Each operation we submit results in a contiguous allocation in the
kernel which can trigger OOM-killing. Tune the queue depths to reduce
the chances of this happening, with a minimal impact on performance.
Bug: http://b/126582877
Test: ./benchmark_device.py with blueline
Change-Id: I4b0f83d54a9a9a69f0ab48322d78fb137c94dd35
Some USB controllers only support 16k writes, so drop down to that.
Bug: http://b/130622010
Test: treehugger
Change-Id: I836d5b2d0d6fcae05f290af11a6a19c6e8d7c1ba
As raw protocol does not allow for splitting err - it has to be redirected to inout.
Before this change it was not done for in-process and all err data was lost.
Bug: 130086616
Test: manual + atest adbd_test
Change-Id: I6cd11c940673d73e2993a6eb23c46d31bd8bf504
It seems like we don't actually always get these events in order, so
demote them to a log and restart the connection instead of aborting.
Bug: http://b/129464137
Test: mma
Change-Id: I7dffbf62f9dea665cf9f9f9e1bd18d444b6905ec
Occasionally, after we open the functionfs endpoint and write our
descriptors to it, we end up in a state where we never receive any
events from ep0, presumably due to a race against some other part of
the system, since adding sleeps reduces the likelihood. Add an
(arbitrarily chosen) timeout for the initial FUNCTIONFS_BIND we read,
and retry on failure.
Bug: http://b/128679787
Test: unplug and replug a blueline 14,000 times with a usb switch board
Change-Id: I4b3d5d552179ee7f3fefcc15f9575f793926f784
Monitor for FUNCTIONFS_UNBIND as well, so that in the case where we get
FUNCTIONFS_BIND, FUNCTIONFS_UNBIND, FUNCTIONFS_BIND, we don't trigger an
assertion failure from seeing two FUNCTIONFS_BINDs.
Bug: http://b/129134256
Test: manual
Change-Id: I80af5f4b833513e932262638b9f8d76bbcb35504
If we fail to submit writes for some reason (e.g. the USB cable was
unplugged), another thread that's waiting on the write mutex can enter
SubmitWrites and attempt to resubmit the writes that we already failed
to submit, leading to a failed assertion of !IoBlock::pending.
Increment writes_submitted_ before actually calling io_submit, so we
skip over these writes and fall through to exit.
Bug: http://b/129134256
Test: manually unplugged a blueline
Change-Id: I2fd0034e45db22c8f637c81039ce686b7aa6a03b
When a function argument takes a std::string, it's inefficient to
pass std::string::c_str(), since that creates an additional copy.
So we change these calling sites.
Test: TreeHugger
Change-Id: I1c0fde7275eb3ebd3baf05ec62581b0243655608
Allow disabling ro.adb.nonblocking_ffs via an ro system property, which
gets overridden by the currently supported persist.adb.nonblocking_ffs.
Bug: http://b/128679787
Test: mma
Change-Id: I28891d2da129e9654bf63ab8a4c330c4d2f5b336
Previously, we were initializing IoBlock::pending when submitting a read,
but leaving it uninitialized when doing a write, which would sometimes
result in an assertion firing when we received a previously allocated
block of memory that happened to have a nonzero value in pending.
Test: treehugger
Change-Id: I434c65f554ceed313fea9cfcc98788908f3ca8bc
As there is no security guarantee when the device is unlocked,
allowing adb root gives us more rooms to debug a USER build images.
Also, this makes it possible to run VTS on a USER build GSI, with
setting ro.debuggable=1 and unlocking the device.
This basically re-lands a reverted change:
https://android-review.googlesource.com/c/platform/system/core/+/437815
Which isn't needed after we moved /sbin/adbd to /system/bin/adbd in
USERDEBUG GSI. But it's still needed for USER build GSI.
Bug: 126493225
Test: unlock a USER build device, check 'adb root' can work
Change-Id: I93f12c8a3fe65c96c947e4602795eadfe591c521
If a device is unplugged, a BIND will happen without an ENABLE until the
device is plugged in, which would previously result in adbd crashing
every 5 seconds due to an attempt to join a thread that hadn't been
started.
