- 23 Aug, 2023 40 commits
-
-
Zhangjin Wu authored
Silence the following warnings reported by the new -Wall -Wextra options with pure assembly code. In file included from sysroot/powerpc/include/stdio.h:13, from nolibc-test.c:13: sysroot/powerpc/include/arch.h: In function '_start': sysroot/powerpc/include/arch.h:192:32: warning: unused variable 'r2' [-Wunused-variable] 192 | register volatile long r2 __asm__ ("r2") = (void *)&TOC - (void *)_start; | ^~ sysroot/powerpc/include/arch.h:187:97: warning: optimization may eliminate reads and/or writes to register variables [-Wvolatile-register-var] 187 | void __attribute__((weak, noreturn, optimize("Os", "omit-frame-pointer"))) __no_stack_protector _start(void) | ^~~~~~ Since only elfv2 ABI requires to save the TOC/GOT pointer to r2 register, when using elfv1 ABI, the old C code is simply ignored by the compiler, but the compiler can not ignore the inline assembly code and will introduce build failure or running segfaults. So, let's further only add the new assembly code for elfv2 ABI with the checking of _CALL_ELF == 2. Link: https://refspecs.linuxfoundation.org/ELF/ppc64/PPC-elf64abi.pdf Link: https://www.llvm.org/devmtg/2014-04/PDFs/Talks/Euro-LLVM-2014-Weigand.pdfSigned-off-by:Zhangjin Wu <falcon@tinylab.org> Signed-off-by:
Willy Tarreau <w@1wt.eu>
-
Zhangjin Wu authored
libc-test is mainly added to compare the behavior of nolibc to the system libc, it is meaningless and error-prone with cross compiling. Let's use HOSTCC instead of CC to avoid wrongly use cross compiler when CROSS_COMPILE is passed or customized. Signed-off-by:
-
Zhangjin Wu authored
This allows to generate smaller text/data/dec size. As the _start_c() function added by crt.h, __stack_chk_init() is called from _start_c() instead of the assembly _start. So, it is able to mark it with static now. Reviewed-by:
Thomas Weißschuh <linux@weissschuh.net> Signed-off-by:
-
Zhangjin Wu authored
After the tests finish, it is valuable to report and summarize with existing test log. This avoid rerun or run the tests again when not necessary. Reviewed-by:
-
Zhangjin Wu authored
Kernel uses ARCH=powerpc for both 32-bit and 64-bit PowerPC, here adds a ppc64 variant for big endian 64-bit PowerPC, users can pass XARCH=ppc64 to test it. The powernv machine of qemu-system-ppc64 is used with powernv_be_defconfig. As the document [1] shows: PowerNV (as Non-Virtualized) is the “bare metal” platform using the OPAL firmware. It runs Linux on IBM and OpenPOWER systems and it can be used as an hypervisor OS, running KVM guests, or simply as a host OS. Notes, - differs from little endian 64-bit PowerPC, vmlinux is used instead of zImage, because big endian zImage [2] only boot on qemu with x-vof=on (added from qemu v7.0) and a fixup patch [3] for qemu v7.0.51: - since the VSX support may be disabled in kernel side, to avoid "illegal instruction" errors due to missing VSX kernel support, let's simply let compiler not generate vector/scalar (VSX) instructions via the '-mno-vsx' option. - as 'man gcc' shows, '-mmultiple' is used to generate code that uses the load multiple word instructions and the store multiple word instructions. Those instructions do not work when the processor is in little-endian mode (except PPC740/PPC750), so, we only enable it for big endian powerpc. - for big endian ppc64, as the help message from arch/powerpc/Kconfig shows, the V2 ABI is standard for 64-bit little-endian, but for big-endian it is less well tested by kernel and toolchain, so, use elfv1 as-is, no need to explicitly ask toolchain to use elfv2 here. [1]: https://qemu.readthedocs.io/en/latest/system/ppc/powernv.html [2]: https://github.com/linuxppc/issues/issues/402 [3]: https://lore.kernel.org/qemu-devel/20220504065536.3534488-1-aik@ozlabs.ru/Suggested-by:
