp(vol->UNC, "\\\\", 2) != 0) {
                                 printk(KERN_WARNING "CIFS: UNC Path 

does not "
"begin with // or

\\\\\n");
                                 goto cifs_parse_mount_err;

Quoted 9 lines · On Wed, 07 Nov 2012 16:20:22 +0100 Federico Sauter…

Here the corrected (and tested) version.

Jeff: I agree with you regarding removing the UNC parameter altogether. Is there any scenario in which an UNC parameter could differ from the specified source parameter?

Regards,
Fred

I suppose someone could provide their own unc= option that differs from the device name. At that point, the device name in /proc/mounts (and in programs like "df") might look like you've mounted one location when in reality you've mounted another.

That doesn't seem like a valid use-case to me though. It's probably best that we discourage such abuse ;).

The tricky part is how to get there from here. The mount helper provides this option pretty much unconditionally, though it generally generates it from the device string. Maybe something like this?

1) fix the kernel to parse a copy of the UNC out of the device string

2) compare that UNC to the one provided by the unc= option

3) if they differ, then printk a warning that mentions that there is a discrepancy and that the kernel is going to use the value of the unc= option for now, but that we'll change that default in two releases.

Something like:

"CIFS VFS: the unc= mount option differs from the device passed in. Using the value of the unc= option. The kernel will begin preferring the value in the device string in 3.10!"

Then, have a patch ready to go for the 3.10 merge window that will make the kernel start ignoring the value of the unc= option. You might even be nice and keep the comparison and warning around for a little while, but switch it to mention that the kernel is going to ignore the value of the unc= option when there is a discrepancy.

In another few releases, we can then get rid of the warning and comparison too.

Sound reasonable?

Quoted 18 lines · Jeff Layton

Quoted 9 lines · On Wed, 07 Nov 2012 16:20:22 +0100 Federico Sauter…

Here the corrected (and tested) version.

Jeff: I agree with you regarding removing the UNC parameter altogether. Is there any scenario in which an UNC parameter could differ from the specified source parameter?

Regards,
Fred

I suppose someone could provide their own unc= option that differs from the device name. At that point, the device name in /proc/mounts (and in programs like "df") might look like you've mounted one location when in reality you've mounted another.

Quoted 2 lines

That doesn't seem like a valid use-case to me though. It's probably best that we discourage such abuse ;).

Would this ever be the case when using a DFS mount?

I'm not sure how
that's handled internally to the kernel (I know Samba doesn't provide support for multiple servers on a DFS root, but Windows does, IIRC). So far as I can tell, in the simple case Samba is a proxy, in the complex case, the share is on another server and the original server is down. Does the kernel use the same device in both cases?

Quoted 25 lines

The tricky part is how to get there from here. The mount helper provides this option pretty much unconditionally, though it generally generates it from the device string. Maybe something like this?

1) fix the kernel to parse a copy of the UNC out of the device string

2) compare that UNC to the one provided by the unc= option

3) if they differ, then printk a warning that mentions that there is a discrepancy and that the kernel is going to use the value of the unc= option for now, but that we'll change that default in two releases.

Something like:

"CIFS VFS: the unc= mount option differs from the device passed in. Using the value of the unc= option. The kernel will begin preferring the value in the device string in 3.10!"

Then, have a patch ready to go for the 3.10 merge window that will make the kernel start ignoring the value of the unc= option. You might even be nice and keep the comparison and warning around for a little while, but switch it to mention that the kernel is going to ignore the value of the unc= option when there is a discrepancy.

Quoted 5 lines

In another few releases, we can then get rid of the warning and comparison too.

Sound reasonable?

Would it be reasonable to deprecate the "UNC" / "path" / "target" options (they all resolve to OPT_UNC from parse_opt_token()) in mount.cifs if the kernel is going to do the heavy lifting? That would break compatibility between mount.cifs and older kernels, though. I guess it doesn't hurt anything to keep passing UNC even if it's not being used, other than users not understanding the unexplained behavior of the kernel deciding what the actual value of UNC is and disregarding their explicit setting of it. After typing that out, it sounds kind of bad.

