Second follow up to https://chromium-review.googlesource.com/c/400015/
The ideal would be that if we fail to start the handler, then we don't
end up passing through our unhandled exception filter at all.
In the case of the non-initial client (i.e. renderers) we can do this by
not setting our UnhandledExceptionFilter until after we know we've
connected successfully (because those connections are synchronous from
its point of view). We also change WaitForNamedPipe in the connection
message to block forever, so as long as the precreated pipe exists,
they'll wait to connect. After the initial client has passed the server
side of that pipe to the handler, the handler has the only handle to it.
So, if the handler has disappeared for whatever reason, pipe-connecting
clients will fail with FILE_NOT_FOUND, and will not stick around in the
connection loop. This means non-initial clients do not need additional
logic to avoid getting stuck in our UnhandledExceptionFilter.
For the initial client, it would be ideal to avoid passing through our
UEF too, but none of the 3 options are great:
1. Block until we find out if we started, and then install the filter.
We don't want to do that, because we don't want to wait.
2. Restore the old filter if it turns out we failed to start. We can't
do that because Chrome disables ::SetUnhandledExceptionFilter()
immediately after StartHandler/SetHandlerIPCPipe returns.
3. Don't install our filter until we've successfully started. We don't
want to do that because we'd miss early crashes, negating the benefit
of deferred startup.
So, we do need to pass through our UnhandledExceptionFilter. I don't
want more Win32 API calls during the vulnerable filter function. So, at
any point during async startup where there's a failure, set a global
atomic that allows the filter function to abort without trying to signal
a handler that's known to not exist.
One further improvement we might want to look at is unexpected
termination of the handler (as opposed to a failure to start) which
would still result in a useless Sleep(60s). This isn't new behaviour,
but now we have a clear thing to do if we detect the handler is gone.
(Also a missing DWORD/size_t cast for the _x64 bots.)
R=mark@chromium.org
BUG=chromium:567850,chromium:656800
Change-Id: I5be831ca39bd8b2e5c962b9647c8bd469e2be878
Reviewed-on: https://chromium-review.googlesource.com/400985
Reviewed-by: Mark Mentovai <mark@chromium.org>
Previously, StartHandler() launched the handler process, then connected
over a pipe to register for crash handling. Instead, the initial client
can create and inherit handles to the handler and pass those handle
values and other data (addresses, etc.) on the command line.
This should improve startup time as there's no need to synchronize with
the process at startup, and allows avoiding a call to CreateProcess()
directly in StartHandler(), which is important for registration for
crash reporting from DllMain().
Incidentally adds new utility functions for string/number conversion and
string splitting.
Note: API change; UseHandler() is removed for all platforms.
BUG=chromium:567850,chromium:656800
Change-Id: I1602724183cb107f805f109674c53e95841b24fd
Reviewed-on: https://chromium-review.googlesource.com/400015
Reviewed-by: Mark Mentovai <mark@chromium.org>
This was done in Chromium’s local copy of Crashpad in 562827afb599. This
change is similar to that one, except more care was taken to avoid
including headers from a .cc or _test.cc when already included by the
associated .h. Rather than using <stddef.h> for size_t, Crashpad has
always used <sys/types.h>, so that’s used here as well.
This updates mini_chromium to 8a2363f486e3a0dc562a68884832d06d28d38dcc,
which removes base/basictypes.h.
e128dcf10122 Remove base/move.h; use std::move() instead of Pass()
8a2363f486e3 Move basictypes.h to macros.h
R=avi@chromium.org
Review URL: https://codereview.chromium.org/1566713002 .
This more-natural spelling doesn’t require Crashpad developers to have
to remember anything special when writing code in Crashpad. It’s easier
to grep for and it’s easier to remove the “compat” part when pre-C++11
libraries are no longer relevant.
R=scottmg@chromium.org
Review URL: https://codereview.chromium.org/1513573005 .
This unifies several things that used a 16-character random string, and
a few other users of random identifiers where it also made sense to use
a 16-character random string.
TEST=crashpad_util_test RandomString.RandomString
R=scottmg@chromium.org
Review URL: https://codereview.chromium.org/1451793002 .
This consolidates all of the twisted casts and comments that discuss how
HANDLEs are really only 32 bits wide even in 64-bit processes on 64-bit
operating systems into a single location.
R=scottmg@chromium.org
Review URL: https://codereview.chromium.org/1422503015 .
Allowing the client to create its own pipe name string caused a race
between client and server. Instead, in this mode, the server now creates
the pipe name along with a pipe, and returns it to its client via a
--handshake-handle. This guarantees that by the time the client gets the
pipe name, the server has already created it.
Ephemeral mode is now implied by --handshake-handle. The --persistent
option is gone. --persistent mode is enabled when using --pipe-name.
BUG=crashpad:69
R=scottmg@chromium.org
Review URL: https://codereview.chromium.org/1432563003 .
