platform/upstream/glibc.git
20 months agocompare_strings.py : Add --gmean flag
Nisha Menon [Mon, 3 Apr 2023 14:11:05 +0000 (10:11 -0400)]
compare_strings.py : Add --gmean flag

To calculate geometric mean for string benchmark results.

Signed-off-by: Nisha Poyarekar <nisha.s.menon@gmail.com>
20 months agox86/dl-cacheinfo: remove unsused parameter from handle_amd
Andreas Schwab [Thu, 9 Feb 2023 13:56:21 +0000 (14:56 +0100)]
x86/dl-cacheinfo: remove unsused parameter from handle_amd

Also replace an unreachable assert with __builtin_unreachable.

20 months agopowerpc: Disable stack protector in early static initialization
Adhemerval Zanella [Mon, 3 Apr 2023 19:10:43 +0000 (16:10 -0300)]
powerpc: Disable stack protector in early static initialization

Similar to fb95c316382679c0826cc8399760977cd95f15c9, also disable
for string-ppc64.c (pulled on rltd as the default string
implementation).

Checked on powerpc64-linux-gnu.

20 months agonptl: Fix tst-cancel30 on sparc64
Adhemerval Zanella [Mon, 3 Apr 2023 17:18:14 +0000 (14:18 -0300)]
nptl: Fix tst-cancel30 on sparc64

As indicated by sparc kernel-features.h, even though sparc64 defines
__NR_pause,  it is not supported (ENOSYS).  Always use ppoll or the
64 bit time_t variant instead.

20 months agomath: Remove the error handling wrapper from fmod and fmodf
Adhemerval Zanella Netto [Mon, 20 Mar 2023 16:01:18 +0000 (13:01 -0300)]
math: Remove the error handling wrapper from fmod and fmodf

The error handling is moved to sysdeps/ieee754 version with no SVID
support.  The compatibility symbol versions still use the wrapper
with SVID error handling around the new code.  There is no new symbol
version nor compatibility code on !LIBM_SVID_COMPAT targets
(e.g. riscv).

The ia64 is unchanged, since it still uses the arch specific
__libm_error_region on its implementation.  For both i686 and m68k,
which provive arch specific implementation, wrappers are added so
no new symbol are added (which would require to change the
implementations).

It shows an small improvement, the results for fmod:

  Architecture     | Input           | master   | patch
  -----------------|-----------------|----------|--------
  x86_64 (Ryzen 9) | subnormals      | 12.5049  | 9.40992
  x86_64 (Ryzen 9) | normal          | 296.939  | 296.738
  x86_64 (Ryzen 9) | close-exponents | 16.0244  | 13.119
  aarch64 (N1)     | subnormal       | 6.81778  | 4.33313
  aarch64 (N1)     | normal          | 155.620  | 152.915
  aarch64 (N1)     | close-exponents | 8.21306  | 5.76138
  armhf (N1)       | subnormal       | 15.1083  | 14.5746
  armhf (N1)       | normal          | 244.833  | 241.738
  armhf (N1)       | close-exponents | 21.8182  | 22.457

Checked on x86_64-linux-gnu, i686-linux-gnu, and aarch64-linux-gnu.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
20 months agomath: Improve fmodf
Adhemerval Zanella Netto [Mon, 20 Mar 2023 16:01:17 +0000 (13:01 -0300)]
math: Improve fmodf

This uses a new algorithm similar to already proposed earlier [1].
With x = mx * 2^ex and y = my * 2^ey (mx, my, ex, ey being integers),
the simplest implementation is:

   mx * 2^ex == 2 * mx * 2^(ex - 1)

   while (ex > ey)
     {
       mx *= 2;
       --ex;
       mx %= my;
     }

With mx/my being mantissa of double floating pointer, on each step the
argument reduction can be improved 8 (which is sizeof of uint32_t minus
MANTISSA_WIDTH plus the signal bit):

   while (ex > ey)
     {
       mx << 8;
       ex -= 8;
       mx %= my;
     }  */

The implementation uses builtin clz and ctz, along with shifts to
convert hx/hy back to doubles.  Different than the original patch,
this path assume modulo/divide operation is slow, so use multiplication
with invert values.

I see the following performance improvements using fmod benchtests
(result only show the 'mean' result):

  Architecture     | Input           | master   | patch
  -----------------|-----------------|----------|--------
  x86_64 (Ryzen 9) | subnormals      | 17.2549  | 12.0318
  x86_64 (Ryzen 9) | normal          | 85.4096  | 49.9641
  x86_64 (Ryzen 9) | close-exponents | 19.1072  | 15.8224
  aarch64 (N1)     | subnormal       | 10.2182  | 6.81778
  aarch64 (N1)     | normal          | 60.0616  | 20.3667
  aarch64 (N1)     | close-exponents | 11.5256  | 8.39685

I also see similar improvements on arm-linux-gnueabihf when running on
the N1 aarch64 chips, where it a lot of soft-fp implementation (for
modulo, and multiplication):

  Architecture     | Input           | master   | patch
  -----------------|-----------------|----------|--------
  armhf (N1)       | subnormal       | 11.6662  | 10.8955
  armhf (N1)       | normal          | 69.2759  | 34.1524
  armhf (N1)       | close-exponents | 13.6472  | 18.2131

Instead of using the math_private.h definitions, I used the
math_config.h instead which is used on newer math implementations.

Co-authored-by: kirill <kirill.okhotnikov@gmail.com>
[1] https://sourceware.org/pipermail/libc-alpha/2020-November/119794.html
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
20 months agomath: Improve fmod
Adhemerval Zanella Netto [Mon, 20 Mar 2023 16:01:16 +0000 (13:01 -0300)]
math: Improve fmod

This uses a new algorithm similar to already proposed earlier [1].
With x = mx * 2^ex and y = my * 2^ey (mx, my, ex, ey being integers),
the simplest implementation is:

   mx * 2^ex == 2 * mx * 2^(ex - 1)

   while (ex > ey)
     {
       mx *= 2;
       --ex;
       mx %= my;
     }

With mx/my being mantissa of double floating pointer, on each step the
argument reduction can be improved 11 (which is sizeo of uint64_t minus
MANTISSA_WIDTH plus the signal bit):

   while (ex > ey)
     {
       mx << 11;
       ex -= 11;
       mx %= my;
     }  */

The implementation uses builtin clz and ctz, along with shifts to
convert hx/hy back to doubles.  Different than the original patch,
this path assume modulo/divide operation is slow, so use multiplication
with invert values.

I see the following performance improvements using fmod benchtests
(result only show the 'mean' result):

  Architecture     | Input           | master   | patch
  -----------------|-----------------|----------|--------
  x86_64 (Ryzen 9) | subnormals      | 19.1584  | 12.5049
  x86_64 (Ryzen 9) | normal          | 1016.51  | 296.939
  x86_64 (Ryzen 9) | close-exponents | 18.4428  | 16.0244
  aarch64 (N1)     | subnormal       | 11.153   | 6.81778
  aarch64 (N1)     | normal          | 528.649  | 155.62
  aarch64 (N1)     | close-exponents | 11.4517  | 8.21306

I also see similar improvements on arm-linux-gnueabihf when running on
the N1 aarch64 chips, where it a lot of soft-fp implementation (for
modulo, clz, ctz, and multiplication):

  Architecture     | Input           | master   | patch
  -----------------|-----------------|----------|--------
  armhf (N1)       | subnormal       | 15.908   | 15.1083
  armhf (N1)       | normal          | 837.525  | 244.833
  armhf (N1)       | close-exponents | 16.2111  | 21.8182

Instead of using the math_private.h definitions, I used the
math_config.h instead which is used on newer math implementations.