-
Zhangjin Wu authored
Kernel uses ARCH=powerpc for both 32-bit and 64-bit PowerPC, here adds a ppc64le variant for little endian 64-bit PowerPC, users can pass XARCH=ppc64le to test it. The powernv machine of qemu-system-ppc64le is used for there is just a working powernv_defconfig. As the document [1] shows: PowerNV (as Non-Virtualized) is the “bare metal” platform using the OPAL firmware. It runs Linux on IBM and OpenPOWER systems and it can be used as an hypervisor OS, running KVM guests, or simply as a host OS. Notes, - since the VSX support may be disabled in kernel side, to avoid "illegal instruction" errors due to missing VSX kernel support, let's simply let compiler not generate vector/scalar (VSX) instructions via the '-mno-vsx' option. - little endian ppc64 prefers elfv2 to elfv1 if the toolchain (e.g. gcc 13.1.0) supports it, let's align with kernel, otherwise, our elfv1 binary will not run on kernel with elfv2 ABI. [1]: https://qemu.readthedocs.io/en/latest/system/ppc/powernv.htmlSuggested-by:
-
Zhangjin Wu authored
Kernel uses ARCH=powerpc for both 32-bit and 64-bit PowerPC, here adds a ppc variant for 32-bit PowerPC and uses it as the default variant of powerpc architecture. Users can pass XARCH=ppc (or ARCH=powerpc) to test 32-bit PowerPC. The default qemu-system-ppc g3beige machine [1] is used to run 32-bit powerpc kernel with pmac32_defconfig. The missing PMACZILOG serial tty and console are enabled in another patch [2]. Note, - zImage doesn't boot due to "qemu-system-ppc: Some ROM regions are overlapping" error, so, vmlinux is used instead. - since the VSX support may be disabled in kernel side, to avoid "illegal instruction" errors due to missing VSX kernel support, let's simply let compiler not generate vector/scalar (VSX) instructions via the '-mno-vsx' option. - as 'man gcc' shows, '-mmultiple' is used to generate code that uses the load multiple word instructions and the store multiple word instructions. Those instructions do not work when the processor is in little-endian mode (except PPC740/PPC750), so, we only enable it for big endian powerpc. [1]: https://qemu.readthedocs.io/en/latest/system/ppc/powermac.html [2]: https://lore.kernel.org/lkml/bb7b5f9958b3e3a20f6573ff7ce7c5dc566e7e32.1690982937.git.tanyuan@tinylab.org/Suggested-by:
-
Zhangjin Wu authored
Most of the CPU architectures have different variants, but kernel usually only accepts parts of them via the ARCH variable, the others should be customized via kernel config files. To simplify testing, a new XARCH variable is added to extend the kernel's ARCH with a few variants of the same architecture, and it is used to customize variant specific variables, at last XARCH is converted to the kernel's ARCH: e.g. make run XARCH=<one of the supported variants> | \ | `-> variant specific variables: | IMAGE, DEFCONFIG, QEMU_ARCH, QEMU_ARGS, CFLAGS ... \ `---> kernel's ARCH XARCH and ARCH are carefully mapped to allow users to pass architecture variants via XARCH or pass architecture via ARCH from cmdline. PowerPC is the first user and also a very good reference architecture of this mapping, it has variants with different combinations of 32-bit/64-bit and bit endian/little endian. To use this mapping, the other architectures can refer to PowerPC, If the target architecture only has one variant, XARCH is simply an alias of ARCH, no additional mapping required. Suggested-by: -
Zhangjin Wu authored
