Hi, Willy
This patchset mainly allows speed up the nolibc test with a minimal kernel config.
As the nolibc supported architectures become more and more, the 'run' test with DEFCONFIG may cost several hours, which is not friendly to develop testing and even for release testing, so, smaller kernel configs may be required, and firstly, we should let nolibc-test work with less kernel config options, this patchset aims to this goal.
This patchset mainly remove the dependency from procfs, tmpfs, net and memfd_create, many failures have been fixed up.
When CONFIG_TMPFS and CONFIG_SHMEM are disabled, kernel will provide a ramfs based tmpfs (mm/shmem.c), it will be used as a choice to fix up some failures and also allow skip less tests.
Besides, it also adds musl support, improves glibc support and fixes up a kernel cmdline passing use case.
This is based on the dev.2023.06.14a branch of linux-rcu [1], all of the supported architectures are tested (with local minimal configs, [5] pasted the one for i386) without failures:
arch/board | result ------------|------------ arm/vexpress-a9 | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/arm-vexpress-a9-nolibc-test.log aarch64/virt | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/aarch64-virt-nolibc-test.log ppc/g3beige | not supported i386/pc | 136 test(s) passed, 3 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/i386-pc-nolibc-test.log x86_64/pc | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/x86_64-pc-nolibc-test.log mipsel/malta | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/mipsel-malta-nolibc-test.log loongarch64/virt | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/loongarch64-virt-nolibc-test.log riscv64/virt | 136 test(s) passed, 3 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/riscv64-virt-nolibc-test.log riscv32/virt | no test log found s390x/s390-ccw-virtio | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/s390x-s390-ccw-virtio-nolibc-test.log
Notes: * The skipped ones are -fstackprotector, chmod_self and chown_self
The -fstackprotector skip is due to gcc version. chmod_self and chmod_self skips are due to procfs not enabled
* ppc/g3beige support is added locally, but not added in this patchset
will send ppc support as a new patchset, it depends on v2 test report patchset [3] and the v5 rv32 support, require changes on Makefile
* riscv32/virt support is still in review, see v5 rv32 support [4]
This patchset doesn't depends on any of my other nolibc patch series, but the new rmdir() routine added in this patchset may be requird to apply the __sysret() from our v4 syscall helper series [2] after that series being merged, currently, we use the old method to let it compile without any dependency.
Here explains all of the patches:
* selftests/nolibc: stat_fault: silence NULL argument warning with glibc selftests/nolibc: gettid: restore for glibc and musl selftests/nolibc: add _LARGEFILE64_SOURCE for musl
The above 3 patches adds musl compile support and improve glibc support.
It is able to build and run nolibc-test with musl libc now, but there are some failures/skips due to the musl its own issues/requirements:
$ sudo ./nolibc-test | grep -E 'FAIL|SKIP' 8 sbrk = 1 ENOMEM [FAIL] 9 brk = -1 ENOMEM [FAIL] 46 limit_int_fast16_min = -2147483648 [FAIL] 47 limit_int_fast16_max = 2147483647 [FAIL] 49 limit_int_fast32_min = -2147483648 [FAIL] 50 limit_int_fast32_max = 2147483647 [FAIL] 0 -fstackprotector not supported [SKIPPED]
musl disabled sbrk and brk for some conflicts with its malloc and the fast version of int types are defined in 32bit, which differs from nolibc and glibc. musl reserved the sbrk(0) to allow get current brk, we added a test for this in the v4 __sysret() helper series [2].
* selftests/nolibc: fix up kernel parameters support
kernel cmdline allows pass two types of parameters, one is without '=', another is with '=', the first one is passed as init arguments, the sencond one is passed as init environment variables.
Our nolibc-test prefer arguments to environment variables, this not work when users add such parameters in the kernel cmdline:
noapic NOLIBC_TEST=syscall
So, this patch will verify the setting from arguments at first, if it is no valid, will try the environment variables instead.
* selftests/nolibc: stat_timestamps: remove procfs dependency
Use '/' instead of /proc/self, or we can add a 'has_proc' condition for this test case, but it is not that necessary to skip the whole stat_timestamps tests for such a subtest binding to /proc/self.
Welcome suggestion from Thomas.
* tools/nolibc: add rmdir() support selftests/nolibc: add a new rmdir() test case
rmdir() routine and test case are added for the coming requirement.
Note, if the __sysret() patchset [2] is applied before us, this patch should be rebased on it and apply the __sysret() helper.
* selftests/nolibc: fix up failures when there is no procfs
call rmdir() to remove /proc completely to rework the checking of /proc, before, the existing of /proc not means the procfs is really mounted.
* selftests/nolibc: rename proc variable to has_proc selftests/nolibc: rename euid0 variable to is_root
align with the has_gettid, has_xxx variables.
* selftests/nolibc: prepare tmpfs and hugetlbfs selftests/nolibc: rename chmod_net to chmod_good selftests/nolibc: link_cross: support tmpfs selftests/nolibc: rename chroot_exe to chroot_file
use file from /tmp instead of file from /proc when there is no procfs this avoid skipping the chmod_net, link_cross, chroot_exe tests
* selftests/nolibc: vfprintf: silence memfd_create() warning selftests/nolibc: vfprintf: skip if neither tmpfs nor hugetlbfs selftests/nolibc: vfprintf: support tmpfs and hugetlbfs
memfd_create from kernel >= v6.2 forcely warn on missing MFD_NOEXEC_SEAL flag, the first one silence it with such flag, for older kernels, use 0 flag as before.
since memfd_create() depends on TMPFS or HUGETLBFS, the second one skip the whole vfprintf instead of simply fail if memfd_create() not work.
the 3rd one futher try the ramfs based tmpfs even when memfd_create() not work.
At last, let's simply discuss about the configs, I have prepared minimal configs for all of the nolibc supported architectures but not sure where should we put them, what about tools/testing/selftests/nolibc/configs ?
Thanks!
Best regards, Zhangjin ---
[1]: https://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git/ [2]: https://lore.kernel.org/linux-riscv/cover.1687187451.git.falcon@tinylab.org/ [3]: https://lore.kernel.org/lkml/cover.1687156559.git.falcon@tinylab.org/ [4]: https://lore.kernel.org/linux-riscv/cover.1687176996.git.falcon@tinylab.org/ [5]: https://pastebin.com/5jq0Vxbz
Zhangjin Wu (17): selftests/nolibc: stat_fault: silence NULL argument warning with glibc selftests/nolibc: gettid: restore for glibc and musl selftests/nolibc: add _LARGEFILE64_SOURCE for musl selftests/nolibc: fix up kernel parameters support selftests/nolibc: stat_timestamps: remove procfs dependency tools/nolibc: add rmdir() support selftests/nolibc: add a new rmdir() test case selftests/nolibc: fix up failures when there is no procfs selftests/nolibc: rename proc variable to has_proc selftests/nolibc: rename euid0 variable to is_root selftests/nolibc: prepare tmpfs and hugetlbfs selftests/nolibc: rename chmod_net to chmod_good selftests/nolibc: link_cross: support tmpfs selftests/nolibc: rename chroot_exe to chroot_file selftests/nolibc: vfprintf: silence memfd_create() warning selftests/nolibc: vfprintf: skip if neither tmpfs nor hugetlbfs selftests/nolibc: vfprintf: support tmpfs and hugetlbfs
tools/include/nolibc/sys.h | 28 ++++ tools/testing/selftests/nolibc/nolibc-test.c | 132 +++++++++++++++---- 2 files changed, 138 insertions(+), 22 deletions(-)
Use another invalid address (void *)1 instead of NULL to silence this compile warning with glibc:
$ make libc-test CC libc-test nolibc-test.c: In function ‘run_syscall’: nolibc-test.c:622:49: warning: null argument where non-null required (argument 1) [-Wnonnull] 622 | CASE_TEST(stat_fault); EXPECT_SYSER(1, stat(NULL, &stat_buf), -1, EFAULT); break; | ^~~~ nolibc-test.c:304:79: note: in definition of macro ‘EXPECT_SYSER2’ 304 | do { if (!cond) pad_spc(llen, 64, "[SKIPPED]\n"); else ret += expect_syserr2(expr, expret, experr1, experr2, llen); } while (0) | ^~~~ nolibc-test.c:622:33: note: in expansion of macro ‘EXPECT_SYSER’ 622 | CASE_TEST(stat_fault); EXPECT_SYSER(1, stat(NULL, &stat_buf), -1, EFAULT); break;
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 486334981e60..99afec93dfae 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -619,7 +619,7 @@ int run_syscall(int min, int max) CASE_TEST(select_stdout); EXPECT_SYSNE(1, ({ fd_set fds; FD_ZERO(&fds); FD_SET(1, &fds); select(2, NULL, &fds, NULL, NULL); }), -1); break; CASE_TEST(select_fault); EXPECT_SYSER(1, select(1, (void *)1, NULL, NULL, 0), -1, EFAULT); break; CASE_TEST(stat_blah); EXPECT_SYSER(1, stat("/proc/self/blah", &stat_buf), -1, ENOENT); break; - CASE_TEST(stat_fault); EXPECT_SYSER(1, stat(NULL, &stat_buf), -1, EFAULT); break; + CASE_TEST(stat_fault); EXPECT_SYSER(1, stat((void *)1, &stat_buf), -1, EFAULT); break; CASE_TEST(stat_timestamps); EXPECT_SYSZR(1, test_stat_timestamps()); break; CASE_TEST(symlink_root); EXPECT_SYSER(1, symlink("/", "/"), -1, EEXIST); break; CASE_TEST(unlink_root); EXPECT_SYSER(1, unlink("/"), -1, EISDIR); break;
As the gettid manpage [1] shows, glibc 2.30 has gettid support, so, let's enable the test for glibc >= 2.30.
gettid works on musl too.
[1]: https://man7.org/linux/man-pages/man2/gettid.2.html
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 99afec93dfae..739c9daa91b6 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -548,6 +548,7 @@ int run_syscall(int min, int max) int tmp; int ret = 0; void *p1, *p2; + int has_gettid = 1;
/* <proc> indicates whether or not /proc is mounted */ proc = stat("/proc", &stat_buf) == 0; @@ -555,6 +556,11 @@ int run_syscall(int min, int max) /* this will be used to skip certain tests that can't be run unprivileged */ euid0 = geteuid() == 0;
+ /* from 2.30, glibc provides gettid() */ +#if defined(__GLIBC_MINOR__) && defined(__GLIBC__) + has_gettid = __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ >= 30); +#endif + for (test = min; test >= 0 && test <= max; test++) { int llen = 0; /* line length */
@@ -564,9 +570,7 @@ int run_syscall(int min, int max) switch (test + __LINE__ + 1) { CASE_TEST(getpid); EXPECT_SYSNE(1, getpid(), -1); break; CASE_TEST(getppid); EXPECT_SYSNE(1, getppid(), -1); break; -#ifdef NOLIBC - CASE_TEST(gettid); EXPECT_SYSNE(1, gettid(), -1); break; -#endif + CASE_TEST(gettid); EXPECT_SYSNE(has_gettid, gettid(), -1); break; CASE_TEST(getpgid_self); EXPECT_SYSNE(1, getpgid(0), -1); break; CASE_TEST(getpgid_bad); EXPECT_SYSER(1, getpgid(-1), -1, ESRCH); break; CASE_TEST(kill_0); EXPECT_SYSZR(1, kill(getpid(), 0)); break;
_GNU_SOURCE Implies _LARGEFILE64_SOURCE in glibc, but in musl, the default configuration doesn't enable _LARGEFILE64_SOURCE.
From include/dirent.h of musl, getdents64 is provided as getdents when _LARGEFILE64_SOURCE is defined.
