Paul Floyd [Sat, 18 Nov 2023 12:02:30 +0000 (13:02 +0100)]
Bug 428364 - Signals inside io_uring_enter not handled
Patch provide by
Andrew Soutar kdebugtraq@andrewsoutar.com
Thanks!
I haven't added the reproducers as they require interrupting
the Valgrind process. Maybe the tests can be adapted to use
fork and have the parent kill the child running io_uring.
Paul Floyd [Sat, 18 Nov 2023 06:34:00 +0000 (07:34 +0100)]
Darwin: remove global variable used in aio_return
This was using a global variable to signify a read success
between pre- and post- aio_return. I don't think that this
was safe. There could be multiple calls to aio_return on
different threads. The global aiocpb_table is safe enough
as the pointers to control block are unique. However if two
threads manipulate was_a_successful_aio_read there's no way
of telling them apart. Rather than trying to make
was_a_successful_aio_read thread safe I did the same as on
FreeBSD and removed post- aio_return and now do everything
in pre- aio_return. No global variable, no thread hazard.
Adds syscall wrappers for aio_readv and aio_writev
Also rewrote the wrappers for aio_read, aio_write and aio_return
as they weren't correctly checking the async memory.
The code is similar th that of Darwin with one exception.
Darwin uses a global variable to communicate between the
pre- and post- aio_return wrappers. I don't think that is
safe when there are multiple aio reads in flight. Instead
I put everything in the pre- aio_return wrapper.
Paul Floyd [Sun, 5 Nov 2023 12:35:01 +0000 (13:35 +0100)]
Bug 476535 - Difference in allocation size for massif/tests/overloaded-new between clang++/libc++ and g++/libstdc++
In the end all I could do was filter the results.
libc++ and libstdc++ allocate different sizes of stuff for their own usr.
That means that when we get to allocating aligned blocks there is some
slop (up to the alignment size) that gets counted. And the amount of
that slop depends on the prior (internal) allocations.
Paul Floyd [Sun, 5 Nov 2023 09:59:43 +0000 (10:59 +0100)]
massif regtest: make verbose tests work on both 32bit and 64bit
Now alloc fns only have the appropriate new overloads which
means 32bit has unsigned size but 64bit has unsigned long.
Filter the unsigned long to unsigned.
Paul Floyd [Sun, 5 Nov 2023 09:06:56 +0000 (09:06 +0000)]
massif regtest: try adding a new expected for overloaded-new
Added a filter for "pool" that's used on Arch Linux
I'm not too confident that the new expected will work for
all architectures or even versions of libstdc++
Paul Floyd [Fri, 3 Nov 2023 20:46:13 +0000 (21:46 +0100)]
FreeBSD: remove some massif regtest expecteds
These were added mainly to have a cleanish slate with GCC on FreeBSD.
Unfortunately FreeBSD doesn't have dependent packages, so no debuginfo
for libstdc++ and the GCC port libstdc++ is stripped. That means that
the filters for internal GCC allocations don't work.
Since I hardly use GCC any more I'm getting rid of these expecteds.
Mark Wielaard [Thu, 2 Nov 2023 18:50:07 +0000 (19:50 +0100)]
Do not run memcheck/tests/x86-linux/scalar when root
The test expects some syscalls to fail, but when run as root
the user actually has permission to do silly things. So skip
the test if we notice we are root.
Mark Wielaard [Mon, 30 Oct 2023 22:30:06 +0000 (23:30 +0100)]
Disable memcpy overlap check and test on amd64 linux
Almost all newer distros have ifunc based memcpy/memmove glibc
implementation which cause false positives. Disable the overlap check
and test on these systems for now.
Andreas Arnez [Thu, 26 Oct 2023 12:00:52 +0000 (14:00 +0200)]
s390x regtest: Activate 128 bit SIMD tests for s390x in vbit-test
The vbit-test test case is currently configured to exclude s390x from the
platforms that execute the 128 bit SIMD irop tests. Since there's no
known issue with those, they can be activated on s390x as well.
Mark Wielaard [Thu, 26 Oct 2023 10:25:44 +0000 (12:25 +0200)]
vg_replace_malloc DELETE should not check size
The DELETE replacement functions check the size argument, but this
doesn't actually exist. Only the DELETE_SIZED replacement functions
get a size (and should check it).