This follows the 64-bit PowerPC ABI [1], refers to the slides: "A new ABI for little-endian PowerPC64 Design & Implementation" [2] and the musl code in arch/powerpc64/crt_arch.h. First, stdu and clrrdi are used instead of stwu and clrrwi for powerpc64. Second, the stack frame size is increased to 32 bytes for powerpc64, 32 bytes is the minimal stack frame size supported described in [2]. Besides, the TOC pointer (GOT pointer) must be saved to r2. This works on both little endian and big endian 64-bit PowerPC. [1]: https://refspecs.linuxfoundation.org/ELF/ppc64/PPC-elf64abi.pdf [2]: https://www.llvm.org/devmtg/2014-04/PDFs/Talks/Euro-LLVM-2014-Weigand.pdfReviewed-by:
-
Zhangjin Wu authored
Both syscall declarations and _start code definition are added for powerpc to nolibc. Like mips, powerpc uses a register (exactly, the summary overflow bit) to record the error occurred, and uses another register to return the value [1]. So, the return value of every syscall declaration must be normalized to match the __sysret() helper, return -value when there is an error, otheriwse, return value directly. Glibc and musl use different methods to check the summary overflow bit, glibc (sysdeps/unix/sysv/linux/powerpc/sysdep.h) saves the cr register to r0 at first, and then check the summary overflow bit in cr0: mfcr r0 r0 & (1 << 28) ? -r3 : r3 --> 10003c14: 7c 00 00 26 mfcr r0 10003c18: 74 09 10 00 andis. r9,r0,4096 10003c1c: 41 82 00 08 beq 0x10003c24 10003c20: 7c 63 00 d0 neg r3,r3 Musl (arch/powerpc/syscall_arch.h) directly checks the summary overflow bit with the 'bns' instruction, it is smaller: /* no summary overflow bit means no error, return value directly */ bns+ 1f /* otherwise, return negated value */ neg r3, r3 1: --> 10000418: 40 a3 00 08 bns 0x10000420 1000041c: 7c 63 00 d0 neg r3,r3 Like musl, Linux (arch/powerpc/include/asm/vdso/gettimeofday.h) uses the same method for do_syscall_2() too. Here applies the second method to get smaller size. [1]: https://man7.org/linux/man-pages/man2/syscall.2.htmlReviewed-by: -
Thomas Weißschuh authored
As discussed with Willy, Paul and Shuah add myself as maintainer for the nolibc subsystem. Link: https://lore.kernel.org/lkml/7afafb6c-9664-44a1-bc8f-d20239db1dd5@paulmck-laptop/Signed-off-by:
Paul E. McKenney <paulmck@kernel.org> Signed-off-by:
-
Thomas Weißschuh authored
It will help the developers to avoid cruft and detect some bugs. Signed-off-by:
-
Thomas Weißschuh authored
Binary size is not important for nolibc-test and some debugging information is nice to have, so don't strip the binary during linking. Signed-off-by:
-
Thomas Weißschuh authored
If read() fails and returns -1 (or returns garbage for some other reason) buf would be accessed out of bounds. Only use the return value of read() after it has been validated. Signed-off-by:
-
Thomas Weißschuh authored
Avoid truncating values before comparing them. As printf in nolibc doesn't support ssize_t add casts to int for printing. Signed-off-by:
-
Thomas Weißschuh authored
These warnings will be enabled later so avoid triggering them. Signed-off-by:
-
Thomas Weißschuh authored
This warning will be enabled later so avoid triggering it. Signed-off-by:
-
Thomas Weißschuh authored
This allows the compiler to generate warnings if they go unused. Functions that are supposed to be used as breakpoints should not be static, so un-statify those if necessary. Signed-off-by:
-
Thomas Weißschuh authored
When warning about unused functions these would be reported by we want to keep them for future use. Suggested-by:
-
Thomas Weißschuh authored
These got copied around as new testcases where created. Signed-off-by:
-
Willy Tarreau authored
