This requires a workaround for the use of __wrap_dladdr, which can't be
used in logalloc-replay. The workaround involves making __wrap_dladdr
expand to dladdr, but that makes the definition ElfLinker.h conflict
with the one in the Android system headers, so we change it to match,
and adjust ElfLinker.cpp accordingly.
And while here, fix the condition in mozglue/misc to match the condition
around including Linker.h in StackWalk.cpp itself.
Differential Revision: https://phabricator.services.mozilla.com/D82648
We still need it for Android < 6.0 (API 23) because otherwise we don't
have a way to hook up mozalloc.
Differential Revision: https://phabricator.services.mozilla.com/D66993
--HG--
extra : moz-landing-system : lando
We still need it for Android < 6.0 (API 23) because otherwise we don't
have a way to hook up mozalloc.
Differential Revision: https://phabricator.services.mozilla.com/D66993
--HG--
extra : moz-landing-system : lando
For now, there is no flag to actually allow it, but this is the
code-side changes to allow the linker being disabled.
Differential Revision: https://phabricator.services.mozilla.com/D54074
--HG--
extra : moz-landing-system : lando
__wrap_dlerror uses a single pointer for all threads, which means one
thread could get the dlerror result from another thread. Normally this
wouldn't cause crashes. However, because dlerror results come from a
per-thread buffer, if a thread exits and our saved dlerror result came
from that thread, the saved pointer could then refer to invalid memory.
The proper way to fix this is to use TLS and have a per-thread pointer
for __wrap_dlerror. However, instead of using up a TLS slot, this patch
keeps the single pointer for custom messages, and fallback to per-thread
dlerror call for system messages. While the race condition still exists,
I think the risk is acceptable. Even when races occur, they should no
longer cause crashes.
MozReview-Commit-ID: 4hGksidjiVz
--HG--
extra : rebase_source : 373000686c426b81ffd7cee88264e89b7a733957
When we modify the debug map, we could be racing with the system linker,
either when we modify the entries or when we change page protection
flags. To fix the race, we need to take the system linker's internal
lock when we perform any kind of modification on the debug map.
One way to hold the system linker lock is to call dl_iterate_phdr, and
perform our actions inside the callback, which is invoked with the
lock being held. However, dl_iterate_phdr is only present on Android
5.0+, and even then, dl_iterate_phdr is only protected by the linker
lock on Android 6.0+.
This means that with this patch, we can only safely modify the debug map
on Android 6.0+, which I think is acceptable for an operation that only
benefits a debugger.
MozReview-Commit-ID: BowBEO8tu8Z
--HG--
extra : amend_source : 837631dfc2ef17b24ffe5778bcb70dc29b7dfc66
We are apparently still crashing even after mprotect() with write flag
returns successfully. This patch reads the flags again after mprotect()
returns, and hopefully the flags will tell the truth of whether the page
is truly writable or not after calling mprotect().
MozReview-Commit-ID: Jsg8vHKFEvJ
--HG--
extra : rebase_source : b028aa0d5cefd50302bfc2502292d9129d202e09
When looping through the debugger helper links during our
dl_iterate_phdr implementation, we effectively race with other threads
dlclose()ing libraries while we're working.
We do have a (rather involved) check in place to ensure that elf headers
are readable. But it turns out in practice, some dlclose() do happen
between the check and the actual read of the elf headers.
Unfortunately, we can't lock the system linker while we're looping, so a
better approach is to only loop through the libraries we loaded, and
rely on the system dl_iterate_phdr to iterate over the (remaining)
system libraries.
Unfortunately (again), Android versions < 5.0 don't have a system
dl_iterate_phdr, so we have to rely on the old iterator when it's not
present.
--HG--
extra : rebase_source : 3fd07589ed1939411ef72f481b7c72f761d53701
We're going to introduce two code paths that need to fill dl_phdr_info
from different iterators, so first move the code to a separate class,
that both code paths will be able to call.
--HG--
extra : rebase_source : a50663ad8d15d4f7a28d7138824003df5edd7f1c
We already dealt with issues around dlsym not resolving weak symbols
with libc in bug 1081034. This fix applies the same workaround to libm,
which solves the isnanf issue.
The previous fix for looking up __isnanf is no longer needed.
The comment with the accompanying change explains things, but the short
version is that clang generates full calls to isnanf, which our
dlsym-based symbol lookup in the custom linker cannot handle correctly.
We therefore need to do extra work for isnanf to find the correct symbol.
Since bug 1307886, we don't actually use szip anymore, and don't even
have the option to package Fennec using it. We can thus remove the
support for loading them, as well as on demand linkage.
The latter might mean we can remove the segfault handler, but it's
unclear whether this is currently working around other issues with
registering signal handlers, so we'll leave that to a followup.
--HG--
extra : rebase_source : ec23cd4e78f259a70f6690adc8dfabb557e8f304
Fill in AArch64 trampoline for Divert(). Even though we're not using
on-demand decompression anymore, I added the AArch64 cases for
completeness.
MozReview-Commit-ID: D91KhHiDo7S
Fix printf macro mismatches where, for example, `PRIxPTR` is defined for
`long` but the ELF `Addr` type is defined as `long long`.
MozReview-Commit-ID: 8hXY1MpHPjS
The bulk of this commit was generated with a script, executed at the top
level of a typical source code checkout. The only non-machine-generated
part was modifying MFBT's moz.build to reflect the new naming.
CLOSED TREE makes big refactorings like this a piece of cake.