Bug: http://b/128255355
Test: manually observed adbd via serial cable while unplugged
Change-Id: I90d6384e7e829df1ffaf73a95f2418a2ff97fb03
1. Don't clear O_CLOEXEC on an fd we're duping to stdin/out/err: dup2
does that for us if oldfd != newfd, and we know the fd can't be
stdin/out/err.
2. Use unique_fd.
3. Remove gotos by breaking the remount logic into its own function.
Test: mma
Change-Id: If9929e7f20de445322fd498e1379a2d4e572cb52
Replace direct logic in support for 'adb remount' with an exec out to
/system/bin/remount to do the heavy lifting. Remount success and
failure strings are reported by the adb remount service in response
to the various reported errors, freeing up the remount command itself
from the legacy of script expectations.
Test: adb-remount-test.sh
Bug: 122602260
Change-Id: I686fa465f463b881bbb38f709d780a95e463be80
Reduce the operation sizes to sizes that are small enough that the
kernel will run the OOM killer to try to make them succeed. Compensate
for the smaller operation sizes by increasing the queue depth.
Bug: http://b/126582877
Test: benchmark_device.py
Change-Id: I4d351f56c8f83fea882614238cc3fec8ba1e9bf9
The kernel does a kmalloc of the write size for each write during
io_submit, which can lead to failure if the kernel's heap is fragmented
such that it can't allocate a contiguous 1MB chunk. Partition writes
into 16 page chunks which are much more likely to succeed.
Bug: http://b/126582877
Test: test_device.py
Change-Id: Ib2d48561594e1d81f1a2f62e34eaa40265ac47c4
* changes:
adb: switch sockets.cpp to ConsumePrefix.
adbd: switch abb to ConsumePrefix.
adb: increment server version.
adb: wait for device to disconnect upon `adb root`.
adb: implement wait-for-disconnect.
adb: tell the client what transport it received.
adbd: switch daemon/services to ConsumePrefix.
adb: switch host_service_to_socket to string_view.
adb: switch handle_host_request to string_view.
adb: switch adb_io.h to string_view.
adb: add helper to consume a prefix on a string_view.
adb: make ParseUint reject garbage at the end by default.
This is corresponding to exec: command of adb and allows for simplifed calls
to binder without shell protocol.
Bug: 111621042
Test: manual
Change-Id: Id6935cd53e351388ecf6d2d15f3a204cb871536a
Switch to the nonblocking FFS implementation that fixes several race
conditions on connect/disconnect.
Test: treehugger
Change-Id: I91fdb9b27152078cb38e610861a05f1fa95aa3a2
Detect kernels that don't support AIO on functionfs endpoints and fall
back to the old legacy implementation.
Test: treehugger
Change-Id: Ice28dcd40c9730ac2dd5e468d065b41c0e076146
The switch to socket_spec_listen broke adbd over TCP, because
socket_spec_listen only listens on localhost.
Bug: http://b/123592649
Test: manual
Change-Id: Id1943ebd7f0059db05ad756fe96189c60ebde337
Background:
We now have two sets of Bionic: the bootstrap Bionic which is at
/system/{lib|bin}/bootstrap for early processes and the default Bionic
which is from the runtime APEX for all the others. In order to give the
same path for Bionic to both categories of processes, the init prepares
two mount namespaces and bind-mount appropriate Bionic files onto the
common mount points under /bionic. For example,
/system/bin/bootstrap/linker is bind-mounted to /bionic/bin/linker for
the early processes. Likewise, /apex/com.android.runtime/bin/linker is
bind-mounted to the same path for rest of the processes.
In addition, in order not to propagate mount events in one mount
namespace to the other namespace, /bionic itself is created as a mount
namespace (via self bind-mount) and its propagation type is set to
private.
Changes required:
This however requires some adjustments to adb sync and remount
mechanism.
For remounting, /bionic path should also be re-mounted for RW, because
it is a RO mount in the beginning. This remounting is done only for the
system-as-root devices where entire / can be re-mounted as RW.
For synching, the sync thread creates a temporary mount namespace where
there is no bind-mount. This ensures that a path that the thread handles
is pointing to the correct file that is expected from the client side.
In addition, push operation to /bionic path is done without unlinking.
This is required because the mount points under /bionic are gone in the
current mount namespace but are still active in other mount namespaces.
If unlinked, the existing mounts on the path are all silently removed.