On i386 (fedora gnu/linux) this causes the following failures:
Bug 432801 - Valgrind 3.16.1 reports a jump based on uninitialized
memory somehow related to clang and signals
Add support for precise computation of SIMD greater-than on
amd64 and x86.
This adds support for 64bit, 16bit, and 8bit to the existing 32bit
support.
The Iop_CmpGT64Ux2 is only supported on PPC32 and PPC64. The above
commit adds a more precise method, expensiveCmpGT, for setting the
vibits for the Iop_CmpGT64Ux2 Iop.
The expected results for the vbit test were not updated to the new more
precise results. This patch updates the expected results for the Iop.
Andreas Arnez [Thu, 19 Oct 2023 12:17:35 +0000 (14:17 +0200)]
s390x regtest: Fix memcheck tests for cu21 and cu42 with Clang
When compiled with Clang, the s390x memcheck tests for cu21 and cu42 yield
different line number information for the inline assemblies. Using gcc,
the inline assemblies do not receive updated line number information
because they are phrased as if they were data, not code. See also the
discussion in https://sourceware.org/bugzilla/show_bug.cgi?id=30206
Fix this by rephrasing the inline assemblies in cu21 and cu42 to normal
instructions, using the respective mnemonics. Since this affects the line
number information, adjust the expected test output files as well.
Andreas Arnez [Tue, 16 May 2023 18:29:33 +0000 (20:29 +0200)]
s390x: Fix memcheck false positives with certain comparisons
Consider this structure definition:
struct s {
unsigned b : 1;
unsigned c : 1;
} x;
Then certain compiler optimizations for a big-endian system may transform
the test
if (x.b || x.c)
...
into a comparison `> 0x3f' of the structure's first byte. Indeed, the
result of this comparison only depends on the two highest bits of the
byte. Thus, even if the lower bits are undefined, memcheck shouldn't
complain, but it does.
For certain cases this can be fixed. Do this by detecting comparisons
like this in the condition code helper for S390_CC_OP_UNSIGNED_COMPARE and
transforming them to a test for the selected bits instead. Add a small
test to verify the fix.
Paul Floyd [Mon, 16 Oct 2023 21:14:39 +0000 (23:14 +0200)]
Bug 475652 - Missing suppression for __wcsncpy_avx2 (strncpy-avx2.S:308)?
Adds a replace strmem wrapper for wcsncpy
Whilst doing this I noticed that
memcheck/tests/filter_memcheck
wasn't filtering memcheck/tests/freebsd/timingsafe properly.
That's because before filtering vg_replace_strmem.c and
vg_replace_malloc.c there's a check that the test exe filename
does not match the line. In this case the test exe was "timingsafe",
but line contained "timingsafe_memcmp", which matched leaving
vg_replace_strmem.c unfiltered. So I just renamed the testcase.
Paul Floyd [Mon, 9 Oct 2023 20:54:30 +0000 (22:54 +0200)]
DHAT: only handle own user reqs, ignore others
Make other tools consistent with this as well
(using memcheck as the model). Also refactored
the DRD user req names to make it clearer which
are Valgrind user reqs, which are DRD public
user reqs and which are DRD internal user reqs.
Paul Floyd [Fri, 29 Sep 2023 20:45:32 +0000 (22:45 +0200)]
coverity: lots of checks that unsigned are >= 0 which is always true
Also put back the isFF flag initialization (used for FreeBSD
non-fixed RO ELF segmentd) . I had intended to delete it but
in the end kept it for traces but had already deleted the init code.
Andreas Arnez [Fri, 29 Sep 2023 14:11:31 +0000 (16:11 +0200)]
s390x: Make z16 machine model known to Valgrind
The z16 machine model hasn't been fully introduced to Valgrind yet. Add
the missing support:
* Add z16 to the list of machine models in each of `libvex.h',
`tests/s390x_features.c', and `s390-check-opcodes.pl'.
* Starting with z16, the "store facility list extended" (STFLE)
instruction can write four instead of three words. Reflect this in the
STFLE-helper and in the `stfle' test case.
* Pass the new STFLE-bits unchanged, except for the vector-packed-decimal
facility, which is unsupported by Valgrind.
