remove_suid bangs on xattrs
Subsystems (3)
Hi everyone,
I'm looking into a 2.6.35 btrfs performance regression, and perf tells me that I'm spending a lot of time hammering on xattrs inside remove_suid. This is pretty surprising because I'm running as root, and my files are not suid. Looking back to this commit:
We've changed remove_suid's semantics from
if (file_is_suid)
try to remove it
To something that always checks to see if we have removal permissions.
Was this intentional? It didn't cause my 2.6.35 regression (that's all my fault) but it does look wrong to me:
mm/filemap.c +10 -4
@@ -1627,12 +1627,18 @@ int __remove_suid(struct dentry *dentry, int kill)int remove_suid (struct dentry *dentry) { -int kill =should_remove_suid (dentry); +int killsuid =should_remove_suid (dentry); +int killpriv =security_inode_need_killpriv (dentry); +int error =0 ; -if (unlikely (kill)) -return __remove_suid (dentry, kill); +if (killpriv <0 ) +return killpriv; +if (killpriv) + error =security_inode_killpriv (dentry); +if (!error && killsuid) + error =__remove_suid (dentry, killsuid); -return 0 ; +return error; }EXPORT_SYMBOL (remove_suid);
-chris
[ sorry, corrected cc list ]
Quoted 48 lines · Chris Mason
Hi everyone,
I'm looking into a 2.6.35 btrfs performance regression, and perf tells me that I'm spending a lot of time hammering on xattrs inside
remove_suid. This is pretty surprising because I'm running as root, and my files are not suid. Looking back to this commit:We've changed remove_suid's semantics from
if (
file_is_suid)
try to remove itTo something that always checks to see if we have removal permissions.
Was this intentional? It didn't cause my 2.6.35 regression (that's all my fault) but it does look wrong to me:
mm/filemap.c +10
-5@@ -1627,12 +1627,18 @@ int __remove_suid(struct dentry *dentry, int kill)int remove_suid (struct dentry *dentry) { -int kill =should_remove_suid (dentry); +int killsuid =should_remove_suid (dentry); +int killpriv =security_inode_need_killpriv (dentry); +int error =0 ; -if (unlikely (kill)) -return __remove_suid (dentry, kill); +if (killpriv <0 ) +return killpriv; +if (killpriv) + error =security_inode_killpriv (dentry); +if (!error && killsuid) + error =__remove_suid (dentry, killsuid); -return 0 ; +return error; }EXPORT_SYMBOL (remove_suid); -chris
Quoted 1 line · Chris Mason
[ sorry, corrected cc list ]
Thanks - sorry for the inconvenience. I'm also cc:ing Andrew Morgan for another opinion.
Quoted 9 lines · Chris Mason
Quoted 8 lines
Hi everyone,
I'm looking into a 2.6.35 btrfs performance regression, and perf tells me that I'm spending a lot of time hammering on xattrs inside
remove_suid. This is pretty surprising because I'm running as root, and my files are not suid. Looking back to this commit:
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=b53767719b6cd8789392ea3e7e2eb7b8906898f0
Quoted 4 lines
Quoted 4 lines
We've changed remove_suid's semantics from
if (
file_is_suid)
try to remove it
(but only if not capable(CAP_FSETID))
Quoted 1 line
Quoted 1 line
To something that always checks to see if we have removal permissions.
(not really - security_inode_need_killpriv() shoudl return <0 only if there was an actual error, and the write needs to be cancelled altogether. It returns 0 if privs don't need to be removed, and >0 if they do.
Quoted 2 lines
Quoted 2 lines
Was this intentional? It didn't cause my 2.6.35 regression (that's all my fault) but it does look wrong to me:
If I'm thinking right, I think the key change we should make is to have CAP_FSETID be honored for maintaining file capabilities.
That would have two (good) results:
1. we should be able to re-arrange the code to check for CAP_FSETID before bothering to check for file capabilities, so we can save the getxattrs which I assume were what you were finding? Even if it wasn't the cause of your performance regression, it should be an improvement.
2. I think it can be seen as a semantic fix. We mostly try to respect suid behavior for file caps, so it will be more consistent to honor CAP_FSETID for file capabilities.
Andrew, what do you think?
