This is a better followup to dc9239dd7b
since I wanted to do fsync-less checkouts in rpm-ostree too, and
replicating the "turn off fsync temporarily" was in retrospect just a
hack.
We can simply add a boolean to the checkout options.
https://github.com/GNOME/ostree/pull/172
Originally, a lot of the `fsync()` calls here were added for the
wrong reason - I was chasing a bug that ended up being the extlinux
bootloader not parsing 64 bit ext4 filesystems. But since it looked
like corruption, I tried adding a lot more `fsync()` calls.
All we should have to do is use `syncfs()`. If that doesn't work,
it's a kernel bug.
I'm making this change because skipping the individual fsyncs can be a
major performance win - it's easier for the FS to optimize, we do more
in parallel, etc.
https://bugzilla.gnome.org/show_bug.cgi?id=757117
A fast way to generate new OSTree content using an existing
tree is to checkout (as hard links), add/replace files, then
call `ostree_repo_scan_hardlinks()`, then commit.
But `ostree_repo_scan_hardlinks()` scans the entire repo, which
can be slow if you have a lot of content.
All we really need is a mapping of (device,inode) -> checksum
just for the objects we checked out, then use that mapping
for commits.
This patch adds API so that callers can create a mapping via
`ostree_repo_devino_cache_new()`, then pass it to
`ostree_repo_checkout_tree_at()` which will populate it, and then
`ostree_repo_write_directory_to_mtree()` can consume it.
I plan to use this in rpm-ostree for package layering work.
Notes:
- The old `ostree_repo_scan_hardlinks()` API still works.
- I tweaked the cache to be a set with the checksum colocated with
the key, to avoid a separate malloc block per entry.
https://github.com/GNOME/ostree/pull/167
This is a continuation of earlier work to drop the individual fsync on
files/directories in favor of relying on `syncfs()` for speed.
As part of that cleanup, I'm porting it to be fd-relative.
I feel relatively confident about this change given that this area of
the code has notable test suite coverage, although that code runs as
non-root.
I'm porting the deployment code to be fd-relative, but part of the
logic was using `GFile` to talk to `OstreeRepoFile` to determine the
"bootcsum" (boot config checksum) before checking out the file tree.
We can avoid having both code paths by checking out the tree first,
then looking at it on the filesystem.
Bison is a well known external dependency, so just require it.
Including the generated content in git means it may or may not
be regenerated based randomly on timestamps, etc.
Also use `$(AM_V_GEN)` so we get prettier output.
Previously we were just ignoring this, which hid a bug in
an earlier commit that generated them.
Also change the `commit` program to use both APIs - this
involves extra code, but not too much.
This way, reverting the fix with this on top caused the test suite to
fail. Adding an active test for this would need a custom test program
using the C API, or adding a cmdline flag to the client, neither of
which quite seemed worth it.
There might be a race here in that we create new symlink files *after*
calling `syncfs`, and they are not guaranteed to end up on disk.
Rework the code so that we create symlinks before, and then only
rename them after (and `fsync()` the directory for good measure).
Additional-fixes-by: Giuseppe Scrivano <gscrivan@redhat.com>
Tested-by: Giuseppe Scrivano <gscrivan@redhat.com>
This still needs verification that we're fixing a real bug; but I'm
fairly confident this won't make the fsync situation worse.
https://bugzilla.gnome.org/show_bug.cgi?id=755595
I was working on a different test, and ended up being very confused at
the behavior where removing the last deployment didn't remove the last
`ostree/X/X/X` ref pointing to its commit.
There's no reason to special case the last undeployment AFAIK, and the
existing code handles this.
xdg-app was hanging for me with v2015.8, but worked with v2015.7.
I narrowed things down to the GMainLoop/context commit, in which
we started pushing a temporary main context for synchronous
requests internally.
That's never really going to work with libsoup - there needs
to be a single main context which works on the socket. Furthermore,
clients couldn't get progress messages that way.