Co-authored-by: kirill <kirill.okhotnikov@gmail.com>
[1] https://sourceware.org/pipermail/libc-alpha/2020-November/119794.html
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
20 months agobenchtests: Add fmodf benchmark
Adhemerval Zanella Netto [Mon, 20 Mar 2023 16:01:15 +0000 (13:01 -0300)]
benchtests: Add fmodf benchmark

1. Subnormals: 128 inputs.
2. Normal numbers with large exponent difference (|x/y| > 2^8):
   1024 inputs between FLT_MIN and FLT_MAX;
3. Close exponents (ey >= -103 and |x/y| < 2^8): 1024 inputs with
   exponents between -10 and 10.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
20 months agobenchtests: Add fmod benchmark
Adhemerval Zanella Netto [Mon, 20 Mar 2023 16:01:14 +0000 (13:01 -0300)]
benchtests: Add fmod benchmark

Add three different dataset, from random floating point numbers:

1. Subnormals: 128 inputs.
2. Normal numbers with large exponent difference (|x/y| > 2^52):
   1024 inputs between DBL_MIN and DBL_MAX;
3. Close exponents (ey >= -907 and |x/y| < 2^52): 1024 inputs with
   exponents between -10 and 10.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
20 months agox86: Set FSGSBASE to active if enabled by kernel
H.J. Lu [Thu, 16 Mar 2023 00:42:54 +0000 (17:42 -0700)]
x86: Set FSGSBASE to active if enabled by kernel

Linux kernel uses AT_HWCAP2 to indicate if FSGSBASE instructions are
enabled.  If the HWCAP2_FSGSBASE bit in AT_HWCAP2 is set, FSGSBASE
instructions can be used in user space.  Define dl_check_hwcap2 to set
the FSGSBASE feature to active on Linux when the HWCAP2_FSGSBASE bit is
set.

Add a test to verify that FSGSBASE is active on current kernels.
NB: This test will fail if the kernel doesn't set the HWCAP2_FSGSBASE
bit in AT_HWCAP2 while fsgsbase shows up in /proc/cpuinfo.
Reviewed-by: Florian Weimer <fweimer@redhat.com>
20 months agox86_64: Fix asm constraints in feraiseexcept (bug 30305)
Florian Weimer [Mon, 3 Apr 2023 15:23:11 +0000 (17:23 +0200)]
x86_64: Fix asm constraints in feraiseexcept (bug 30305)

The divss instruction clobbers its first argument, and the constraints
need to reflect that.  Fortunately, with GCC 12, generated code does
not actually change, so there is no externally visible bug.

Suggested-by: Jakub Jelinek <jakub@redhat.com>
Reviewed-by: Noah Goldstein <goldstein.w.n@gmail.com>
20 months agomanual: Document __wur usage under _FORTIFY_SOURCE
Siddhesh Poyarekar [Mon, 3 Apr 2023 14:20:04 +0000 (10:20 -0400)]
manual: Document __wur usage under _FORTIFY_SOURCE

The __warn_unused_result__ attribute is only enabled when fortification
is enabled.  Mention that in the document.  The rationale for this is
essentially to mitigate against CWE-252:

[1] https://cwe.mitre.org/data/definitions/252.html

Signed-off-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Reviewed-by: Florian Weimer <fweimer@redhat.com>
20 months agohurd: Microoptimize _hurd_self_sigstate ()
Sergey Bugaev [Sun, 19 Mar 2023 15:10:14 +0000 (18:10 +0300)]
hurd: Microoptimize _hurd_self_sigstate ()

When THREAD_GETMEM is defined with inline assembly, the compiler may not
optimize away the two reads of _hurd_sigstate. Help it out a little bit
by only reading it once. This also makes for a slightly cleaner code.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-32-bugaevc@gmail.com>

20 months agohurd: Add vm_param.h for x86_64
Sergey Bugaev [Sun, 19 Mar 2023 15:10:12 +0000 (18:10 +0300)]
hurd: Add vm_param.h for x86_64

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-30-bugaevc@gmail.com>

20 months agohurd: Implement _hurd_longjmp_thread_state for x86_64
Sergey Bugaev [Sun, 19 Mar 2023 15:10:11 +0000 (18:10 +0300)]
hurd: Implement _hurd_longjmp_thread_state for x86_64

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-29-bugaevc@gmail.com>

20 months agohtl: Implement thread_set_pcsptp for x86_64
Sergey Bugaev [Sun, 19 Mar 2023 15:10:05 +0000 (18:10 +0300)]
htl: Implement thread_set_pcsptp for x86_64

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-23-bugaevc@gmail.com>

20 months agox86_64: Add rtld-stpncpy & rtld-strncpy
Sergey Bugaev [Sun, 19 Mar 2023 15:10:04 +0000 (18:10 +0300)]
x86_64: Add rtld-stpncpy & rtld-strncpy

Just like the other existing rtld-str* files, this provides rtld with
usable versions of stpncpy and strncpy.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-22-bugaevc@gmail.com>

20 months agohtl: Add tcb-offsets.sym for x86_64
Sergey Bugaev [Sun, 19 Mar 2023 15:10:03 +0000 (18:10 +0300)]
htl: Add tcb-offsets.sym for x86_64

The source code is the same as sysdeps/i386/htl/tcb-offsets.sym, but of
course the produced tcb-offsets.h will be different.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-21-bugaevc@gmail.com>

20 months agohurd: Move a couple of signal-related files to x86
Sergey Bugaev [Sun, 19 Mar 2023 15:10:02 +0000 (18:10 +0300)]
hurd: Move a couple of signal-related files to x86

These do not need any changes to be used on x86_64.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-20-bugaevc@gmail.com>

20 months agohurd: Use uintptr_t for register values in trampoline.c
Sergey Bugaev [Sun, 19 Mar 2023 15:09:58 +0000 (18:09 +0300)]
hurd: Use uintptr_t for register values in trampoline.c

This is more correct, if only because these fields are defined as having
the type unsigned int in the Mach headers, so casting them to a signed
int and then back is suboptimal.

