SIGHUP handling depends on the local socket fd being fully closed in
order to trigger an event on its counterpart. Release the local socket
fd inside of Subprocess when returning it to ensure this.
Bug: http://b/29565233
Change-Id: I13b17bcddf0b396a5f4880f9e410fbbf24d9370d
This commit fixes two somewhat related issues in shell_service.
- The fd returned by StartSubprocess is owned by a unique_fd
contained in the Subprocess object, but also gets closed by the
caller. Resolve this by duping the returned file descriptor.
- A Subprocess object can be destroyed immediately after its initial
construction in StartSubprocess if we're sufficiently unlucky.
Split up the fork/exec and "start management thread" steps, so that
we can safely do everything we need to do on the Subprocess before
handing it over to the thread that'll eventually destroy it.
Also includes squashed patches from AOSP master that allow for use of
unique_fd inside adb.
Bug: http://b/29254462
Change-Id: Id9cf0b7e7a7293bee7176919edc758597691c636
(cherry picked from commit c0e6e40cc9)
(cherry picked from commit 54c72aaccc)
(cherry picked from commit 2c5d1d7cd9)
(cherry picked from commit 2a7b86337f)
(cherry picked from commit 13ea01db45)
(cherry picked from commit 344778da41)
Before this, adb will fail to start for the second user who tries because
/tmp/adb.log already exists and isn't writable by the second user.
Also allow $TMPDIR to override the use of /tmp.
Bug: http://b/29201686
Bug: https://code.google.com/p/android/issues/detail?id=211420
(cherry picked from commit d89a6c2285)
Change-Id: Ibb1369e8043feb9774593e317dcd7bb93323fd89
When non-interactive sessions exit via Ctrl+C, adbd sends SIGHUP
to the child process to let it know to exit. However, adbd was not
closing the pipes to the child process, so if the subprocess ignored
SIGHUP and continued writing it could fill up the pipe and block
forever while adbd waits for it to exit.
This CL adds the necessary calls to close the subprocess pipe after
sending SIGHUP.
Bug: 28981563
Change-Id: I318e322e563241052648361172f4859c297837fb
(cherry picked from commit f2aa186c7b)
Emulator console now requires authentication; this means
'adb emu ...' commands silently fail because of it.
This CL adds an 'auth <token>' command to each user command,
making sure it won't be silently ignored.
Bug: https://code.google.com/p/android/issues/detail?id=211233
Change-Id: Id9ca4999fd2e6393cc88278eaf444243e13c0ec0
close_all_sockets was assuming that all registered local sockets used
local_socket_close as their close function. However, this is not true
for JDWP sockets.
Bug: http://b/28347842
Change-Id: I40a1174845cd33f15f30ce70828a7081cd5a087e
(cherry picked from commit 53eb31d87c)
sockets.cpp was branching on whether a socket close function was
local_socket_close in order to avoid a potential deadlock if the socket
list lock was held while closing a peer socket.
Bug: http://b/28347842
Change-Id: I5e56f17fa54275284787f0f1dc150d1960256ab3
(cherry picked from commit 9b587dec6d)
Our version of mingw doesn't support std::mutex or
std::recursive_mutex, so implement our own using the Windows primitives.
Bug: http://b/28347842
Change-Id: I4e1d56a89bc5fcb5f859bf5014343697a4a85b77
(cherry picked from commit 903b749f78)
Android Wear has unfortunately been using port 5601 for years, which
falls into the range of ports we were previously polling for Android
emulators. Reduce the maximum number of emulators we can support so
that 5601 no longer falls within our range.
Bug: http://b/26468076
Change-Id: I931809cfa412122f4781eebe0164facab12c95f0
adbd on device kicks the usb connection when usb IO fails. But to notify
adb on host to reconnect it, adbd needs to close ep0 to reset the usb
connection. Otherwise, adb on host cann't connect the device unless the
usb cable is reconnected or adb host server is restarted.
This can be tested by using `adb reconnect device` command.