Previously, crashpad_handler made its own receive right, and transferred
a corresponding send right to its client. There are two advantages to
making the receive right in the client:
- It is possible to monitor the receive right for a port-destroyed
notificaiton in the client, allowing the handler to be restarted if
it dies.
- For the future run-from-launchd mode (bug crashpad:25), the handler
will obtain its receive right from the bootstrap server instead of
making its own. Having the handler get its receive right from
different sources allows more code to be shared than if it were to
sometimes get a receive right and sometimes make a receive right and
transfer a send right.
This includes a restructuring in crashpad_client_mac.cc that will make
it easier to give it an option to restart crashpad_handler if it dies.
The handler starting logic should all behave the same as before.
BUG=crashpad:68
R=rsesek@chromium.org
Review URL: https://codereview.chromium.org/1409073013 .
This allows better code sharing in crashpad_handler’s main(). It doesn’t
look like much of an improvement now, but a separate change will cause
the Mac ExceptionHandlerServer() to be constructed with an argument. It
will be beneficial for Mac and Windows to be able to share the Run()
call.
R=scottmg@chromium.org
Review URL: https://codereview.chromium.org/1402333004 .
Fixes two incorrect usages of ssize_t/off_t being implicitly converted
to bool. As such, I think it's worth the cost of the additional !! on
BOOL returning Win32 functions.
R=mark@chromium.org
Review URL: https://codereview.chromium.org/1408123006 .
Capture the memory for the loader lock (can be inspected by !cs), as
well as all locks that were created with .DebugInfo which can be viewed
with !locks.
e.g.
0:000> !cs ntdll!LdrpLoaderLock
-----------------------------------------
Critical section = 0x778d6410 (ntdll!LdrpLoaderLock+0x0)
DebugInfo = 0x778d6b6c
NOT LOCKED
LockSemaphore = 0x0
SpinCount = 0x04000000
0:000> !locks -v
CritSec ntdll!RtlpProcessHeapsListLock+0 at 778d7620
LockCount NOT LOCKED
RecursionCount 0
OwningThread 0
EntryCount 0
ContentionCount 0
CritSec +7a0248 at 007a0248
LockCount NOT LOCKED
RecursionCount 0
OwningThread 0
EntryCount 0
ContentionCount 0
CritSec crashy_program!g_critical_section_with_debug_info+0 at 01342c48
LockCount NOT LOCKED
RecursionCount 0
OwningThread 0
EntryCount 0
ContentionCount 0
CritSec crashy_program!crashpad::`anonymous namespace'::g_test_critical_section+0 at 01342be0
WaiterWoken No
LockCount 0
RecursionCount 1
OwningThread 34b8
EntryCount 0
ContentionCount 0
*** Locked
Scanned 4 critical sections
R=mark@chromium.org
BUG=crashpad:52
Review URL: https://codereview.chromium.org/1392093003 .
Windows requires the connection to the handler to do anything, so it
can't really be implemented or tested without CrashpadClient and the
connection machinery.
R=mark@chromium.org
BUG=crashpad:53
Review URL: https://codereview.chromium.org/1356383002 .
The pipe handle was being leaked on connections (oops!). On XP this
resulted in the next test's CreateNamedPipe to fail, because the
previous one still existed (because all handles were not closed). More
recent OSs are more forgiving so I got away with the buggy code.
R=mark@chromium.org
BUG=crashpad:50
Review URL: https://codereview.chromium.org/1337953003 .
PROCESS_ALL_ACCESS was changed in later SDKs and the newer value fails
when run on XP with ERROR_ACCESS_DENIED. Use the old value to maintain
compatibility with XP.
R=mark@chromium.org
BUG=crashpad:50
Review URL: https://codereview.chromium.org/1337133002 .
This replaces the registration server, and adds dispatch to a delegate
on crash requests.
(As you are already aware) we went around in circles on trying to come
up with a slightly-too-fancy threading design. All of them seemed to
have problems when it comes to out of order events, and orderly
shutdown, so I've gone back to something not-too-fancy.
Two named pipe instances (that clients connect to) are created. These
are used only for registration (which should take <1ms), so 2 should be
sufficient to avoid any waits. When a client registers, we duplicate
an event to it, which is used to signal when it wants a dump taken.
The server registers threadpool waits on that event, and also on the
process handle (which will be signalled when the client process exits).
These requests (in particular the taking of the dump) are serviced
on the threadpool, which avoids us needing to manage those threads,
but still allows parallelism in taking dumps. On process termination,
we use an IO Completion Port to post a message back to the main thread
to request cleanup. This complexity is necessary so that we can
unregister the threadpool waits without being on the threadpool, which
we need to do synchronously so that we can be sure that no further
callbacks will execute (and expect to have the client data around
still).
In a followup, I will readd support for DumpWithoutCrashing -- I don't
think it will be too difficult now that we have an orderly way to
clean up client records in the server.
R=cpu@chromium.org, mark@chromium.org, jschuh@chromium.org
BUG=crashpad:1,crashpad:45
Review URL: https://codereview.chromium.org/1301853002 .
