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 $checksum.done or something
before it's processed. (Or really, rework how we do caching more
intelligently in general).
This fixes the issue that interrupted pulls failed with such
webservers, although repeated attempts would eventually succeed
because we'd unlink files that failed to pull.
Related: https://bugzilla.redhat.com/show_bug.cgi?id=1207292
The use case for non-default sysroots that I know of are:
1) The current test suite
2) Installers (Anaconda)
3) Inspecting VM disks
For 2) and 3), it'll quickly be obvious if they're not running as
root, and these are more obscure cases. We want to allow 1), and this
is a simple way to do it.
https://bugzilla.gnome.org/show_bug.cgi?id=747164
Checkout was one of the first complex code paths I tried to convert to
*at(). I ended up keeping both, because I hit the "xattrs for a
symlink" problem. Later, Florian gave me a workaround, and we started
using it here, but the GFile * parameters weren't deleted. They're
not used, so do so now.
Starting down the path of not using libgsystem. The main win here
will be code sharing between ostree/rpm-ostree as well as going down
the path of not using GFile * for local files.
The previous diff algorithm was file tree based, and only looked
at modified files that lived at the same path.
However, components like the Linux kernel have versioned
subdirectories, e.g. /usr/lib/modules/$kver/.../ext4.ko. We want to
be able to detect these "modified renames" so that we can compute
diffs (rollsum, bsdiff).
This does an rsync-style prepared delta basically. On my test data,
it shaves ~6MB of uncompressed data. Not a huge amount, but I expect
this to be more useful for things like binaries which embed data, etc.
There's still some silliness here, but there is now only one opcode
open-splice-and-close, that writes a single chunk from the payload.
This is really all we need for metadata, and small content objects are
also fine with this.
We get some deduplication between content objects by creating a
dictionary for (uid,gid,mode) tuples and xattrs.
This still keeps the operation/payload code in, so we could do
rollsums in a future update easily.
I was hitting a bug in libguestfs/guestmount/FUSE where it blew up
with EINVAL on directories containing lots of files (more than
32000?). We really want to use prefixed subdirs just like the real
objects/ directory does.
This allows us to share more code between the paths, is more
efficient, etc.
gnome-continuous uses the ostree_repo_scan_hardlinks() mode to
avoid re-checksumming everything. However, when I ported the commit
code to use openat() and friends, this optimization was lost.
Re add it. The difference is about 15s versus 5 minutes.
This follows up from the previous commit; now that pull knows how to
do the efficient link() or copy for local files, we can just have
pull-local call into ostree_repo_pull().
As part of this:
- pull() can also accept a file:/// URI instead
of a remote name (since pull local supports anonymous pulls)
- pull() knows an "override-remote-name" option, since pull-local
supported writing a ref out even if there wasn't a remote with
that name
It's always been suboptimal to have both pull and pull-local; as we go
beyond the raw object data into things like deltas and summary files,
the logic to perform e.g. mirroring should only be in one place.
This will be used by Pulp's OSTree content plugin at least to perform
promotions.
When doing a pull --mirror from an archive-z2 repository into another
archive-z2 repository, currently we gunzip/checksum/gzip each content
object. The re-gzip process in particular is fairly expensive.
This does assume that the upstream content is trusted and correct.
It'd be nice in the future to do at least a CRC check, if not the full
checksum. (Could we append CRC data to the end of filez objects?)
We could also choose to only do this optimization if fetching over
TLS.
before: 1626 metadata, 20320 content objects fetched; 299634 KiB transferred in 62 seconds
after : 1626 metadata, 20320 content objects fetched; 299634 KiB transferred in 11 seconds
For future delta work where we do more interesting things than just
"tar of new objects", this lays the groundwork for doing streaming
writes into content objects.
It's also more efficient, as we avoid many intermediate allocations
and virtual calls. Just a single `g_output_stream_write_all` for the
splice case.
Conflicts:
src/libostree/ostree-repo-private.h
src/libostree/ostree-repo-static-delta-processing.c
We could just make everything relative to this, but the objects/ and
tmp/ are accessed very often, so I think it's worth holding individual
fds.
This fd can cover everything else: refs, deltas, etc.
subscription-manager has a daemon that runs in a confined domain,
and it doesn't have permission to write usr_t, which is the default
label of /ostree/deploy/$osname/deploy.
A better long term fix is probably to move the origin file into the
deployment root as /etc/ostree/origin.conf or so.
In the meantime, let's ensure the .origin files are labeled as
configuration.
If an object already existed and we somehow tried to pull it, the
caller would still expect a returned checksum.
This appears to happen with static deltas for some reason; we might be
including duplicate metadata objects. Regardless, this is a bug that
should be fixed.
We have a chain of checksums from the root up until here. While doing
checksums of the objects individually would be a good redundancy check
for test cases and the like, when doing a pull there's no good reason
to burn cycles on SHA256.
This caused deadlocks and/or EMFILE due to the interaction between
threads and fds. What we really want here is a better pull-based
model for parsing content objects.
Another idea would be to change static deltas so that content objects
have a special opcode that includes their metadata first, and then do
rollsums etc. only over actual content.
See projectatomic/rpm-ostree#42 for rationale. There are two high
level use cases:
- If the OS comes unconfigured, this is a way to point it at a repo of your choice.
- To switch between repositories while keeping the same branch easily.
At some point, we might want to expose a uniform way to refer
to deployments by an index. At the moment undeploy is the only
command that does.
I plan to introduce another command which optionally takes an index,
so prepare a helper function for this.
This is a noticeable cleanup, and fixes another big user of GFile* in
performance/security sensitive codepaths.
I'm specifically making this change because the static deltas code was
leaking temporary files, and cleaning that up nicely would be best if
we were fd relative.
This will be used for a later change to use openat() for the fetching
code. Note that we drop the code to use mmap() - it was an attempt to
avoid keeping a fd open, but we do correctly close anyways.
Regression from 86764dbf00
This function is kind of fiendish now that we have 3 cases, each of
which want to be optimized somewhat to only load what's necessary
(e.g. don't open the file if we don't have an output for stream
requested).
Clean things up so that BARE_USER and BARE are separate conditionals
that share as much as possible, and fix the bug that asserted we
were in BARE mode.
I tested this by running test-basic-user.sh by hand.
Some use cases for checkouts don't need to fsync during checkout.
Installer programs for example will just do a global fsync at the end.
In the future, the default "ostree admin" core could also be
rearchitected to only do a transaction commit right before reboot, and
do the fsync then.
https://bugzilla.gnome.org/show_bug.cgi?id=742482
This is just an efficiency optimization. We're getting fairly close
to all of the hot code paths using `*at()`.
Note that we end up maintaining a half-duplicate code path set here,
because we still need to support commits from an arbitrary GFile *,
which in a possible common case is an OSTree commit.
I think it's worth it though.
rpm-ostree had code to check for this, which didn't actually work.
I don't see a no backwards compatibility concern in changing this, as
it's unlikely a caller would try to sensibly disambiguate FAILED.
Colin Walters