Skip to contentSkip to conversation

New: hugetlbfs scrambles mode when sticky bit is set

Bug 213785 Andrew Morton 4 participants 5 messages Jul 20, 2021 – Jul 21, 2021 1 message not in the archive Memory management

lore
Share & tools
lore

Andrew Morton

author maintainer
#1 lore raw
In reply to a message not in the archive

(switched to email. Please respond via emailed reply-to-all, not via the bugzilla web interface).

Quoted 47 lines · On Mon, 19 Jul 2021 15:34:00 +0000 [email protected] wrote:

https://bugzilla.kernel.org/show_bug.cgi?id=213785

            Bug ID: 213785
           Summary: hugetlbfs scrambles mode when sticky bit is set
           Product: Memory Management
           Version: 2.5
    Kernel Version: >4.19 (certainly >= 5.4)
          Hardware: All
                OS: Linux
              Tree: Mainline
            Status: NEW
          Severity: normal
          Priority: P1
         Component: Other
          Assignee: [email protected]
          Reporter: [email protected]
        Regression: Yes

I noticed the following strange behaviour with hugetlbfs mounts for kernel versions something north of 4.19, but certainly since 5.4:

when the mode= option of the mount has the sticky bit set, the mode of the mount point gets scrambled.

To test the following command line can be used:
mkdir -p /tmp/tlbtest && mount -t hugetlbfs -o mode=.... none /tmp/tlbtest && stat -c 'mode: %04a' /tmp/tlbtest

For kernel versions <= 4.19 or if the first byte of mode is 0, stat outputs what was set using "-o mode".

But on newer kernel versions if the first byte of mode is 1 the output of stat is as follows (input -> output):
1700 -> 1244
1750 -> 1326
1770 -> 1352
1775 -> 1357
1777 -> 1361

The behaviour is reproducible across different kernel versions and architectures (5.4.47-amd64, 5.9.8-amd64, 5.10.40-ppc64el, 5.12.15-amd64).

Signature · 4 lines
You may reply to this email to add a comment.

You are receiving this mail because:
You are the assignee for the bug.
#2 lore raw

Add David, Al

Quoted 29 lines · Andrew Morton
Quoted 29 lines

https://bugzilla.kernel.org/show_bug.cgi?id=213785

Summary: hugetlbfs scrambles mode when sticky bit is set
Product: Memory Management
Version: 2.5
Kernel Version: >4.19 (certainly >= 5.4)
Regression: Yes

I noticed the following strange behaviour with hugetlbfs mounts for kernel versions something north of 4.19, but certainly since 5.4:

when the mode= option of the mount has the sticky bit set, the mode of the mount point gets scrambled.

To test the following command line can be used:
mkdir -p /tmp/tlbtest && mount -t hugetlbfs -o mode=.... none /tmp/tlbtest && stat -c 'mode: %04a' /tmp/tlbtest

For kernel versions <= 4.19 or if the first byte of mode is 0, stat outputs what was set using "-o mode".

But on newer kernel versions if the first byte of mode is 1 the output of stat is as follows (input -> output):
1700 -> 1244
1750 -> 1326
1770 -> 1352
1775 -> 1357
1777 -> 1361

Quoted 2 lines
Quoted 2 lines

The behaviour is reproducible across different kernel versions and architectures (5.4.47-amd64, 5.9.8-amd64, 5.10.40-ppc64el, 5.12.15-amd64).

I took a quick look and believe a change in behavior was caused by commit 32021982a324 "Convert the hugetlbfs to use the fs_context during mount".

Prior to the commit, code processing the mode option used match_octal() to convert the command line string to a numeric value. Since match_octal expects a string representing an octal value, it does not require a leading '0'. As a result, prior to this commit the argument 'mode=1700' would result in a mode value of 01700. After the commit one must precede octal values with 0. So, mode=1700 would result in a mode value of 03244 (& 01777U) = 1244.

If my analysis is correct, I am not sure how to proceed. IMO, the current behavior is 'more correct'. However, until v5.1 a preceeding 0 was not required when specifying mode for hugetlbfs. So, this was certainly a change in behavior. Suggestions?

Signature · 1 line
Mike Kravetz
#3 lore raw
Quoted 47 lines · Mike Kravetz

Add David, Al

Quoted 29 lines · Andrew Morton
Quoted 29 lines

https://bugzilla.kernel.org/show_bug.cgi?id=213785

Summary: hugetlbfs scrambles mode when sticky bit is set
Product: Memory Management
Version: 2.5
Kernel Version: >4.19 (certainly >= 5.4)
Regression: Yes

I noticed the following strange behaviour with hugetlbfs mounts for kernel versions something north of 4.19, but certainly since 5.4:

when the mode= option of the mount has the sticky bit set, the mode of the mount point gets scrambled.