#if defined(_LARGEFILE64_SOURCE) ... #define getdents64 getdents #endif
Let's define _LARGEFILE64_SOURCE to fix up this compile error:
tools/testing/selftests/nolibc/nolibc-test.c: In function ‘test_getdents64’: tools/testing/selftests/nolibc/nolibc-test.c:453:8: warning: implicit declaration of function ‘getdents64’; did you mean ‘getdents’? [-Wimplicit-function-declaration] 453 | ret = getdents64(fd, (void *)buffer, sizeof(buffer)); | ^~~~~~~~~~ | getdents /usr/bin/ld: /tmp/ccKILm5u.o: in function `test_getdents64': nolibc-test.c:(.text+0xe3e): undefined reference to `getdents64' collect2: error: ld returned 1 exit status
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 739c9daa91b6..d43116553288 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -1,6 +1,7 @@ /* SPDX-License-Identifier: GPL-2.0 */
#define _GNU_SOURCE +#define _LARGEFILE64_SOURCE
/* libc-specific include files * The program may be built in 3 ways:
kernel parameters allow pass two types of strings, one type is like 'noapic', another type is like 'panic=5', the first type is passed as arguments of the init program, the second type is passed as environment variables of the init program.
when users pass kernel parameters like this:
noapic NOLIBC_TEST=syscall
our nolibc-test program will ignore the NOLIBC_TEST environment variable.
Let's verify the first type (from arguments), if it is invalid, get the test setting from NOLIBC_TEST environment variable instead.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 15 ++++++++++++++- 1 file changed, 14 insertions(+), 1 deletion(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index d43116553288..ebec948ec808 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -951,6 +951,7 @@ int main(int argc, char **argv, char **envp) int ret = 0; int err; int idx; + int len, valid = 0; char *test;
environ = envp; @@ -969,7 +970,19 @@ int main(int argc, char **argv, char **envp) * syscall:5-15[:.*],stdlib:8-10 */ test = argv[1]; - if (!test) + + /* verify the test setting from argv[1] */ + if (test) { + for (idx = 0; test_names[idx].name; idx++) { + len = strlen(test_names[idx].name); + if (strncmp(test, test_names[idx].name, len) == 0) { + valid = 1; + break; + } + } + } + + if (!valid) test = getenv("NOLIBC_TEST");
if (test) {
Since it is not really necessary to use /proc/self here, instead of adding a condition check, we use the always existing '/' path instead of /proc/self, this eventually let it work without procfs.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index ebec948ec808..2ef44176f7a9 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -520,7 +520,7 @@ static int test_stat_timestamps(void) if (sizeof(st.st_atim.tv_sec) != sizeof(st.st_atime)) return 1;
- if (stat("/proc/self/", &st)) + if (stat("/", &st)) return 1;
if (st.st_atim.tv_sec != st.st_atime || st.st_atim.tv_nsec > 1000000000)
Hi, Thomas
I'm preparing a revision for this series, in the past days, when I was working on testing our new 'minimal' kernel config support for all of the architectures, the time cost (and wait) is really appreciable and the repeated develop and test is really a big pain, I can also image when you was working on stack-protector and Willy was working on lots of old features ;-)
As you explained before, I knew the idea of using '/proc/self' here is important to not using a fixed-time file, besides our proposed method (make sure it at least not fail, just skip for !procfs):
- CASE_TEST(stat_timestamps); EXPECT_SYSZR(1, test_stat_timestamps()); break; + CASE_TEST(stat_timestamps); EXPECT_SYSZR(proc, test_stat_timestamps()); break;
To further avoid skip it for !procfs (I don't mean relaly disable it for the default tinyconfig support, which need more discuss, at least provide the possibility to pass without procfs), do you like this change? it doesn't depend on 'proc' now.
- if (stat("/proc/self/", &st)) + if (stat("/proc/self/", &st) && stat("/init", &st) && stat("/", &st))
The "/init" is compiled for 'run' target every time, so, the time stamp should be dynamic enough, for libc-test, the /proc/self should be always there (if still not enough, we can reuse the init file list here), the "/" here is only for the worst-case scene ;-)
Thanks, Zhangjin
Since it is not really necessary to use /proc/self here, instead of adding a condition check, we use the always existing '/' path instead of /proc/self, this eventually let it work without procfs.
Signed-off-by: Zhangjin Wu falcon@tinylab.org
tools/testing/selftests/nolibc/nolibc-test.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index ebec948ec808..2ef44176f7a9 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -520,7 +520,7 @@ static int test_stat_timestamps(void) if (sizeof(st.st_atim.tv_sec) != sizeof(st.st_atime)) return 1;
- if (stat("/proc/self/", &st))
- if (stat("/", &st)) return 1;
if (st.st_atim.tv_sec != st.st_atime || st.st_atim.tv_nsec > 1000000000) -- 2.25.1
Hi Zhangjin,
On 2023-06-28 21:59:22+0800, Zhangjin Wu wrote:
I'm preparing a revision for this series, in the past days, when I was working on testing our new 'minimal' kernel config support for all of the architectures, the time cost (and wait) is really appreciable and the repeated develop and test is really a big pain, I can also image when you was working on stack-protector and Willy was working on lots of old features ;-)
To be honest I almost never built a kernel. Most of the time I tested my stuff with qemu-user. This made the dev-cycle really fast, especially with a binfmt setup that launches foreign binaries automatically with qemu-user.
As you explained before, I knew the idea of using '/proc/self' here is important to not using a fixed-time file, besides our proposed method (make sure it at least not fail, just skip for !procfs):
- CASE_TEST(stat_timestamps); EXPECT_SYSZR(1, test_stat_timestamps()); break; + CASE_TEST(stat_timestamps); EXPECT_SYSZR(proc, test_stat_timestamps()); break;To further avoid skip it for !procfs (I don't mean relaly disable it for the default tinyconfig support, which need more discuss, at least provide the possibility to pass without procfs), do you like this change? it doesn't depend on 'proc' now.
- if (stat("/proc/self/", &st)) + if (stat("/proc/self/", &st) && stat("/init", &st) && stat("/", &st))The "/init" is compiled for 'run' target every time, so, the time stamp should be dynamic enough, for libc-test, the /proc/self should be always there (if still not enough, we can reuse the init file list here), the "/" here is only for the worst-case scene ;-)
Both aproaches seem fine. Just skipping on !proc seems good enough.
As for enabling proc in the test configs I just tested a plain tinyconfig vs one with CONFIG_PROC_FS enabled:
tinyconfig: 375.06user 53.21system 2:05.80elapsed tinyconfig + CONFIG_PROC_FS: 397.77user 56.84system 2:09.24elapsed
The overhead seems acceptable.
Note as for disabling memfd:
It seems currently MEMFD_CREATE is hardwired to only be enabled when either TMPFS or HUGETLBFS is enabled.
But the memfd code and syscalls seem to work perfectly fine with those options disabled. I'll send a patch to fix up the Kconfigs to enable that usecase.
Thanks, Zhangjin
Since it is not really necessary to use /proc/self here, instead of adding a condition check, we use the always existing '/' path instead of /proc/self, this eventually let it work without procfs.
Signed-off-by: Zhangjin Wu falcon@tinylab.org
tools/testing/selftests/nolibc/nolibc-test.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index ebec948ec808..2ef44176f7a9 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -520,7 +520,7 @@ static int test_stat_timestamps(void) if (sizeof(st.st_atim.tv_sec) != sizeof(st.st_atime)) return 1;
- if (stat("/proc/self/", &st))
- if (stat("/", &st)) return 1;
if (st.st_atim.tv_sec != st.st_atime || st.st_atim.tv_nsec > 1000000000) -- 2.25.1
Hi Zhangjin,
On 2023-06-28 21:59:22+0800, Zhangjin Wu wrote:
I'm preparing a revision for this series, in the past days, when I was working on testing our new 'minimal' kernel config support for all of the architectures, the time cost (and wait) is really appreciable and the repeated develop and test is really a big pain, I can also image when you was working on stack-protector and Willy was working on lots of old features ;-)
To be honest I almost never built a kernel. Most of the time I tested my stuff with qemu-user. This made the dev-cycle really fast, especially with a binfmt setup that launches foreign binaries automatically with qemu-user.
Yeah, qemu-user-static + binfmt_misc work perfectly, but my host kernel is not that new, so, I'm still a little worried about that there may be some hidden issues.
As you explained before, I knew the idea of using '/proc/self' here is important to not using a fixed-time file, besides our proposed method (make sure it at least not fail, just skip for !procfs):
- CASE_TEST(stat_timestamps); EXPECT_SYSZR(1, test_stat_timestamps()); break; + CASE_TEST(stat_timestamps); EXPECT_SYSZR(proc, test_stat_timestamps()); break;To further avoid skip it for !procfs (I don't mean relaly disable it for the default tinyconfig support, which need more discuss, at least provide the possibility to pass without procfs), do you like this change? it doesn't depend on 'proc' now.
- if (stat("/proc/self/", &st)) + if (stat("/proc/self/", &st) && stat("/init", &st) && stat("/", &st))The "/init" is compiled for 'run' target every time, so, the time stamp should be dynamic enough, for libc-test, the /proc/self should be always there (if still not enough, we can reuse the init file list here), the "/" here is only for the worst-case scene ;-)
Both aproaches seem fine. Just skipping on !proc seems good enough.
To get less skips, let's use the second method, just updated my local patches ;-)
As for enabling proc in the test configs I just tested a plain tinyconfig vs one with CONFIG_PROC_FS enabled:
tinyconfig: 375.06user 53.21system 2:05.80elapsed tinyconfig + CONFIG_PROC_FS: 397.77user 56.84system 2:09.24elapsed
The overhead seems acceptable.
Yeah, only one option is ok, but "multiple options x multiple architectures x multiple repeated runs", that is 'huge' ;-)
Note as for disabling memfd:
It seems currently MEMFD_CREATE is hardwired to only be enabled when either TMPFS or HUGETLBFS is enabled.
But the memfd code and syscalls seem to work perfectly fine with those options disabled. I'll send a patch to fix up the Kconfigs to enable that usecase.
Good catch!
but for the vfprintf test cases, It is able to open a file from tmpfs directly. If no tmpfs, use the default ramfs (initramfs uses) instead, this will also avoid the new flags trying (to silence the warning).
static int expect_vfprintf(int llen, size_t c, const char *expected, const char *fmt, ...) { + static const char *tmpfile = "/tmp/nolibc-vfprintf"; + struct stat stat_buf; int ret, fd, w, r; char buf[100]; FILE *memfile; va_list args;
- fd = memfd_create("vfprintf", 0); + if (stat("/tmp/.", &stat_buf)) { + pad_spc(llen, 64, "[SKIPPED]\n"); + return 0; + } + + fd = open(tmpfile, O_CREAT | O_TRUNC | O_RDWR, 0755); ... + unlink(tmpfile); ...
tmpfs is mounted (in another patch) like procfs in prepare() for pid==1.
I plan to use this method in the revision, do you like this?
memfd_create() was designed to do this work, but in current stage, opening tmpfile ourselves may be better.
Thanks, Zhangjin
On 2023-06-30 05:23:35+0800, Zhangjin Wu wrote:
Hi Zhangjin,
On 2023-06-28 21:59:22+0800, Zhangjin Wu wrote:
I'm preparing a revision for this series, in the past days, when I was working on testing our new 'minimal' kernel config support for all of the architectures, the time cost (and wait) is really appreciable and the repeated develop and test is really a big pain, I can also image when you was working on stack-protector and Willy was working on lots of old features ;-)
To be honest I almost never built a kernel. Most of the time I tested my stuff with qemu-user. This made the dev-cycle really fast, especially with a binfmt setup that launches foreign binaries automatically with qemu-user.