Quoted 31 lines

Quoted 25 lines

The tricky part is how to get there from here. The mount helper provides this option pretty much unconditionally, though it generally generates it from the device string. Maybe something like this?

1) fix the kernel to parse a copy of the UNC out of the device string

2) compare that UNC to the one provided by the unc= option

3) if they differ, then printk a warning that mentions that there is a discrepancy and that the kernel is going to use the value of the unc= option for now, but that we'll change that default in two releases.

Something like:

"CIFS VFS: the unc= mount option differs from the device passed in. Using the value of the unc= option. The kernel will begin preferring the value in the device string in 3.10!"

Then, have a patch ready to go for the 3.10 merge window that will make the kernel start ignoring the value of the unc= option. You might even be nice and keep the comparison and warning around for a little while, but switch it to mention that the kernel is going to ignore the value of the unc= option when there is a discrepancy.

Quoted 5 lines

In another few releases, we can then get rid of the warning and comparison too.

Sound reasonable?

Quoted 9 lines

Would it be reasonable to deprecate the "UNC" / "path" / "target" options (they all resolve to OPT_UNC from parse_opt_token()) in mount.cifs if the kernel is going to do the heavy lifting? That would break compatibility between mount.cifs and older kernels, though. I guess it doesn't hurt anything to keep passing UNC even if it's not being used, other than users not understanding the unexplained behavior of the kernel deciding what the actual value of UNC is and disregarding their explicit setting of it. After typing that out, it sounds kind of bad.

Eventually, yes...we'd stop the mount helper from providing a unc= option as well. That said, we'll have to contend with the old kernel + new userspace problem for quite a while, so that can't happen at the same time.

The first step is to get the kernel to where it doesn't actually need the option. Once that's done we can formulate a plan to stop providing it.

Quoted 41 lines · Jeff Layton

Quoted 31 lines

Quoted 25 lines

The tricky part is how to get there from here. The mount helper provides this option pretty much unconditionally, though it generally generates it from the device string. Maybe something like this?

1) fix the kernel to parse a copy of the UNC out of the device string

2) compare that UNC to the one provided by the unc= option

3) if they differ, then printk a warning that mentions that there is a discrepancy and that the kernel is going to use the value of the unc= option for now, but that we'll change that default in two releases.

Something like:

"CIFS VFS: the unc= mount option differs from the device passed in. Using the value of the unc= option. The kernel will begin preferring the value in the device string in 3.10!"

Then, have a patch ready to go for the 3.10 merge window that will make the kernel start ignoring the value of the unc= option. You might even be nice and keep the comparison and warning around for a little while, but switch it to mention that the kernel is going to ignore the value of the unc= option when there is a discrepancy.

Quoted 5 lines

In another few releases, we can then get rid of the warning and comparison too.

Sound reasonable?

Quoted 9 lines

Would it be reasonable to deprecate the "UNC" / "path" / "target" options (they all resolve to OPT_UNC from parse_opt_token()) in mount.cifs if the kernel is going to do the heavy lifting? That would break compatibility between mount.cifs and older kernels, though. I guess it doesn't hurt anything to keep passing UNC even if it's not being used, other than users not understanding the unexplained behavior of the kernel deciding what the actual value of UNC is and disregarding their explicit setting of it. After typing that out, it sounds kind of bad.

Quoted 9 lines

Eventually, yes...we'd stop the mount helper from providing a unc= option as well. That said, we'll have to contend with the old kernel + new userspace problem for quite a while, so that can't happen at the same time.

The first step is to get the kernel to where it doesn't actually need the option. Once that's done we can formulate a plan to stop providing it.

I came across this issue precisely because I was using a mount helper different than samba (busybox, to be precise.) Thus, the unc parameter was being passed in a slightly different way and thus an error message resulted. This is to say that, if we radically changed the unc/path/target options interface, it would most likely break existing mount implementations.

I suggest that we just copy the device name as UNC in the parse_options function and ignore whatever the user had provided explicitly without issuing any warnings. I personally like warnings, but the current situation is that this option is implicitly passed by the mount helper, and thus it would most likely confuse users more than it helps them.

Please let me know what you think. I can quickly provide a tested patch, if you agree to the course of action.