After freeing the head of the perfstat block list, we must set it to NULL
to indicate that new blocks must be allocated. Previously, we risked
segfaults by trying to use a freed list.
Signed-off-by: David Sheets <dsheets@docker.com>
This adds two new control channel requests, START_PERFSTAT and
STOP_PERFSTAT, that begin and end performance statistics collection
respectively. START_PERFSTAT takes the name of the mountpoint on which
to enable statistics collection and returns the current monotonic time.
STOP_PERFSTAT takes the name of the mountpoint on which to disable
statistics collection and returns the current monotonic time and the
message timings that have been collected.
Performance statistics are collected in a linked list of blocks. pthread
locks are used for synchronization.
Signed-off-by: David Sheets <dsheets@docker.com>
This will potentially leave stuff in lost+found but may be able to
recover from more filesystem corruption.
See #736
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
Move to the development track of `containerd` not the legacy 0.2
branch. The commands have changed a bit.
This does increase the image size as we are bundling the Docker
copy and our copy, and the new one is larger as it is growing features.
Hopefully Docker will shrink eventually. Also we may replace `ctr`
with a library.
Fix#1029
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
Add basic database setup for AWS, GCP although these are not yet used
by the setup code but will be useful later.
Currently each is gated by `mobyplatform` but this can be removed once we
construct Moby per platform, and once these are containerised so they
are not run from `openrc`.
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
Previously we allocated `sizeof(struct ring) + size`. This patch
allocates `sizeof(struct ring)` and then `size` for the payload
separately. Hopefully the payload will be better aligned.
Signed-off-by: David Scott <dave.scott@docker.com>
This patch adds a simple ring buffer implementation and uses it to buffer
the reads and writes to/from the AF_HYPERV socket and tap file descriptor.
This removes the need to perform small reads and writes for the per-packet
headers and allows a read on the Hyper-V socket to block at the same time
as a write to the tap device (and vice-versa)
The configuration in the init.d script is:
- a max message size (individual read or write) of 8192. Experimentally
this seems to be the largest completely reliable size across the Windows
versions we can support. Messages of length 16384 sometimes fail.
- a buffer size of 256KiB in each direction.
Single stream TCP throughput as measured by iperf increases modestly, by
another 100Mbit/sec.
Signed-off-by: David Scott <dave.scott@docker.com>
In #1057 removed the 9p paths from the database mount
but that broke some users accessing them directly. Fix these.
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
These never got an interface on any platform, and I don't think they
ever will, we can increase global limits or you can set something
with a privileged container. Can add back later if required.
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
Use a bind mount for OSX and Windows for 9p filesystem.
This makes it easier to use a different database source,
and to share database into system containers.
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
C version of #1006
Note that I switched all the C builds to use -O2 and that meant
that the compiler found some more warnings so I also fixed these
up. The possibly undefined ones were harmless, the aliasing one
is now more correct.
As these are small programs, the caching from `docker build` makes
no real difference, and worst case compile time is much better.
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
Now there is an Alpine 3.5 variant of the Go 1.7 images, use this.
fix#972
Note updated the containers/binfmt image as this will be converted
to go-compile shortly, at which point alpine-build-go can be removed.
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
As the build requires networking, is non repeatable as dependencies
may change, makes sense to make it a static package for now.
Plan is to rewrite in Go anyway at some point see #467
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
Go code is really fast to compile so we do not really need to use the
cache features of `docker build`. So make a compile container instead.
This can also output a build context and Dockerfile if you want to do
a build.
For reference, an uncached `docker build` of our Go code takes about
7s, a cached one 1.2s, and this takes 1.7s, so the best case is a little
worse, but we save a lot of images, and the worst case is better.
This is mainly designed to make the nested builds for containerd
containers simpler too. Will add a variant for the C code as well.
Also add `-static` to the flags so we always make static executables,
which was omitted previously.
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
