Running `ostree commit --tree=ref=a --tree=dir=b` involves reading the
whole of a into an `OstreeMutableTree` before composing `b` on top. This
is inefficient if a is a complete rootfs and b is just touching one file.
We process O(size of a + size of b) directories rather than
O(number of touched dirs).
This commit makes `ostree commit` more efficient at composing multiple
directories together. With `ostree_mutable_tree_fill_empty_from_dirtree`
we create a lazy `OstreeMutableTree` which only reads the underlying
information from disk when needed. We don't need to read all the
subdirectories just to get the checksum of a tree already checked into the
repo.
This provides great speedups when composing a rootfs out of multiple other
rootfs as we do in our build system. We compose multiple containers
together with:
ostree commit --tree=ref=base-rootfs --tree=ref=container1 --tree=ref=container2
and it is much faster now.
As a test I ran
time ostree --repo=... commit --orphan --tree=ref=big-rootfs --tree=dir=modified_etc
Where modified_etc contained a modified sudoers file under /etc. I used
`strace` to count syscalls and I seperatly took timing measurements. To
test with a cold cache I ran
sync && echo 3 | sudo tee /proc/sys/vm/drop_caches
Results:
| | Before | After |
| -------------------- | ------ | ----- |
| Time (cold cache) | 8.1s | 0.12s |
| Time (warm cache) | 3.7s | 0.08s |
| `openat` calls | 53589 | 246 |
| `fgetxattr` calls | 78916 | 0 |
I'm not sure if this will have some negative interaction with the
`_ostree_repo_commit_modifier_apply` which is short-circuited here. I
think it was disabled for `--tree=ref=` anyway, but I'm not certain. All
the tests pass anyway.
I originally implemented this in terms of the `OstreeRepoFile` APIs, but
it was *way* less efficient, opening and reading many files unnecessarily.
Closes: #1643
Approved by: cgwalters
This implements a TODO item from
`ostree_mutable_tree_get_contents_checksum`. We now no-longer invalidate
the dirtree contents checksum at `get_contents_checksum` time - we
invalidate it when the mtree is modified. This is implemented by keeping
a pointer to the parent directory in each `OstreeMutableTree`. This gives
us stronger invariants on `contents_checksum`.
For even stronger guarantees about invariants we could make
`ostree_repo_write_mtree` or similar a member of `OstreeMutableTree` and
remove `ostree_mutable_tree_set_metadata_checksum`.
I think I've fixed a bug here too. We now invalidate parent's contents
checksum when our metadata checksum changes, whereas we didn't before.
Closes: #1655
Approved by: cgwalters
A prelude to my understanding. Unfortunately `OstreeMutableTree` provides
little encapsulation, as each member has setters† so it's difficult to come
up with a list of invariants.
† `files` and `subdirs` only have getters, but the getters return mutable
references to the internals, so we still can't reason about invariants.
Closes: #1645
Approved by: jlebon
Much like the (optional) initramfs at
`/usr/lib/ostree-boot/initramfs-<SHA256>` or
`/usr/lib/modules/$kver/initramfs` you can now optionally include a
flattened devicetree (.dtb) file alongside the kernel at
`/usr/lib/ostree-boot/devicetree-<SHA256>` or
`/usr/lib/modules/$kver/devicetree`.
This is useful for embedded ARM systems which need the devicetree file
loaded by the bootloader for the kernel to discover and initialise
hardware. See https://en.wikipedia.org/wiki/Device_tree for more
information.
This patch was mostly produced by copy-pasting code for initramfs handling
and renaming `s/initramfs/devicetree/g`. It's not beautiful, but it is
fairly straightforward.
It may be useful to extend device-tree support in a number ways in the
future. Device trees dependant on many details of the hardware they
support. This makes them unlike kernels, which may support many different
hardware variants as long as the instruction-set matches. This means that
a ostree tree created with a device-tree in this manner will only boot on
a single model of hardware. This is sufficient for my purposes, but may
not be for others'.
I've tested this on my NVidia Tegra TK1 device which has u-boot running
in syslinux-compatible mode.
Closes: #1411
Approved by: cgwalters
The bootloader spec says:
> `devicetree` refers to the binary device tree to use when executing the
> kernel. This also shall be a path relative to the `$BOOT` directory. This
> key is optional. Example:
> `6a9857a393724b7a981ebb5b8495b9ea/3.8.0-2.fc19.armv7hl/tegra20-paz00.dtb`
This is necessary for booting my NVidia Tegra TK1 device. It uses u-boot
with syslinux compatibility. In the syslinux files that come with the
device this is called `FDT`, but u-boot treats `FDT and `DEVICETREE` as
synonyms.
See also: [f43c401 in u-boot].
[f43c401 in u-boot]: http://git.denx.de/?p=u-boot.git;a=commit;h=f43c401b72bb0db43ab0b55c4a79e1f4889d3aa2Closes: #1411
Approved by: cgwalters
If you want cleanup, but don't want to prune the repo. Pruning can
be quite expensive so ostree admin deploy can be much faster without
pruning.
Closes: #1418
Approved by: cgwalters
In the next commit I will add --no-prune which will affect cleaning. By
doing this refactor we avoid having to add a NO_PRUNE flag.
Closes: #1418
Approved by: cgwalters
You'd expect
ostree commit --tree=ref=A --tree=ref=B
to produce a commit with the union of the trees given. Instead you'd get
a commit with the contents of just the latter commit. This was due to an
optimisation where we'd skip filling out the `files` and `subdirs`
members of the mtree, just filling in the metadata instead. This backfires
becuase this same code relies on checking the `files` and `subdirs` members
itself to work out whether the mtree is empty.
