New get_sdev_from_fd helper first gets mnt_id from fd using fdinfo and
then converts mnt_id to sdev using mountinfo.
By default mnt_id to sdev conversion only works for mounts in mntinfo.
If parse_mountinfo argument is true, will also parse current process
mountinfo when looking for mount sdev, this should be used only with
temporary mounts just created by criu in current mntns.
v3: add argument to parse self mountinfo for auxiliary mounts
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
Only place where we used __open_mountpoint with non -1 mnt_fd is
open_mountpoint. Let's use check_mountpoint_fd for this case, so that we
now can remove mnt_id argument. Also now __open_mountpoint actually
always does open.
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
GNUTLS_SHUT_RDWR sends an alert containing a close request and waits for
the peer to reply with the same message.
Signed-off-by: Andrei Vagin <avagin@gmail.com>
We need to be sure that page-server doesn't wait for a new command when we
call gnutls_bye() that sends an alert containing a close request.
Signed-off-by: Andrei Vagin <avagin@gmail.com>
This commit simply makes copies of SOCK_STREAM unix socket tests and uses
SOCK_SEQPACKET instead.
Signed-off-by: Bui Quang Minh <minhquangbui99@gmail.com>
We have multiple options which are valid only on restore or only on dump
or in any other specific criu mode, so it would be useful to have info
about current mode in opts so that we can validate other options against
current mode.
Plan is to use it for mount-v2 option as it is only valid on restore,
and this would make handling of different types mountpoints much easier.
Realization is a bit different for general code and rpc:
- When criu mode is set from main() we just parse mode from argv[optind]
just after parse_options() found optind of the command. Note that
opts.mode is available before check_options().
- For rpc service we reset opts.mode to CR_SWRK each time we restart
cr_service_work(), in the original service process we still have
CR_SERVICE to differentiate between them, and each request handling
function which does setup_opts_from_req sets opts.mode in accordance
with the processed request type. And it is also available before
check_options().
Now in check_options we can add filters on one mode only options.
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
Several lines above if (optind >= argc) we go to usage label and fail,
thus we don't need to check (optind < argc) here as it is always true.
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
Looks like in commit [1] we've non-intentionally added this tmp file to
git, let's remove it.
Fixes: 01ee29702 ("s390:zdtm: Enable zdtm for s390") [1]
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
zdtm_ct.c:44:12: error: function declaration isn’t a prototype [-Werror=strict-prototypes]
44 | static int create_timens()
Signed-off-by: Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com>
In most cases we run tests as:
./test/zdtm.py run -a
But it's also possible to run tests from root makefile:
make test
In this case, if criu tree have no ./test/umount2 binary
built we get the error like:
make[3]: *** No rule to make target 'umount2'. Stop.
It's worth to mention this "3". That's because we have
build process tree like this:
make -> make -> make -> zdtm.py -> make umount2
and also we have MAKEFLAGS variable set to:
build=-r -R -f ...
And that's bad because "-r" option means no builtin
rules and -R means no builtin variables. That makes
`make umount2` not working. Let's just cleanup this
variable to make things work properly.
Fixes: #1699https://github.com/checkpoint-restore/criu/issues/1699
Signed-off-by: Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com>
In contrast to the CLI it is not possible to do a single pre-dump via
RPC and thus libcriu. In cr-service.c pre-dump always goes into a
pre-dump loop followed by a final dump. runc already works around this
to only do a single pre-dump by killing the CRIU process waiting for the
message for the final dump.
Trying to implement pre-dump in crun via libcriu it is not as easy to
work around CRIU's pre-dump loop expectations as with runc that directly
talks to CRIU via RPC.
We know that LXC/LXD also does single pre-dumps using the CLI and runc
also only does single pre-dumps by misusing the pre-dump loop interface.
With this commit it is possible to trigger a single pre-dump via RPC and
libcriu without misusing the interface provided via cr-service.c. So
this commit basically updates CRIU to the existing use cases.
The existing pre-dump loop still sounds like a very good idea, but so
far most tools have decided to implement the pre-dump loop themselves.
With this change we can implement pre-dump in crun to match what is
currently implemented in runc.
Signed-off-by: Adrian Reber <areber@redhat.com>
We added cross-compile tests with testing debian release to be able to
replicate the error reported in #1653, however, installing build
dependencies in this release currently fails with the following error:
libc6-dev:armhf : Breaks: libc6-dev-armhf-cross (< 2.33~) but 2.32-1cross4 is to be installed
This is not something we can fix, therefore using the debian unstable
release (instead of testing) could be more reliable option for our CI.
This would still replicate the problem reported in #1653.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
The --timeout option was introduced in [1] to prevent criu dump from
being able to hang indefinitely and allow users to adjust the time limit
in seconds for collecting tasks during the dump operation.
[1] d0ff730
Signed-off-by: Radostin Stoyanov <radostin@redhat.com>
Fixes: e2e8be37 ("x86/compel/fault-inject: Add a fault-injection for corrupting extended regset")
Signed-off-by: Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com>
Since
e2e8be37 ("x86/compel/fault-inject: Add a fault-injection for corrupting extended regset")
we doing fault-injection test for C/R of threads register set by filling tasks
xsave structures with the garbage. But there are some features for which that's not
safe. It leads to failures like described in #1635
In this particular case we meet the problem with PKRU feature, the problem
that after corrupting pkru registers we may restrict access to some vma areas,
so, after that process with the parasite injected get's segfault and crashes.
Let's manually specify which features is save to fill with the garbage by
keeping proper XFEATURE_MASK_FAULTINJ mask value.