Bug: 25935458
Change-Id: I80979d6029e60b9cfd218f7b9b5201810238ec07
Problem: For devices using /dev/usb-ffs/adb, Run
`while true; do adb reconnect device; sleep 1; done`. And the
device soon becomes offline. The adbd log shows that calling
adb_read(h->bulk_out) in usb_ffs_read() gets EOVERFLOW error.
Reason: When kicking a transport using usb-ffs, /dev/usb-ffs/adb/ep0
is not closed, and the device will not notify a usb connection reset
to host. So the host will continue to send unfinished packets even
if a new transport is started on device. The unfinished packets may
not have the same size as what is expected on device, so adbd on
device gets EOVERFLOW error. At the worst case, adbd has to create new
transports for each unfinished packet.
Fixes:
The direct fix is to make the usb connection reset when kicking transports,
as in https://android-review.googlesource.com/#/c/211267/1. And I think
we can make following improvements beside that.
1. Close a file that is used in other threads isn't safe. Because the file
descriptor may be reused to open other files, and other threads may operate
on the wrong file. So use dup2(dummy_fd) to replace close() in kick function,
and really close the file descriptor after the read/write threads exit.
2. Open new usb connection after usb_close() instead of after
usb_kick(). After usb_kick(), the transport may still exist and
reader/writer for the transport may be still running. But after
usb_close(), the previous transport is guaranteed to be destroyed.
Bug: 25935458
Change-Id: I1eff99662d1bf1cba66af7e7142f4c0c4d82c01b
(cherry picked from commit 005bf1e05b)
On mac, if the adb server kicks a transport on some error, mac usb driver
will not report a new usb device. So instead of relying on mac usb driver
to report new usb devices, this CL uses a loop to search for usb devices
not exist before. Note that this is also the behavior on windows and linux
host.
`adb reconnect` can be used to verity this CL.
Bug: 25935458
Change-Id: I890e0eb1fae173f2e7a0c962ededa294d821e015
(cherry picked from commit 48d4c0c42a)
Fix broken kick_transport test, and make it not access atransport
internal variables.
Bug: 25935458
Change-Id: I91b4d32a222b2f369f801bbe3903acac9c8ea4f7
(cherry picked from commit 7f27490e7f)
128 maximum FDs is a pretty low limit, which can easily be exhausted by
port forwarding. Bump the maximum up to 2048, and add a test that checks
whether we can actually use a few hundred sockets.
Bug: https://code.google.com/p/android/issues/detail?id=12141
Bug: http://b/28246942
Change-Id: Ia4a2ff776e8e58ec13378756f19d80392679ece9
(cherry picked from commit b31e17107c)
The recent `adb root` changes are incompatible with older versions of
the server. Bump the version number to force the server to restart.
Bug: http://b/28194507
Change-Id: I970806e3b68c1f8e3273a4b1f0ecc4aca5086be9
(cherry picked from commit 057095d207)
Add reconnect command for debugging. `reconnect` kicks a transport
from the host side, `reconnect device` kicks a transport from
the device side. They can be used to produce transport errors.
Bug: 25935458
Change-Id: I47daa338796b561941e7aba44a51a6dd117d1e98
(cherry picked from commit 1f4ec19e49)
Otherwise, `adb root` while the device is in recovery will hang.
Bug: http://b/28168212
Change-Id: Ibc8038b1745139fb505e466c74c605b9df474c8c
(cherry picked from commit 48a348183f)
Allow waiting for a device in any state.
Bug: http://b/28168212
Change-Id: I1876ecd70089ca88f2da5de4182e13724ec50501
(cherry picked from commit 86441c31fe)
The help for wait-for-* documents this as working, but we only had a
special case for 'wait-for-device'. Make the others work as well.
Bug: http://b/28170823
Change-Id: I57d70f8b9a383947c4f761919bfbd8515a013520
(cherry picked from commit 08881e72f4)
Dumpstate now supports zipped bugreport, whose output is more complete
than the flat-file bugreports provided prior to N. As such, adb now has
a 'adb bugreport <ZIP_FILE>' name whose implementation:
- Calls the new bugreportz binary.
- Parses its output, which in case of success is the path of the .zip
file.
- Pulls the device file and renames it according to the command-line
argument.
