Paul Floyd [Tue, 23 Dec 2025 20:22:17 +0000 (21:22 +0100)]
Darwin mmap: use ARG5 for VG_(di_notify_mmap)
The fix for bug280965 (a long long time ago) changed VG_(di_notify_mmap)
so that instead of opening a new fd to peek at the ELF header it would
reuse the same mmap fd with pread() to avoid resetting fcntl flags.
Except that the patch was applied to mmap in syswrap-generic but not
to the mmap in syswrap-darwin. So the problem remained there on Darwin
for 14 years.
Paul Floyd [Mon, 22 Dec 2025 09:19:30 +0000 (10:19 +0100)]
Helgrind regtest: add asserts to cond_timedwait_test.c
This test fails on Darwin, It looks like Darwin uses another mutex
and it's complaining that the condition mutex is not locked when
the unlock call is made. I made these changes when checking that the
testcase has the same behaviour on FreeBSD.
Paul Floyd [Fri, 19 Dec 2025 05:58:51 +0000 (06:58 +0100)]
Bug 258140 - Valgrind on OS X always reports some memory "still reachable"
Just needed one more suppression for the minimal test in the bugzilla
item. I expect that there will be a seemingly endless number of leaks
in Darwin (different libs, later versions).
Martin Cermak [Tue, 16 Dec 2025 08:52:29 +0000 (09:52 +0100)]
Wrap new lsm_list_modules syscall
The lsm_list_modules syscall returns a list of the active LSM ids. On
success this function returns the number of @ids array elements. This
value may be zero if there are no LSMs active. If @size is insufficient
to contain the return data -E2BIG is returned and @size is set to the
minimum required size. In all other cases a negative value indicating
the error is returned.
Paul Floyd [Sun, 14 Dec 2025 19:23:10 +0000 (20:23 +0100)]
Darwin DRD: redirs for semaphore functions
Like 2802b31ae644301ee641dcae20d1f84373615d54 for Helgrind.
Use the DRD script supported_sem_init in Helgrind tests that use
semaphores, it's clearer than just using tests/os_test
Paul Floyd [Sun, 14 Dec 2025 17:21:19 +0000 (18:21 +0100)]
Darwin helgrind: redirs for semaphore functions
The semaphore functions are not in the pthread library (similar to FreeBSD).
Change the redirs to use libsystem_kernel.dylib
Unfortunately this doesn't help much. Darwin has deprecated these functions
so the tests will compile (with a warning) but at runtime they all fail
with errors like ENOSYS and EBADF.
Paul Floyd [Sun, 14 Dec 2025 12:31:18 +0000 (13:31 +0100)]
Darwin fdleaks: fix handling of recvmsg cmsg
In the Darwin headers the macro __DARWIN_ALIGN32 is used for
the CMSG macros. We were using just ALIGN which looks like
it uses __DARWIN_ALIGN for 8 byte alignment. Using the
wrong alignment gate a wrong count of fds and check_cmsg_for_fds
was only calling ML_(record_fd_open_named) for one fd. The result
was fd errors when writing to a second fd that should have been
recorded.
Paul Floyd [Sun, 14 Dec 2025 09:05:04 +0000 (10:05 +0100)]
Darwin regtest: update none ioctl_moans
This testcase needs IOCTL_REQUEST_BASE to be set to some bogus
value with the platform dependent "no direction" flag set.
That flag wasn't being set for Darwin and so the testcase was
failing.
Andreas Arnez [Fri, 5 Dec 2025 17:01:22 +0000 (18:01 +0100)]
s390x: Simplify single multiplications
So far the single-multiplication instructions are implemented by taking
the low half of a widening multiply. This pattern is then recognized by
the instruction selector, which combines it to a single multiplication
again.
This is unnecessarily complicated, since single-multiply operations can be
expressed directly by Iop_Mul32 and Iop_Mul64 instead. So do this,
simplifying the code generation.