Yeah, qemu-user-static + binfmt_misc work perfectly, but my host kernel is not that new, so, I'm still a little worried about that there may be some hidden issues.
qemu-user shouldn't have specific requirements for the host kernel. Could you elaborate?
As you explained before, I knew the idea of using '/proc/self' here is important to not using a fixed-time file, besides our proposed method (make sure it at least not fail, just skip for !procfs):
- CASE_TEST(stat_timestamps); EXPECT_SYSZR(1, test_stat_timestamps()); break; + CASE_TEST(stat_timestamps); EXPECT_SYSZR(proc, test_stat_timestamps()); break;To further avoid skip it for !procfs (I don't mean relaly disable it for the default tinyconfig support, which need more discuss, at least provide the possibility to pass without procfs), do you like this change? it doesn't depend on 'proc' now.
- if (stat("/proc/self/", &st)) + if (stat("/proc/self/", &st) && stat("/init", &st) && stat("/", &st))The "/init" is compiled for 'run' target every time, so, the time stamp should be dynamic enough, for libc-test, the /proc/self should be always there (if still not enough, we can reuse the init file list here), the "/" here is only for the worst-case scene ;-)
Both aproaches seem fine. Just skipping on !proc seems good enough.
To get less skips, let's use the second method, just updated my local patches ;-)
As for enabling proc in the test configs I just tested a plain tinyconfig vs one with CONFIG_PROC_FS enabled:
tinyconfig: 375.06user 53.21system 2:05.80elapsed tinyconfig + CONFIG_PROC_FS: 397.77user 56.84system 2:09.24elapsed
The overhead seems acceptable.
Yeah, only one option is ok, but "multiple options x multiple architectures x multiple repeated runs", that is 'huge' ;-)
In your other patchset you mentioned a few options that were needed. But if we can drop CONFIG_NET completely, reduce CONFIG_TMPFS to CONFIG_MEMFD I would like to reevaluate the overall impact.
Note as for disabling memfd:
It seems currently MEMFD_CREATE is hardwired to only be enabled when either TMPFS or HUGETLBFS is enabled.
But the memfd code and syscalls seem to work perfectly fine with those options disabled. I'll send a patch to fix up the Kconfigs to enable that usecase.
Good catch!
but for the vfprintf test cases, It is able to open a file from tmpfs directly. If no tmpfs, use the default ramfs (initramfs uses) instead, this will also avoid the new flags trying (to silence the warning).
static int expect_vfprintf(int llen, size_t c, const char *expected, const char *fmt, ...) { + static const char *tmpfile = "/tmp/nolibc-vfprintf"; + struct stat stat_buf;
This should not be static.
int ret, fd, w, r; char buf[100]; FILE *memfile; va_list args; - fd = memfd_create("vfprintf", 0); + if (stat("/tmp/.", &stat_buf)) { + pad_spc(llen, 64, "[SKIPPED]\n"); + return 0; + }
Instead of checking with stat() here it would be nicer to check the result of open() below.
+ + fd = open(tmpfile, O_CREAT | O_TRUNC | O_RDWR, 0755);
Seems like a good usecase for open("/tmp", O_TMPFILE | O_EXCL | ...)
... + unlink(tmpfile);
... and drop this.
fd = memfd_open(); + if (fd == -1) + fd = open("/tmp", O_TMPFILE | O_EXCL | O_CREAT | O_RDWR, 0600); + if (fd == -1) + skip()
May be enough.
...tmpfs is mounted (in another patch) like procfs in prepare() for pid==1.
I plan to use this method in the revision, do you like this?
memfd_create() was designed to do this work, but in current stage, opening tmpfile ourselves may be better.
I'm not really a fan of having so much fallback code for features that should always be available. But if we can keep it straightforward it's probably fine. Or Willy has another opinion :-)
A reverse operation of mkdir is meaningful, add rmdir() here.
This is required by nolibc-test to remove /proc if CONFIG_PROC_FS not enabled.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/include/nolibc/sys.h | 28 ++++++++++++++++++++++++++++ 1 file changed, 28 insertions(+)
diff --git a/tools/include/nolibc/sys.h b/tools/include/nolibc/sys.h index 856249a11890..8ddfd9185da6 100644 --- a/tools/include/nolibc/sys.h +++ b/tools/include/nolibc/sys.h @@ -716,6 +716,34 @@ int mkdir(const char *path, mode_t mode) return ret; }
+/* + * int rmdir(const char *path); + */ + +static __attribute__((unused)) +int sys_rmdir(const char *path) +{ +#ifdef __NR_rmdir + return my_syscall1(__NR_rmdir, path); +#elif defined(__NR_unlinkat) + return my_syscall3(__NR_unlinkat, AT_FDCWD, path, AT_REMOVEDIR); +#else + return -ENOSYS; +#endif +} + +static __attribute__((unused)) +int rmdir(const char *path) +{ + int ret = sys_rmdir(path); + + if (ret < 0) { + SET_ERRNO(-ret); + ret = -1; + } + return ret; +} +
/* * int mknod(const char *path, mode_t mode, dev_t dev);
A new rmdir_blah test case is added to remove a non-existing /blah, which expects failure with ENOENT errno.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 2ef44176f7a9..9673c338d42e 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -612,6 +612,7 @@ int run_syscall(int min, int max) CASE_TEST(lseek_m1); EXPECT_SYSER(1, lseek(-1, 0, SEEK_SET), -1, EBADF); break; CASE_TEST(lseek_0); EXPECT_SYSER(1, lseek(0, 0, SEEK_SET), -1, ESPIPE); break; CASE_TEST(mkdir_root); EXPECT_SYSER(1, mkdir("/", 0755), -1, EEXIST); break; + CASE_TEST(rmdir_blah); EXPECT_SYSER(1, rmdir("/blah"), -1, ENOENT); break; CASE_TEST(open_tty); EXPECT_SYSNE(1, tmp = open("/dev/null", 0), -1); if (tmp != -1) close(tmp); break; CASE_TEST(open_blah); EXPECT_SYSER(1, tmp = open("/proc/self/blah", 0), -1, ENOENT); if (tmp != -1) close(tmp); break; CASE_TEST(poll_null); EXPECT_SYSZR(1, poll(NULL, 0, 0)); break;
When there is no procfs support, the /proc is not mountable, but the /proc directory has been created in the prepare() stage and therefore, the checking of /proc in the run_syscall() stage will be always true and at last will fail all of the procfs dependent test cases.
To solve this issue, one method is checking /proc/self instead of /proc, another one is removing the /proc directory completely (when there is really no procfs support), we apply the second method to avoid mislead the users.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 9673c338d42e..e5a218ef8a2d 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -928,8 +928,11 @@ int prepare(void)
/* try to mount /proc if not mounted. Silently fail otherwise */ if (stat("/proc/.", &stat_buf) == 0 || mkdir("/proc", 0755) == 0) { - if (stat("/proc/self", &stat_buf) != 0) - mount("/proc", "/proc", "proc", 0, 0); + if (stat("/proc/self", &stat_buf) != 0) { + /* If not mountable, remove /proc completely to avoid misuse */ + if (mount("none", "/proc", "proc", 0, 0) != 0) + rmdir("/proc"); + } }
return 0;
More conditional variables will be added, let's rename them with more meaningful 'has_' or 'is_' prefix.
This converts proc to has_proc to reflect whether or not proc/ is mounted.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index e5a218ef8a2d..929d43b95acc 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -544,15 +544,15 @@ int run_syscall(int min, int max) struct timezone tz; struct stat stat_buf; int euid0; - int proc; + int has_proc; int test; int tmp; int ret = 0; void *p1, *p2; int has_gettid = 1;
- /* <proc> indicates whether or not /proc is mounted */ - proc = stat("/proc", &stat_buf) == 0; + /* <has_proc> indicates whether or not /proc is mounted */ + has_proc = stat("/proc", &stat_buf) == 0;
/* this will be used to skip certain tests that can't be run unprivileged */ euid0 = geteuid() == 0; @@ -582,12 +582,12 @@ int run_syscall(int min, int max) CASE_TEST(chdir_root); EXPECT_SYSZR(1, chdir("/")); break; CASE_TEST(chdir_dot); EXPECT_SYSZR(1, chdir(".")); break; CASE_TEST(chdir_blah); EXPECT_SYSER(1, chdir("/blah"), -1, ENOENT); break; - CASE_TEST(chmod_net); EXPECT_SYSZR(proc, chmod("/proc/self/net", 0555)); break; - CASE_TEST(chmod_self); EXPECT_SYSER(proc, chmod("/proc/self", 0555), -1, EPERM); break; - CASE_TEST(chown_self); EXPECT_SYSER(proc, chown("/proc/self", 0, 0), -1, EPERM); break; + CASE_TEST(chmod_net); EXPECT_SYSZR(has_proc, chmod("/proc/self/net", 0555)); break; + CASE_TEST(chmod_self); EXPECT_SYSER(has_proc, chmod("/proc/self", 0555), -1, EPERM); break; + CASE_TEST(chown_self); EXPECT_SYSER(has_proc, chown("/proc/self", 0, 0), -1, EPERM); break; CASE_TEST(chroot_root); EXPECT_SYSZR(euid0, chroot("/")); break; CASE_TEST(chroot_blah); EXPECT_SYSER(1, chroot("/proc/self/blah"), -1, ENOENT); break; - CASE_TEST(chroot_exe); EXPECT_SYSER(proc, chroot("/proc/self/exe"), -1, ENOTDIR); break; + CASE_TEST(chroot_exe); EXPECT_SYSER(has_proc, chroot("/proc/self/exe"), -1, ENOTDIR); break; CASE_TEST(close_m1); EXPECT_SYSER(1, close(-1), -1, EBADF); break; CASE_TEST(close_dup); EXPECT_SYSZR(1, close(dup(0))); break; CASE_TEST(dup_0); tmp = dup(0); EXPECT_SYSNE(1, tmp, -1); close(tmp); break; @@ -608,7 +608,7 @@ int run_syscall(int min, int max) CASE_TEST(link_root1); EXPECT_SYSER(1, link("/", "/"), -1, EEXIST); break; CASE_TEST(link_blah); EXPECT_SYSER(1, link("/proc/self/blah", "/blah"), -1, ENOENT); break; CASE_TEST(link_dir); EXPECT_SYSER(euid0, link("/", "/blah"), -1, EPERM); break; - CASE_TEST(link_cross); EXPECT_SYSER(proc, link("/proc/self/net", "/blah"), -1, EXDEV); break; + CASE_TEST(link_cross); EXPECT_SYSER(has_proc, link("/proc/self/net", "/blah"), -1, EXDEV); break; CASE_TEST(lseek_m1); EXPECT_SYSER(1, lseek(-1, 0, SEEK_SET), -1, EBADF); break; CASE_TEST(lseek_0); EXPECT_SYSER(1, lseek(0, 0, SEEK_SET), -1, ESPIPE); break; CASE_TEST(mkdir_root); EXPECT_SYSER(1, mkdir("/", 0755), -1, EEXIST); break;
More conditional variables will be added, let's rename them with more meaningful 'has_' or 'is_' prefix.