Fixes: e2e8be37 ("x86/compel/fault-inject: Add a fault-injection for corrupting extended regset")
https://github.com/checkpoint-restore/criu/issues/1635
Signed-off-by: Alexander Mikhalitsyn <alexander.mikhalitsyn@virtuozzo.com>
We try to disable time namespace based testing for kernels older than
5.11. But we fail to come up with the correct if condition.
This changes (major <= 5) to (major < 5). There are no kernels with
major > 5 so currently the time namespace based are never run. This
should finally change it to run time namespace based tests on kernel
versions newer than 5.10.
Signed-off-by: Adrian Reber <areber@redhat.com>
The version of ps in Alpine image by default is very limited.
It is based on the one from busybox and doesn't support options
such as '-p'.
Signed-off-by: Radostin Stoyanov <radostin@redhat.com>
Since commit 83301b5367a98 ("af_unix: Set TCP_ESTABLISHED for datagram sockets
too") in Linux kernel, SOCK_DGRAM unix sockets can have TCP_ESTABLISHED state
when connected. So we need to fix checks that assume SOCK_DRAM sockets cannot
have TCP_ESTABLISHED state.
Signed-off-by: Bui Quang Minh <minhquangbui99@gmail.com>
The function run_tcp_server() was the last place CRIU was still using
the IPv4 only function inet_ntoa(). It was only used during a print, so
that it did not really break anything, but with this commit the output
is now no longer:
Accepted connection from 0.0.0.0:58396
but correctly displaying the IPv6 address
Accepted connection from ::1:58398
if connecting via IPv6.
Signed-off-by: Adrian Reber <areber@redhat.com>
An issue with dumping deleted reg files in overlayfs:
After deleting a file originated from lower layer in merged dir,
fstat() on the /proc/$pid/map_files symlink returns st_nlink=1, while
linkat() fails with errno ENOENT.
Signed-off-by: langyenan <ianlang@tencent.com>
Looking at CI logs there are often messages like:
"[WARNING] Option --keep-going is more useful when running multiple tests"
This commit removes '--keep-going' from single zdtm test runs.
Signed-off-by: Adrian Reber <areber@redhat.com>
Starting with gcc-11, Debian's armhf compiler no longer builds with
a default -mfpu= option. Instead it enables the FPU via an extension
to the -march flag (--with-arch=armv7-a+fp). criu's Makefile explicitly
passes its own -march=armv7-a setting, which overrides the +fp default,
so we end up with no FPU:
cc1: error: '-mfloat-abi=hard': selected architecture lacks an FPU
Signed-off-by: Radostin Stoyanov <radostin@redhat.com>
Debian testing has newer compiler version and running
cross compilation tests would allow us to catch any compilation
errors early.
Signed-off-by: Radostin Stoyanov <radostin@redhat.com>
The current debian stable release is Bullseye, not Buster. However, we
can use the 'stable' release instead. This would allow the CI to
automatically pick up updates in the future.
Signed-off-by: Radostin Stoyanov <radostin@redhat.com>
When we declare struct and at the same time declare variable pointer of
this struct type, it looks like clang-format threats "*" as a
multiplication operator instead of indirection (pointer declaration)
operator and puts spaces on both sides, which looks wrong.
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
This is a test for "ghost/mount: allocate remounted_rw in shmem to get
info from other processes" patch, without the patch test fails with:
############# Test zdtm/static/mntns_ghost01 FAIL at result check ##############
Test output: ================================
16:15:19.607: 5: ERR: mntns_ghost01.c:95: open for write on rofs -> 7 (errno = 11 (Resource temporarily unavailable))
16:15:19.607: 4: FAIL: mntns_ghost01.c:121: Test died (errno = 11 (Resource temporarily unavailable))
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
Previousely I din't mention this case because we had bad error handling
in ghost cleanup path.
Without these patch but with proper error handling for unlink we have an
error in mntns_ghost01 test:
Error (criu/files-reg.c:2269): Failed to unlink the remap file:
Read-only file system
Cherry-picked from Virtuozzo criu:
https://src.openvz.org/projects/OVZ/repos/criu/commits/151c859e1
Changes: check lookup_mnt_id return for NULL
Fixes: fd0a3cd9ef ("mount: remount ro mounts writable before
ghost-file restore")
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
Previousely remounted_rw was not shared between all processes on
restore, thus cleanup didn't got this info from rfi_remap and these
mounts were wrongly left writable after restore.
Cherry-picked from Virtuozzo criu:
https://src.openvz.org/projects/OVZ/repos/criu/commits/3a1a592e7
Fixes: fd0a3cd9ef ("mount: remount ro mounts writable before
ghost-file restore")
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
If unlinkat fails it means that fs is in "corrupted" state - spoiled
with non-unlinked auxiliary directories.
While on it add fixme note as this function can be racy and BUG_ON if
path contains double slashes.
Cherry-picked from Virtuozzo criu:
https://src.openvz.org/projects/OVZ/repos/criu/commits/b7b4e69fd
Changes: simplify while loop condition, remove confusing FIXME, remove
excess !count check in favour of while loop condition check
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
1) On error paths need to close fd and unlock mutex.
2) Make rfi_remap return special return code to identify EEXIST from
linkat_hard, all other errors should be reported up.
3) Report unlinkat error as criu should not corrupt fs.
Cherry-picked from Virtuozzo criu:
https://src.openvz.org/projects/OVZ/repos/criu/commits/fe1d0be14
Changes: use close_safe(), fix order in "Fake %s -> %s link" error
message.
Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>
Always wait() for forked child processes. It avoid zombie processes in
containers that don't have an init process reaping orphans.
Signed-off-by: Nicolas Viennot <Nicolas.Viennot@twosigma.com>