To test the following command line can be used:
mkdir -p /tmp/tlbtest && mount -t hugetlbfs -o mode=.... none /tmp/tlbtest && stat -c 'mode: %04a' /tmp/tlbtest

For kernel versions <= 4.19 or if the first byte of mode is 0, stat outputs what was set using "-o mode".

But on newer kernel versions if the first byte of mode is 1 the output of stat is as follows (input -> output):
1700 -> 1244
1750 -> 1326
1770 -> 1352
1775 -> 1357
1777 -> 1361

Quoted 2 lines
Quoted 2 lines

The behaviour is reproducible across different kernel versions and architectures (5.4.47-amd64, 5.9.8-amd64, 5.10.40-ppc64el, 5.12.15-amd64).

I took a quick look and believe a change in behavior was caused by commit 32021982a324 "Convert the hugetlbfs to use the fs_context during mount".

Prior to the commit, code processing the mode option used match_octal() to convert the command line string to a numeric value. Since match_octal expects a string representing an octal value, it does not require a leading '0'. As a result, prior to this commit the argument 'mode=1700' would result in a mode value of 01700. After the commit one must precede octal values with 0. So, mode=1700 would result in a mode value of 03244 (& 01777U) = 1244.

Quoted 4 lines

If my analysis is correct, I am not sure how to proceed. IMO, the current behavior is 'more correct'. However, until v5.1 a preceeding 0 was not required when specifying mode for hugetlbfs. So, this was certainly a change in behavior. Suggestions?

Probably should follow what shmem does:

mm/shmem.c: fsparam_u32oct("mode", Opt_mode),

(also several other filesystems)

#5 lore raw

Re: [Bug 213785] New: hugetlbfs: interprets mode as decimal (was: hugetlbfs scrambles mode when sticky bit is set)

In reply to Mike Kravetz

Dear Mike,

Quoted 3 lines · On Tue, 20 Jul 2021 21:38 -0700 Mike Kravetz wrote:

I took a quick look and believe a change in behavior was caused by commit 32021982a324 "Convert the hugetlbfs to use the fs_context during mount".

Thanks for taking a look at this issue and pointing to the commit that introduced this change.
It looks credible.

Quoted 10 lines

Prior to the commit, code processing the mode option used match_octal() to convert the command line string to a numeric value. Since match_octal expects a string representing an octal value, it does
not require a leading '0'. As a result, prior to this commit the argument 'mode=1700' would result in a mode value of 01700. After the
commit one must precede octal values with 0. So, mode=1700 would result
in a mode value of 03244 (& 01777U) = 1244.

I've been racking my brain trying to figure out what was going on, but the solution is so simple…

Quoted 5 lines

If my analysis is correct, I am not sure how to proceed. IMO, the current behavior is 'more correct'. However, until v5.1 a preceeding 0
was not required when specifying mode for hugetlbfs. So, this was certainly a change in behavior. Suggestions?

Generally, I would agree that interpreting integers as decimals is the correct way, but in the case of file modes I cannot remember having used or seen modes in anything but octal over all the years I've been working with Linux.

As I understand Documentation/admin-guide/mm/hugetlbpage.rst the mode option should be interpreted as octal:

Quoted 1 line

The mode option sets [the mode] to value & 01777. This value is given

in octal. By default the value 0755 is picked.

Furthermore, I had a look at the POSIX documentation for chmod(1)
(https://pubs.opengroup.org/onlinepubs/9699919799/utilities/chmod.html)
which also interprets all integers as octals.

So my suggestion would be to convert the mode option to a fsparam_u32oct and if possible backport this change to the supported

Quoted 1 line

=5.1 kernel versions.

Regards,
Dennis

#4 lore raw
In reply to Matthew Wilcox
Quoted 23 lines · Matthew Wilcox
Quoted 17 lines · Mike Kravetz

Add David, Al

Quoted 1 line · Andrew Morton
Quoted 1 line

https://bugzilla.kernel.org/show_bug.cgi?id=213785

I took a quick look and believe a change in behavior was caused by commit 32021982a324 "Convert the hugetlbfs to use the fs_context during mount".

Prior to the commit, code processing the mode option used match_octal() to convert the command line string to a numeric value. Since match_octal expects a string representing an octal value, it does not require a leading '0'. As a result, prior to this commit the argument 'mode=1700' would result in a mode value of 01700. After the commit one must precede octal values with 0. So, mode=1700 would result in a mode value of 03244 (& 01777U) = 1244.

Quoted 4 lines

If my analysis is correct, I am not sure how to proceed. IMO, the current behavior is 'more correct'. However, until v5.1 a preceeding 0 was not required when specifying mode for hugetlbfs. So, this was certainly a change in behavior. Suggestions?

Quoted 6 lines

Probably should follow what shmem does:

mm/shmem.c: fsparam_u32oct("mode", Opt_mode),

(also several other filesystems)

Thanks Matthew!

That looks to be exactly what is needed.

Signature · 1 line
Mike Kravetz