We can move this information into ProtocolState and save having two
virtual functions for every protocol. Moving some bits out of the
codegen'd IPC code is a nice bonus, though we keep the strange setup
where toplevel protocols have two mChannel member variables.
ProtocolName() is only used for producing error messages and annotating
crash reports. But examining actual crash reports that would have used
the result of ProtocolName() indicates that we can always tell what the
erroring protocol is due to the stack backtrace. So having this virtual
function around just provides duplicate information, and it takes up too
much space in the vtable besides. Let's get rid of it.
lower.py generates repetitious:
SetManager(...);
Register(...); // Or RegisterID.
SetIPCChannel(...);
calls, which are moderately sized, given that the above call sequence
requires virtual calls in several places. Instead of codegenning this
sequence, let's consolidate the sequence into IProtocol and change the
code generator to call into the consolidated function instead.
This function is only overriden in two places, both of which go away
after early beta is done. We shouldn't be paying for its vtable entry
after that point.
The reasoning here is the same as for the protocol register/lookup
functions: these functions are all basic functionality that should not
be overriden by subclasses.
This functionality is base functionality for top-level and non-toplevel
protocols; nobody overrides this stuff, so it's safe to move into
ProtocolState.
IProtocol, which is inherited by every generated IPDL protocol and every
concrete protocol implementation in-tree, has a number of virtual
methods that are only relevant when distinguishing between top-level
protocols (IToplevelProtocol) and managed protocols (everything else).
These virtual methods require pointers in every protocol's vtable, which
is wasteful, and it's also somewhat confusing that many methods exist
but don't really need to be overridable in any useful way.
Let's clean this up, by creating a ProtocolState class to hold methods
that solely differ between top-level protocols and everything else.
This commit does that work and moves Shmem-related methods into this
class as a proof that this can be done in a reasonable way.
I've also manually verified that no other references to HANDLE_EINTR are
wrapping a close() in any less syntactically obvious way.
MozReview-Commit-ID: 3KkBwFIhEIq
--HG--
extra : rebase_source : 4e79a70b3be22a7721b6f85b19ee5a31c98df456
This patch was reviewed in parts, however the intermediate states would not build:
Bug 1443954 - Part 3A: Strip pointers from the argument to WriteParam and WriteIPDLParam before selecting the ParamTraits impl, r=froydnj
Bug 1443954 - Part 3B: Move nsIAlertNotification serialization to the refcounted system, r=bz
Bug 1443954 - Part 3C: Move geolocation serialization to the refcounted system, r=bz
Bug 1443954 - Part 3D: Move nsIInputStream serialization to the refcounted system, r=baku
Bug 1443954 - Part 3E: Move BlobImpl serialization to the refcounted system, r=baku
Bug 1443954 - Part 3F: Correctly implement ParamTraits for actors after the ParamTraits changes, r=froydnj
The current code is a bit of a mess. This patch does the following.
- Changes the processing from backwards to forwards. This avoids the need for
all the `found` booleans, because if a flag is present multiple times, the
last one will naturally override.
- Tightens up the checking. It now doesn't use assertions, but instead returns
false if any of the options are missing arguments, or have malformed
arguments, or any of the mandatory flags are missing. (It assumes that
-appdir and -profile are optional.)
- Renames the loop variable `idx` as `i`.
- Changes `!strcmp(...)` to `strcmp(...) == 0`, because I find that clearer.
- Avoids a redundant nsCString when handling -appdir.
The patch also tweaks GeckoChildProcessHost::mGroupId, which was buggy. It
holds the appModelUserId argument, which XRE_InitChildProcess() always expects
is present in the command. But it's only set to a non-empty value in
InitWindowsGroupID(), which is only called for plugin processes. So in lots of
cases the appModelUserId argument was missing, and a different argument would
be interpreted as the appModelUserId argument (seemingly without noticeable ill
effect).
The patch changes things to mGroupId defaults to "-", which means it's always
present in the command.
Note: all this explains why the old code for ContentProcess::Init() started
processing from argument aArgc, instead of aArgc-1 as you might expect -- it
had to read one extra arg in order to see the argument following -appdir,
because XRE_InitChildProcess() was decrementing aArgc for the appModelUserId
argument even when that argument wasn't present. The new code for
ContentProcess::Init() doesn't have to read past aArgc-1 because the mGroupId
fix ensures the appModelUserId argument is always present.
MozReview-Commit-ID: 8a8k6ABYMgo
--HG--
extra : rebase_source : 70695125ee26e67af3337119f4dfc293a0dab74c
Once we've figured out that some task needs to be reposted, there's no
reason to continue scanning the list to find other tasks that need to be
reposted, since the logic in this function just requires one task that
needs to be reposted.