This commit removes the optimisation, fixing the bug. Maybe there's a way
to keep the optimisation and still fix the bug but it's not obvious to
me.
Closes: #581
Approved by: cgwalters
musl libc's implementation of `realpath` works by opening the path and then
doing a lookup in `/proc/self/fd` to find the canonical path. This fails
if `/proc` is not mounted. This causes problems for us if
`ostree-prepare-root` is `init` as `/proc` won't be mounted.
We have to mount `/proc` anyway for `/proc/cmdline` so this fix just
expands the scope over which `/proc` is mounted to include both our
`realpath` calls.
See also:
* http://www.openwall.com/lists/musl/2016/06/08/2 and
* http://git.musl-libc.org/cgit/musl/tree/src/misc/realpath.c?id=e738b8cbe64b6dd3ed9f47b6d4cd7eb2c422b38dCloses: #485
Approved by: cgwalters
I've seen it fail with musl which needs `/proc` to be mounted for it to
work. The error messages we're rather confusing before. At least this
now points to the right location.
Closes: #485
Approved by: cgwalters
This test previously depended on manually building ostree-prepare-root.
Since 42dab85 we've been able to build static binaries with the usual
autotools build-system. This change reflects the fact that
`ostree-prepare-root` is built into $srcdir rather than `src/switchroot`
where I was building manually.
This test now passes with `./configure --with-static-compiler=gcc` (glibc)
but still fails with `./configure --with-static-compiler=musl-gcc` (musl).
Closes: #485
Approved by: cgwalters
Was failing with error:
src/switchroot/ostree-prepare-root.c:30:20: fatal error: config.h: No such file or directory
compilation terminated.
Reported by and fix provided by @gatispaeglis.
Closes: #485
Approved by: cgwalters
This simplifies the build system by removing the need for
`libswitchroot-mountutil.la`.
Original idea by @cgwalters in #477.
Closes: #478
Approved by: cgwalters
The `warn()` libc extension has exactly the same behaviour as our own
`perrorv` function, but is available in (at least) glibc and musl. As an
added bonus the similar function `err()` which will exit with an error
code afterwards.
This implementation is tidier and allows us to get rid of our own
`perrorv`. It paves the way to removing `ostree-mount-util.c` to simplify
the build scripts.
Original idea by @cgwalters in #477.
Closes: #478
Approved by: cgwalters
If the `--with-static-compiler=musl-gcc` configure flag is given.
ostree-prepare-root can be used as init in a system without a populated
/lib. To support this use case we need to link statically as we will be
unable to locate libc.so at run time if it's not installed in /lib.
We support building ostree-prepare-root with a different compiler to the
rest of ostree so we can use musl rather than glibc. This reduces the size
of the executable significantly: from ~700K -> ~30K. We have to use
`_SCRIPTS` here to get autotools to install this as an executable but
without generating rules to make it itself which we have specified
manually.
See https://lists.gnu.org/archive/html/help-gnu-utils/2007-01/msg00007.html
for advice on using autotools in this manner.
Closes: #477
Approved by: cgwalters
POSIX and GNU define conflicting versions of `strerror_r`. The GNU
version returns the string but doesn't necessilary write into buf.
The POSIX version writes into buf and returns the length but doesn't
necessilary append a terminate the string with a NUL if it's too long
to fit in buf.
This commit fixes building ostree-prepare-root with musl libc. The
stripped static build with musl on my machine is 30K vs. 724K with glibc
static and 11K with glibc shared.
Closes: #477
Approved by: cgwalters
This allows ostree-prepare-root outside of the initramfs context where the
real rootfs is already mounted at /. We can't use `mount --move` in this
case because we would be trying to move / into a subdirectory of itself.
Closes: #403
Approved by: cgwalters
...for simplicity. This way we don't need to keep concatenating
deploy_path to everything. We can just refer relative to the current
working directory.
We need to do this after bind-mounting it over itself otherwise our cwd
is still on the non-bind-mounted filesystem below.
Closes: #403
Approved by: cgwalters
Typically we have our ready made-up up root at
`/sysroot/ostree/deploy/.../` (`deploy_path`) and the real rootfs at
`/sysroot` (`root_mountpoint`). We want to end up with our made-up root
at `/sysroot/` and the real rootfs under `/sysroot/sysroot` as systemd
will be responsible for moving `/sysroot` to `/`.
We need to do this in 3 moves to avoid trying to move `/sysroot` under
itself:
1. `/sysroot/ostree/deploy/...` -> `/sysroot.tmp`
2. `/sysroot` -> `/sysroot.tmp/sysroot`
3. `/sysroot.tmp` -> `/sysroot`
This is a refactoring to group all these operations together so I can
implement an alternative in terms of `pivot_root`.
Closes: #403
Approved by: cgwalters
This supports running ostree on embedded platforms without an initrd.
Specificially I'm trying to do bringup on an NVidia Tegra based Jetson TK1
dev board.
Closes: #403
Approved by: cgwalters
These tests use unshare and mount to prepare a fake initrd/early boot
directory structure so we can then test ostree-prepare-root.
Things that are tested:
* Running in an initrd environment
* Running without initrd
* /var and /sysroot being mounted correctly
* /usr being mounted read-only
Things not tested (yet):
* Running as init - this could be accomplished by unsharing the pid
namespace too.
* mounting/unmounting `/proc` if `/proc/cmdline` isn't available
* Persistent overlayfs for `/usr`
* Probably other things
The tests are basic but can be extended in the future as we do more work
on `ostree-prepare-root`.
These tests must be run as root as they require the ability to `mount`
and to `unshare` the mount namespace. Perhaps in the future we can use
user namespaces for this test once they are more widely available.
Closes: #403
Approved by: cgwalters
William Manley