+CC linux-api
On Tue, Jun 29, 2021 at 10:10 PM Gabriel Krisman Bertazi
[off-list ref] wrote:
Hi,
This is the third version of the FAN_FS_ERROR patches. The main change
in this version is the inode information being reported through an FID
record, which means it requires the group to be created with
FAN_REPORT_FID. It indeed simplifies a lot the FAN_FS_ERROR patch
itself.
I am glad that you took this path.
Uniformity across the UAPI is important.
This change raises the question of how we report non-inode errors. On
one hand, we could omit the FID report, but then fsid would also be
ommited. I chose to report these kind of errors against the root
inode.
There are other option to consider.
To avoid special casing error events in fanotify event read code,
it would is convenient to use a non-zero length FID, but you can
use a 8 bytes zero buffer as NULL-FID
If I am not mistaken, that amounts to 64 bytes of event_len
including the event_metadata and both records which is pretty
nicely aligned.
All 3 handle_type options below are valid options:
1. handle_type FILEID_ROOT
2. handle_type FILEID_INVALID
3. handle_type FILEID_INO32_GEN (i.e. ino=0;gen=0)
The advantage of option #3 is that the monitoring program
does not need to special case the NULL_FID case when
parsing the FID to informative user message.
The other changes in this iteration were made to attend to Amir
feedback. Thank you again for your very detailed input. It is really
appreciated.
This was tested with LTP for regressions, and also using the sample on
the last patch, with a corrupted image. I can publish the bad image
upon request.
Just to set expectations, we now have an official standard for fanotify [1]
where we require an LTP test and man page update patch before merge
of UAPI changes.
That should not stop us from continuing the review process - it's just
a heads up, but I think that we are down to implementation details in
the review anyway and that the UAPI (give or take root inode) is
pretty much clear at this point, so spreading the review of UAPI to
wider audience is not a bad idea.
w.r.t man page update, I know you have created the admin-guide book,
but it's not the same. For linux-api reviewers, reviewing the changed to
fanotify man pages is good way to make sure we did not miss any corners.
w.r.t LTP test, I don't think that using a corrupt image will be a good way
for an LTP test. LTP tests can prepare and mount an ext4 loop image.
Does ext4 have some debugging method to inject an error?
Because that would be the best way IMO.
If it doesn't, you can implement this in ext4 and use it in the test if that
debug file exists - skip the test otherwise - it's common practice.
Thanks,
Amir.
[1] https://lore.kernel.org/linux-fsdevel/YMKv1U7tNPK955ho@google.com/
On Wed 30-06-21 08:10:39, Amir Goldstein wrote:
quoted
This change raises the question of how we report non-inode errors. On
one hand, we could omit the FID report, but then fsid would also be
ommited. I chose to report these kind of errors against the root
inode.
There are other option to consider.
Yeah, so reporting against root inode has the disadvantage that in
principle you don't know whether the error really happened on the root
inode or whether the event is in fact without an inode. So some information
is lost here. Maybe the set of errors that can happen without an inode and
the set of errors that can happen with an inode are disjoint, so no
information is actually lost but then does reporting root inode actually
bring any benefit? So I agree reporting root inode is not ideal.
To avoid special casing error events in fanotify event read code,
it would is convenient to use a non-zero length FID, but you can
use a 8 bytes zero buffer as NULL-FID
If I am not mistaken, that amounts to 64 bytes of event_len
including the event_metadata and both records which is pretty
nicely aligned.
All 3 handle_type options below are valid options:
1. handle_type FILEID_ROOT
2. handle_type FILEID_INVALID
3. handle_type FILEID_INO32_GEN (i.e. ino=0;gen=0)
The advantage of option #3 is that the monitoring program
does not need to special case the NULL_FID case when
parsing the FID to informative user message.
I actually like #2 more. #1 has similar problems as I outlined above for
reporting root dir. The advantage that userspace won't have to special case
FILEID_INO32_GEN FID in #3 is IMHO a dream - if you want a good message,
you should report the problem was on a superblock, not some just some
zeroes instead of proper inode info. Even more if it was on a real inode,
good reporter will e.g. try to resolve it to a path.
