Skip to contentSkip to conversation

Handling ':' on filenames

RFC PATCH In review Eduardo Habkost 6 participants 8 messages Mar 6, 2009 – Mar 10, 2009

lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 Handling ':' on filenames

This patch fixes this issue:

  $ qemu-img create -f qcow2 /tmp/a:b 1G
  $ qemu-system-x86_64 -hda qcow2:/tmp/a:b
  qemu: could not open disk image /tmp/a:b
  $

Based on a suggestion by Daniel Berrange.

However, this is still just a workaround. The semantics of filenames containing colon characters (and how this can be escaped, avoided, or worked around) are not very clear.

Going further, what if we stop using "protocol:filename" strings internally, except where the user interface or external data really requires this format?

---
 block.c |    8 ++++++++
 1 files changed, 8 insertions(+), 0 deletions(-)
block.c +8 -0
index 7c744c7..04488d6 100644
@@ -236,6 +236,14 @@ static BlockDriver *find_protocol(const char *filename)
         is_windows_drive_prefix(filename))
         return &bdrv_raw;
 #endif
+
+    /* Protocol name will never start with a slash.
+     * This allows the user to specify absolute filenames
+     * containing a ":" character.
+     */
+    if (*filename == '/')
+        return &bdrv_raw;
+
     p = strchr(filename, ':');
     if (!p)
         return &bdrv_raw;
Signature · 4 lines
1.6.1

-- 
Eduardo
#1 lore raw
Quoted 6 lines · Eduardo Habkost

This patch fixes this issue:

  $ qemu-img create -f qcow2 /tmp/a:b 1G
  $ qemu-system-x86_64 -hda qcow2:/tmp/a:b
  qemu: could not open disk image /tmp/a:b
  $

This patch looks good to me with the following comments:

1. I think the example above is wrong.
a. There is no protocol "qcow2"
b. We do not want to use "qcow2:/tmp/a:b" as filenames (been there).
We want to use -drive file=/tmp/a:b,format=qcow2 (and similar "qemu-img info -f ").
c. But, without this patch, even when using appropriate -drive, we'd
still get the same result (could not open disk image). Qemu considers "/tmp/a" to be a protocol, which does not exist, so Qemu fails to open the image.
With this patch, Qemu opens the file successfully.

Quoted 10 lines

Based on a suggestion by Daniel Berrange.

However, this is still just a workaround. The semantics of filenames containing colon characters (and how this can be escaped, avoided, or worked around) are not very clear.

Going further, what if we stop using "protocol:filename" strings internally, except where the user interface or external data really requires this format?

2. Do you mean something like adding a '-drive protocol=xxx' option ?

Quoted 11 lines
---
 block.c |    8 ++++++++
 1 files changed, 8 insertions(+), 0 deletions(-)
block.c +0 -0
index 7c744c7..04488d6 100644

block.c static BlockDriver *find_protocol(const char *filename) · line 245

Quoted 10 lines
block.c +7 -0
@@ -236,6 +236,14 @@ static BlockDriver *find_protocol(const char *filename)
         is_windows_drive_prefix(filename))
         return &bdrv_raw;
 #endif
+
+    /* Protocol name will never start with a slash.
+     * This allows the user to specify absolute filenames
+     * containing a ":" character.
+     */
+    if (*filename == '/')
+        return &bdrv_raw;

3. The patch limits protocols names to not start with '/' (full paths). I think we should apply the same logic to relative paths, so protocol names would not start with '.' as well (no protocol starts with '.' today):

+ if ((*filename == '/') || (*filename == '.'))
+ return &bdrv_raw;

4. Note that this patch does not limit image formats to raw, it just says "use no protocols for full/relative paths".

block.c static BlockDriver *find_protocol(const char *filename) · line 249

Quoted 4 lines
block.c +1 -0
+
     p = strchr(filename, ':');
     if (!p)
         return &bdrv_raw;

Regards,
Uri.

#2 lore raw
Quoted 3 lines · Uri Lublin

b. We do not want to use "qcow2:/tmp/a:b" as filenames (been there).
We want to use -drive file=/tmp/a:b,format=qcow2 (and similar "qemu-img info -f ").

I can't help but feel that any approach to extracting filenames from the middle of arguments will always have problems.

I suppose a syntax like this:

-drive file=/tmp/file,with,commas:and:colons ,format=qcow2

(i.e. with the next parameter to -drive starting in a new argument)... wouldn't win many supporters though?

Cheers,

Signature · 1 line
Stuart Brady
#3 lore raw

Excerpts from Stuart Brady's message of Dom Mar 08 08:49:41 -0300 2009:

Quoted 8 lines · Uri Lublin
Quoted 3 lines

b. We do not want to use "qcow2:/tmp/a:b" as filenames (been there).
We want to use -drive file=/tmp/a:b,format=qcow2 (and similar "qemu-img info -f ").

I can't help but feel that any approach to extracting filenames from the middle of arguments will always have problems.

Quoted 6 lines

I suppose a syntax like this:

-drive file=/tmp/file,with,commas:and:colons ,format=qcow2

(i.e. with the next parameter to -drive starting in a new argument)... wouldn't win many supporters though?