Also, remove an extra reassignment of uesp -- this is another remnant of
the ecx kludge.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-16-bugaevc@gmail.com>

20 months agohurd: Move rtld-strncpy-c.c out of mach/hurd/
Sergey Bugaev [Sun, 19 Mar 2023 15:09:57 +0000 (18:09 +0300)]
hurd: Move rtld-strncpy-c.c out of mach/hurd/

There's nothing Mach- or Hurd-specific about it; any port that ends
up with rtld pulling in strncpy will need this.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-15-bugaevc@gmail.com>

20 months agohurd: More 64-bit integer casting fixes
Sergey Bugaev [Sun, 19 Mar 2023 15:09:55 +0000 (18:09 +0300)]
hurd: More 64-bit integer casting fixes

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-13-bugaevc@gmail.com>

20 months agomach, hurd: Drop __libc_lock_self0
Sergey Bugaev [Sun, 19 Mar 2023 15:09:54 +0000 (18:09 +0300)]
mach, hurd: Drop __libc_lock_self0

This was used for the value of libc-lock's owner when TLS is not yet set
up, so THREAD_SELF can not be used. Since the value need not be anything
specific -- it just has to be non-NULL -- we can just use a plain
constant, such as (void *) 1, for this. This avoids accessing the symbol
through GOT, and exporting it from libc.so in the first place.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-12-bugaevc@gmail.com>

20 months agostdio-common: Fix building when !IS_IN (libc)
Sergey Bugaev [Sun, 19 Mar 2023 15:09:53 +0000 (18:09 +0300)]
stdio-common: Fix building when !IS_IN (libc)

In this case, _itoa_word () is already defined inline in the header (see
sysdeps/generic/_itoa.h), and the second definition causes an error.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-11-bugaevc@gmail.com>

20 months agohurd: Fix _hurd_setup_sighandler () signature
Sergey Bugaev [Sun, 19 Mar 2023 15:09:52 +0000 (18:09 +0300)]
hurd: Fix _hurd_setup_sighandler () signature

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-10-bugaevc@gmail.com>

20 months agohurd: Disable O_TRUNC and FS_RETRY_MAGICAL in rtld
Sergey Bugaev [Sun, 19 Mar 2023 15:09:51 +0000 (18:09 +0300)]
hurd: Disable O_TRUNC and FS_RETRY_MAGICAL in rtld

hurd/lookup-retry.c is compiled into rtld, the dynamic linker/loader. To
avoid pulling in file_set_size, file_utimens, tty/ctty stuff, more
string/memory code (memmove, strncpy, strcpy), and more strtoul/itoa
code, compile out support for O_TRUNC and FS_RETRY_MAGICAL when building
hurd/lookup-retry.c for rtld. None of that functionality is useful to
rtld during startup anyway. Keep support for FS_RETRY_MAGICAL("/"),
since that does not pull in much, and is required for following absolute
symlinks.

The large number of extra code being pulled into rtld was noticed by
reviewing librtld.map & elf/librtld.os.map in the build tree.

It is worth noting that once libc.so is loaded, the real __open, __stat,
etc. replace the minimal versions used initially by rtld -- this is
especially important in the Hurd port, where the minimal rtld versions
do not use the dtable and just pass real Mach port names as fds. Thus,
once libc.so is loaded, rtld will gain access to the full
__hurd_file_name_lookup_retry () version, complete with FS_RETRY_MAGICAL
support, which is important in case the program decides to
dlopen ("/proc/self/fd/...") or some such.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-9-bugaevc@gmail.com>

20 months agohurd: Fix file name in #error
Sergey Bugaev [Sun, 19 Mar 2023 15:09:50 +0000 (18:09 +0300)]
hurd: Fix file name in #error

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-8-bugaevc@gmail.com>

20 months agohurd: Swap around two function calls
Sergey Bugaev [Sun, 19 Mar 2023 15:09:49 +0000 (18:09 +0300)]
hurd: Swap around two function calls

...to keep `sigexc' port initialization in one place, and match what the
comments say.

No functional change.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-7-bugaevc@gmail.com>

20 months agohurd: Remove __hurd_threadvar_stack_{offset,mask}
Sergey Bugaev [Sun, 19 Mar 2023 15:09:48 +0000 (18:09 +0300)]
hurd: Remove __hurd_threadvar_stack_{offset,mask}

Noone is or should be using __hurd_threadvar_stack_{offset,mask}, we
have proper TLS now. These two remaining variables are never set to
anything other than zero, so any code that would try to use them as
described would just dereference a zero pointer and crash. So remove
them entirely.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230319151017.531737-6-bugaevc@gmail.com>

20 months agohurd: Make exception subcode a long
Sergey Bugaev [Sun, 19 Mar 2023 15:09:47 +0000 (18:09 +0300)]
hurd: Make exception subcode a long

On EXC_BAD_ACCESS, exception subcode is used to pass the faulting memory
address, so it needs to be (at least) pointer-sized. Thus, make it into
a long. This matches the corresponding change in GNU Mach.
Message-Id: <20230319151017.531737-5-bugaevc@gmail.com>

20 months agotime: Fix strftime(3) API regarding nullability
Alejandro Colomar [Sun, 12 Mar 2023 00:08:10 +0000 (01:08 +0100)]
time: Fix strftime(3) API regarding nullability

strftime(3) doesn't accept null pointers in any of the parameters.

Cc: Paul Eggert <eggert@cs.ucla.edu>
Signed-off-by: Alejandro Colomar <alx@kernel.org>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoUpdate arm libm-tests-ulps
Adhemerval Zanella [Thu, 30 Mar 2023 13:44:33 +0000 (13:44 +0000)]
Update arm libm-tests-ulps

For the next test from cf7ffdd8a5f6da55397e10b3860062944312824c.

21 months agogetlogin_r: fix missing fallback if loginuid is unset (bug 30235)
Andreas Schwab [Wed, 15 Mar 2023 10:44:24 +0000 (11:44 +0100)]
getlogin_r: fix missing fallback if loginuid is unset (bug 30235)

When /proc/self/loginuid is not set, we should still fall back to using
the traditional utmp lookup, instead of failing right away.

21 months agomemalign: Support scanning for aligned chunks.
DJ Delorie [Wed, 29 Mar 2023 04:18:40 +0000 (00:18 -0400)]
memalign: Support scanning for aligned chunks.

This patch adds a chunk scanning algorithm to the _int_memalign code
path that reduces heap fragmentation by reusing already aligned chunks
instead of always looking for chunks of larger sizes and splitting
them.  The tcache macros are extended to allow removing a chunk from
the middle of the list.

The goal is to fix the pathological use cases where heaps grow
continuously in workloads that are heavy users of memalign.

Note that tst-memalign-2 checks for tcache operation, which
malloc-check bypasses.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agomalloc: Use C11 atomics on memusage
Adhemerval Zanella [Fri, 11 Mar 2022 16:53:11 +0000 (13:53 -0300)]
malloc: Use C11 atomics on memusage

Checked on x86_64-linux-gnu.

Reviewed-by: DJ Delorie <dj@redhat.com>
21 months agoRemove --enable-tunables configure option
Adhemerval Zanella Netto [Thu, 23 Mar 2023 13:13:51 +0000 (10:13 -0300)]
Remove --enable-tunables configure option

And make always supported.  The configure option was added on glibc 2.25
and some features require it (such as hwcap mask, huge pages support, and
lock elisition tuning).  It also simplifies the build permutations.