euid0 is converted to is_root to indicate whether or not running as root.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 929d43b95acc..69f59a395746 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -543,7 +543,7 @@ int run_syscall(int min, int max) struct timeval tv; struct timezone tz; struct stat stat_buf; - int euid0; + int is_root; int has_proc; int test; int tmp; @@ -555,7 +555,7 @@ int run_syscall(int min, int max) has_proc = stat("/proc", &stat_buf) == 0;
/* this will be used to skip certain tests that can't be run unprivileged */ - euid0 = geteuid() == 0; + is_root = geteuid() == 0;
/* from 2.30, glibc provides gettid() */ #if defined(__GLIBC_MINOR__) && defined(__GLIBC__) @@ -585,7 +585,7 @@ int run_syscall(int min, int max) CASE_TEST(chmod_net); EXPECT_SYSZR(has_proc, chmod("/proc/self/net", 0555)); break; CASE_TEST(chmod_self); EXPECT_SYSER(has_proc, chmod("/proc/self", 0555), -1, EPERM); break; CASE_TEST(chown_self); EXPECT_SYSER(has_proc, chown("/proc/self", 0, 0), -1, EPERM); break; - CASE_TEST(chroot_root); EXPECT_SYSZR(euid0, chroot("/")); break; + CASE_TEST(chroot_root); EXPECT_SYSZR(is_root, chroot("/")); break; CASE_TEST(chroot_blah); EXPECT_SYSER(1, chroot("/proc/self/blah"), -1, ENOENT); break; CASE_TEST(chroot_exe); EXPECT_SYSER(has_proc, chroot("/proc/self/exe"), -1, ENOTDIR); break; CASE_TEST(close_m1); EXPECT_SYSER(1, close(-1), -1, EBADF); break; @@ -607,7 +607,7 @@ int run_syscall(int min, int max) CASE_TEST(ioctl_tiocinq); EXPECT_SYSZR(1, ioctl(0, TIOCINQ, &tmp)); break; CASE_TEST(link_root1); EXPECT_SYSER(1, link("/", "/"), -1, EEXIST); break; CASE_TEST(link_blah); EXPECT_SYSER(1, link("/proc/self/blah", "/blah"), -1, ENOENT); break; - CASE_TEST(link_dir); EXPECT_SYSER(euid0, link("/", "/blah"), -1, EPERM); break; + CASE_TEST(link_dir); EXPECT_SYSER(is_root, link("/", "/blah"), -1, EPERM); break; CASE_TEST(link_cross); EXPECT_SYSER(has_proc, link("/proc/self/net", "/blah"), -1, EXDEV); break; CASE_TEST(lseek_m1); EXPECT_SYSER(1, lseek(-1, 0, SEEK_SET), -1, EBADF); break; CASE_TEST(lseek_0); EXPECT_SYSER(1, lseek(0, 0, SEEK_SET), -1, ESPIPE); break;
Let's make the /tmp and /hugetlb directories and mount tmpfs and hugetlbfs respectively, they will be used to check the existing of tmpfs and hugetlbfs.
We will also use tmpfs to let some test cases run without procfs.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 69f59a395746..8b587961e46a 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -935,6 +935,18 @@ int prepare(void) } }
+ /* try to mount /tmp if not mounted. Silently fail otherwise */ + if (stat("/tmp/.", &stat_buf) == 0 || mkdir("/tmp", 0755) == 0) { + if (mount("none", "/tmp", "tmpfs", 0, 0) != 0) + rmdir("/tmp"); + } + + /* try to mount /hugetlb if not mounted. Silently fail otherwise */ + if (stat("/hugetlb/.", &stat_buf) == 0 || mkdir("/hugetlb", 0755) == 0) { + if (mount("none", "/hugetlb", "hugetlbfs", 0, 0) != 0) + rmdir("/hugetlb"); + } + return 0; }
When CONFIG_NET is not enabled, there would be no /proc/self/net, let's use /tmp/blah in such case and rename chmod_net to chmod_good.
This allows to test chmod_good with tmpfs even when procfs is not there.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 14 +++++++++++++- 1 file changed, 13 insertions(+), 1 deletion(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 8b587961e46a..eca0070151b6 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -550,10 +550,22 @@ int run_syscall(int min, int max) int ret = 0; void *p1, *p2; int has_gettid = 1; + int has_tmpdir = 0; + char *tmpdir = NULL;
/* <has_proc> indicates whether or not /proc is mounted */ has_proc = stat("/proc", &stat_buf) == 0;
+ /* <has_tmpdir> indicates whether or not /tmp/blah is there */ + if (stat("/proc/self/net", &stat_buf) == 0) { + tmpdir = "/proc/self/net"; + has_tmpdir = 1; + } else if (stat("/tmp/.", &stat_buf) == 0) { + tmpdir = "/tmp/blah"; + if (mkdir(tmpdir, 0755) == 0) + has_tmpdir = 1; + } + /* this will be used to skip certain tests that can't be run unprivileged */ is_root = geteuid() == 0;
@@ -582,7 +594,7 @@ int run_syscall(int min, int max) CASE_TEST(chdir_root); EXPECT_SYSZR(1, chdir("/")); break; CASE_TEST(chdir_dot); EXPECT_SYSZR(1, chdir(".")); break; CASE_TEST(chdir_blah); EXPECT_SYSER(1, chdir("/blah"), -1, ENOENT); break; - CASE_TEST(chmod_net); EXPECT_SYSZR(has_proc, chmod("/proc/self/net", 0555)); break; + CASE_TEST(chmod_good); EXPECT_SYSZR(has_tmpdir, chmod(tmpdir, 0555)); break; CASE_TEST(chmod_self); EXPECT_SYSER(has_proc, chmod("/proc/self", 0555), -1, EPERM); break; CASE_TEST(chown_self); EXPECT_SYSER(has_proc, chown("/proc/self", 0, 0), -1, EPERM); break; CASE_TEST(chroot_root); EXPECT_SYSZR(is_root, chroot("/")); break;
When CONFIG_NET is not enabled, there would be no /proc/self/net, use /tmp/blah in such case.
This allows to test link_cross with tmpfs even when procfs is not there.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index eca0070151b6..c900564219e8 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -620,7 +620,7 @@ int run_syscall(int min, int max) CASE_TEST(link_root1); EXPECT_SYSER(1, link("/", "/"), -1, EEXIST); break; CASE_TEST(link_blah); EXPECT_SYSER(1, link("/proc/self/blah", "/blah"), -1, ENOENT); break; CASE_TEST(link_dir); EXPECT_SYSER(is_root, link("/", "/blah"), -1, EPERM); break; - CASE_TEST(link_cross); EXPECT_SYSER(has_proc, link("/proc/self/net", "/blah"), -1, EXDEV); break; + CASE_TEST(link_cross); EXPECT_SYSER(has_tmpdir, link(tmpdir, "/blah"), -1, EXDEV); break; CASE_TEST(lseek_m1); EXPECT_SYSER(1, lseek(-1, 0, SEEK_SET), -1, EBADF); break; CASE_TEST(lseek_0); EXPECT_SYSER(1, lseek(0, 0, SEEK_SET), -1, ESPIPE); break; CASE_TEST(mkdir_root); EXPECT_SYSER(1, mkdir("/", 0755), -1, EEXIST); break;
When there is no procfs, let's use tmpfs and create a tmp file for chroot_exe test.
Since chroot_exe is mainly testing the not directory case (ENOTDIR), so, rename it to chroot_file may be better.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 15 ++++++++++++++- 1 file changed, 14 insertions(+), 1 deletion(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index c900564219e8..19e4ef5ce578 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -547,11 +547,14 @@ int run_syscall(int min, int max) int has_proc; int test; int tmp; + int fd; int ret = 0; void *p1, *p2; int has_gettid = 1; int has_tmpdir = 0; + int has_tmpfile = 0; char *tmpdir = NULL; + char *tmpfile = NULL;
/* <has_proc> indicates whether or not /proc is mounted */ has_proc = stat("/proc", &stat_buf) == 0; @@ -560,10 +563,20 @@ int run_syscall(int min, int max) if (stat("/proc/self/net", &stat_buf) == 0) { tmpdir = "/proc/self/net"; has_tmpdir = 1; + + tmpfile = "/proc/self/exe"; + has_tmpfile = 1; } else if (stat("/tmp/.", &stat_buf) == 0) { tmpdir = "/tmp/blah"; if (mkdir(tmpdir, 0755) == 0) has_tmpdir = 1; + + tmpfile = "/tmp/dummy"; + fd = open(tmpfile, O_CREAT); + if (fd != -1) { + has_tmpfile = 1; + close(fd); + } }
/* this will be used to skip certain tests that can't be run unprivileged */ @@ -599,7 +612,7 @@ int run_syscall(int min, int max) CASE_TEST(chown_self); EXPECT_SYSER(has_proc, chown("/proc/self", 0, 0), -1, EPERM); break; CASE_TEST(chroot_root); EXPECT_SYSZR(is_root, chroot("/")); break; CASE_TEST(chroot_blah); EXPECT_SYSER(1, chroot("/proc/self/blah"), -1, ENOENT); break; - CASE_TEST(chroot_exe); EXPECT_SYSER(has_proc, chroot("/proc/self/exe"), -1, ENOTDIR); break; + CASE_TEST(chroot_file); EXPECT_SYSER(has_tmpfile, chroot(tmpfile), -1, ENOTDIR); break; CASE_TEST(close_m1); EXPECT_SYSER(1, close(-1), -1, EBADF); break; CASE_TEST(close_dup); EXPECT_SYSZR(1, close(dup(0))); break; CASE_TEST(dup_0); tmp = dup(0); EXPECT_SYSNE(1, tmp, -1); close(tmp); break;
pass MFD_NOEXEC_SEAL flag to memfd_create() to silence this kernel warning inserted in the middle of the whole test result:
memfd_create() without MFD_EXEC nor MFD_NOEXEC_SEAL, pid=1
The mixed test result looks this:
Running test 'vfprintf' 0 emptymemfd_create() without MFD_EXEC nor MFD_NOEXEC_SEAL, pid=1 'init' "" = "" [OK]
From v6.2, MFD_NOEXEC_SEAL must be passed for the non-executable memfd.
Since MFD_NOEXEC_SEAL is a whole new flag, to avoid adding ugly #ifdef macros, let's use magic number here directly.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 19e4ef5ce578..8b1ce9911c5c 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -774,7 +774,17 @@ static int expect_vfprintf(int llen, size_t c, const char *expected, const char FILE *memfile; va_list args;
- fd = memfd_create("vfprintf", 0); + /* silence warning for kernel >= v6.2: + * + * "memfd_create() without MFD_EXEC nor MFD_NOEXEC_SEAL, pid=<pid>" + * + * try MFD_NOEXEC_SEAL (0x0008U) flag for kernels >= v6.2, error means + * the kernel is too old and require old flags + */ + fd = memfd_create("vfprintf", 0x0008U); + if (fd == -1) + fd = memfd_create("vfprintf", 0); + if (fd == -1) { pad_spc(llen, 64, "[FAIL]\n"); return 1;
As fs/Kconfig shows, MEMFD_CREATE depends on TMPFS or HUGETLBFS:
config MEMFD_CREATE def_bool TMPFS || HUGETLBFS
Let's skip vfprintf test if they are not there.
The /tmp and /hugetlb directories have been created to mount tmpfs and hugetlbfs respectively, if they are not enabled in kernel configuration, neither /tmp nor /hugetlb will be created.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 8b1ce9911c5c..85fa64746cde 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -769,11 +769,22 @@ int run_stdlib(int min, int max)
static int expect_vfprintf(int llen, size_t c, const char *expected, const char *fmt, ...) { + struct stat stat_buf; int ret, fd, w, r; + int tmpfs = 0, hugetlbfs = 0; char buf[100]; FILE *memfile; va_list args;
+ /* memfd_create depends on tmpfs or hugetlbfs */ + tmpfs = stat("/tmp/.", &stat_buf) == 0; + hugetlbfs = stat("/hugetlb/.", &stat_buf) == 0; + + if (!tmpfs && !hugetlbfs) { + pad_spc(llen, 64, "[SKIPPED]\n"); + return 0; + } + /* silence warning for kernel >= v6.2: * * "memfd_create() without MFD_EXEC nor MFD_NOEXEC_SEAL, pid=<pid>"
When CONFIG_TMPFS not enabled, kernel will use the ramfs based tmpfs instead, but memfd_create doesn't work with such tmpfs, for this type of tmpfs, let's use it instead of memfd_create.