For *other* internal uses where we added APIs that talk to the remote
repo, we cleanly push a temporary main context.
(Note that I kind of snuck in a change here around the GError handling
in pulls that isn't strictly related but came up in testing)
I noticed xdg-app was looping trying to fetch 1427 refs. We
don't want to do that unless asked to.
(And also, we need to make static delta requests async)
There's no reason to keep them hidden. I have a hard policy that
OSTree should *not* be used to carry secrets. Things like host ssh
private keys should be set up out of band by an OS-external
configuration mechanism such as kickstart, cloud-init, etc.
We also assume that hiding binaries is not very useful as most
attackers would be able to find them on the Internet or (for
subscribed content) acting as a customer.
This fixes a bug with mirroring because we changed to take the
unmodified upstream objects rather than uncompress <-> recompress.
https://bugzilla.gnome.org/show_bug.cgi?id=748959
First of all, what we were doing with having GMainLoop in the internal
APIs is wrong. Synchronous APIs should always create their own main
context and not iterate the caller's. Doing the latter creates
potential for evil reentrancy issues. Sync API should block, async
API is for not blocking.
Now that's out of the way, fix the pull code to do the clean
```
while (termination_condition (state))
g_main_context_iteration (mainctx, TRUE);
```
model for looping. This is a lot easier to understand and ultimately
more reliable than having other code call `g_main_loop_quit()`, as the
loop condition is in exactly one place.
We can also remove the idle source which only fired once.
Note we have to add a hack here to discard the synchronous session and
create a new one which we only use async.
https://bugzilla.gnome.org/show_bug.cgi?id=753336
I'd like to migrate content from the GNOME wiki, as frankly the wiki
is crap. Markdown in git is better in every way.
Start by fleshing out the README.md to be more useful.
libsoup will cache sessions, so it might be the case that we get a
reused session when pulling from the same repo multiple times in one
process.
In this case we were leaking signal connections, which caused
callbacks into freed memory with bad consequences.
Fix it by tying the signal connection to the object lifetime.
I did a quick audit pass through the pull code. What I focused on the
most is the case where `gpg-verify-summary=true`, and in particular
where `gpg-verify=false` too. This should be a valid and secure
configuration.
The primary change here is to error out very quickly if either
`summary` or `summary.sig` are 404. Previously, we'd only error out
if we were processing deltas.
Expand the existing test case to cover this, plus invalid summary and
invalid sig. (The test case was failing with current git master too).
Just noticed this while debugging something else. We don't want to
commit the whole test dir, just the workdir.
Trying to commit the repo itself is potentially subject to race
conditions at least.
Having undefined (but in practice rarely changing) ordering for
`readdir()` ended up screwing us over for bootloader config
generation; see https://bugzilla.redhat.com/show_bug.cgi?id=1226520
Let's make things significantly more likely to fail more quickly in
the future if similar bugs are introduced. We accomplish this by
introducing a little `LD_PRELOAD` library that randomizes the results
of `readdir()`.
We need to check that it's 'ay'. Also reuse the existing validation
function to check it's 32 bytes rather than potentially crashing with
assertion.
Just noticed this during a code review.
Having undefined ordering (but in practice rarely changing)
ordering for `readdir()` ended up screwing us over with respect
to bootloader config file read ordering.
Let's make things significantly more likely to fail more quickly in
the future if similar bugs are introduced. We accomplish this by
introducing a little `LD_PRELOAD` library that randomizes the results
of `readdir()`.
One can run this on a machine to validate things. I'd like to
get this plugged into the actual OSTree tests as soon as we can
figure out how to sanely run grub2-generate as non-root in
our test suite.
Alternatively, this script can easily be run on a real install.
I haven't done a full dig through the history, but it seems quite
possible right now we've been relying on inode enumeration
order for generating bootloader configuration.
Most of the time, newer inodes (i.e. later written files) will win.
But that's obviously not reliable.
