rofiles: Add --copyup option #1382
Conversation
|
Depends: #1378 |
cgwalters
force-pushed
the
rofiles-copyup
branch
from
689fe28 to
6e06ee7
Compare
| @@ -286,6 +333,9 @@ callback_truncate (const char *path, off_t size) | |||
| static int | |||
| callback_utimens (const char *path, const struct timespec tv[2]) | |||
| { | |||
| /* This one isn't write-verified, we support changing times | |||
| * even for hardlinked files. | |||
There was a problem hiding this comment.
Can you expand on this? Don't we need to keep bare objects at OSTREE_TIMESTAMP to be consistent?
There was a problem hiding this comment.
It's an interesting point...gets a bit into a philosophical question of whether rofiles-fuse is for "anything using hardlinks" or whether it's really specific to ostree. The "anything using hardlinks" case might be a backup system or non-os-data object store was my thought originally.
For example I could probably have "object semantics" for my ~/Documents which has a bunch of random PDFs and .txt and LibreOffice files etc. that should really never be overwritten in place. OTOH I've been thinking about storing all of that in git.
So...hmm. It's not a new issue here and we should think about changing this, but I worry a bit about the backcompat issues. In practice nothing we use rofiles-fuse for should be touching timestamps for files it didn't create at the time I would guess.
There was a problem hiding this comment.
For example I could probably have "object semantics" for my ~/Documents which has a bunch of random PDFs and .txt and LibreOffice files etc. that should really never be overwritten in place. OTOH I've been thinking about storing all of that in git.
I've been using git-annex for this for some time now. Good news, it's in fedora.
There was a problem hiding this comment.
Right, git-annex is likely an excellent example of a possible user of rofiles-fuse semantics, specifically its direct mode. Though I didn't look closely.
src/rofiles-fuse/main.c
Outdated
| if (opt_copyup) | ||
| { | ||
| (void) close (fd); | ||
| fd = openat (basefd, path, finfo->flags & ~O_TRUNC, mode); |
There was a problem hiding this comment.
Minor: we can just stop filtering out O_TRUNC in the copy-up case, right? And only pay the ftruncate call in the non-copy-up case.
There was a problem hiding this comment.
BTW I wish I'd seen this comment while working on this patchset on the plane O_TRUNC was the source of bugs.
Sadly https://sourceware.org/bugzilla/show_bug.cgi?id=22089 is I think going to actually force us to cave here. Even if we got the glibc patch in today, we need to support the RHEL glibc. See also discussion about fish as part of the general Fedora tracker. This is basically needed to unblock rpm-ostree unified core 🌐: coreos/rpm-ostree#729 Closes: ostreedev#1377
cgwalters
force-pushed
the
rofiles-copyup
branch
from
a638c02 to
276f54d
Compare
cgwalters
force-pushed
the
rofiles-copyup
branch
from
276f54d to
ab634a2
Compare
|
Consumer: coreos/rpm-ostree#1171 |
src/rofiles-fuse/main.c
Outdated
| @@ -385,7 +385,12 @@ do_open (const char *path, mode_t mode, struct fuse_file_info *finfo) | |||
| if (opt_copyup) | |||
| { | |||
| (void) close (fd); | |||
| fd = openat (basefd, path, finfo->flags & ~O_TRUNC, mode); | |||
| /* Note that unlike the initial open, we will pass through | |||
| * O_TRUNC. More ideally in this copyup case we'd avoid copying | |||
There was a problem hiding this comment.
Hmm, so then we don't need to ftruncate down below again, right? Shouldn't the if condition at 399 be changed to
if (!opt_copyup && (finfo->flags & O_TRUNC))
?
| * the whole file in the first place, but eh. It's not like we're | ||
| * high performance anyways. | ||
| */ | ||
| fd = openat (basefd, path, finfo->flags & ~(O_EXCL|O_CREAT), mode); |
There was a problem hiding this comment.
Why are we filtering out O_EXCL and O_CREAT here? This is something that should've been caught by the first openat, right? Or is this just to not have a TOCTOU race on the file existing?
There was a problem hiding this comment.
We're internally reopening the file we just created. So if the user provided O_EXCL we need to filter that out. We might as well drop O_CREAT too.
|
⚡ Test exempted: merge already tested. |
Sadly https://sourceware.org/bugzilla/show_bug.cgi?id=22089 is I think going to
actually force us to cave here. Even if we got the glibc patch in today, we need
to support the RHEL glibc. See also discussion about fish as part of the general
Fedora tracker.
This is basically needed to unblock rpm-ostree unified core 🌐:
coreos/rpm-ostree#729
Closes: #1377