Changes from v1:
 * Remove glibc.rtld.dynamic_sort changes, it is orthogonal and needs
   more discussion.
 * Cleanup more code.
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
21 months agoRemove --disable-experimental-malloc option
Adhemerval Zanella [Tue, 28 Mar 2023 18:46:34 +0000 (15:46 -0300)]
Remove --disable-experimental-malloc option

It is the default since 2.26 and it has bitrotten over the years,
By using it multiple malloc tests fails:

  FAIL: malloc/tst-memalign-2
  FAIL: malloc/tst-memalign-2-malloc-hugetlb1
  FAIL: malloc/tst-memalign-2-malloc-hugetlb2
  FAIL: malloc/tst-memalign-2-mcheck
  FAIL: malloc/tst-mxfast-malloc-hugetlb1
  FAIL: malloc/tst-mxfast-malloc-hugetlb2
  FAIL: malloc/tst-tcfree2
  FAIL: malloc/tst-tcfree2-malloc-hugetlb1
  FAIL: malloc/tst-tcfree2-malloc-hugetlb2

Checked on x86_64-linux-gnu.

Reviewed-by: DJ Delorie <dj@redhat.com>
21 months agoAllow building with --disable-nscd again
Flavio Cruz [Tue, 28 Mar 2023 13:16:17 +0000 (10:16 -0300)]
Allow building with --disable-nscd again

The change 88677348b4de breaks the build with undefiend references to
the NSCD functions.

21 months agosystem: Add "--" after "-c" for sh (BZ #28519)
Joe Simmons-Talbott [Wed, 22 Mar 2023 18:04:30 +0000 (14:04 -0400)]
system: Add "--" after "-c" for sh (BZ #28519)

Prevent sh from interpreting a user string as shell options if it
starts with '-' or '+'.  Since the version of /bin/sh used for testing
system() is different from the full-fledged system /bin/sh add support
to it for handling "--" after "-c".  Add a testcase to ensure the
expected behavior.

Signed-off-by: Joe Simmons-Talbott <josimmon@redhat.com>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoposix: Fix some crashes in wordexp [BZ #18096]
Julian Squires [Wed, 22 Mar 2023 16:39:57 +0000 (14:09 -0230)]
posix: Fix some crashes in wordexp [BZ #18096]

Without these fixes, the first three included tests segfault (on a
NULL dereference); the fourth aborts on an assertion, which is itself
unnecessary.

Signed-off-by: Julian Squires <julian@cipht.net>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoLoongArch: ldconfig: Add comments for using EF_LARCH_OBJABI_V1
caiyinyu [Tue, 28 Mar 2023 01:19:53 +0000 (09:19 +0800)]
LoongArch: ldconfig: Add comments for using EF_LARCH_OBJABI_V1

We added Adhemerval Zanella's comment to explain the reason for
using EF_LARCH_OBJABI_V1.

21 months agoelf: Take into account ${sysconfdir} in elf/tst-ldconfig-p.sh
Romain Geissler [Sun, 26 Mar 2023 19:25:58 +0000 (19:25 +0000)]
elf: Take into account ${sysconfdir} in elf/tst-ldconfig-p.sh

Take into account ${sysconfdir} in elf/tst-ldconfig-p.sh.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoFix tst-glibc-hwcaps-prepend-cache with custom configure prefix value
Romain Geissler [Sun, 26 Mar 2023 20:00:16 +0000 (20:00 +0000)]
Fix tst-glibc-hwcaps-prepend-cache with custom configure prefix value

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoFix tst-ldconfig-ld_so_conf-update with custom configure prefix value
Romain Geissler [Sun, 26 Mar 2023 20:00:15 +0000 (20:00 +0000)]
Fix tst-ldconfig-ld_so_conf-update with custom configure prefix value

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agosupport: introduce support_sysconfdir_prefix
Romain Geissler [Sun, 26 Mar 2023 20:00:14 +0000 (20:00 +0000)]
support: introduce support_sysconfdir_prefix

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoRemove set-hooks.h from generic includes
Adhemerval Zanella Netto [Tue, 27 Dec 2022 21:11:45 +0000 (18:11 -0300)]
Remove set-hooks.h from generic includes

The hooks mechanism uses symbol sets for running lists of functions,
which requires either extra linker directives to provide any hardening
(such as RELRO) or additional code (such as pointer obfuscation via
mangling with random value).

Currently only hurd uses set-hooks.h so we remove it from the generic
includes.  The generic implementation uses direct function calls which
provide hardening and good code generation, observability and debugging
without the need for extra linking options or special code handling.

Reviewed-by: Carlos O'Donell <carlos@redhat.com>
21 months agoRemove --with-default-link configure option
Adhemerval Zanella Netto [Tue, 27 Dec 2022 21:11:44 +0000 (18:11 -0300)]
Remove --with-default-link configure option

Now that there is no need to use a special linker script to hardening
internal data structures, remove the --with-default-link configure
option and associated definitions.
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
21 months agolibio: Remove the usage of __libc_IO_vtables
Adhemerval Zanella Netto [Tue, 27 Dec 2022 21:11:43 +0000 (18:11 -0300)]
libio: Remove the usage of __libc_IO_vtables

Instead of using a special ELF section along with a linker script
directive to put the IO vtables within the RELRO section, the libio
vtables are all moved to an array marked as data.relro (so linker
will place in the RELRO segment without the need of extra directives).

To avoid static linking namespace issues and including all vtable
referenced objects, all required function pointers are set to weak alias.

Checked on x86_64-linux-gnu, i686-linux-gnu, and aarch64-linux-gnu.

Reviewed-by: Carlos O'Donell <carlos@redhat.com>
21 months agolibio: Do not autogenerate stdio_lim.h
Adhemerval Zanella Netto [Mon, 6 Mar 2023 18:53:55 +0000 (15:53 -0300)]
libio: Do not autogenerate stdio_lim.h

Instead define the required fields in system dependend files.  The only
system dependent definition is FILENAME_MAX, which should match POSIX
PATH_MAX, and it is obtained from either kernel UAPI or mach headers.
Currently set pre-defined value from current kernels.

It avoids a circular dependendy when including stdio.h in
gen-as-const-headers files.

Checked on x86_64-linux-gnu and i686-linux-gnu
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
21 months agoMove libc_freeres_ptrs and libc_subfreeres to hidden/weak functions
Adhemerval Zanella Netto [Tue, 27 Dec 2022 21:11:42 +0000 (18:11 -0300)]
Move libc_freeres_ptrs and libc_subfreeres to hidden/weak functions

They are both used by __libc_freeres to free all library malloc
allocated resources to help tooling like mtrace or valgrind with
memory leak tracking.

The current scheme uses assembly markers and linker script entries
to consolidate the free routine function pointers in the RELRO segment
and to be freed buffers in BSS.

This patch changes it to use specific free functions for
libc_freeres_ptrs buffers and call the function pointer array directly
with call_function_static_weak.

It allows the removal of both the internal macros and the linker
script sections.

Checked on x86_64-linux-gnu, i686-linux-gnu, and aarch64-linux-gnu.
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
21 months agobenchtests: Move libmvec benchtest inputs to benchtests directory
Joe Ramsay [Fri, 24 Mar 2023 12:10:10 +0000 (12:10 +0000)]
benchtests: Move libmvec benchtest inputs to benchtests directory

This allows other targets to use the same inputs for their own libmvec
microbenchmarks without having to duplicate them in their own
subdirectory.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
21 months agostdio-common: tests: don't double-define _FORTIFY_SOURCE
Sam James [Tue, 21 Feb 2023 09:27:26 +0000 (09:27 +0000)]
stdio-common: tests: don't double-define _FORTIFY_SOURCE

Exactly the same as 35bcb08eaa953c9b8bef6ab2486dc4361e1f26c0.

