Instead of merging unfinished requests with child's IOVs we queued them
into parent's IOV list. Fix it.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
Commit 9cb20327aa ("return to epoll_wait after completing forks") was only
half way there. Adding the other half.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
It is possible that when pages request from the remove source arrive, part
of the memory range covered by the request would be already gone because of
madvise(MADV_DONTNEED), mremap() etc.
Ensure we are not trying to uffd_copy more than we are allowed.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
If we get fork() event just before transferring last IOV of the parent
process, continuing to background fetch after completing fork event
handling will cause lazy-pages daemon to exit and nothing will monitor the
child process memory.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
Since the memory mapping is now split between ->iovs and ->reqs lists, any
update to memory layout should take into account both lists.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
Instead of recalculating required for lazy_pages_info->buf when copying
IOVs at fork() time, keep the size of the buffer in the lazy_pages_info
struct.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
When we return from epoll_run_rfds with positive return value it means that
event handling loop was interrupted because the event should be handled
outside of that loop. Is always the case with UFFD_EVENT_FORK.
It may happen that the event occurred after we've completed the memory
transfer and we are on the way to successful return from the
handle_requests() function, but instead of returning 0 we will return the
positive value we've got from epoll_run_rfds.
Explicitly assigning return value of complete_forks() fixes this issue.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
With userfaultfd we cannot reliably service process_vm_readv calls. The
maps007 test that uses these calls passed previously by sheer luck.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
In the current model we haven't started the background page transfer until
POLL_TIMEOUT time has elapsed since the last uffd or socket event. If the
restored process will do memory access one in (POLL_TIMEOUT - eplsilon) the
filling of its memory can take ages.
This patch changes them model in the following way:
* poll for the events indefinitely until the restore is complete
* the restore completion event causes reset of the poll timeout to zero and
* starts the background transfers
* after each transfer we return to check if there are any uffd events to
handle
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
Currently, once we get to transfer pages in the "background", we try to
fetch the entire IOV at once. For large IOVs this may impact #PF latency
for the #PF events occurred during the transfer.
Let's add a simple heuristic for controlling size of the background
transfers. Initially, the transfer will be limited to some default value.
Every time we transfer a chunk we increase the transfer size until it
reaches a pre-defined maximal size. A page fault event resets the
background transfer size to its initial value.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
The complete_forks function presumes that it always has a work to do
because we assume that fork event is the only case when we drop out of
epoll_run_rfds with positive return value.
Teach complete_forks to bail out when there is no pending forks to process
to allow exiting epoll_run_rfds for different reasons.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
First check if there are pages we need to transfer and only afterwards
check if there are outstanding requests. Also, instead checking 'bool
remaining' to see if there is more work to do we can simply check if all
the lpi's have been already serviced.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
The intention is to use this function for transferring all the pages that
didn't cause a #PF.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
The function anyway pick the next page range to transfer it's just doing it
in very simple FIFO manner.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
We already have a queue for the requested memory ranges which contains
'lp_req' objects. These objects hold the same information as the lazy_iov:
start address of the range, end address and the address that the range had
at the dump time.
Rather than keep this information twice and use double bookkeeping, we can
extract the requested range from lpi->iovs and move it to lpi->reqs.
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
Instead of relying on length of various lists add a boolean variable to
lazy_pages_info to make it clean when the process has exited
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
Currently zdtm doesn't detect when restore failed, if it is executed
with strace. With this patch, fake-restore.sh creates a test file, and
zdtm is able to distinguish when restore failed.
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
The get() method requires a key and now we are using an index. That
will never work correctly as it is now.
Acked-by: Adrian Reber <adrian@lisas.de>
Reported-by: Adrian Reber <adrian@lisas.de>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
Currently we restore all sockets in the root mount namespace, because we
were not able to get any information about a mount point where a socket
is bound. It is obviously incorrect in some cases.