The -drive option has a way to escape colons. For the filename "/tmp/a,b,c", you can use:

-drive file=/tmp/a,,b,,c,format=qcow2

Not pretty, but it seems to work.

Signature · 1 line
Eduardo
#4 lore raw
Quoted 6 lines · Eduardo Habkost

The -drive option has a way to escape colons. For the filename "/tmp/a,b,c", you can use:

-drive file=/tmp/a,,b,,c,format=qcow2

Not pretty, but it seems to work.

I wonder how you quote filenames from the monitor.

In my own management scripts I've taken to making a symlink with a "safe" name to the real image file, and passing the safe name to KVM, so that it's unambiguous. Causes a bit of fun with qcow2 files, which refer to their backing files relative to the symlink not the target file.

-- Jamie

#5 lore raw
In reply to Uri Lublin

Excerpts from kvm_autotest's message of Dom Mar 08 08:28:31 -0300 2009:
<snip>

Quoted 25 lines
Quoted 9 lines
---
 block.c |    8 ++++++++
 1 files changed, 8 insertions(+), 0 deletions(-)
block.c +0 -0
index 7c744c7..04488d6 100644

block.c static BlockDriver *find_protocol(const char *filename) · line 245

Quoted 10 lines
block.c +7 -0
@@ -236,6 +236,14 @@ static BlockDriver *find_protocol(const char *filename)
         is_windows_drive_prefix(filename))
         return &bdrv_raw;
 #endif
+
+    /* Protocol name will never start with a slash.
+     * This allows the user to specify absolute filenames
+     * containing a ":" character.
+     */
+    if (*filename == '/')
+        return &bdrv_raw;

3. The patch limits protocols names to not start with '/' (full paths). I think we should apply the same logic to relative paths, so protocol names would not start with '.' as well (no protocol starts with '.' today):

Quoted 2 lines

+ if ((*filename == '/') || (*filename == '.'))
+ return &bdrv_raw;

Is ':' a valid character for Windows filenames? In this case, we may want to check for backslashes also.

The "./filename" approach for relative paths seems nice to complete the fix, but I still see this is a workaround to a more fundamental problem.

Quoted 12 lines

4. Note that this patch does not limit image formats to raw, it just says "use no protocols for full/relative paths".

block.c static BlockDriver *find_protocol(const char *filename) · line 249

Quoted 4 lines
block.c +1 -0
+
     p = strchr(filename, ':');
     if (!p)
         return &bdrv_raw;

Regards,
Uri.

Signature · 1 line
Eduardo
#6 lore raw
Quoted 13 lines · Eduardo Habkost
Quoted 13 lines

block.c static BlockDriver *find_protocol(const char *filename) · line 245

Quoted 7 lines
block.c +7 -0
+
+    /* Protocol name will never start with a slash.
+     * This allows the user to specify absolute filenames
+     * containing a ":" character.
+     */
+    if (*filename == '/')
+        return &bdrv_raw;

3. The patch limits protocols names to not start with '/' (full paths). I think we should apply the same logic to relative paths, so protocol names would not start with '.' as well (no protocol starts with '.' today):

Quoted 5 lines
Quoted 2 lines

+ if ((*filename == '/') || (*filename == '.'))
+ return &bdrv_raw;

Is ':' a valid character for Windows filenames? In this case, we may want to check for backslashes also.

I guess so; full paths are (at least used to be) something like "C:\some-dir\some-file"

Amit

#7 lore raw
In reply to Uri Lublin

Excerpts from kvm_autotest's message of Dom Mar 08 08:28:31 -0300 2009:

Quoted 8 lines
Quoted 6 lines · Eduardo Habkost

This patch fixes this issue:

  $ qemu-img create -f qcow2 /tmp/a:b 1G
  $ qemu-system-x86_64 -hda qcow2:/tmp/a:b
  qemu: could not open disk image /tmp/a:b
  $
Quoted 7 lines

This patch looks good to me with the following comments:

1. I think the example above is wrong.
a. There is no protocol "qcow2"
b. We do not want to use "qcow2:/tmp/a:b" as filenames (been there).
We want to use -drive file=/tmp/a:b,format=qcow2 (and similar "qemu-img info -f ").

Oops. Ignore that, I've pasted the wrong command-line. That was when I was testing the behaviour of qemu when I specified other things before the colon. The error appears when I do this:

 $ qemu-system-x86_64 -hda /tmp/a:b
Quoted 6 lines

c. But, without this patch, even when using appropriate -drive, we'd
still get the same result (could not open disk image). Qemu considers "/tmp/a" to be a protocol, which does not exist, so Qemu fails to open the image.
With this patch, Qemu opens the file successfully.

Yes.

Quoted 7 lines
Quoted 7 lines

Based on a suggestion by Daniel Berrange.

However, this is still just a workaround. The semantics of filenames containing colon characters (and how this can be escaped, avoided, or worked around) are not very clear.

Quoted 5 lines
Quoted 3 lines

Going further, what if we stop using "protocol:filename" strings internally, except where the user interface or external data really requires this format?

2. Do you mean something like adding a '-drive protocol=xxx' option ?

Where possible, yes.

Signature · 1 line
Eduardo