Fix this by sorting the returned configuration internally.
When I was introducing the `_UNLOCKED` flag, I only audited
subcommands of `ostree admin`, but I missed that `ostree admin
instutil` also used the option parsing. Those are only used by
Anaconda today so we can ignore them for locking purposes.
Also, the usage help generation was grabbing the lock unnecessarily.
External daemons like rpm-ostree want push notification any time a
change is made by an external entity. inotify provides notification,
but a problem is there's no easy way to monitor all of the refs.
In the past, there has been discussion of opt-in recursive timestamps:
https://lkml.org/lkml/2013/4/5/307
But in today's world, let's just bump the mtime on the repo itself, as
a central inotify point.
Closes: https://github.com/GNOME/ostree/pull/111
When I removed the `transaction` symlink, that made this test start
failing. Fix it by doing `chmod` on `repo/objects`, which is what the
core `ostree_repo_is_writable()` looks at.
The previous commit introduced locking for `ostree admin deploy`, but
we do expect people to possibly accidentally do e.g.
`ostree admin upgrade` concurrently.
Using consistent locking in the admin commands will help rpm-ostree.
Closes: https://github.com/GNOME/ostree/pull/110
The blocking locking API wasn't sufficient for use in the rpm-ostree
daemon; it really wants to know if the lock is held, then continue to
do other things (like service DBus requests), and get notification
when the lock is available.
We also add an async variant that can be called if the lock is not
available.
Implement a higher level "loop until lock is available" method in the
`ostree admin` commandline.
This originally was a way that we detected the case where a pull was
interrupted. Later, we added `.commitpartial` files which also cover
this case.
See also https://github.com/GNOME/ostree/pull/85
We still want to honor their existence (and unlink them) in case an
old version of ostree was in use, but I believe it's safe to stop
creating them now.
The only case where this would break is if you have a version of
ostree that predates commitpartial in your rollback history, but such
old versions are no longer in use by operating systems I support at
least.
Closes: https://github.com/GNOME/ostree/pull/100
An OSTree user noticed that `ostree fsck` would produce `missing
object` errors in the case of interrupted pulls.
It's possible to do e.g. `ostree pull --subpath=/usr/share/rpm ...`,
which gets you just that portion of the commit. The use case for this
was being able to see what changes would appear in an update before
actually downloading all of it.
(I think this would be better covered by static deltas, but those
aren't final yet, and `--subpath` predates it)
Further, `.commitpartial` is used as a successor to the `transaction`
symlink for more precise knowledge in the case where a pull was
interrupted that we needed to resume scanning.
So it makes sense for `ostree fsck` to be aware of it.
First, this is just a general continuation of the `GFile -> openat`
transition.
Second, it's preparatory work for fsck to gain awareness of partial
commits.
If a system administrator happens to type `ostree admin upgrade`
multiple times, currently that will lead to a potentially corrupted
system.
I originally attempted to do locking *internally* in `libostree`, but
that didn't work out because currently a number of the commands
perform multi-step operations that all need to be serialized. All of
the current code in `ostree admin deploy` is an example.
Therefore, allow callers to perform locking, as most of the higher
level logic is presently implemented there.
At some point, we can revisit having internal locking, but it will be
difficult. A more likely approach would be similar to Java's approach
with concurrency on iterators - a "fail fast" method.
These fsyncs were added for what turned out to be a fairly bogus
reason; I was hitting read errors from extlinux after upgrades and out
of conservatisim tried adding fsync calls, but the *actual* problem
was that extlinux didn't support 64 bit ext4. Now that at least for
Project Atomic hosts we're just targeting grub2, we can drop these
fsync calls and rely on `syncfs()` being both faster and catching any
errors.
For some sort of crazy reason, the `sync()` system call doesn't
actually return an error code, even though from what I can tell in the
kernel it wouldn't be terribly hard to add.