Quoted 22 lines
Quoted 22 lines
mm/filemap.c +9
-3@@ -1627,12 +1627,18 @@ int __remove_suid(struct dentry *dentry, int kill)int remove_suid (struct dentry *dentry) { -int kill =should_remove_suid (dentry); +int killsuid =should_remove_suid (dentry); +int killpriv =security_inode_need_killpriv (dentry); +int error =0 ; -if (unlikely (kill)) -return __remove_suid (dentry, kill); +if (killpriv <0 ) +return killpriv; +if (killpriv) + error =security_inode_killpriv (dentry); +if (!error && killsuid) + error =__remove_suid (dentry, killsuid);
Quoted 6 lines
mm/filemap.c int __remove_suid(struct dentry *dentry, int kill)
Quoted 6 lines
mm/filemap.c +1
-2-return 0 ; +return error; }EXPORT_SYMBOL (remove_suid); -chris
thanks,
-serge
Quoted 15 lines · Serge E. Hallyn
Quoted 1 line · Chris Mason
[ sorry, corrected cc list ]
Thanks - sorry for the inconvenience. I'm also cc:ing Andrew Morgan for another opinion.
Quoted 9 lines · Chris Mason
Quoted 8 lines
Hi everyone,
I'm looking into a 2.6.35 btrfs performance regression, and perf tells me that I'm spending a lot of time hammering on xattrs inside
remove_suid. This is pretty surprising because I'm running as root, and my files are not suid. Looking back to this commit:
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=b53767719b6cd8789392ea3e7e2eb7b8906898f0
Quoted 6 lines
Quoted 4 lines
Quoted 4 lines
We've changed remove_suid's semantics from
if (
file_is_suid)
try to remove it(but only if not
capable(CAP_FSETID))
I disagree. I think the relevant capability test should be with respect to CAP_SETFCAP.
Since this is the capability that allows you to put a capability on a file, it should be the one to retain it if the file is modified.
Quoted 20 lines
Quoted 1 line
Quoted 1 line
To something that always checks to see if we have removal permissions.
(not really -
security_inode_need_killpriv()shoudl return <0 only if there was an actual error, and the write needs to be cancelled altogether. It returns 0 if privs don't need to be removed, and >0 if they do.Quoted 2 lines
Quoted 2 lines
Was this intentional? It didn't cause my 2.6.35 regression (that's all my fault) but it does look wrong to me:
If I'm thinking right, I think the key change we should make is to have
CAP_FSETIDbe honored for maintaining file capabilities.That would have two (good) results:
1. we should be able to re-arrange the code to check for
CAP_FSETIDbefore bothering to check for file capabilities, so we can save the getxattrs which I assume were what you were finding? Even if it wasn't the cause of your performance regression, it should be an improvement.
Quoted 6 lines
2. I think it can be seen as a semantic fix. We mostly try to respect suid behavior for file caps, so it will be more consistent to honor
CAP_FSETIDfor file capabilities.Andrew, what do you think?
I think the test should be with respect to CAP_SETFCAP, but I agree with the rest of your comments.
Lots of small writes to 'any' file also tends to bang on this code. I've been wondering if it might make sense to cache, in the inode, that a file does *not* have any capabilities associated with it. That way the kernel wouldn't need to look up the xattrs twice for the same incapable file - which is, by far, the common case.
Cheers
Andrew
Quoted 27 lines
Quoted 27 lines
Quoted 27 lines
mm/filemap.c +10
-4@@ -1627,12 +1627,18 @@ int __remove_suid(struct dentry *dentry, int kill) int remove_suid(struct dentry *dentry) { - int kill = should_remove_suid(dentry); + int killsuid = should_remove_suid(dentry); + int killpriv = security_inode_need_killpriv(dentry); + int error = 0; - if (unlikely(kill)) - return __remove_suid(dentry, kill); + if (killpriv < 0) + return killpriv; + if (killpriv) + error = security_inode_killpriv(dentry); + if (!error && killsuid) + error = __remove_suid(dentry, killsuid); - return 0; + return error; } EXPORT_SYMBOL(remove_suid);
Quoted 5 lines
Quoted 1 line
Quoted 1 line
-chris
thanks,
-serge
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 22 lines · Andrew G. Morgan
Quoted 20 lines · Serge E. Hallyn
Quoted 1 line · Chris Mason
[ sorry, corrected cc list ]
Thanks - sorry for the inconvenience. I'm also cc:ing Andrew Morgan for another opinion.
Quoted 13 lines · Chris Mason
Quoted 12 lines
Hi everyone,
I'm looking into a 2.6.35 btrfs performance regression, and perf tells me that I'm spending a lot of time hammering on xattrs inside
remove_suid. This is pretty surprising because I'm running as root, and my files are not suid. Looking back to this commit:
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=b53767719b6cd8789392ea3e7e2eb7b8906898f0We've changed remove_suid's semantics from
if (
file_is_suid)
try to remove it
Quoted 7 lines
Quoted 1 line
(but only if not
capable(CAP_FSETID))I disagree. I think the relevant capability test should be with respect to
CAP_SETFCAP.Since this is the capability that allows you to put a capability on a file, it should be the one to retain it if the file is modified.