If using -D_FORITFY_SOURCE=3 (in my case, I've patched GCC to add
=3 instead of =2 (we've done =2 for years in Gentoo)), building
glibc tests will fail on tst-bz11319-fortify2 like:
```
<command-line>: error: "_FORTIFY_SOURCE" redefined [-Werror]
<built-in>: note: this is the location of the previous definition
cc1: all warnings being treated as errors
```

It's just because we're always setting -D_FORTIFY_SOURCE=2
rather than unsetting it first. If F_S is already 2, it's harmless,
but if it's another value (say, 1, or 3), the compiler will bawk.

(I'm not aware of a reason this couldn't be tested with =3,
but the toolchain support is limited for that (too new), and we want
to run the tests everywhere possible.)

As Siddhesh noted previously, we could implement some fallback
logic to determine the maximal F_S value supported by the toolchain,
which is a bit easier now that autoconf-archive has been updated for F_S=3
(https://github.com/autoconf-archive/autoconf-archive/pull/269), but let's
revisit this if it continues to crop up.

Signed-off-by: Sam James <sam@gentoo.org>
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
21 months agoLoongArch: ldconfig: Ignore EF_LARCH_OBJABI_V1 in shared objects
Xi Ruoyao [Sun, 26 Mar 2023 11:13:34 +0000 (19:13 +0800)]
LoongArch: ldconfig: Ignore EF_LARCH_OBJABI_V1 in shared objects

Binutils 2.40 sets EF_LARCH_OBJABI_V1 for shared objects:

    $ ld --version | head -n1
    GNU ld (GNU Binutils) 2.40
    $ echo 'int dummy;' > dummy.c
    $ cc dummy.c -shared
    $ readelf -h a.out | grep Flags
    Flags:                             0x43, DOUBLE-FLOAT, OBJ-v1

We need to ignore it in ldconfig or ldconfig will consider all shared
objects linked by Binutils 2.40 "unsupported".  Maybe we should stop
setting EF_LARCH_OBJABI_V1 for shared objects, but Binutils 2.40 is
already released and we cannot change it.

21 months ago_dl_map_object_from_fd: Remove unnecessary debugger notification in error path
Andreas Schwab [Thu, 23 Mar 2023 15:18:50 +0000 (16:18 +0100)]
_dl_map_object_from_fd: Remove unnecessary debugger notification in error path

After commit ed3ce71f5c ("elf: Move la_activity (LA_ACT_ADD) after
_dl_add_to_namespace_list() (BZ #28062)") it is no longer necessary to
reset the debugger state in the error case, since the debugger
notification only happens after no more errors can occur.

21 months agohppa: Drop 16-byte pthread lock alignment
John David Anglin [Sun, 26 Mar 2023 21:16:22 +0000 (21:16 +0000)]
hppa: Drop 16-byte pthread lock alignment

Linux threads were removed about 12 years ago and the current
nptl implementation only requires 4-byte alignment for pthread
locks.

The 16-byte alignment causes various issues. For example in
building ignition-msgs, we have:

/usr/include/google/protobuf/map.h:124:37: error: static assertion failed
  124 |   static_assert(alignof(value_type) <= 8, "");
      |                 ~~~~~~~~~~~~~~~~~~~~^~~~

This is caused by the 16-byte pthread lock alignment.

Signed-off-by: John David Anglin <dave.anglin@bell.net>
21 months agoMinor: don't call _dl_debug_update (which can have side effects) inside assert
Paul Pluzhnikov [Sat, 25 Mar 2023 21:27:01 +0000 (21:27 +0000)]
Minor: don't call _dl_debug_update (which can have side effects) inside assert

21 months agox86: Don't check PREFETCHWT1 in tst-cpu-features-cpuinfo.c
DJ Delorie [Fri, 10 Mar 2023 03:32:54 +0000 (22:32 -0500)]
x86: Don't check PREFETCHWT1 in tst-cpu-features-cpuinfo.c

Don't check PREFETCHWT1 against /proc/cpuinfo since kernel doesn't report
PREFETCHWT1 in /proc/cpuinfo.

Reviewed-by: Noah Goldstein <goldstein.w.n@gmail.com>
21 months agoDeclare wcstofN, wcstofNx for C2x
Joseph Myers [Tue, 14 Mar 2023 18:11:27 +0000 (18:11 +0000)]
Declare wcstofN, wcstofNx for C2x

WG14 accepted the changes in N3105 to define wcstofN and wcstofNx
functions for C2x.  Thus enable those for C2x (given also __GLIBC_USE
(IEC_60559_TYPES_EXT) and support for the relevant _FloatN / _FloatNx
type) rather than only for __USE_GNU.

Tested for x86_64.

21 months agoUpdate printf %b/%B C2x support
Joseph Myers [Tue, 14 Mar 2023 16:58:35 +0000 (16:58 +0000)]
Update printf %b/%B C2x support

WG14 recently accepted two additions to the printf/scanf %b/%B
support: there are now PRIb* and SCNb* macros in <inttypes.h>, and
printf %B is now an optional feature defined in normative text,
instead of recommended practice, with corresponding PRIB* macros that
can also be used to test whether that optional feature is supported.
See N3072 items 14 and 15 for details (those changes were accepted,
some other changes in that paper weren't).

Add the corresponding PRI* macros to glibc and update one place in the
manual referring to %B as recommended.  (SCNb* should naturally be
added at the same time as the corresponding scanf %b support.)

Tested for x86_64 and x86.

21 months agoARC: run child from the separate start block in __clone
Pavel Kozlov [Thu, 2 Mar 2023 16:10:19 +0000 (20:10 +0400)]
ARC: run child from the separate start block in __clone

For better debug experience use separate code block with extra
cfi_* directives to run child (same as in __clone3).

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoARC: Add the clone3 wrapper
Pavel Kozlov [Thu, 2 Mar 2023 16:10:18 +0000 (20:10 +0400)]
ARC: Add the clone3 wrapper

Use the clone3 wrapper on ARC. It doesn't care about stack alignment.
All callers should provide an aligned stack.
It follows the internal signature:

extern int clone3 (struct clone_args *__cl_args, size_t __size,
 int (*__func) (void *__arg), void *__arg);

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoLoongArch: Add get_rounding_mode.
caiyinyu [Wed, 8 Mar 2023 07:55:15 +0000 (15:55 +0800)]
LoongArch: Add get_rounding_mode.

21 months agoLoongArch: Add support for ldconfig.
caiyinyu [Tue, 7 Mar 2023 12:23:57 +0000 (20:23 +0800)]
LoongArch: Add support for ldconfig.