Also because we will presumably have more filesystems supporting this in
the future, normal inodes can be reported with other handle types anyway.
So IMO #2 is the most sensible option.
w.r.t LTP test, I don't think that using a corrupt image will be a good way
for an LTP test. LTP tests can prepare and mount an ext4 loop image.
Does ext4 have some debugging method to inject an error?
Because that would be the best way IMO.
If it doesn't, you can implement this in ext4 and use it in the test if that
debug file exists - skip the test otherwise - it's common practice.
Ext4 does not have an error injection facility. Not sure if we want to
force Gabriel into creating one just for these LTP tests. Actually creating
ext4 images with known problems (through mke2fs and debugfs) should be
rather easy and we could then test we get expected error notifications
back...
Honza
--
Jan Kara [off-list ref]
SUSE Labs, CR
On Thu, Jul 8, 2021 at 2:32 PM Jan Kara [off-list ref] wrote:
On Wed 30-06-21 08:10:39, Amir Goldstein wrote:
quoted
quoted
This change raises the question of how we report non-inode errors. On
one hand, we could omit the FID report, but then fsid would also be
ommited. I chose to report these kind of errors against the root
inode.
There are other option to consider.
Yeah, so reporting against root inode has the disadvantage that in
principle you don't know whether the error really happened on the root
inode or whether the event is in fact without an inode. So some information
is lost here. Maybe the set of errors that can happen without an inode and
the set of errors that can happen with an inode are disjoint, so no
information is actually lost but then does reporting root inode actually
bring any benefit? So I agree reporting root inode is not ideal.
quoted
To avoid special casing error events in fanotify event read code,
it would is convenient to use a non-zero length FID, but you can
use a 8 bytes zero buffer as NULL-FID
If I am not mistaken, that amounts to 64 bytes of event_len
including the event_metadata and both records which is pretty
nicely aligned.
All 3 handle_type options below are valid options:
1. handle_type FILEID_ROOT
2. handle_type FILEID_INVALID
3. handle_type FILEID_INO32_GEN (i.e. ino=0;gen=0)
The advantage of option #3 is that the monitoring program
does not need to special case the NULL_FID case when
parsing the FID to informative user message.
I actually like #2 more. #1 has similar problems as I outlined above for
reporting root dir. The advantage that userspace won't have to special case
FILEID_INO32_GEN FID in #3 is IMHO a dream - if you want a good message,
you should report the problem was on a superblock, not some just some
zeroes instead of proper inode info. Even more if it was on a real inode,
good reporter will e.g. try to resolve it to a path.
Also because we will presumably have more filesystems supporting this in
the future, normal inodes can be reported with other handle types anyway.
So IMO #2 is the most sensible option.
I am perfectly fine with #2, but just FYI, there is no ambiguity around using
FILEID_ROOT - it is a special "application level" type used by nfsd to describe
a handle to the root of an export (I'm not in which protocol versions).
The inode itself does not even need to be the root of the filesystem
(it seldom is)
nfsd keeps a dentry with elevated refcount per export in order to "decode"
those export root handles.
The filesystem itself will never return this value, so when reporting
errors on a
specific inode and the specific inode is the filesystem root directory then
the type will be the same as the native filesystem type and not FILEID_ROOT.
For this reason I thought it would be safe to use FILEID_ROOT for reporting
events on sb without inode and FILEID_INVALID for reporting failure to encode
fh, for example when not guessing max_fh_len correctly.
But it's just nice to have.
quoted
w.r.t LTP test, I don't think that using a corrupt image will be a good way
for an LTP test. LTP tests can prepare and mount an ext4 loop image.
Does ext4 have some debugging method to inject an error?
Because that would be the best way IMO.
If it doesn't, you can implement this in ext4 and use it in the test if that
debug file exists - skip the test otherwise - it's common practice.
Ext4 does not have an error injection facility. Not sure if we want to
force Gabriel into creating one just for these LTP tests. Actually creating
ext4 images with known problems (through mke2fs and debugfs) should be
rather easy and we could then test we get expected error notifications
back...
Ok.
Thanks,
Amir.