Recent fix ceb528fe ("selftests/nolibc: avoid gaps in test numbers") had the annoying side effect of always returning skipped tests, which are normally supposed to happen only when certain features are missing to run the test (missing kernel options, toolchain not supporting stack-protector etc). As such there are now always warnings. Let's modify the test to not use the condition and instead use a ternary expression to check the result. Fixes: ceb528fe ("selftests/nolibc: avoid gaps in test numbers") Cc: Thomas WeiÃ<9F>schuh <linux@weissschuh.net> Signed-off-by:
-
Thomas Weißschuh authored
Otherwise both gcc and clang may generate warnings about type mismatches: sysroot/mips/include/string.h:12:14: warning: mismatch in argument 1 type of built-in function 'malloc'; expected 'unsigned int' [-Wbuiltin-declaration-mismatch] 12 | static void *malloc(size_t len); | ^~~~~~ The compiler provides __SIZE_TYPE__ as the type that corresponds to size_t (typically "long unsigned int" or "unsigned int"). It was verified to be available at least since gcc-3.4 and clang-3.8, so from now on we'll use this definition for size_t. Signed-off-by: -
Thomas Weißschuh authored
getauxval() returns an unsigned long but the overall type of the ternary operator needs to be signed. Signed-off-by:
-
Thomas Weißschuh authored
This warning will be enabled later so avoid triggering it. Signed-off-by:
-
Thomas Weißschuh authored
It's documented as returning int which is also implemented by glibc and musl, so adopt that return type. Signed-off-by:
-
Thomas Weißschuh authored
Nobody needs it, get rid of it. Signed-off-by:
-
Yuan Tan authored
Add a test case of pipe that sends and receives message in a single process. Suggested-by:
Thomas Weißschuh <thomas@t-8ch.de> Suggested-by:
Yuan Tan <tanyuan@tinylab.org> Reviewed-by:
-
Yuan Tan authored
According to manual page [1], posix spec [2] and source code like arch/mips/kernel/syscall.c, for historic reasons, the sys_pipe() syscall on some architectures has an unusual calling convention. It returns results in two registers which means there is no need for it to do verify the validity of a userspace pointer argument. Historically that used to be expensive in Linux. These days the performance advantage is negligible. Nolibc doesn't support the unusual calling convention above, luckily Linux provides a generic sys_pipe2() with an additional flags argument from 2.6.27. If flags is 0, then pipe2() is the same as pipe(). So here we use sys_pipe2() to implement the pipe(). pipe2() is also provided to allow users to use flags argument on demand. [1]: https://man7.org/linux/man-pages/man2/pipe.2.html [2]: https://pubs.opengroup.org/onlinepubs/9699919799/functions/pipe.htmlSuggested-by:
-
Zhangjin Wu authored
The other tests use 1 as failure, mmap_munmap_good uses -1 as failure, let's fix up this. Signed-off-by:
-
Thomas Weißschuh authored
If the test description is longer than the status alignment the parameter 'n' to putcharn() would lead to a signed underflow that then gets converted to a very large unsigned value. This in turn leads out-of-bound writes in memset() crashing the application. The failure case of EXPECT_PTRER() used in "mmap_bad" exhibits this exact behavior. Fixes: 29f5540b ("selftests/nolibc: add EXPECT_PTREQ, EXPECT_PTRNE and EXPECT_PTRER") Signed-off-by:
-
Ryan Roberts authored
Add a minimal implementation of setvbuf(), which error checks the mode argument (as required by spec) and returns. Since nolibc never buffers output, nothing needs to be done. The kselftest framework recently added a call to setvbuf(). As a result, any tests that use the kselftest framework and nolibc cause a compiler error due to missing function. This provides an urgent fix for the problem which is preventing arm64 testing on linux-next. Example: clang --target=aarch64-linux-gnu -fintegrated-as -Werror=unknown-warning-option -Werror=ignored-optimization-argument -Werror=option-ignored -Werror=unused-command-line-argument --target=aarch64-linux-gnu -fintegrated-as -fno-asynchronous-unwind-tables -fno-ident -s -Os -nostdlib \ -include ../../../../include/nolibc/nolibc.h -I../..