fexecve overwrites argv[0] for cases in which the exec'd binary should
close the fd itself rather than the calling process. This makes it
inappropriate for use in the VM. Instead the VM should use
execveat, however it hasn't been added to the C Library everywhere,
yet.
fixes#30971
Change-Id: I82085f24ded2f0b846d9193ca55903686e1f3585
Reviewed-on: https://dart-review.googlesource.com/10205
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Fuchsia requires the ability to sandbox Isolates w.r.t. file IO.
When a new Isolate starts, Fuchsia will pass the Isolate an object
called a namespace. We can translate the namespace object into a
file descriptor suitable for passing to the *at() family of
POSIX file system calls. The file system calls will then
have visibility only into the specified namespace.
We also plumb Namespaces through on all the other platforms as well to
make the change easier to test and so that in the future we can
implement e.g. per-isolate cwds.
This change adds a new internal class to dart:io called _Namespace,
which is implemented in a patch file. See:
sdk/lib/io/namespace_impl.dart
runtime/bin/namespace_patch.dart
The embedder can set up a non-default namespace by calling
_Namespace._setupNamespace during Isolate setup.
Instances of _Namespace have a native field that holds a pointer
to a native Namespace object. See:
runtime/bin/namespace.h
Calls from e.g. file_impl.dart are now also passed a
_Namespace object. The implementations in e.g. file.cc and
file_linux.cc then extract the namespace, and use it to compute a
file descriptor and path suitable for passing to e.g. openat().
related US-313
R=asiva@google.com, rmacnak@google.com
Review-Url: https://codereview.chromium.org/3007703002 .
Moved the responsibility to provide MaxRSS and CurrentRSS from the VM to
the embedder.
The embedder can opt-it by setting a Dart_GetEmbedderInformation
callback using the public Dart_SetEmbedderInformationCallback API.
The implementation of the Dart_GetEmbedderInformation should mandatory
fill the version field and the ones it is able to fill.
The name field must reference a constant C style string, it will not be
freed by the VM code.
R=zra@google.com
Review-Url: https://codereview.chromium.org/2996803002 .
Like HOST_ARCH_*, HOST_OS_* describes the OS the VM is running on, which may be different from the OS the VM is generating code for during AOT compilation.
Currently we conflate the two when emitting AOT as assembly, and we get away with it because Flutter only uses assembly for targeting iOS and one can only target iOS from a Mac, but we expect to use assembly for Android as well so native tools can unwind Dart frames.
R=zra@google.com
Review-Url: https://codereview.chromium.org/2750843003 .
Somtimes execve() fails and then we fall back to exit(1) in the fork()ed but not
exeve()ed subprocess.
One of the issues is e.g. that we don't want to run LSAN in the fork()ed but not
execve()ed subprocess, the results will be bogus (plus the output of
LSAN is not read on the other side (i.e. the unix pipe is not drained) which
blocks the subprocess indefinitely).
R=vegorov@google.com
Review-Url: https://codereview.chromium.org/2632483002 .
About this CL:
The only purpose of `if (fork() == 0) exit(0)` is to wake up
a thread in the parent process which might be blocked on `wait()`.
There is no need to run atexit() handlers in the `fork()`ed child.
This is a *workaround attempt* for a deadlocked `free()` call inside the
processing of atexit handlers in glibc.
(Side note: There might be better ways of notifying the thread, like sending a
signal to the particular pthread with `pthread_kill` which would make the
`wait()` syscall be interrupted.)
About the issue:
It is still unclear why, in this particular case, the tcmalloc locks should
be hold during the `exit()` call:
* via a static initializer tcmalloc uses
`pthread_atfork(before=ObtainAllLocks(),
after_parent=ReleaseAllLocks(),
after_child=ReleaseAllLocks())`
to register locking & unlocking around `fork()`
* glibc's `fork()` runs the either `after_parent` or `after_child` handlers
(unconditionally) which should free the locks
* the `exit()` call later should be free to malloc/free
The [BUG] describes more in detail how we can hit a tcmalloc deadlock in a
different situation (it's a linux kernel bug).
Namely, if the linux kernel runs OOM during `fork()` and therefore fails to
set the new thread-id. The glibc code hits an assert and tries to allocate
memory before `after_parent`/`after_child` handlers were executed which
deadlocks.
BUG=https://github.com/dart-lang/sdk/issues/28246R=vegorov@google.com
Review-Url: https://codereview.chromium.org/2618723002 .
Now, failing to allocate a buffer for stderr or stdout will generate
an ENOMEM OSError exception.
Also previously, read() returning an error would cause buffers for
stderr and stdout to leak. After this change, they'll be reclaimed.
R=asiva@google.com
Review-Url: https://codereview.chromium.org/2596543002 .
Previously we would compile in implementaitions of native calls for
IO functions that would never be used. This CL provides implementations
that throw a Dart exception if they're called by mistake. It also uses
a DART_IO_DISABLED preprocessor define to clean up the build files and
check that we're not including code we shouldn't.
R=iposva@google.com, johnmccutchan@google.com
Review URL: https://codereview.chromium.org/1839463002 .
Also some style cleanups.
Decided to fix these leaks by using Dart_ScopeAllocate
instead of malloc and new. Leaks are noted in the CL.
I haven't finished looking over all the code in
//runtime/bin yet, but this CL was getting big.
Review URL: https://codereview.chromium.org/1781883002 .
asiva