In 4.10 kernel, we added the SIOCUNIXFILE ioctl for unix sockets. This
ioctl opens a file to which a socket is bound and returns a file
descriptor.
This new ioctl allows us to get mnt_id by reading fdinfo, and mnt_id
is enough to find a proper mount point and a mount namespace.
The logic of this patch is straight forward. On dump, we save mnt_id for
sockets, on restore we find a mount namespace by mnt_id and restore this
socket in its mount namespace.
Acked-by: Cyrill Gorcunov <gorcunov@openvz.org>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
unix_process_name() are called when sockets are being collected,
but at this moment we don't have socket descriptors.
A socket descriptor is reuired to get mnt_id, what will allow to resolve
a socket path in its mount namespace.
Acked-by: Cyrill Gorcunov <gorcunov@openvz.org>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
This ioctl opens a file to which a socket is bound and
returns a file descriptor. This file descriptor can be used to get
mnt_id and a file path.
Acked-by: Cyrill Gorcunov <gorcunov@openvz.org>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
The USK_CALLBACK flag means that a socket is externel and will be
restored by a plugin. open_unixsk_standalone should not be called to
these sockets.
$ make -C test/others/unix-callback/ run
...
(00.109338) 7471: sk unix: Opening standalone socket (id 0xd ino 0 peer 0x63b)
(00.109376) 7471: Error (criu/sk-unix.c:1128): sk unix: BUG at criu/sk-unix.c:1128
Acked-by: Cyrill Gorcunov <gorcunov@openvz.org>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
Unix file sockets have to be restored in proper mount namespaces.
Acked-by: Cyrill Gorcunov <gorcunov@openvz.org>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
When walking over unix sockets make sure the
queuer is present before accessing it.
https://jira.sw.ru/browse/PSBM-82796
Reported-by: Vitaly Ostrosablin <vostrosablin@virtuozzo.com>
Signed-off-by: Cyrill Gorcunov <gorcunov@virtuozzo.com>
Reviewed-by: Kirill Tkhai <ktkhai@virtuozzo.com>
It is not used, probably was committed by mistake.
Fixes: 2d093a1702 ("travis: add a job to test on the fedora rawhide")
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
In Ubuntu Bionic for armhf, clang is compiled for armv8l rather than
armv7l (as it was and still is for gcc) and so it uses armv8 by default.
This breaks compilation of tests using smp_mb():
> error: instruction requires: data-barriers
The fix is to add "-march=armv7-a" to CFLAGS which we already do,
except not for the tests.
Signed-off-by: Kir Kolyshkin <kolyshkin@gmail.com>
The newly introduced output of the CRIU and kernel version does not
happen when running CRIU under RPC. This moves the print_versions()
function util.c and calls it from cr-service.c
Signed-off-by: Adrian Reber <areber@redhat.com>
This patch makes restore_one_inotify() to request specific
watch descriptor number instead of iterating in (possible)
long-duration loop if system supports it.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Currently, one feature is supported. Add possibility
for a test to depend on several features.
v2: Delete excess "if" as suggested by Andrey Vagin.
Rename variables to decrise patch size.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
We need this to save backward compatibility with old images.
It is zero only if it is absent in an image file.
Signed-off-by: Andrei Vagin <avagin@virtuozzo.com>
tun test in nested net ns wrapper.
Signed-off-by: Andrew Vagin <avagin@virtuozzo.com>
ktkhai: Makefile hunks
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
This adds new tunfile_entry::ns_id field and populates
it in standard socket way. Restore uses this ns_id
to choose correct namespace. Note, we could completelly
skip set_netns() on restore in case of !has_ns_id, but
using top_net_ns invents some definite behaviour.
Signed-off-by: Andrew Vagin <avagin@virtuozzo.com>
ktkhai: comment written/code movings
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
LAZY_PAGES_SK_OF is need only once for every process,
and it's not frequently used, so we can place it
to fdstore.
https://travis-ci.org/tkhai/criu/builds/343405755
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