BUG: 27653204
Change-Id: I7169fe157c77bbef1684d0cb4e43095d95ddf2b8
ADB local transport for the emulator based on qemu pipes uses a socket
transport. Before this CL, multiple connection requests accepted by the
qemu pipe device would result in ADB transport with the same serial.
The register function would fail as a result, and all subsequent
connections would fail.
Test:
while true; do adb kill-server; adb devices; done
Doesn't fail for > 10 minutes.
This CL replaces an earlier CL
(I0fdcf2694516151c5f8f8e1580648b940679c981) that was unsafe for real
devices.
BUG=27441661
Change-Id: I7d801b175f3bee10fc7e0ab1b12d5623984371b9
forkpty only returns an fd on the master side; trying to keep track of
it on both sides is incorrect.
Bug: http://b/27504427
Change-Id: Ia4a454d8490c77738f9c3350a4236b3d99c8758b
(cherry picked from commit fcb063ce37)
If we return before waiting for the other side to finish, it'll get sent
SIGHUP.
Bug: http://b/27414382
Change-Id: I93514d2242ed0d7aa93a6ec6836270dc0783506d
(cherry picked from commit d26abed750)
The directory name should be based off of the local path, not the remote
path.
Change-Id: I75b089b8734e9dbf8e466b1e00ea18549fd101bb
(cherry picked from commit 89ec3a8d0f)
This matches scp's behavior when pulling a directory that collides
with a symlink to a directory.
Bug: http://b/27362811
Change-Id: I0936d1ad48f13e24cd382e8e8400cc752bac3b66
(cherry picked from commit 1e611a33d5)
Currently targeting a device by serial requires matching the serial
number exactly. This CL relaxes the matching rules for local transports
to ignore protocol prefixes and make the port optional:
[tcp:|udp:]<hostname>[:port]
The purpose of this is to allow a user to set ANDROID_SERIAL to
something like "tcp:100.100.100.100" and have it work for both fastboot
and adb (assuming the device comes up at 100.100.100.100 in both
modes).
This CL also adds some unit tests for the modified functions to make
sure they work as expected.
Bug: 27340240
Change-Id: I006e0c70c84331ab44d05d0a0f462d06592eb879
(cherry picked from commit 3f902aad5b)
On linux we archive both Linux and Windows binaries.
Bug: 27315911
Change-Id: I17d25cd1ea611a9c917f1e78b47f21a8959eea20
(cherry picked from commit 71edfc8b9a)
Previously, for `adb pull $remote $local`, we would do the equivalent of
mkdir -p on `dirname $local`. This patch changes the behavior to only
creating directories that are being pulled, like scp.
Bug: http://b/27362811
Change-Id: I79f975ee9f2d9bc9e8be6a7c4f2de6d7ae2d2d23
(cherry picked from commit 71728ca300)
Currently adb only realizes a TCP transport has gone away when it tries
to send a packet, which caused problems in particular for `adb reboot`
since no packets are sent, leading to the client hanging until Ctrl+C.
This CL turns on TCP keepalive packets to send 1 packet every second,
allowing up to 10 failures before disconnecting. Using built-in TCP
functionality turns out to be much cleaner in this case than trying to
implement our own keepalive packets at the application layer, and
should be more lightweight since it's all done in the TCP stack.
Bug: http://b/23093474
Change-Id: Ifb41cbb85b9752a9f394e1eed3c6ac4da47a4e4d
(cherry picked from commit bfd9503d1d)
It does happens that the adb_write only writes to the
qemu pipe partially which throws host side's adb backend
into confusion and crashes.
This CL replaces adb_write with WriteFdExactly;
adb_read with ReadFdExactly.
(cherry picked from commit f66c5938be)
Change-Id: I684f5df79b1e3f00b4b7a2452c2712a73c15973c
Move the fdevent for the framework authentication connection out of
atransport into its own static variable in adb_auth_client, since its
lifetime is completely unrelated to that of the USB connection.
Bug: http://b/27297963
Change-Id: Ie6180d0b59d133120c5755e239e76ab33ed3cc1d
(cherry picked from commit 9f48611554)