Also, since Iop_Mul32 and Iop_Mul64 haven't really been used before, the
instruction selector generates sub-optimal code for them. Fix that.
Paul Floyd [Thu, 11 Dec 2025 08:06:38 +0000 (09:06 +0100)]
Darwin regtest: update a few expecteds
One predated correct redirection of the C++ library operator new
One has a few line number changes plus I removed an assert that is firing
(asserts firing in tests is never a good thing). And the last one is just
line number changes.
Paul Floyd [Thu, 11 Dec 2025 08:03:08 +0000 (09:03 +0100)]
Darwin libcproc: fix for VG_(read_millisecond_timer)
This function had some peculiar workaround for how the result
was handled. That peculiarity seems to have gone away (some time
before Darwin 17). Now use 'normal' handling for the result for
Darwin 17+.
This was causing some DRD timed mutex testcases to fail.
Paul Floyd [Wed, 10 Dec 2025 07:58:25 +0000 (08:58 +0100)]
Darwin regtest: numerous bits of Darwin 18 cleanup
Add a few more ignore-fn options to the massif tests for the
extra leaks that are in Darwin 18.
Add some suppressions to a couple of memcheck tests that use
num-callers of 2 or 3. The shorter callsteack means some of the
many Darwin default suppressions don't match which was causing
these tests to fail.
Add more DRD suppressions for getaddrinfo. There seems to be a never
ending amount of conflicts in dyld.
Support pmaxsb, pminsb, pmaxuw, pminuw, pmaxud and pminud instructions
in guest_x86_toIR.c and host_x86_isel.c. Move test functions from
amd64/sse4-64.c to sse4-common.h and update none/tests/x86/sse4-x86.c
and sse4-x86.stdout.exp to test all min/max instructions.
Paul Floyd [Tue, 9 Dec 2025 06:33:49 +0000 (07:33 +0100)]
Darwin suppressions: merge old suppression files to darwin-legacy.supp
Merge files for Darwin 10 to 16 into a legacy file. Rename the Darwin 17
file as just darwin.supp. I'll add suppressions to that file, at least for
10.14 and 10.15. When I get to Darwin 20 and the dyld shared cache
we'll see, that might need a new set of suppression files.
Remove freebsd-helgrind-dbgrtld.supp. I thought that it would be useful
but it wasn't. When I merged the Darwin suppression files I wrote a little
duplication detection tool (only for exact duplicates). This found one
duplicate in solaris12.supp that I've removed.
Martin Cermak [Mon, 8 Dec 2025 12:20:25 +0000 (13:20 +0100)]
Wrap new lsm_get_self_attr and lsm_set_self_attr syscalls
Create a system call lsm_get_self_attr() to provide the security
module maintained attributes of the current process.
Create a system call lsm_set_self_attr() to set a security
module maintained attribute of the current process.
Historically these attributes have been exposed to user space via
entries in procfs under /proc/self/attr.
Declare lsm_get_self_attr and lsm_set_self_attr wrappers in
priv_syswrap-linux.h and hook it for {amd64,arm,arm64,mips64,\
ppc32,ppc64,riscv64,s390x,x86}-linux.
Paul Floyd [Mon, 8 Dec 2025 07:24:46 +0000 (08:24 +0100)]
Darwin suppressions: start cleaning up the suppression files
This commit moves DRD suppressions from darwin17.supp to darwin10-drd.supp.
It also removes some pthread lib suppressions. There's an obj suppression
at the top level for libsystem_pthread.dylib that should cover all that.
I want to avoid having 20+ Darwin suppression files (and rising with
every new version). I'd like to have something a bit more like FreeBSD,
with one suppression file for memcheck, one for DRD and one for Helgrind.
That might be a bit optimistic (especially when DSC support lands). So
what I'll do to begin with is to merge all of the Darwin <= 16 memcheck
files into one darwin-legacy.supp, split out a new darwin-helgrind.supp
file and rename darwin10-drd.supp to just darwin-drd.supp. That leaves
darwin9-drd.supp which I'll either rename or merge.