At the same time, let's support hugetlbfs too if there is one.
Signed-off-by: Zhangjin Wu falcon@tinylab.org --- tools/testing/selftests/nolibc/nolibc-test.c | 8 ++++++++ 1 file changed, 8 insertions(+)
diff --git a/tools/testing/selftests/nolibc/nolibc-test.c b/tools/testing/selftests/nolibc/nolibc-test.c index 85fa64746cde..bff72365a158 100644 --- a/tools/testing/selftests/nolibc/nolibc-test.c +++ b/tools/testing/selftests/nolibc/nolibc-test.c @@ -796,6 +796,14 @@ static int expect_vfprintf(int llen, size_t c, const char *expected, const char if (fd == -1) fd = memfd_create("vfprintf", 0);
+ /* memfd_create not work with ramfs based tmpfs, try tmpfs and hugetlbfs in order instead */ + if (fd == -1) { + if (tmpfs) + fd = open("/tmp/vfprintf", O_CREAT | O_TRUNC | O_RDWR); + else if (hugetlbfs) + fd = open("/hugetlb/vfprintf", O_CREAT | O_TRUNC | O_RDWR); + } + if (fd == -1) { pad_spc(llen, 64, "[FAIL]\n"); return 1;
Hi Zhangjin,
some general comments for the whole series.
On 2023-06-21 20:52:30+0800, Zhangjin Wu wrote:
Hi, Willy
This patchset mainly allows speed up the nolibc test with a minimal kernel config.
As the nolibc supported architectures become more and more, the 'run' test with DEFCONFIG may cost several hours, which is not friendly to develop testing and even for release testing, so, smaller kernel configs may be required, and firstly, we should let nolibc-test work with less kernel config options, this patchset aims to this goal.
This patchset mainly remove the dependency from procfs, tmpfs, net and memfd_create, many failures have been fixed up.
When CONFIG_TMPFS and CONFIG_SHMEM are disabled, kernel will provide a ramfs based tmpfs (mm/shmem.c), it will be used as a choice to fix up some failures and also allow skip less tests.
Did you look into how much this duplicates from the kernels already existing "tinyconfig" and "kvm_guest.config" functionality?
And it would be interesting how much impact the enablement of procfs, tmpfs, net and memfd_create has in constrast to the minimal configuration. It seems unfortunate to me to complicate the testsuite to handle such uncommon scenarios.
Besides, it also adds musl support, improves glibc support and fixes up a kernel cmdline passing use case.
This is based on the dev.2023.06.14a branch of linux-rcu [1], all of the supported architectures are tested (with local minimal configs, [5] pasted the one for i386) without failures:
arch/board | result ------------|------------ arm/vexpress-a9 | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/arm-vexpress-a9-nolibc-test.log aarch64/virt | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/aarch64-virt-nolibc-test.log ppc/g3beige | not supported i386/pc | 136 test(s) passed, 3 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/i386-pc-nolibc-test.log x86_64/pc | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/x86_64-pc-nolibc-test.log mipsel/malta | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/mipsel-malta-nolibc-test.log loongarch64/virt | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/loongarch64-virt-nolibc-test.log riscv64/virt | 136 test(s) passed, 3 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/riscv64-virt-nolibc-test.log riscv32/virt | no test log founds390x/s390-ccw-virtio | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/s390x-s390-ccw-virtio-nolibc-test.log
Notes:
The skipped ones are -fstackprotector, chmod_self and chown_self
The -fstackprotector skip is due to gcc version. chmod_self and chmod_self skips are due to procfs not enabled
ppc/g3beige support is added locally, but not added in this patchset
will send ppc support as a new patchset, it depends on v2 test report patchset [3] and the v5 rv32 support, require changes on Makefile
riscv32/virt support is still in review, see v5 rv32 support [4]
This patchset doesn't depends on any of my other nolibc patch series, but the new rmdir() routine added in this patchset may be requird to apply the __sysret() from our v4 syscall helper series [2] after that series being merged, currently, we use the old method to let it compile without any dependency.
Here explains all of the patches:
selftests/nolibc: stat_fault: silence NULL argument warning with glibc selftests/nolibc: gettid: restore for glibc and musl selftests/nolibc: add _LARGEFILE64_SOURCE for musl
The above 3 patches adds musl compile support and improve glibc support.
It is able to build and run nolibc-test with musl libc now, but there are some failures/skips due to the musl its own issues/requirements:
$ sudo ./nolibc-test | grep -E 'FAIL|SKIP' 8 sbrk = 1 ENOMEM [FAIL] 9 brk = -1 ENOMEM [FAIL] 46 limit_int_fast16_min = -2147483648 [FAIL] 47 limit_int_fast16_max = 2147483647 [FAIL] 49 limit_int_fast32_min = -2147483648 [FAIL] 50 limit_int_fast32_max = 2147483647 [FAIL] 0 -fstackprotector not supported [SKIPPED]
musl disabled sbrk and brk for some conflicts with its malloc and the fast version of int types are defined in 32bit, which differs from nolibc and glibc. musl reserved the sbrk(0) to allow get current brk, we added a test for this in the v4 __sysret() helper series [2].
We could add new macros
#define UINT_MAX(t) (~(t)0) #define SINT_MAX(t) (((t)1 << (sizeof(t) * 8 - 2)) - (t)1 + ((t)1 << (sizeof(t) * 8 - 2)))
to get whatever is appropriate for the respective type.
selftests/nolibc: fix up kernel parameters support
kernel cmdline allows pass two types of parameters, one is without '=', another is with '=', the first one is passed as init arguments, the sencond one is passed as init environment variables.
Our nolibc-test prefer arguments to environment variables, this not work when users add such parameters in the kernel cmdline:
noapic NOLIBC_TEST=syscall
So, this patch will verify the setting from arguments at first, if it is no valid, will try the environment variables instead.
This would be much simpler as:
test = getenv("NOLIBC_TEST"); if (!test) test = argv[1];
It changes the semantics a bit, but it doesn't seem to be an issue. (Maybe gated behind getpid() == 1).
selftests/nolibc: stat_timestamps: remove procfs dependency
Use '/' instead of /proc/self, or we can add a 'has_proc' condition for this test case, but it is not that necessary to skip the whole stat_timestamps tests for such a subtest binding to /proc/self.
Welcome suggestion from Thomas.
As above, I think the impact of depending on CONFIG_PROC_FS is justifiable.
The usage of /proc/self was actually intentional. This file has a timestamp of the start of the referenced process. So each invocation of nolibc-test tests a new timestamp.
In contrast if nolibc-test is invocated from a prebaked filesystem the timestamp of "/" will always be fixed, reducing the chance to find errors.
tools/nolibc: add rmdir() support selftests/nolibc: add a new rmdir() test case
rmdir() routine and test case are added for the coming requirement.
Note, if the __sysret() patchset [2] is applied before us, this patch should be rebased on it and apply the __sysret() helper.
selftests/nolibc: fix up failures when there is no procfs
call rmdir() to remove /proc completely to rework the checking of /proc, before, the existing of /proc not means the procfs is really mounted.
selftests/nolibc: rename proc variable to has_proc selftests/nolibc: rename euid0 variable to is_root
align with the has_gettid, has_xxx variables.
selftests/nolibc: prepare tmpfs and hugetlbfs selftests/nolibc: rename chmod_net to chmod_good selftests/nolibc: link_cross: support tmpfs selftests/nolibc: rename chroot_exe to chroot_file
use file from /tmp instead of file from /proc when there is no procfs this avoid skipping the chmod_net, link_cross, chroot_exe tests
selftests/nolibc: vfprintf: silence memfd_create() warning selftests/nolibc: vfprintf: skip if neither tmpfs nor hugetlbfs selftests/nolibc: vfprintf: support tmpfs and hugetlbfs
memfd_create from kernel >= v6.2 forcely warn on missing MFD_NOEXEC_SEAL flag, the first one silence it with such flag, for older kernels, use 0 flag as before.
Given this is only a problem when nolibc-test is PID1 and printing to the system console, we could also just disable warnings on the system console through syscall() or /proc/sys/kernel/printk.
It would also avoid cluttering the tests for limited gain.
since memfd_create() depends on TMPFS or HUGETLBFS, the second one skip the whole vfprintf instead of simply fail if memfd_create() not work.
the 3rd one futher try the ramfs based tmpfs even when memfd_create() not work.
At last, let's simply discuss about the configs, I have prepared minimal configs for all of the nolibc supported architectures but not sure where should we put them, what about tools/testing/selftests/nolibc/configs ?
Thanks!
Best regards, Zhangjin
Zhangjin Wu (17): selftests/nolibc: stat_fault: silence NULL argument warning with glibc selftests/nolibc: gettid: restore for glibc and musl selftests/nolibc: add _LARGEFILE64_SOURCE for musl selftests/nolibc: fix up kernel parameters support selftests/nolibc: stat_timestamps: remove procfs dependency tools/nolibc: add rmdir() support selftests/nolibc: add a new rmdir() test case selftests/nolibc: fix up failures when there is no procfs selftests/nolibc: rename proc variable to has_proc selftests/nolibc: rename euid0 variable to is_root selftests/nolibc: prepare tmpfs and hugetlbfs selftests/nolibc: rename chmod_net to chmod_good selftests/nolibc: link_cross: support tmpfs selftests/nolibc: rename chroot_exe to chroot_file selftests/nolibc: vfprintf: silence memfd_create() warning selftests/nolibc: vfprintf: skip if neither tmpfs nor hugetlbfs selftests/nolibc: vfprintf: support tmpfs and hugetlbfs
tools/include/nolibc/sys.h | 28 ++++ tools/testing/selftests/nolibc/nolibc-test.c | 132 +++++++++++++++---- 2 files changed, 138 insertions(+), 22 deletions(-)
-- 2.25.1
Hi, Thomas
Hi Zhangjin,
some general comments for the whole series.
On 2023-06-21 20:52:30+0800, Zhangjin Wu wrote:
Hi, Willy
This patchset mainly allows speed up the nolibc test with a minimal kernel config.
As the nolibc supported architectures become more and more, the 'run' test with DEFCONFIG may cost several hours, which is not friendly to develop testing and even for release testing, so, smaller kernel configs may be required, and firstly, we should let nolibc-test work with less kernel config options, this patchset aims to this goal.
This patchset mainly remove the dependency from procfs, tmpfs, net and memfd_create, many failures have been fixed up.
When CONFIG_TMPFS and CONFIG_SHMEM are disabled, kernel will provide a ramfs based tmpfs (mm/shmem.c), it will be used as a choice to fix up some failures and also allow skip less tests.
Did you look into how much this duplicates from the kernels already existing "tinyconfig" and "kvm_guest.config" functionality?
Very good question and suggestion, thanks. it is just between tinyconfig and kvm_guest.config, the former is not enough, the later provides more. tinyconfig may be a very good base for us.
Just tested some architectures, based on tinyconfig, seems we only need to enable very few options, for example, TTY, PRINTK, CONSOLE related and target virtual board related options, but it requires more time to just list the required options.
The 'minimal' ones I have prepared were shrunk from the 'defconfig', now we need to add options from 'tinyconfig', with allnoconfig, it should be smaller and therefore faster ;-)
Based on my local powerpc porting, I have prepared some changes like this:
# extra kernel configs by architecture EXTCONFIG_powerpc = --enable SERIAL_PMACZILOG --enable CONFIG_SERIAL_PMACZILOG_CONSOLE EXTCONFIG = --set-str CONFIG_INITRAMFS_SOURCE $(CURDIR)/initramfs $(EXTCONFIG_$(ARCH))
...
menuconfig: $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) menuconfig
extconfig: $(Q)$(srctree)/scripts/config --file $(srctree)/.config $(EXTCONFIG) $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) olddefconfig
kernel: initramfs extconfig $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) $(IMAGE_NAME)
'menuconfig' is added for development, for example, find why something not work and add the missing options.