In order to prevent the unwanted situation, the moint points are not
unlinked but truncated to 0. This however is not a significant problem
because the files that serve as mount points do not carry any
useful information (i.e. the content is meaningless).
Bug: 879416
Test: adb sync
adb push <random_file> /bionic/bin/linker64
adb push <random_file> /system/bin/bootstrap/bin/linker64
system/core/fs_mgr/tests/adb-remount-test.sh
Change-Id: Id87dc9ee7ec5c43d06b54969b55e2cb394329317
And deprecate one more old style function that is not used after this
change.
Test: boot, disable and enable verity
Change-Id: Id509f479850120352b4ea4dc3b6c40f6e8e2e53e
vsock is a socket address family for communicating into and out of
virtual machines. Addresses have a port and CID. The CID is unique to
each virtual machine on the computer. The VM host always has CID 2.
http://man7.org/linux/man-pages/man7/vsock.7.html
Inside the android guest, the adb daemon hosts a vsock server with
VMADDR_CID_ANY, automatically using the guest CID. The adb server
can now connect to addresses of the form vsock:cid:port, where the CID
must be specified and the port defaults to 5555.
This is a significant speed improvement for ADB connections in
Cuttlefish, with 150-200 MB/s for `adb push` and 100-150 MB/s for
`adb pull`. It also allows removing some proxying steps from Cuttlefish,
simplifying the full connection path, and removes a dependency on the
unstable ivshmem protocol.
Commands tested against a Cuttlefish VM with CID 3:
adb connect vsock:3:5555
adb -s vsock:3:5555 shell
adb disconnect vsock:3:5555
Supporting "adb disconnect" and "adb -s" required modifying some of the
parts that parse addresses / serials.
push/pull trials with native adb vsock support in cuttlefish:
100m: 1 file pushed. 297.6 MB/s (104857600 bytes in 0.336s)
100m: 1 file pushed. 270.3 MB/s (104857600 bytes in 0.370s)
100m: 1 file pushed. 271.7 MB/s (104857600 bytes in 0.368s)
100m: 1 file pushed. 250.5 MB/s (104857600 bytes in 0.399s)
100m: 1 file pushed. 277.1 MB/s (104857600 bytes in 0.361s)
100m: 1 file pushed. 263.5 MB/s (104857600 bytes in 0.379s)
100m: 1 file pushed. 242.6 MB/s (104857600 bytes in 0.412s)
100m: 1 file pushed. 271.8 MB/s (104857600 bytes in 0.368s)
100m: 1 file pushed. 267.1 MB/s (104857600 bytes in 0.374s)
/data/local/tmp/100m: 1 file pulled. 212.8 MB/s (104857600 bytes in 0.470s)
/data/local/tmp/100m: 1 file pulled. 236.7 MB/s (104857600 bytes in 0.423s)
/data/local/tmp/100m: 1 file pulled. 201.2 MB/s (104857600 bytes in 0.497s)
/data/local/tmp/100m: 1 file pulled. 255.6 MB/s (104857600 bytes in 0.391s)
/data/local/tmp/100m: 1 file pulled. 199.6 MB/s (104857600 bytes in 0.501s)
/data/local/tmp/100m: 1 file pulled. 214.6 MB/s (104857600 bytes in 0.466s)
/data/local/tmp/100m: 1 file pulled. 254.2 MB/s (104857600 bytes in 0.393s)
/data/local/tmp/100m: 1 file pulled. 212.5 MB/s (104857600 bytes in 0.471s)
/data/local/tmp/100m: 1 file pulled. 218.9 MB/s (104857600 bytes in 0.457s)
/data/local/tmp/100m: 1 file pulled. 223.6 MB/s (104857600 bytes in 0.447s)
Bug: 121166534
Change-Id: I50f21fb5c9acafb8daa789df4e28c9e1bbbbf2ef
Test: adb connect/shell/disconnect
Preparatory step for testing adb on GCE on non-linux hosts: instead of
pointing them at a device (emulated or otherwise), point them at adbd
running on a linux host instead.
Test: adbd & adb connect localhost:5555; adb -e wait-for-device shell
Change-Id: Ib22d51a4fc9e6e68f71bf1b3b9b2e1b0bd844760