\ -static -ffreestanding -Wall za-fork.c build/kselftest/arm64/fp/za-fork-asm.o -o build/kselftest/arm64/fp/za-fork In file included from <built-in>:1: In file included from ./../../../../include/nolibc/nolibc.h:97: In file included from ./../../../../include/nolibc/arch.h:25: ./../../../../include/nolibc/arch-aarch64.h:178:35: warning: unknown attribute 'optimize' ignored [-Wunknown-attributes] void __attribute__((weak,noreturn,optimize("omit-frame-pointer"))) __no_stack_protector _start(void) ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ In file included from za-fork.c:12: ../../kselftest.h:123:2: error: call to undeclared function 'setvbuf'; ISO C99 and later do not support implicit function declarations [-Wimplicit-function-declaration] setvbuf(stdout, NULL, _IOLBF, 0); ^ ../../kselftest.h:123:24: error: use of undeclared identifier '_IOLBF' setvbuf(stdout, NULL, _IOLBF, 0); ^ 1 warning and 2 errors generated. Signed-off-by:Ryan Roberts <ryan.roberts@arm.com> Reported-by:
Linux Kernel Functional Testing <lkft@linaro.org> Link: https://lore.kernel.org/linux-kselftest/CA+G9fYus3Z8r2cg3zLv8uH8MRrzLFVWdnor02SNr=rCz+_WGVg@mail.gmail.com/Reviewed-by:
Mark Brown <broonie@kernel.org> Signed-off-by:
-
Zhangjin Wu authored
As the head comment of nolibc-test.c shows, it can be built in 3 ways: $(CC) -nostdlib -include /path/to/nolibc.h => NOLIBC already defined $(CC) -nostdlib -I/path/to/nolibc/sysroot => _NOLIBC_* guards are present $(CC) with default libc => NOLIBC* never defined Only last two of them are tested currently, let's allow test the first one too. This may help to find issues about using nolibc.h to build programs. it derives from this change: commit 3a8039e2 ("tools/nolibc: Fix build of stdio.h due to header ordering") Usage: // test with sysroot by default $ make run-user // test without sysroot, using nolibc.h directly $ make run-user NOLIBC_SYSROOT=0 Signed-off-by: -
Zhangjin Wu authored
It is able to run nolibc-test directly without qemu-user when the target machine is the same as the host machine. Sometimes, the result running locally may help a lot when the qemu-user package is too old. When the target machine differs from the host machine, it is also able to run nolibc-test directly with qemu-user-static + binfmt_misc. Link: https://lore.kernel.org/lkml/ZKutZwIOfy5MqedG@1wt.eu/Signed-off-by:
-
Zhangjin Wu authored
The startup code is critical to get the right argc, argv, envp/environ and _auxv, let's add a startup test group and the corresponding testcases. The "environ" test case is also moved from the stdlib test group to this new startup test group and it is renamed to "environ_envp". Since argv0 has been used by many other test cases, let's add testcases to gurantee it too. Signed-off-by:
-
Zhangjin Wu authored
4 new pointer compare macros are added, they are similar to the integer compare macros. Signed-off-by:
-
Zhangjin Wu authored
move most of the _start operations to _start_c(), include the stackprotector initialization. Signed-off-by:
-
Zhangjin Wu authored
move most of the _start operations to _start_c(), include the stackprotector initialization. Signed-off-by:
-
Zhangjin Wu authored
move most of the _start operations to _start_c(), include the stackprotector initialization. Signed-off-by:
-
Zhangjin Wu authored
move most of the _start operations to _start_c(), include the stackprotector initialization. Also clean up the instructions in delayed slots. Signed-off-by:
-
Zhangjin Wu authored
move most of the _start operations to _start_c(), include the stackprotector initialization. Signed-off-by:
-