'extconfig' is added to enable additional options (before, based on defconfig) to let nolibc-test happy (for powerpc, add missing console options which has been added as modules in default config).
Based on your suggestion, this may be a good new target:
tinyconfig: $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) mrproper tinyconfig prepare
And this one, use 'allnoconfig' instead of 'olddefconfig':
extconfig: $(Q)$(srctree)/scripts/config --file $(srctree)/.config $(EXTCONFIG) $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) KCONFIG_ALLCONFIG=$(srctree)/.config allnoconfig
So, the new 'tinyconfig' may function as the smallest test environment, for faster compile and as a boundary test of nolibc-test itself.
But again, still need time to list the minimally required options, if they are few, listing them in the EXTCONFIG_<ARCH> line may be acceptable, but if the options are 'huge', standalone nolibc.config may be required, let's wait for one or two days.
And it would be interesting how much impact the enablement of procfs, tmpfs, net and memfd_create has in constrast to the minimal configuration.
For the test speed (mainly kernel compile) itself, when for one architecture on a very good test host, the time cost increment is very little (see below), but it does save some, especially when for lots of architectures ;-)
Comparing the rv64 testing speed on a Ubuntu 20.04 over '4G Mem + 4 Cores of i7-8550U CPU @ 1.80GHz' Virtual Machne, this time include:
* nolibc-test sysroot install + build * kernel config + build * qemu boot with opensbi + u-boot + kernel v6.4-rcx
Testing results:
'minimal':
arch/board | result ------------|------------ riscv64/virt | 136 test(s) passed, 3 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/riscv64-virt-nolibc-test.log
LOG: see all results for all boards in /labs/linux-lab/logging/nolibc/nolibc-test.log
real 1m57.395s user 4m50.002s sys 1m0.866s
'minimal' + procfs, net, shmem/tmpfs, devtmpfs/devmtmpfs_mount ...:
arch/board | result ------------|------------ riscv64/virt | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/riscv64-virt-nolibc-test.log
LOG: see all results for all boards in /labs/linux-lab/logging/nolibc/nolibc-test.log
real 2m17.812s user 6m4.695s sys 1m7.061s
It did save 20s (~17.1%) for us, not too much, but really faster.
It seems unfortunate to me to complicate the testsuite to handle such uncommon scenarios.
Yeah, such a config is not common, but as explained above, beside the compile speedup improvement, it is really a good boundary test environment for nolibc-test itself to make sure it work (no failure, less skips) at an extremely worst-case scene, although our changes looks many, but every one is as simple as CLOC ;-)
And that also means, nolibc is able to run with a very 'tiny' kernel config and users could reuse our config fragments and add their own for their embedded devices.
Besides, it also adds musl support, improves glibc support and fixes up a kernel cmdline passing use case.
This is based on the dev.2023.06.14a branch of linux-rcu [1], all of the supported architectures are tested (with local minimal configs, [5] pasted the one for i386) without failures:
arch/board | result ------------|------------ arm/vexpress-a9 | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/arm-vexpress-a9-nolibc-test.log aarch64/virt | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/aarch64-virt-nolibc-test.log ppc/g3beige | not supported i386/pc | 136 test(s) passed, 3 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/i386-pc-nolibc-test.log x86_64/pc | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/x86_64-pc-nolibc-test.log mipsel/malta | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/mipsel-malta-nolibc-test.log loongarch64/virt | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/loongarch64-virt-nolibc-test.log riscv64/virt | 136 test(s) passed, 3 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/riscv64-virt-nolibc-test.log riscv32/virt | no test log founds390x/s390-ccw-virtio | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/s390x-s390-ccw-virtio-nolibc-test.log
(snipped)
It is able to build and run nolibc-test with musl libc now, but there are some failures/skips due to the musl its own issues/requirements:
$ sudo ./nolibc-test | grep -E 'FAIL|SKIP' 8 sbrk = 1 ENOMEM [FAIL] 9 brk = -1 ENOMEM [FAIL] 46 limit_int_fast16_min = -2147483648 [FAIL] 47 limit_int_fast16_max = 2147483647 [FAIL] 49 limit_int_fast32_min = -2147483648 [FAIL] 50 limit_int_fast32_max = 2147483647 [FAIL] 0 -fstackprotector not supported [SKIPPED]musl disabled sbrk and brk for some conflicts with its malloc and the fast version of int types are defined in 32bit, which differs from nolibc and glibc. musl reserved the sbrk(0) to allow get current brk, we added a test for this in the v4 __sysret() helper series [2].
We could add new macros
#define UINT_MAX(t) (~(t)0) #define SINT_MAX(t) (((t)1 << (sizeof(t) * 8 - 2)) - (t)1 + ((t)1 << (sizeof(t) * 8 - 2)))
to get whatever is appropriate for the respective type.
They work perfectly, thanks:
/* for fast int test cases in stdlib test, musl use 32bit fast int */ #undef UINT_MAX #define UINT_MAX(t) (~(t)0) #undef SINT_MAX #define SINT_MAX(t) (((t)1 << (sizeof(t) * 8 - 2)) - (t)1 + ((t)1 << (sizeof(t) * 8 - 2))) #undef SINT_MIN #define SINT_MIN(t) (-SINT_MAX(t) - 1)
...
CASE_TEST(limit_int_fast16_min); EXPECT_EQ(1, INT_FAST16_MIN, (int_fast16_t) SINT_MIN(int_fast16_t)); break; CASE_TEST(limit_int_fast16_max); EXPECT_EQ(1, INT_FAST16_MAX, (int_fast16_t) SINT_MAX(int_fast16_t)); break; CASE_TEST(limit_uint_fast16_max); EXPECT_EQ(1, UINT_FAST16_MAX, (uint_fast16_t) UINT_MAX(uint_fast16_t)); break; CASE_TEST(limit_int_fast32_min); EXPECT_EQ(1, INT_FAST32_MIN, (int_fast32_t) SINT_MIN(int_fast32_t)); break; CASE_TEST(limit_int_fast32_max); EXPECT_EQ(1, INT_FAST32_MAX, (int_fast32_t) SINT_MAX(int_fast32_t)); break; CASE_TEST(limit_uint_fast32_max); EXPECT_EQ(1, UINT_FAST32_MAX, (uint_fast32_t) UINT_MAX(uint_fast32_t)); break;
To avoid overriding the existing macros, perhaps we should add something like UINT_TYPE_MAX(t), SINT_TYPE_MAX(t) and SINT_TYPE_MIN(t) ?
selftests/nolibc: fix up kernel parameters support
kernel cmdline allows pass two types of parameters, one is without '=', another is with '=', the first one is passed as init arguments, the sencond one is passed as init environment variables.
Our nolibc-test prefer arguments to environment variables, this not work when users add such parameters in the kernel cmdline:
noapic NOLIBC_TEST=syscall
So, this patch will verify the setting from arguments at first, if it is no valid, will try the environment variables instead.
This would be much simpler as:
test = getenv("NOLIBC_TEST"); if (!test) test = argv[1];
It changes the semantics a bit, but it doesn't seem to be an issue. (Maybe gated behind getpid() == 1).
Cool suggestion, it looks really better:
if (getpid() == 1) { prepare(); /* kernel cmdline may pass: "noapic NOLIBC_TEST=syscall", * to init program: * * "noapic" as arguments, * "NOLIBC_TEST=syscall" as environment variables, * * to avoid getting null test in this case, parsing * environment variables at first. */ test = getenv("NOLIBC_TEST"); if (!test) test = argv[1]; } else { /* for normal nolibc-test program, prefer arguments */ test = argv[1]; if (!test) test = getenv("NOLIBC_TEST"); }
selftests/nolibc: stat_timestamps: remove procfs dependency
Use '/' instead of /proc/self, or we can add a 'has_proc' condition for this test case, but it is not that necessary to skip the whole stat_timestamps tests for such a subtest binding to /proc/self.
Welcome suggestion from Thomas.
As above, I think the impact of depending on CONFIG_PROC_FS is justifiable.
The usage of /proc/self was actually intentional. This file has a timestamp of the start of the referenced process. So each invocation of nolibc-test tests a new timestamp.
In contrast if nolibc-test is invocated from a prebaked filesystem the timestamp of "/" will always be fixed, reducing the chance to find errors.
Yeah, it is resonable to reserve this as /proc/self, for a balance, to avoid failure in the worst-case scene, let's add a 'has_proc' condition there instead, just like the other /proc/self based test cases do.
- tools/nolibc: add rmdir() support selftests/nolibc: add a new rmdir() test case
(snipped)
selftests/nolibc: vfprintf: silence memfd_create() warning selftests/nolibc: vfprintf: skip if neither tmpfs nor hugetlbfs selftests/nolibc: vfprintf: support tmpfs and hugetlbfs
memfd_create from kernel >= v6.2 forcely warn on missing MFD_NOEXEC_SEAL flag, the first one silence it with such flag, for older kernels, use 0 flag as before.
Given this is only a problem when nolibc-test is PID1 and printing to the system console, we could also just disable warnings on the system console through syscall() or /proc/sys/kernel/printk.
Ok, I did think about disabling console for this call, but I was worried about the requirement of root (euid0) to do so, limiting it under PID1 may solve the root permission issue, but still need to find the right syscall to avoid the dependency of /proc/sys/kernel/printk, otherwise, to avoid failure for !procfs, the whole vfprintf will be skipped for such a warning, to be honest, it looks not a good direction.
It would also avoid cluttering the tests for limited gain.
Hmm, if consider the more code lines about disabling/enabling console and the dependency of /proc/sys/kernel/printk, I do prefer current change.
But I'm also interested in how the other applications developers to treat this warning, from the new kernel version side, we should use the latest non executable flags for security, but to let applications work with old kernels, we must support old flags, checking the kernel versions may be another choice.
Perhaps it's time for us to add the 'uname()' for nolibc, but the version comparing may be not that easy when we are in c context ;-)
https://www.man7.org/linux/man-pages/man2/uname.2.html
So, the current method may be still a 'balanced' solution, it tries supported flags from new kernel to old kernel to get a better and working memfd_create() without the version checking, is this cleaner?
int i; /* kernel >= v6.2 require MFD_NOEXEC_SEAL (0x0008U), but older ones not support this flag */ unsigned int flags[2] = {0x0008U, 0};
for (i = 0; i < 2; i ++) { /* try supported flags from new kernels to old kernels */ fd = memfd_create("vfprintf", flags[i]); if (fd != -1) break; }
if (fd == -1) { ... }
Thanks very much.
Best regards, Zhangjin
since memfd_create() depends on TMPFS or HUGETLBFS, the second one skip the whole vfprintf instead of simply fail if memfd_create() not work.
the 3rd one futher try the ramfs based tmpfs even when memfd_create() not work.
At last, let's simply discuss about the configs, I have prepared minimal configs for all of the nolibc supported architectures but not sure where should we put them, what about tools/testing/selftests/nolibc/configs ?
Thanks!