21 months agolinux: fix ntp_gettime abi break (BZ# 30156)
Kacper Piwiński [Thu, 9 Mar 2023 10:38:59 +0000 (11:38 +0100)]
linux: fix ntp_gettime abi break (BZ# 30156)

Between versions v2.11 and v2.12 struct ntptimeval got new fields.
That wasn't a problem because new function ntp_gettimex was created
(and made default) to support new struct.  Old ntp_gettime was not
using new fields so it was safe to call with old struct
definition.  Then commits 5613afe9e3dff and b6ad64b907a (added for
64 bit time_t support), ntp_gettime start setting new fields.

Sets fields manually to maintain compatibility with v2.11 struct
definition.

Resolves #30156

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoelf: Add missing dependency between resolvfail and testobj1.so
Arsen Arsenović [Tue, 7 Mar 2023 10:29:35 +0000 (11:29 +0100)]
elf: Add missing dependency between resolvfail and testobj1.so

It was possible to run this test individually and have it fail because
it can't find testobj1.so.  This patch adds that dependency, to prevent
such issues.

Reviewed-by: Florian Weimer <fweimer@redhat.com>
21 months agoelf: Add -z lazy to some more tests
Arsen Arsenović [Tue, 7 Mar 2023 10:29:34 +0000 (11:29 +0100)]
elf: Add -z lazy to some more tests

Some toolchains, such as that used on Gentoo Hardened, set -z now out of
the box.  This trips up a couple of tests.

Reviewed-by: Florian Weimer <fweimer@redhat.com>
21 months agoBenchtests: Remove simple_str(r)chr
Wilco Dijkstra [Fri, 3 Mar 2023 16:10:55 +0000 (16:10 +0000)]
Benchtests: Remove simple_str(r)chr

Instead of benchmarking slow byte oriented loops, include the optimized generic
strchr and strrchr implementation.  Adjust iteration count to reduce benchmark
time.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoBenchtests: Remove simple_str(n)casecmp
Wilco Dijkstra [Fri, 3 Mar 2023 13:21:05 +0000 (13:21 +0000)]
Benchtests: Remove simple_str(n)casecmp

Remove the slow byte oriented loops.  Adjust iteration count to reduce
benchmark time.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoBenchtests: Remove simple_memcmp
Wilco Dijkstra [Fri, 3 Mar 2023 13:04:00 +0000 (13:04 +0000)]
Benchtests: Remove simple_memcmp

Remove the slow byte oriented simple_memcmp.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoBenchtests: Remove simple_strcspn/strpbrk/strsep
Wilco Dijkstra [Fri, 3 Mar 2023 13:03:19 +0000 (13:03 +0000)]
Benchtests: Remove simple_strcspn/strpbrk/strsep

Remove simple_strcspn/strpbrk/strsep which are significantly slower than the
generic implementations.  Also remove oldstrsep and oldstrtok since they are
practically identical to the generic implementation.  Adjust iteration count
to reduce benchmark time.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoBenchtests: Remove memchr_strnlen
Wilco Dijkstra [Fri, 3 Mar 2023 12:57:49 +0000 (12:57 +0000)]
Benchtests: Remove memchr_strnlen

Remove memchr_strnlen since it is now the same as generic_strnlen.  Adjust
iteration count to reduce benchmark time.  Keep memchr_strlen since the
generic strlen does not use memchr.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoBenchtests: Remove simple_mem(r)chr
Wilco Dijkstra [Fri, 3 Mar 2023 12:54:45 +0000 (12:54 +0000)]
Benchtests: Remove simple_mem(r)chr

Instead of benchmarking slow byte oriented loops, include the optimized
generic memchr/memrchr implementation.  Adjust iteration count to reduce
benchmark time.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoBenchtests: Remove simple_strcpy_chk
Wilco Dijkstra [Fri, 3 Mar 2023 12:44:46 +0000 (12:44 +0000)]
Benchtests: Remove simple_strcpy_chk

Remove the slow byte oriented simple_strcpy_chk and simple_stpcpy_chk.
Adjust iteration count to increase benchmark time.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoBenchtests: Remove simple_str(n)cmp
Wilco Dijkstra [Fri, 3 Mar 2023 12:40:22 +0000 (12:40 +0000)]
Benchtests: Remove simple_str(n)cmp

Instead of benchmarking slow byte oriented loops, include the optimized generic
strcmp/strncmp implementation.  Adjust iteration count to reduce benchmark time.

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agomalloc: Fix transposed arguments in sysmalloc_mmap_fallback call
Robert Morell [Tue, 7 Mar 2023 13:14:45 +0000 (10:14 -0300)]
malloc: Fix transposed arguments in sysmalloc_mmap_fallback call

git commit 0849eed45daa ("malloc: Move MORECORE fallback mmap to
sysmalloc_mmap_fallback") moved a block of code from sysmalloc to a
new helper function sysmalloc_mmap_fallback(), but 'pagesize' is used
for the 'minsize' argument and 'MMAP_AS_MORECORE_SIZE' for the
'pagesize' argument.

Fixes: 0849eed45daa ("malloc: Move MORECORE fallback mmap to sysmalloc_mmap_fallback")
Signed-off-by: Robert Morell <rmorell@nvidia.com>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agort: fix shm_open not set ENAMETOOLONG when name exceeds {_POSIX_PATH_MAX}
abushwang [Tue, 7 Mar 2023 12:16:20 +0000 (20:16 +0800)]
rt: fix shm_open not set ENAMETOOLONG when name exceeds {_POSIX_PATH_MAX}

according to man-pages-posix-2017, shm_open() function may fail if the length
of the name argument exceeds {_POSIX_PATH_MAX} and set ENAMETOOLONG

Signed-off-by: abushwang <abushwangs@gmail.com>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agoposix: Ensure the initial signal disposition for tst-spawn7
Adhemerval Zanella Netto [Tue, 7 Mar 2023 16:31:52 +0000 (13:31 -0300)]
posix: Ensure the initial signal disposition for tst-spawn7

To avoid possible failure if any parent set any initial signal
disposition as SIG_IGN (for instance if the testcase is issued
with nohup).

Checked on x86_64-linux-gnu.
Tested-by: Florian Weimer <fweimer@redhat.com>
Reviewed-by: Florian Weimer <fweimer@redhat.com>
21 months agohurd: fix build of tst-system.c
Adam Yi [Wed, 8 Mar 2023 08:11:47 +0000 (03:11 -0500)]
hurd: fix build of tst-system.c

We made tst-system.c depend on pthread, but that requires linking with
$(shared-thread-library). It does not fail under Linux because the
variable expands to nothing under Linux, but it fails for Hurd.

I tested verified via cross-compiling that "make check" now works
for Hurd.

Signed-off-by: Adam Yi <ayi@janestreet.com>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agox86: Fix bug about glibc.cpu.hwcaps.
caiyinyu [Tue, 28 Feb 2023 08:21:41 +0000 (16:21 +0800)]
x86: Fix bug about glibc.cpu.hwcaps.

Recorded in [BZ #30183]:

1. export GLIBC_TUNABLES=glibc.cpu.hwcaps=-AVX512
2. Add  _dl_printf("p -- %s\n", p); just before switch(nl) in
   sysdeps/x86/cpu-tunables.c
3. compiled and run ./testrun.sh /usr/bin/ls
you will get:

p -- -AVX512
p -- LC_ADDRESS=en_US.UTF-8
p -- LC_NUMERIC=C
...