We're not using named shared memory, and supporting only anonymous
shared memory allows using other backends that are more compatible
with preventing a process from accessing any shared memory it wasn't
explicitly granted (i.e., sandboxing).
Specifically: SharedMemory::Open is removed; SharedMemory::Create no
longer takes a name, no longer has the open_existing option which doesn't
apply to anonymous memory, and no longer supports read-only memory
(anonymous memory which can never have been written isn't very useful).
This patch also fixes some comments in what remains of SharedMemory::Create.
MozReview-Commit-ID: 4kBrURtxq20
--HG--
extra : rebase_source : f6b1fb2fc79b6e9cdd251b3d9041036c0be503f9
They are not yet fully async because ContentParent::InitInternal calls
OtherPid(), which will block until the process is spawned. Deferring the calls
to OtherPid() will be a subject of a follow up patch.
MozReview-Commit-ID: 4TFkMpdQtRw
--HG--
extra : rebase_source : 3e7567679ae04aa4c04ea6f6c146e70417e7ce05
This will let us manipulate it from multiple threads in a future patch.
MozReview-Commit-ID: 2AOgho8SEX9
--HG--
extra : rebase_source : 49e7d82b38e6f7b7eb9f6cb7c61a2b41d34f2bdc
This patch replaces the large -intPrefs/-boolPrefs/-stringPrefs flags with
a short-lived, anonymous, shared memory segment that is used to pass the early
prefs.
Removing the bloat from the command line is nice, but more important is the
fact that this will let us pass more prefs at content process start-up, which
will allow us to remove the early/late prefs split (bug 1436911).
Although this mechanism is only used for prefs, it's conceivable that it could
be used for other data that must be received very early by children, and for
which the command line isn't ideal.
Notable details:
- Much of the patch deals with the various platform-specific ways of passing
handles/fds to children.
- Linux and Mac: we use a fixed fd (8) in combination with the new
GeckoChildProcessHost::AddFdToRemap() function (which ensures the child
won't close the fd).
- Android: like Linux and Mac, but the handles get passed via "parcels" and
we use the new SetPrefsFd() function instead of the fixed fd.
- Windows: there is no need to duplicate the handle because Windows handles
are system-wide. But we do use the new
GeckoChildProcessHost::AddHandleToShare() function to add it to the list of
inheritable handles. We also ensure that list is processed on all paths
(MOZ_SANDBOX with sandbox, MOZ_SANDBOX without sandbox, non-MOZ_SANDBOX) so
that the handles are marked as inheritable. The handle is passed via the
-prefsHandle flag.
The -prefsLen flag is used on all platforms to indicate the size of the
shared memory segment.
- The patch also moves the serialization/deserialization of the prefs in/out of
the shared memory into libpref, which is a better spot for it. (This means
Preferences::MustSendToContentProcesses() can be removed.)
MozReview-Commit-ID: 8fREEBiYFvc
--HG--
extra : rebase_source : 7e4c8ebdbcd7d74d6bd2ab3c9e75a6a17dbd8dfe
This changes the lifecycle and API for TextureReadLock to fix file descriptor exhaustion
crashes. These changes are partially superficial and mostly align the API of TextureReadLocks
with their actual usage.
The changes are:
1. Create the TextureReadLock in the TextureClient constructor so it's available before IPC creation
a. This is superficial as EnableReadLock was always called before IPC creation
2. Send the ReadLockDescriptor in the PTextureConstructor message and close the file handle
3. Receive the ReadLockDescriptor in TextureHost and close the file handle
4. Send a boolean flag in layer transactions if the texture is read locked instead of a descriptor
5. Use a boolean flag in TextureHost to determine if the ReadLock must be unlocked instead of a nullptr
I believe that we can remove the InitReadLocks code from LayerTransaction as that was added to
prevent file descriptor limits in IPDL messages and is no longer needed with this change. But
that is a non-essential change and this patch is already big enough.
MozReview-Commit-ID: DzHujrOQejH
--HG--
extra : rebase_source : 3bdd7c9bc8edfdc386faad8a9e59ad7dc18ed91d
Switch the order of the IPC FD argument and the crash FD argument in
e10s calls, because the IPC FD is the primary FD, and the crash FD
should be grouped with the crash annotation FD.
MozReview-Commit-ID: CAVyYAIIBPm
--HG--
extra : rebase_source : 596f590443f727d1a79582202eed122f79ae85cf