Best regards, Zhangjin
Zhangjin Wu (17): selftests/nolibc: stat_fault: silence NULL argument warning with glibc selftests/nolibc: gettid: restore for glibc and musl selftests/nolibc: add _LARGEFILE64_SOURCE for musl selftests/nolibc: fix up kernel parameters support selftests/nolibc: stat_timestamps: remove procfs dependency tools/nolibc: add rmdir() support selftests/nolibc: add a new rmdir() test case selftests/nolibc: fix up failures when there is no procfs selftests/nolibc: rename proc variable to has_proc selftests/nolibc: rename euid0 variable to is_root selftests/nolibc: prepare tmpfs and hugetlbfs selftests/nolibc: rename chmod_net to chmod_good selftests/nolibc: link_cross: support tmpfs selftests/nolibc: rename chroot_exe to chroot_file selftests/nolibc: vfprintf: silence memfd_create() warning selftests/nolibc: vfprintf: skip if neither tmpfs nor hugetlbfs selftests/nolibc: vfprintf: support tmpfs and hugetlbfs
tools/include/nolibc/sys.h | 28 ++++ tools/testing/selftests/nolibc/nolibc-test.c | 132 +++++++++++++++---- 2 files changed, 138 insertions(+), 22 deletions(-)
-- 2.25.1
On 2023-06-23 02:45:59+0800, Zhangjin Wu wrote:
some general comments for the whole series.
On 2023-06-21 20:52:30+0800, Zhangjin Wu wrote:
Hi, Willy
This patchset mainly allows speed up the nolibc test with a minimal kernel config.
As the nolibc supported architectures become more and more, the 'run' test with DEFCONFIG may cost several hours, which is not friendly to develop testing and even for release testing, so, smaller kernel configs may be required, and firstly, we should let nolibc-test work with less kernel config options, this patchset aims to this goal.
This patchset mainly remove the dependency from procfs, tmpfs, net and memfd_create, many failures have been fixed up.
When CONFIG_TMPFS and CONFIG_SHMEM are disabled, kernel will provide a ramfs based tmpfs (mm/shmem.c), it will be used as a choice to fix up some failures and also allow skip less tests.
Did you look into how much this duplicates from the kernels already existing "tinyconfig" and "kvm_guest.config" functionality?
Very good question and suggestion, thanks. it is just between tinyconfig and kvm_guest.config, the former is not enough, the later provides more. tinyconfig may be a very good base for us.
Just tested some architectures, based on tinyconfig, seems we only need to enable very few options, for example, TTY, PRINTK, CONSOLE related and target virtual board related options, but it requires more time to just list the required options.
The 'minimal' ones I have prepared were shrunk from the 'defconfig', now we need to add options from 'tinyconfig', with allnoconfig, it should be smaller and therefore faster ;-)
Based on my local powerpc porting, I have prepared some changes like this:
# extra kernel configs by architecture EXTCONFIG_powerpc = --enable SERIAL_PMACZILOG --enable CONFIG_SERIAL_PMACZILOG_CONSOLE EXTCONFIG = --set-str CONFIG_INITRAMFS_SOURCE $(CURDIR)/initramfs $(EXTCONFIG_$(ARCH)) ...also menuconfig: $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) menuconfig extconfig: $(Q)$(srctree)/scripts/config --file $(srctree)/.config $(EXTCONFIG) $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) olddefconfig kernel: initramfs extconfig $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) $(IMAGE_NAME)
'menuconfig' is added for development, for example, find why something not work and add the missing options.
'extconfig' is added to enable additional options (before, based on defconfig) to let nolibc-test happy (for powerpc, add missing console options which has been added as modules in default config).
Based on your suggestion, this may be a good new target:
tinyconfig: $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) mrproper tinyconfig prepareAnd this one, use 'allnoconfig' instead of 'olddefconfig':
extconfig: $(Q)$(srctree)/scripts/config --file $(srctree)/.config $(EXTCONFIG) $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) KCONFIG_ALLCONFIG=$(srctree)/.config allnoconfigSo, the new 'tinyconfig' may function as the smallest test environment, for faster compile and as a boundary test of nolibc-test itself.
But again, still need time to list the minimally required options, if they are few, listing them in the EXTCONFIG_<ARCH> line may be acceptable, but if the options are 'huge', standalone nolibc.config may be required, let's wait for one or two days.
FYI there are many more tests in tools/testing/selftests/ that need custom configs to run. Maybe we can reuse some of their configuration machinery. (And qemu machinery maybe)
And it would be interesting how much impact the enablement of procfs, tmpfs, net and memfd_create has in constrast to the minimal configuration.
For the test speed (mainly kernel compile) itself, when for one architecture on a very good test host, the time cost increment is very little (see below), but it does save some, especially when for lots of architectures ;-)
Comparing the rv64 testing speed on a Ubuntu 20.04 over '4G Mem + 4 Cores of i7-8550U CPU @ 1.80GHz' Virtual Machne, this time include:
- nolibc-test sysroot install + build
- kernel config + build
- qemu boot with opensbi + u-boot + kernel v6.4-rcx
Testing results:
'minimal':
arch/board | result ------------|------------ riscv64/virt | 136 test(s) passed, 3 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/riscv64-virt-nolibc-test.logLOG: see all results for all boards in /labs/linux-lab/logging/nolibc/nolibc-test.log real 1m57.395s user 4m50.002s sys 1m0.866s
'minimal' + procfs, net, shmem/tmpfs, devtmpfs/devmtmpfs_mount ...:
arch/board | result ------------|------------ riscv64/virt | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/riscv64-virt-nolibc-test.logLOG: see all results for all boards in /labs/linux-lab/logging/nolibc/nolibc-test.log real 2m17.812s user 6m4.695s sys 1m7.061s
It did save 20s (~17.1%) for us, not too much, but really faster.
It seems unfortunate to me to complicate the testsuite to handle such uncommon scenarios.
Yeah, such a config is not common, but as explained above, beside the compile speedup improvement, it is really a good boundary test environment for nolibc-test itself to make sure it work (no failure, less skips) at an extremely worst-case scene, although our changes looks many, but every one is as simple as CLOC ;-)
And that also means, nolibc is able to run with a very 'tiny' kernel config and users could reuse our config fragments and add their own for their embedded devices.
It would mean that nolibc-test is able to run on *really* trimmed down systems, which seems of limited use. If the testsuite has more dependencies it would not stop nolibc itself to run on them.
As for the CONFIG_NET dependency, which I would guess is one of the more expensive configs to enable:
link_cross can be easily adapted to instead use /proc/self.
chmod_net relies on /proc/$PID/net accepting chmod(). It is the only file in /proc/$PID/ that works that way.
Maybe its a kernel bug anyways and we shouldn't rely on it anyways? I'm taking a look.
Besides, it also adds musl support, improves glibc support and fixes up a kernel cmdline passing use case.
This is based on the dev.2023.06.14a branch of linux-rcu [1], all of the supported architectures are tested (with local minimal configs, [5] pasted the one for i386) without failures:
arch/board | result ------------|------------ arm/vexpress-a9 | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/arm-vexpress-a9-nolibc-test.log aarch64/virt | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/aarch64-virt-nolibc-test.log ppc/g3beige | not supported i386/pc | 136 test(s) passed, 3 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/i386-pc-nolibc-test.log x86_64/pc | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/x86_64-pc-nolibc-test.log mipsel/malta | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/mipsel-malta-nolibc-test.log loongarch64/virt | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/loongarch64-virt-nolibc-test.log riscv64/virt | 136 test(s) passed, 3 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/riscv64-virt-nolibc-test.log riscv32/virt | no test log founds390x/s390-ccw-virtio | 138 test(s) passed, 1 skipped, 0 failed. See all results in /labs/linux-lab/logging/nolibc/s390x-s390-ccw-virtio-nolibc-test.log
(snipped)
It is able to build and run nolibc-test with musl libc now, but there are some failures/skips due to the musl its own issues/requirements:
$ sudo ./nolibc-test | grep -E 'FAIL|SKIP' 8 sbrk = 1 ENOMEM [FAIL] 9 brk = -1 ENOMEM [FAIL] 46 limit_int_fast16_min = -2147483648 [FAIL] 47 limit_int_fast16_max = 2147483647 [FAIL] 49 limit_int_fast32_min = -2147483648 [FAIL] 50 limit_int_fast32_max = 2147483647 [FAIL] 0 -fstackprotector not supported [SKIPPED]musl disabled sbrk and brk for some conflicts with its malloc and the fast version of int types are defined in 32bit, which differs from nolibc and glibc. musl reserved the sbrk(0) to allow get current brk, we added a test for this in the v4 __sysret() helper series [2].
We could add new macros
#define UINT_MAX(t) (~(t)0) #define SINT_MAX(t) (((t)1 << (sizeof(t) * 8 - 2)) - (t)1 + ((t)1 << (sizeof(t) * 8 - 2)))
to get whatever is appropriate for the respective type.
They work perfectly, thanks:
/* for fast int test cases in stdlib test, musl use 32bit fast int */ #undef UINT_MAX #define UINT_MAX(t) (~(t)0) #undef SINT_MAX #define SINT_MAX(t) (((t)1 << (sizeof(t) * 8 - 2)) - (t)1 + ((t)1 << (sizeof(t) * 8 - 2))) #undef SINT_MIN #define SINT_MIN(t) (-SINT_MAX(t) - 1) ... CASE_TEST(limit_int_fast16_min); EXPECT_EQ(1, INT_FAST16_MIN, (int_fast16_t) SINT_MIN(int_fast16_t)); break; CASE_TEST(limit_int_fast16_max); EXPECT_EQ(1, INT_FAST16_MAX, (int_fast16_t) SINT_MAX(int_fast16_t)); break; CASE_TEST(limit_uint_fast16_max); EXPECT_EQ(1, UINT_FAST16_MAX, (uint_fast16_t) UINT_MAX(uint_fast16_t)); break; CASE_TEST(limit_int_fast32_min); EXPECT_EQ(1, INT_FAST32_MIN, (int_fast32_t) SINT_MIN(int_fast32_t)); break; CASE_TEST(limit_int_fast32_max); EXPECT_EQ(1, INT_FAST32_MAX, (int_fast32_t) SINT_MAX(int_fast32_t)); break; CASE_TEST(limit_uint_fast32_max); EXPECT_EQ(1, UINT_FAST32_MAX, (uint_fast32_t) UINT_MAX(uint_fast32_t)); break;To avoid overriding the existing macros, perhaps we should add something like UINT_TYPE_MAX(t), SINT_TYPE_MAX(t) and SINT_TYPE_MIN(t) ?
They should only be visible inside nolibc-test.c I think. But yes the UINT_MAX naming is bad.
Also when going away from testing constants maybe we can get back some test strength by validating the sizeof() of the datatypes.
<snip>
selftests/nolibc: vfprintf: silence memfd_create() warning selftests/nolibc: vfprintf: skip if neither tmpfs nor hugetlbfs selftests/nolibc: vfprintf: support tmpfs and hugetlbfs
memfd_create from kernel >= v6.2 forcely warn on missing MFD_NOEXEC_SEAL flag, the first one silence it with such flag, for older kernels, use 0 flag as before.
Given this is only a problem when nolibc-test is PID1 and printing to the system console, we could also just disable warnings on the system console through syscall() or /proc/sys/kernel/printk.
Ok, I did think about disabling console for this call, but I was worried about the requirement of root (euid0) to do so, limiting it under PID1 may solve the root permission issue, but still need to find the right syscall to avoid the dependency of /proc/sys/kernel/printk, otherwise, to avoid failure for !procfs, the whole vfprintf will be skipped for such a warning, to be honest, it looks not a good direction.
This should work:
syslog(__NR_syslog, 6 /* SYSLOG_ACTION_CONSOLE_OFF */);
It would also avoid cluttering the tests for limited gain.
Hmm, if consider the more code lines about disabling/enabling console and the dependency of /proc/sys/kernel/printk, I do prefer current change.
It should really only be the single line above.
But I'm also interested in how the other applications developers to treat this warning, from the new kernel version side, we should use the latest non executable flags for security, but to let applications work with old kernels, we must support old flags, checking the kernel versions may be another choice.
I know that systemd does it the same way as you proposed it, with non-negligible code overhead.
But for nolibc-test I really don't see any security issue.