The function, TUNABLE_CALLBACK (set_hwcaps)
(tunable_val_t *valp), checks far more than it should and it
should stop at end of "-AVX512".

21 months agoposix: Fix system blocks SIGCHLD erroneously [BZ #30163]
Adam Yi [Tue, 7 Mar 2023 12:30:02 +0000 (07:30 -0500)]
posix: Fix system blocks SIGCHLD erroneously [BZ #30163]

Fix bug that SIGCHLD is erroneously blocked forever in the following
scenario:

1. Thread A calls system but hasn't returned yet
2. Thread B calls another system but returns

SIGCHLD would be blocked forever in thread B after its system() returns,
even after the system() in thread A returns.

Although POSIX does not require, glibc system implementation aims to be
thread and cancellation safe. This bug was introduced in
5fb7fc96350575c9adb1316833e48ca11553be49 when we moved reverting signal
mask to happen when the last concurrently running system returns,
despite that signal mask is per thread. This commit reverts this logic
and adds a test.

Signed-off-by: Adam Yi <ayi@janestreet.com>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
21 months agogshadow: Matching sgetsgent, sgetsgent_r ERANGE handling (bug 30151)
Florian Weimer [Tue, 21 Feb 2023 08:20:28 +0000 (09:20 +0100)]
gshadow: Matching sgetsgent, sgetsgent_r ERANGE handling (bug 30151)

Before this change, sgetsgent_r did not set errno to ERANGE, but
sgetsgent only check errno, not the return value from sgetsgent_r.
Consequently, sgetsgent did not detect any error, and reported
success to the caller, without initializing the struct sgrp object
whose address was returned.

This commit changes sgetsgent_r to set errno as well.  This avoids
similar issues in applications which only change errno.

Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
21 months agoUpdate kernel version to 6.2 in header constant tests
Joseph Myers [Mon, 6 Mar 2023 15:13:22 +0000 (15:13 +0000)]
Update kernel version to 6.2 in header constant tests

This patch updates the kernel version in the tests tst-mman-consts.py,
tst-mount-consts.py and tst-pidfd-consts.py to 6.2.  (There are no new
constants covered by these tests in 6.2 that need any other header
changes, and the removed MAP_VARIABLE for hppa was addressed
separately.)

Tested with build-many-glibcs.py.

21 months agoarm: Remove __builtin_arm_uqsub8 usage on string-fza.h
Adhemerval Zanella Netto [Tue, 28 Feb 2023 18:23:25 +0000 (15:23 -0300)]
arm: Remove __builtin_arm_uqsub8 usage on string-fza.h

The __builtin_arm_uqsub8 is an internal GCC builtin which might change
in future release (the correct way is to include "arm_acle.h" and use
__uqsub8 ()).  Since not all compilers support it, just use the
inline assembler instead.

Checked on armv7a-linux-gnueabihf.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
21 months agoalpha: Remove strncmp optimization
Adhemerval Zanella Netto [Tue, 28 Feb 2023 17:24:00 +0000 (14:24 -0300)]
alpha: Remove strncmp optimization

The generic implementation already cover word access along with
cmpbge for both aligned and unaligned, so use it instead.

Checked qemu static for alpha-linux-gnu.

21 months agopowerpc: Remove powerpc64 strncmp variants
Adhemerval Zanella Netto [Tue, 28 Feb 2023 17:23:59 +0000 (14:23 -0300)]
powerpc: Remove powerpc64 strncmp variants

The default, and power7 implementation just adds word aligned
access when inputs have the same aligment.  The unaligned case
is still done by byte operations.

This is already covered by the generic implementation, which also add
the unaligned input optimization.

Checked on powerpc64-linux-gnu built without multi-arch for powerpc64,
power7, power8, and power9 (build for le).
Reviewed-by: Rajalakshmi Srinivasaraghavan <rajis@linux.ibm.com>
21 months agopowerpc: Remove strncmp variants
Adhemerval Zanella Netto [Tue, 28 Feb 2023 17:23:58 +0000 (14:23 -0300)]
powerpc: Remove strncmp variants

The default, power4, and power7 implementation just adds word aligned
access when inputs have the same aligment.  The unaligned case
is still done by byte operations.

This is already covered by the generic implementation, which also add
the unaligned input optimization.

Checked on powerpc-linux-gnu built without multi-arch for powerpc,
power4, and power7.
Reviewed-by: Rajalakshmi Srinivasaraghavan <rajis@linux.ibm.com>
21 months agoC2x scanf binary constant handling
Joseph Myers [Thu, 2 Mar 2023 19:10:37 +0000 (19:10 +0000)]
C2x scanf binary constant handling

C2x adds binary integer constants starting with 0b or 0B, and supports
those constants for the %i scanf format (in addition to the %b format,
which isn't yet implemented for scanf in glibc).  Implement that scanf
support for glibc.

As with the strtol support, this is incompatible with previous C
standard versions, in that such an input string starting with 0b or 0B
was previously required to be parsed as 0 (with the rest of the input
potentially matching subsequent parts of the scanf format string).
Thus this patch adds 12 new __isoc23_* functions per long double
format (12, 24 or 36 depending on how many long double formats the
glibc configuration supports), with appropriate header redirection
support (generally very closely following that for the __isoc99_*
scanf functions - note that __GLIBC_USE (DEPRECATED_SCANF) takes
precedence over __GLIBC_USE (C2X_STRTOL), so the case of GNU
extensions to C89 continues to get old-style GNU %a and does not get
this new feature).  The function names would remain as __isoc23_* even
if C2x ends up published in 2024 rather than 2023.

When scanf %b support is added, I think it will be appropriate for all
versions of scanf to follow C2x rules for inputs to the %b format
(given that there are no compatibility concerns for a new format).

Tested for x86_64 (full glibc testsuite).  The first version was also
tested for powerpc (32-bit) and powerpc64le (stdio-common/ and wcsmbs/
tests), and with build-many-glibcs.py.

21 months agoFix stringop-overflow warning in test-strncat.
Stefan Liebler [Tue, 28 Feb 2023 12:48:34 +0000 (13:48 +0100)]
Fix stringop-overflow warning in test-strncat.

Starting with commit
b2c474f8de4c92bfe7435853a96805ec32d68dfa
"x86: Fix strncat-avx2.S reading past length [BZ #30065]"

Building on s390 the test fails due warnings like:

In function ‘do_one_test’,
    inlined from ‘do_overflow_tests’ at test-strncat.c:175:7:
test-strncat.c:31:18: error: ‘strnlen’ specified bound [42949665464294967295] exceeds maximum object size 2147483647 [-Werror=stringop-overflow=]
   31 | # define STRNLEN strnlen
      |                  ^
test-strncat.c:83:16: note: in expansion of macro ‘STRNLEN’
   83 |   size_t len = STRNLEN (src, n);
|                ^~~~~~~

In all werror cases, the call to strnlen (.., SIZE_MAX) is inlined.
Therefore this patch just marks the do_one_test function as noinline.

Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
21 months agonis: Fix stringop-truncation warning with -O3 in nis_local_host.
Stefan Liebler [Tue, 28 Feb 2023 12:37:35 +0000 (13:37 +0100)]
nis: Fix stringop-truncation warning with -O3 in nis_local_host.

