Skip to contentSkip to conversation

evm: stop avoidably reading i_writecount in evm_file_release

PATCH In review Mateusz Guzik 3 participants 5 messages Aug 6, 2024 – Oct 10, 2024

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 evm: stop avoidably reading i_writecount in evm_file_release

The EVM_NEW_FILE flag is unset if the file already existed at the time of open and this can be checked without looking at i_writecount.

Not accessing it reduces traffic on the cacheline during parallel open of the same file and drop the evm_file_release routine from second place to bottom of the profile.

---

The context is that I'm writing a patch which removes one lockref get/put cycle on parallel open. An operational WIP reduces ping-pong in that area and made do_dentry_open skyrocket along with evm_file_release, due to i_writecount access. With the patch they go down again and apparmor takes the rightful first place.

The patch accounts for about 5% speed up at 20 cores running open3 from will-it-scale on top of the above wip. (the apparmor + lockref thing really don't scale, that's next)

I would provide better measurements, but the wip is not ready (as the description suggests) and I need evm out of the way for the actual patch.

 security/integrity/evm/evm_main.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
security/integrity/evm/evm_main.c +2 -1
index 62fe66dd53ce..309630f319e2 100644
@@ -1084,7 +1084,8 @@ static void evm_file_release(struct file *file)
 	if (!S_ISREG(inode->i_mode) || !(mode & FMODE_WRITE))
 		return;
 
-	if (iint && atomic_read(&inode->i_writecount) == 1)
+	if (iint && iint->flags & EVM_NEW_FILE &&
+	    atomic_read(&inode->i_writecount) == 1)
 		iint->flags &= ~EVM_NEW_FILE;
 }
 
Signature · 1 line
2.43.0
Reviewed-by on v1 1/1
#1 lore raw
Quoted 2 lines · Mateusz Guzik

The EVM_NEW_FILE flag is unset if the file already existed at the time of open and this can be checked without looking at i_writecount.

Agreed. EVM_NEW_FILE is not going to be set during the open(), only before, in evm_post_path_mknod().

Looks good to me.

Thanks

Roberto

Quoted 38 lines

Not accessing it reduces traffic on the cacheline during parallel open of the same file and drop the evm_file_release routine from second place to bottom of the profile.

---

The context is that I'm writing a patch which removes one lockref get/put cycle on parallel open. An operational WIP reduces ping-pong in that area and made do_dentry_open skyrocket along with evm_file_release, due to i_writecount access. With the patch they go down again and apparmor takes the rightful first place.

The patch accounts for about 5% speed up at 20 cores running open3 from will-it-scale on top of the above wip. (the apparmor + lockref thing really don't scale, that's next)

I would provide better measurements, but the wip is not ready (as the description suggests) and I need evm out of the way for the actual patch.

 security/integrity/evm/evm_main.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
security/integrity/evm/evm_main.c +2 -1
index 62fe66dd53ce..309630f319e2 100644
@@ -1084,7 +1084,8 @@ static void evm_file_release(struct file *file)
 	if (!S_ISREG(inode->i_mode) || !(mode & FMODE_WRITE))
 		return;
 
-	if (iint && atomic_read(&inode->i_writecount) == 1)
+	if (iint && iint->flags & EVM_NEW_FILE &&
+	    atomic_read(&inode->i_writecount) == 1)
 		iint->flags &= ~EVM_NEW_FILE;
 }
 
#2 lore raw
Quoted 5 lines · Roberto Sassu
Quoted 2 lines · Mateusz Guzik

The EVM_NEW_FILE flag is unset if the file already existed at the time of open and this can be checked without looking at i_writecount.

Quoted 8 lines

Agreed. EVM_NEW_FILE is not going to be set during the open(), only before, in evm_post_path_mknod().

Looks good to me.

Thanks

thanks for the review

are there plans to pick this up for this merge window?

Quoted 42 lines

Roberto

Quoted 38 lines

Not accessing it reduces traffic on the cacheline during parallel open of the same file and drop the evm_file_release routine from second place to bottom of the profile.

---

The context is that I'm writing a patch which removes one lockref get/put cycle on parallel open. An operational WIP reduces ping-pong in that area and made do_dentry_open skyrocket along with evm_file_release, due to i_writecount access. With the patch they go down again and apparmor takes the rightful first place.

The patch accounts for about 5% speed up at 20 cores running open3 from will-it-scale on top of the above wip. (the apparmor + lockref thing really don't scale, that's next)

I would provide better measurements, but the wip is not ready (as the description suggests) and I need evm out of the way for the actual patch.

 security/integrity/evm/evm_main.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
security/integrity/evm/evm_main.c +2 -1
index 62fe66dd53ce..309630f319e2 100644
@@ -1084,7 +1084,8 @@ static void evm_file_release(struct file *file)
      if (!S_ISREG(inode->i_mode) || !(mode & FMODE_WRITE))
              return;
 
-     if (iint && atomic_read(&inode->i_writecount) == 1)
+     if (iint && iint->flags & EVM_NEW_FILE &&
+         atomic_read(&inode->i_writecount) == 1)
              iint->flags &= ~EVM_NEW_FILE;
 }
 
Signature · 1 line
Mateusz Guzik <mjguzik gmail.com>
#3 lore raw
Quoted 12 lines · Mateusz Guzik
Quoted 10 lines · Roberto Sassu
Quoted 2 lines · Mateusz Guzik

The EVM_NEW_FILE flag is unset if the file already existed at the time of open and this can be checked without looking at i_writecount.

Agreed. EVM_NEW_FILE is not going to be set during the open(), only before, in evm_post_path_mknod().

Looks good to me.

Quoted 7 lines
Quoted 3 lines

Thanks

thanks for the review

are there plans to pick this up for this merge window?

Welcome! Sorry, Mimi wants to do few more checks. It is more likely that we will pick this for the next version.

Roberto

Quoted 44 lines
Quoted 42 lines

Roberto

Quoted 38 lines

Not accessing it reduces traffic on the cacheline during parallel open of the same file and drop the evm_file_release routine from second place to bottom of the profile.

---

The context is that I'm writing a patch which removes one lockref get/put cycle on parallel open. An operational WIP reduces ping-pong in that area and made do_dentry_open skyrocket along with evm_file_release, due to i_writecount access. With the patch they go down again and apparmor takes the rightful first place.

The patch accounts for about 5% speed up at 20 cores running open3 from will-it-scale on top of the above wip. (the apparmor + lockref thing really don't scale, that's next)

I would provide better measurements, but the wip is not ready (as the description suggests) and I need evm out of the way for the actual patch.

 security/integrity/evm/evm_main.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
security/integrity/evm/evm_main.c +2 -1
index 62fe66dd53ce..309630f319e2 100644
@@ -1084,7 +1084,8 @@ static void evm_file_release(struct file *file)
      if (!S_ISREG(inode->i_mode) || !(mode & FMODE_WRITE))
              return;
 
-     if (iint && atomic_read(&inode->i_writecount) == 1)
+     if (iint && iint->flags & EVM_NEW_FILE &&
+         atomic_read(&inode->i_writecount) == 1)
              iint->flags &= ~EVM_NEW_FILE;
 }
 
maintainer
#4 lore raw
In reply to Mateusz Guzik
Quoted 8 lines · Mateusz Guzik

The EVM_NEW_FILE flag is unset if the file already existed at the time of open and this can be checked without looking at i_writecount.

Not accessing it reduces traffic on the cacheline during parallel open of the same file and drop the evm_file_release routine from second place to bottom of the profile.

Sorry for the delay. It's now queued in next-integrity.

thanks,

Mimi