# The main substitution.
find . -name '*.cpp' -o -name '*.cc' -o -name '*.h' -o -name '*.mm' -o -name '*.idl'| \
xargs perl -p -i -e '
s/nsRefPtr\.h/RefPtr\.h/g; # handle includes
s/nsRefPtr ?</RefPtr</g; # handle declarations and variables
'
# Handle a special friend declaration in gfx/layers/AtomicRefCountedWithFinalize.h.
perl -p -i -e 's/::nsRefPtr;/::RefPtr;/' gfx/layers/AtomicRefCountedWithFinalize.h
# Handle nsRefPtr.h itself, a couple places that define constructors
# from nsRefPtr, and code generators specially. We do this here, rather
# than indiscriminantly s/nsRefPtr/RefPtr/, because that would rename
# things like nsRefPtrHashtable.
perl -p -i -e 's/nsRefPtr/RefPtr/g' \
mfbt/nsRefPtr.h \
xpcom/glue/nsCOMPtr.h \
xpcom/base/OwningNonNull.h \
ipc/ipdl/ipdl/lower.py \
ipc/ipdl/ipdl/builtin.py \
dom/bindings/Codegen.py \
python/lldbutils/lldbutils/utils.py
# In our indiscriminate substitution above, we renamed
# nsRefPtrGetterAddRefs, the class behind getter_AddRefs. Fix that up.
find . -name '*.cpp' -o -name '*.h' -o -name '*.idl' | \
xargs perl -p -i -e 's/nsRefPtrGetterAddRefs/RefPtrGetterAddRefs/g'
if [ -d .git ]; then
git mv mfbt/nsRefPtr.h mfbt/RefPtr.h
else
hg mv mfbt/nsRefPtr.h mfbt/RefPtr.h
fi
--HG--
rename : mfbt/nsRefPtr.h => mfbt/RefPtr.h
This commit was generated using the following script, executed at the
top level of a typical source code checkout.
# Don't modify select files in mfbt/ because it's not worth trying to
# tease out the dependencies currently.
#
# Don't modify anything in media/gmp-clearkey/0.1/ because those files
# use their own RefPtr, defined in their own RefCounted.h.
find . -name '*.cpp' -o -name '*.h' -o -name '*.mm' -o -name '*.idl'| \
grep -v 'mfbt/RefPtr.h' | \
grep -v 'mfbt/nsRefPtr.h' | \
grep -v 'mfbt/RefCounted.h' | \
grep -v 'media/gmp-clearkey/0.1/' | \
xargs perl -p -i -e '
s/mozilla::RefPtr/nsRefPtr/g; # handle declarations in headers
s/\bRefPtr</nsRefPtr</g; # handle local variables in functions
s#mozilla/RefPtr.h#mozilla/nsRefPtr.h#; # handle #includes
s#mfbt/RefPtr.h#mfbt/nsRefPtr.h#; # handle strange #includes
'
# |using mozilla::RefPtr;| is OK; |using nsRefPtr;| is invalid syntax.
find . -name '*.cpp' -o -name '*.mm' | xargs sed -i -e '/using nsRefPtr/d'
# RefPtr.h used |byRef| for dealing with COM-style outparams.
# nsRefPtr.h uses |getter_AddRefs|.
# Fixup that mismatch.
find . -name '*.cpp' -o -name '*.h'| \
xargs perl -p -i -e 's/byRef/getter_AddRefs/g'
This conversion was done with the script:
find . -name '*.cpp' -o -name '*.h' -o -name '*.mm' -o -name '*.idl' | \
egrep -v 'cairo-win32-refptr.h|RefPtr.h|TestRefPtr.cpp' | \
xargs sed -i -e 's/mozilla::TemporaryRef</already_AddRefed</g' \
-e 's/TemporaryRef</already_AddRefed</g'
Manual fixups were performed in the following instances:
- We handled mfbt/RefPtr.h manually so as to not convert TemporaryRef itself
into already_AddRefed.
- The following files had explicit Move() calls added to make up for the lack
of a copy constructor on already_AddRefed:
dom/base/ImageEncoder.cpp
dom/media/MediaTaskQueue.{h,cpp}
dom/media/webaudio/PannerNode.cpp
- A redundant overload for MediaTaskQueue::Dispatch was deleted.
- A few manual fixups were required in mfbt/tests/TestRefPtr.cpp.
- Comments, using declarations, and forward declarations relating to
TemporaryRef in dom/canvas/ and gfx/layers/ were changed to refer to
already_AddRefed.
Having this implicit conversion means that we can silently do extra
refcounting when it's completely unnecessary. It's also an obstacle to
making RefPtr more nsRefPtr-like, so let's get rid of it.
This allows to resolve weak symbols from some Android device's libc that
dlsym() won't. This is effectively an alternative fix to bug 791419, without
requiring wrapping symbols.
The new class is kind of like SystemElf, but using our linker's own symbol
resolution. This also adds some initialization from ELF program headers that
weren't done previously for self_elf, as well as registration as for CustomElf
instances.
In order to avoid adding more dlsym overhead than there already is, resolve
symbols directly in the library containing the linker. (GetSymbolPtr is
essentially free ; dlsym makes the system linker compule a ElfHash itself,
and that's quite expensive to do on all symbols)
This also paves the way for direct symbol resolution in all system libraries.
It's necessary to delay it because for the second part, we need to call dlopen,
and until recently bionic's linker dead-locked when using dlopen from a static
initializer.