When building with -O3 on s390x/x86_64, I get this stringop-truncation warning
which leads to a build fail:

In function ‘nis_local_host’,
    inlined from ‘nis_local_host’ at nis_local_names.c:147:1:
nis_local_names.c:171:11: error: ‘strncpy’ output may be truncated copying between 0 and 1023 bytes from a string of length 1024 [-Werror=stringop-truncation]
171 |           strncpy (cp, nis_local_directory (), NIS_MAXNAMELEN - len -1);
       |           ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

We can just ignore this warning as the hostname + '.' + directory-name + '\0' always fits
in __nishostname with length of (NIS_MAXNAMELEN + 1) as there is the runtime check above.
Furthermore as we already know the length of the directory-name, we can also just use
memcpy to copy the directory-name inclusive the NUL-termination.

Note: This werror was introduced with commit
32c7acd46401530fdbd4e98508c9baaa705f8b53
"Replace rawmemchr (s, '\0') with strchr"

Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
21 months agosupport: use 64-bit time_t (bug 30111)
Andreas Schwab [Tue, 28 Feb 2023 13:37:18 +0000 (10:37 -0300)]
support: use 64-bit time_t (bug 30111)

Ensure to use 64-bit time_t in the test infrastructure.

21 months agoLoongArch: Update libm-test-ulps.
caiyinyu [Tue, 28 Feb 2023 12:23:01 +0000 (20:23 +0800)]
LoongArch: Update libm-test-ulps.

21 months agoLoongArch: Further refine the condition to enable static PIE
Xi Ruoyao [Mon, 27 Feb 2023 11:08:09 +0000 (19:08 +0800)]
LoongArch: Further refine the condition to enable static PIE

Before GCC r13-2728, it would produce a normal dynamic-linked executable
with -static-pie.  I mistakely believed it would produce a static-linked
executable, so failed to detect the breakage.  Then with Binutils 2.40
and (vanilla) GCC 12, libc_cv_static_pie_on_loongarch is mistakenly
enabled and cause a building failure with "undefined reference to
_DYNAMIC".

Fix the issue by disabling static PIE if -static-pie creates something
with a INTERP header.

21 months agohurd: Fix some broken indentation
Sergey Bugaev [Wed, 1 Mar 2023 16:23:55 +0000 (19:23 +0300)]
hurd: Fix some broken indentation

Also, fix a couple of typos. No functional change.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230301162355.426887-2-bugaevc@gmail.com>

21 months agohurd: Remove the ecx kludge
Sergey Bugaev [Wed, 1 Mar 2023 16:23:54 +0000 (19:23 +0300)]
hurd: Remove the ecx kludge

"We don't need it any more"

The INTR_MSG_TRAP macro in intr-msg.h used to play little trick with
the stack pointer: it would temporarily save the "real" stack pointer
into ecx, while setting esp to point to just before the message buffer,
and then invoke the mach_msg trap. This way, INTR_MSG_TRAP reused the
on-stack arguments laid out for the containing call of
_hurd_intr_rpc_mach_msg (), passing them to the mach_msg trap directly.

This, however, required special support in hurdsig.c and trampoline.c,
since they now had to recognize when a thread is inside the piece of
code where esp doesn't point to the real tip of the stack, and handle
this situation specially.

Commit 1d20f33ff4fb634310f27493b7b87d0b20f4a0b0 has removed the actual
temporary change of esp by actually re-pushing mach_msg arguments onto
the stack, and popping them back at end. It did not, however, deal with
the rest of "the ecx kludge" code in other files, resulting in potential
crashes if a signal arrives in the middle of pushing arguments onto the
stack.

Fix that by removing "the ecx kludge". Instead, when we want a thread
to skip the RPC, but cannot make just make it jump to after the trap
since it's not done adjusting the stack yet, set the SYSRETURN register
to MACH_SEND_INTERRUPTED (as we do anyway), and rely on the thread
itself for detecting this case and skipping the RPC.

This simplifies things somewhat and paves the way for a future x86_64
port of this code.

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230301162355.426887-1-bugaevc@gmail.com>

22 months agoAdd AArch64 HWCAP2 values from Linux 6.2 to bits/hwcap.h
Joseph Myers [Tue, 28 Feb 2023 15:57:40 +0000 (15:57 +0000)]
Add AArch64 HWCAP2 values from Linux 6.2 to bits/hwcap.h

Linux 6.2 adds three new AArch64 HWCAP2 values; add them to glibc's
AArch64 bits/hwcap.h.

Tested with build-many-glibcs.py for aarch64-linux-gnu.

22 months agocrypt: Remove invalid end of page test badsalttest
Adhemerval Zanella [Mon, 27 Feb 2023 12:34:07 +0000 (09:34 -0300)]
crypt: Remove invalid end of page test badsalttest

The input argument passes an invalid string without a NUL terminator
on crypt settings inputs, which might lead to invalid OOB on strncmp.

Implementations only assume there is a NUL terminator if the string is
shorter than the specified size, so strings don't need to always be NUL
terminated (stratcliff.c has tests for this).

Also adapt the code to use libsupport.

Checked on arm-linux-gnuabihf.

22 months agoS390: Fix _FPU_SETCW/GETCW when compiling with Clang [BZ #30130]
Andreas Arnez [Tue, 28 Feb 2023 12:48:06 +0000 (13:48 +0100)]
S390: Fix _FPU_SETCW/GETCW when compiling with Clang [BZ #30130]

The _FPU_SETCW and _FPU_GETCW macros are defined with inline assemblies.
They use the sfpc and efpc instructions, respectively.  But both contain
a spurious second operand that leads to a compile error with Clang.
Removing this operand works both with gcc/gas (since binutils 2.18) as
well as with clang/llvm.

22 months agos390x: Regenerate ULPs.
Stefan Liebler [Tue, 28 Feb 2023 09:38:25 +0000 (10:38 +0100)]
s390x: Regenerate ULPs.

Needed due to recent commits:
- "added pair of inputs for hypotf in binary32"
commit ID cf7ffdd8a5f6da55397e10b3860062944312824c

- "update auto-libm-test-out-hypot"
commit ID 3efbf11fdf15ed991d2c41743921c524a867e145

22 months agoAdd Arm HWCAP values from Linux 6.2 to bits/hwcap.h
Joseph Myers [Tue, 28 Feb 2023 00:07:59 +0000 (00:07 +0000)]
Add Arm HWCAP values from Linux 6.2 to bits/hwcap.h

Linux 6.2 adds six new Arm HWCAP values and two new HWCAP2 values; add
them to glibc's Arm bits/hwcap.h, with corresponding dl-procinfo.c and
dl-procinfo.h updates.

Tested with build-many-glibcs.py for arm-linux-gnueabi.

22 months agohtl: Add pthreadtypes-arch.h for x86_64
Sergey Bugaev [Tue, 21 Feb 2023 21:19:32 +0000 (00:19 +0300)]
htl: Add pthreadtypes-arch.h for x86_64

Signed-off-by: Sergey Bugaev <bugaevc@gmail.com>
Message-Id: <20230221211932.296459-5-bugaevc@gmail.com>