Paul Floyd [Sun, 7 Dec 2025 21:11:19 +0000 (22:11 +0100)]
Darwin: make fixup_macho_loadcmds.c less version dependent
Remove a load of DARWIN_VERS checks. It's a pointless maintenance
burden. Add suppport for SDK >= 10.4.6. First step to adding
macOS 10.14 Mojave support.
Paul Floyd [Sun, 7 Dec 2025 20:59:40 +0000 (21:59 +0100)]
FreeBSD and Darwin: clean up fake sigreturn syscall numbers
Firstly make them a bit more similar. Secondly, switch Darwin
from using __NR_MAXSYSCALL which changes with every Darwin version
to using 1000 like FreeBSD.
Change NEWS for macOS 10.13, it has now progressed beyond
'preliminary'.
Paul Floyd [Sun, 7 Dec 2025 08:52:04 +0000 (09:52 +0100)]
Darwin syscall: initial wrapper for mach_voucher_extract_attr_recipe
It doesn't do any checking yet. However, with this change running
TextEditor with --tool=none gets as far as drawing a white box on
the screen. Then it fails because there is a missing mach trap handler
for
Paul Floyd [Sun, 7 Dec 2025 07:50:32 +0000 (08:50 +0100)]
Darwin mach_msg_host: move assign_port_name to post
There's a GrP comment saying that assign_port_name should only be called on success.
And indeed when it is called in the PRE and it can't find the info for the port
then the code will assert.
Paul Floyd [Sun, 7 Dec 2025 06:54:23 +0000 (07:54 +0100)]
Darwin stacktraces: add a hack to prevent segfault when dereferencing bp
Can now turn a couple more helgrind tests back on.
Need to get to the bottom of the bp dereferencing issue. There is a test
already so it us supposed to be safe.
Paul Floyd [Sat, 6 Dec 2025 09:50:38 +0000 (10:50 +0100)]
FreeBSD client stack: add an assert to check the stringtable doesn't get overwritten
There's already an assert that the stringsize calculated matches the
stringsize writte. This adds a check that the pointer table area
does not overwrite the stringtable (that is, that the NULL pointer after the
last auxv entry pointer does not overwrite the first string [either the
interpreter or argv[0])
Paul Floyd [Fri, 5 Dec 2025 20:59:34 +0000 (21:59 +0100)]
Darwin: also clean up client stack creation
Make stringbase and strtab be based off clstack_end + 1
rather than clstack_end. clstack_end is the address of the
last byte of the stack, not one past the end of the stack.
That means that it is not word aligned. Adding 1 makes
the calculation word aligned.
Depending on the length of the string table with rounding
it could happen that the executable_path pointer would
overwrite the start of argv[0] in the string table.
I think that that there are supposed to be 0-16 bytes between
the end of the pointers and the string table (or is that
ELF?).
Paul Floyd [Fri, 5 Dec 2025 12:47:19 +0000 (13:47 +0100)]
Darwin: no need to turn off alignment assert in leak check
Mixed up 2 files with previous commit. This also contains
a missing proto and the previous commit fixes issues with badly formed
segments. Needs more investigation.
Florian Krohm [Wed, 3 Dec 2025 22:26:57 +0000 (22:26 +0000)]
Change the data type of libvex_Backend::emit.
Both the s390 and arm insn emitters need to know the host's hardware
capabilities. Today, these are provided by means of global variables
s390_host_hwcaps and arm_hwcaps. To eliminate that kludge the emit
function is changed. Instead of passing VexEndness we now pass a pointer
to VexArchInfo which provides both the endianess and hardware capabilities.
Those global variables will be removed in a followup patch.
Now that the stfle insn is available we can use it to tell uas how
many double words are needed to store all facility bits. Hence,
S390_NUM_FACILITY_DW can go.