I'm ok with that.
Quoted 24 lines
Quoted 24 lines
Quoted 1 line
Quoted 1 line
To something that always checks to see if we have removal permissions.
(not really -
security_inode_need_killpriv()shoudl return <0 only if there was an actual error, and the write needs to be cancelled altogether. It returns 0 if privs don't need to be removed, and >0 if they do.Quoted 2 lines
Quoted 2 lines
Was this intentional? It didn't cause my 2.6.35 regression (that's all my fault) but it does look wrong to me:
If I'm thinking right, I think the key change we should make is to have
CAP_FSETIDbe honored for maintaining file capabilities.That would have two (good) results:
1. we should be able to re-arrange the code to check for
CAP_FSETIDbefore bothering to check for file capabilities, so we can save the getxattrs which I assume were what you were finding? Even if it wasn't the cause of your performance regression, it should be an improvement.2. I think it can be seen as a semantic fix. We mostly try to respect suid behavior for file caps, so it will be more consistent to honor
CAP_FSETIDfor file capabilities.
Quoted 5 lines
Quoted 2 lines
Andrew, what do you think?
I think the test should be with respect to
CAP_SETFCAP, but I agree with the rest of your comments.
I also point out, with some shame, that on first reading Chris' email, I had to look at that code more than I should have needed to to recall the details. So I think the function could stand some cleaning up/ simplifying.
I'll try to take a look at this next week.
Quoted 5 lines
Lots of small writes to 'any' file also tends to bang on this code. I've been wondering if it might make sense to cache, in the inode, that a file does *not* have any capabilities associated with it. That way the kernel wouldn't need to look up the xattrs twice for the same incapable file - which is, by far, the common case.
That could also be shared with a new (old) optional xattr-free file-backed-filecaps mount option :)
thanks,
-serge
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 9 lines · Andreas Dilger
Quoted 5 lines · Andrew G. Morgan
Lots of small writes to 'any' file also tends to bang on this code. I've been wondering if it might make sense to cache, in the inode, that a file does *not* have any capabilities associated with it. That way the kernel wouldn't need to look up the xattrs twice for the same incapable file - which is, by far, the common case.
That would be a blessing. I see a steady stream of
getxattr("security.capability")requests, and being able to disable this
Do you think it would help at all to add a S_NO_POSIXCAPS to i_flags, and set that the first time we find that getxattr("security.capability") finds no capabilities? I.e. are these requests frequently for the same inode, or always for new ones?
Quoted 2 lines
(possibly even in the superblock with a flag) would avoid expensive RPCs on a network filesystem.
Hmm, as it is, the get_vfs_caps_from_disk() does not get called if MNT_NOSUID. But the cap_inode_need_killpriv() does, so a quick way to reduce that # for you would be to pass the inode to security_inode_need_killpriv (so it can get to mnt), and have that check for MNT_NOSUID, and then you can mount your network fs's with MNT_NOSUID... Would that help you?
-serge
Quoted 8 lines · Serge E. Hallyn
Quoted 7 lines · Andreas Dilger
Quoted 5 lines · Andrew G. Morgan
Lots of small writes to 'any' file also tends to bang on this code. I've been wondering if it might make sense to cache, in the inode, that a file does *not* have any capabilities associated with it. That way the kernel wouldn't need to look up the xattrs twice for the same incapable file - which is, by far, the common case.
Quoted 8 lines
Quoted 2 lines
That would be a blessing. I see a steady stream of
getxattr("security.capability")requests, and being able to disable thisDo you think it would help at all to add a
S_NO_POSIXCAPStoi_flags, and set that the first time we find thatgetxattr("security.capability")finds no capabilities? I.e. are these requests frequently for the same inode, or always for new ones?
That would be useful, or as you suggest a MNT_* flag.
Quoted 9 lines
Quoted 2 lines
(possibly even in the superblock with a flag) would avoid expensive RPCs on a network filesystem.
Hmm, as it is, the
get_vfs_caps_from_disk()does not get called ifMNT_NOSUID. But thecap_inode_need_killpriv()does, so a quick way to reduce that # for you would be to pass the inode tosecurity_inode_need_killpriv(so it can get to mnt), and have that check forMNT_NOSUID, and then you can mount your network fs's withMNT_NOSUID... Would that help you?
Except there are users that do use SUID binaries on network filesystems (e.g. root or /usr filesystems). Something more like MNT_NOCAPXATTR would be better.
Cheers, Andreas