Andreas Arnez [Fri, 29 Sep 2023 14:11:31 +0000 (16:11 +0200)]
s390x regtest: Adjust vec2_float for Clang
The vec2_float test case doesn't compile with Clang, due to two issues:
* Clang doesn't correctly support `long double' vectors, as documented in
this GitHub issue: https://github.com/llvm/llvm-project/issues/61291
* llvm-as behaves differently from gas for the reserved field in the
`.insn vrr' directive. Instead of inserting the given data from the
first operand, it initializes the reserved field with zeros.
Andreas Arnez [Fri, 29 Sep 2023 14:11:31 +0000 (16:11 +0200)]
s390x regtest: Additional test case fixes
While fixing Clang compilation issues, I noticed additional problematic
constructs in s390x test cases, some of which are uncovered with compiler
optimization options such as -O3:
* `bic.c' relies on an int to be sign-extended when passing it in a
register to an inline-assembly.
* `cksm.c' clobbers a register in an inline assembly without specifying
that.
* `cu12.c' and `cu14.c' read the variable `pattern1' beyond its size.
* `mvst.c' modifies input operands in inline assemblies.
* `srnmb.c' relies on Valgrind's error message for a bad rounding mode to
point to the line number where get_rounding_mode() is called, not to the
function itself.
* `stfle.c' has a code path with an uninitialized variable that may cause
a compiler warning.
* `stmg.c' modifies r15 in an inline assembly and doesn't expect the
compiler to use r15 as the base address for an input operand.
* Various inline assemblies use the "=m" constraint even when the
instruction only supports more restricted address operands.
Fix all of the above. For consistency and clarity, avoid the "=m"
constraint even when it would be applicable.
Andreas Arnez [Fri, 29 Sep 2023 14:11:31 +0000 (16:11 +0200)]
s390x regtest: Refactor mul test cases for Clang support
The mul test case crashes when compiled with clang because of register
clashes in some inline assemblies. Obviously clang does not treat
register clobbers as "early clobbers", but sometimes uses the same
registers for inputs as well.
Rewrite the inline assemblies in mul.h to avoid this issue. Also, reduce
the code duplication in the process. Adjust mul_GE.c accordingly. Keep
the expected output for all mul test cases the same as before.
Andreas Arnez [Fri, 29 Sep 2023 14:11:31 +0000 (16:11 +0200)]
s390x regtest: Adjust op00 and op_exception tests for Clang
As part of fixing Bug 465782, adjust the op00 and op_exception tests.
The op00 test case doesn't compile with clang because the .hword assembler
directive is not known to llvm-as. When replacing it with .short, the
test case fails because clang/llvm-as generates different line number
information from gcc/gas. Using gcc, Valgrind's error message locates the
unrecognised instruction
at 0x........: main (op00.c:3)
whereas with clang, it shows:
at 0x........: main (op00.c:4)
The asm statement in op00.c is indeed located at line 4, so when compiling
with gcc, the indicated line number is just wrong. This is because gas
only updates line number information for instruction directives, not for
directives like `.short' as used here. See also the discussion in
https://sourceware.org/bugzilla/show_bug.cgi?id=30206
The same applies to the op_exception test case, where `.long' is used to
emit illegal instructions.
Make these test cases independent from this difference, by replacing the
`.short' and `.long' directives by equivalent `.insn' directives. Also
drop the unnecessary duplication of op00.stderr.exp.
Andreas Arnez [Fri, 29 Sep 2023 14:11:31 +0000 (16:11 +0200)]
s390x regtest: Adjust various test cases for Clang
There are various issues when compiling the s390x test cases with
Clang (see Bug 465782):
* Bugs in inline assemblies tolerated/obscured by GCC:
- "d" (or "r") constraint instead of "a" where a register from 1-15 is
needed
- mix-up between vector register and general register in inline assembly
- too unspecific constraints such as "m" or "g"
- use of input operand for output
- missing register clobber declaration
* Missing Clang features:
- "%R" (register) and "%O" (offset) modifiers with address constraints
in inline assemblies
- variable-length array within a struct
* Use of the -mlong-double-128 command line option -- has been redundant
for some time with GCC, and Clang only supports 128-bit long double
* Bad definition of _FPU_SETCW/_FPU_GETCW in <fpu_control.h>
(See https://sourceware.org/bugzilla/show_bug.cgi?id=30130)
* Use of pow() function without linking libm -- GCC optimizes the call
away, while Clang keeps a reference to pow()
Perform appropriate changes to the test cases such that they can be
compiled both with Clang and with GCC.
git://sourceware.org / valgrind.git / log