Regardless though, it is better for userspace apps to use `syncfs()`
to avoid flushing filesystems unrelated to what they want to sync. In
the case of OSTree, this does matter - for example you might have a
network mount point backing your database, and we don't want to block
upgrades on syncing it.
This change is safe because we're doing syncfs in *addition* to the
previous global `sync()` (a revision from an earlier patch).
Now because OSTree only touches the `/` mount point which covers the
repository, the deployment roots (including their copy of `/etc`), as
well as `/boot`, we should at some point later be able to drop the
`sync()` call. Note that on initial system installs we do relabel
`/var` but that shouldn't happen at ostree time - any new directories
are taken care of via `systemd-tmpfiles` on boot.
This way external programs like rpm-ostree can do fd-relative
operations on the deployment directories, like inspecting the RPM
database.
Closes: https://github.com/GNOME/ostree/pull/91
First, git doesn't do this, and whatever Linus thinks is right or
something.
Second specifically to OSTree, it's quite common to not have
intermediate commits. If one wants to reset a ref in order to prune
data after a deployment, the parentage check will fail.
Closes: https://github.com/GNOME/ostree/pull/87
I was looking at https://bugzilla.gnome.org/show_bug.cgi?id=738954
which wants us to ensure we chown() the refs. As part of that,
I did a generic conversion to use `*at()` (which naturally gives
us more low level control so we can call `fchown` etc.
This patch also sneaks in a change to respect the repo's
`disable_fsync` flag - if fsync is not set, then we never
`fdatasync()` (unlike the `g_file_replace_contents()` default. Also
unlike it, if fsync is enabled, we *always* sync even if the file
didn't exist.
This will be used by rpm-ostree to unset the immutable bit temporarily
in order to do package layering. We could add an API to deploy a tree
without the immutable bit, but this is simpler.
rpm-ostree currently uses ostree_repo_checkout_tree(), which as a side
effect will use the uncompressed objects cache by default. This is
rather annoying if you're using rpm-ostree on a server-side
repository, because if you then rsync the repo, you'll be syncing out
the uncompressed objects unless you exclude them.
We added the ability to disable the uncompressed cache in the
repository config to fix this, but it's better to allow application
control over this. The uncompressed cache will in some future version
become opt in as well.
This new API further:
- Drops the `GFile` usage in favor of `openat` APIs
- Improves ergonomics by avoiding callers having to query the source
`GFileInfo` (and carry around a copy of `OSTREE_GIO_FAST_QUERYINFO`)
- Has a more extensible options structure
Per the comment, I rather crudely have the `ostree checkout` builtin
call both APIs to ensure some testing coverage.
However, I'd like to in the future have easier-to-set-up testing code
that calls `libtest.sh` to set up dummy data.
It's valid for the remote server to say 200 OK and give us the entire
file instead of a 206 Partial Content, and in that case we should blow
away the previous cached data, rather than blindly appending to it and
thus creating multiple copies of the data inside the file.
This problem primarily occurs when we do have the complete file, and
we're interrupted, then try again, where the new process didn't record
the download was already complete. We do a range request for bytes
past the end, and some web servers (e.g. Akamai) will return 200 OK
with the whole content again, rather than a 416 Requested Range Not
Satisfiable.
Thus we could also fix this by saner caching strategy - since we know
the file is complete, rename it again to $checksum.done or something
before it's processed. (Or really, rework how we do caching more
intelligently in general).
This fixes the issue that interrupted pulls failed with such
webservers, although repeated attempts would eventually succeed
because we'd unlink files that failed to pull.
Related: https://bugzilla.redhat.com/show_bug.cgi?id=1207292
The use case for non-default sysroots that I know of are:
1) The current test suite
2) Installers (Anaconda)
3) Inspecting VM disks
For 2) and 3), it'll quickly be obvious if they're not running as
root, and these are more obscure cases. We want to allow 1), and this
is a simple way to do it.
https://bugzilla.gnome.org/show_bug.cgi?id=747164
Colin Walters