Perhaps it's time for us to add the 'uname()' for nolibc, but the version comparing may be not that easy when we are in c context ;-)
Please no :-)
So, the current method may be still a 'balanced' solution, it tries supported flags from new kernel to old kernel to get a better and working memfd_create() without the version checking, is this cleaner?
int i; /* kernel >= v6.2 require MFD_NOEXEC_SEAL (0x0008U), but older ones not support this flag */
It is not required, only desired. The functionality still works as expected. I don't think the "old" way can ever stop working as it would break userspace ABI.
unsigned int flags[2] = {0x0008U, 0};
for (i = 0; i < 2; i ++) {
Loops like this should use ARRAY_SIZE() to calculate the termination condition.
/* try supported flags from new kernels to old kernels */ fd = memfd_create("vfprintf", flags[i]); if (fd != -1) break;}
if (fd == -1) { ... }
<snip>
Thomas
On 2023-06-24 08:52:55+0200, Thomas Weißschuh wrote:
As for the CONFIG_NET dependency, which I would guess is one of the more expensive configs to enable:
link_cross can be easily adapted to instead use /proc/self.
chmod_net relies on /proc/$PID/net accepting chmod(). It is the only file in /proc/$PID/ that works that way.
Maybe its a kernel bug anyways and we shouldn't rely on it anyways? I'm taking a look.
It indeed seems to be a kernel bug. The following patch aligns /proc/$PID/net with all the other /proc/$PID stuff.
diff --git a/fs/proc/proc_net.c b/fs/proc/proc_net.c index a0c0419872e3..8c5e9abf4380 100644 --- a/fs/proc/proc_net.c +++ b/fs/proc/proc_net.c @@ -320,6 +320,7 @@ static int proc_tgid_net_getattr(struct mnt_idmap *idmap,
const struct inode_operations proc_net_inode_operations = { .lookup = proc_tgid_net_lookup, + .setattr = proc_setattr, .getattr = proc_tgid_net_getattr, };
I'm not entirely sure about the process to synchronize the application of the fix in the procfs tree and the fix/removal of the testcase in the nolibc tree so we avoid broken states. Or if this would technically be a (relevant) break of userspace ABI and therefore has to stay as it is.
Any ideas?
Thomas
Hi, Thomas
On 2023-06-23 02:45:59+0800, Zhangjin Wu wrote:
some general comments for the whole series.
On 2023-06-21 20:52:30+0800, Zhangjin Wu wrote:
Hi, Willy
This patchset mainly allows speed up the nolibc test with a minimal kernel config.
(snip)
Based on my local powerpc porting, I have prepared some changes like this:
# extra kernel configs by architecture EXTCONFIG_powerpc = --enable SERIAL_PMACZILOG --enable CONFIG_SERIAL_PMACZILOG_CONSOLE EXTCONFIG = --set-str CONFIG_INITRAMFS_SOURCE $(CURDIR)/initramfs $(EXTCONFIG_$(ARCH)) ...also menuconfig: $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) menuconfig extconfig: $(Q)$(srctree)/scripts/config --file $(srctree)/.config $(EXTCONFIG) $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) olddefconfig kernel: initramfs extconfig $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) $(IMAGE_NAME)
'menuconfig' is added for development, for example, find why something not work and add the missing options.
'extconfig' is added to enable additional options (before, based on defconfig) to let nolibc-test happy (for powerpc, add missing console options which has been added as modules in default config).
Based on your suggestion, this may be a good new target:
tinyconfig: $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) mrproper tinyconfig prepareAnd this one, use 'allnoconfig' instead of 'olddefconfig':
extconfig: $(Q)$(srctree)/scripts/config --file $(srctree)/.config $(EXTCONFIG) $(Q)$(MAKE) -C $(srctree) ARCH=$(KARCH) CC=$(CC) CROSS_COMPILE=$(CROSS_COMPILE) KCONFIG_ALLCONFIG=$(srctree)/.config allnoconfigSo, the new 'tinyconfig' may function as the smallest test environment, for faster compile and as a boundary test of nolibc-test itself.
But again, still need time to list the minimally required options, if they are few, listing them in the EXTCONFIG_<ARCH> line may be acceptable, but if the options are 'huge', standalone nolibc.config may be required, let's wait for one or two days.
FYI there are many more tests in tools/testing/selftests/ that need custom configs to run. Maybe we can reuse some of their configuration machinery. (And qemu machinery maybe)
Yeah, thanks, these files may be very good references:
$ find tools/testing/selftests/ -name "*config" | grep qemu tools/testing/selftests/wireguard/qemu/arch/aarch64.config tools/testing/selftests/wireguard/qemu/arch/aarch64_be.config tools/testing/selftests/wireguard/qemu/arch/arm.config tools/testing/selftests/wireguard/qemu/arch/armeb.config tools/testing/selftests/wireguard/qemu/arch/i686.config tools/testing/selftests/wireguard/qemu/arch/m68k.config tools/testing/selftests/wireguard/qemu/arch/mips.config tools/testing/selftests/wireguard/qemu/arch/mips64.config tools/testing/selftests/wireguard/qemu/arch/mips64el.config tools/testing/selftests/wireguard/qemu/arch/mipsel.config tools/testing/selftests/wireguard/qemu/arch/powerpc.config tools/testing/selftests/wireguard/qemu/arch/powerpc64.config tools/testing/selftests/wireguard/qemu/arch/powerpc64le.config tools/testing/selftests/wireguard/qemu/arch/riscv32.config tools/testing/selftests/wireguard/qemu/arch/riscv64.config tools/testing/selftests/wireguard/qemu/arch/s390x.config tools/testing/selftests/wireguard/qemu/arch/um.config tools/testing/selftests/wireguard/qemu/arch/x86_64.config tools/testing/selftests/wireguard/qemu/debug.config tools/testing/selftests/wireguard/qemu/kernel.config
And I have prepared most of them, just left 2-3 architectures.
And it would be interesting how much impact the enablement of procfs, tmpfs, net and memfd_create has in constrast to the minimal configuration.
(snip)
It did save 20s (~17.1%) for us, not too much, but really faster.
It seems unfortunate to me to complicate the testsuite to handle such uncommon scenarios.
Yeah, such a config is not common, but as explained above, beside the compile speedup improvement, it is really a good boundary test environment for nolibc-test itself to make sure it work (no failure, less skips) at an extremely worst-case scene, although our changes looks many, but every one is as simple as CLOC ;-)
And that also means, nolibc is able to run with a very 'tiny' kernel config and users could reuse our config fragments and add their own for their embedded devices.
It would mean that nolibc-test is able to run on *really* trimmed down systems, which seems of limited use. If the testsuite has more dependencies it would not stop nolibc itself to run on them.
As for the CONFIG_NET dependency, which I would guess is one of the more expensive configs to enable:
link_cross can be easily adapted to instead use /proc/self.
Yes, we can simply use /proc/self or proc/self/cmdline.
chmod_net relies on /proc/$PID/net accepting chmod(). It is the only file in /proc/$PID/ that works that way.
Something like /tmp/xxx has been used in our patchset, to get a chmodable file, tmpfs is a good choice, even when the TMPFS (and SHMEM) option is disabled, a ramfs based tmpfs will be provided by kernel.
Maybe its a kernel bug anyways and we shouldn't rely on it anyways? I'm taking a look.
Good catch, I have been also interested in such a curious difference, because all of the other proc interfaces are not chmodable, not sure if it is 'intentional', let's discuss in your patch thread, perhaps we should cc the /proc/self/net maintainers/authors ;-)
(snipped)
It is able to build and run nolibc-test with musl libc now, but there are some failures/skips due to the musl its own issues/requirements:
$ sudo ./nolibc-test | grep -E 'FAIL|SKIP' 8 sbrk = 1 ENOMEM [FAIL] 9 brk = -1 ENOMEM [FAIL] 46 limit_int_fast16_min = -2147483648 [FAIL] 47 limit_int_fast16_max = 2147483647 [FAIL] 49 limit_int_fast32_min = -2147483648 [FAIL] 50 limit_int_fast32_max = 2147483647 [FAIL] 0 -fstackprotector not supported [SKIPPED]musl disabled sbrk and brk for some conflicts with its malloc and the fast version of int types are defined in 32bit, which differs from nolibc and glibc. musl reserved the sbrk(0) to allow get current brk, we added a test for this in the v4 __sysret() helper series [2].
We could add new macros
#define UINT_MAX(t) (~(t)0) #define SINT_MAX(t) (((t)1 << (sizeof(t) * 8 - 2)) - (t)1 + ((t)1 << (sizeof(t) * 8 - 2)))
to get whatever is appropriate for the respective type.
They work perfectly, thanks:
/* for fast int test cases in stdlib test, musl use 32bit fast int */ #undef UINT_MAX #define UINT_MAX(t) (~(t)0) #undef SINT_MAX #define SINT_MAX(t) (((t)1 << (sizeof(t) * 8 - 2)) - (t)1 + ((t)1 << (sizeof(t) * 8 - 2))) #undef SINT_MIN #define SINT_MIN(t) (-SINT_MAX(t) - 1) ... CASE_TEST(limit_int_fast16_min); EXPECT_EQ(1, INT_FAST16_MIN, (int_fast16_t) SINT_MIN(int_fast16_t)); break; CASE_TEST(limit_int_fast16_max); EXPECT_EQ(1, INT_FAST16_MAX, (int_fast16_t) SINT_MAX(int_fast16_t)); break; CASE_TEST(limit_uint_fast16_max); EXPECT_EQ(1, UINT_FAST16_MAX, (uint_fast16_t) UINT_MAX(uint_fast16_t)); break; CASE_TEST(limit_int_fast32_min); EXPECT_EQ(1, INT_FAST32_MIN, (int_fast32_t) SINT_MIN(int_fast32_t)); break; CASE_TEST(limit_int_fast32_max); EXPECT_EQ(1, INT_FAST32_MAX, (int_fast32_t) SINT_MAX(int_fast32_t)); break; CASE_TEST(limit_uint_fast32_max); EXPECT_EQ(1, UINT_FAST32_MAX, (uint_fast32_t) UINT_MAX(uint_fast32_t)); break;To avoid overriding the existing macros, perhaps we should add something like UINT_TYPE_MAX(t), SINT_TYPE_MAX(t) and SINT_TYPE_MIN(t) ?
They should only be visible inside nolibc-test.c I think.
Yeah, I did so.
But yes the UINT_MAX naming is bad.
Also when going away from testing constants maybe we can get back some test strength by validating the sizeof() of the datatypes.
It seems easier.
<snip>
selftests/nolibc: vfprintf: silence memfd_create() warning selftests/nolibc: vfprintf: skip if neither tmpfs nor hugetlbfs selftests/nolibc: vfprintf: support tmpfs and hugetlbfs
memfd_create from kernel >= v6.2 forcely warn on missing MFD_NOEXEC_SEAL flag, the first one silence it with such flag, for older kernels, use 0 flag as before.
Given this is only a problem when nolibc-test is PID1 and printing to the system console, we could also just disable warnings on the system console through syscall() or /proc/sys/kernel/printk.
Ok, I did think about disabling console for this call, but I was worried about the requirement of root (euid0) to do so, limiting it under PID1 may solve the root permission issue, but still need to find the right syscall to avoid the dependency of /proc/sys/kernel/printk, otherwise, to avoid failure for !procfs, the whole vfprintf will be skipped for such a warning, to be honest, it looks not a good direction.
This should work:
syslog(__NR_syslog, 6 /* SYSLOG_ACTION_CONSOLE_OFF */);
Thanks, will try it like this:
syslog(__NR_syslog, 6 /* SYSLOG_ACTION_CONSOLE_OFF */); memfd_create() syslog(__NR_syslog, 7 /* SYSLOG_ACTION_CONSOLE_ON */);
Putting it in prepare() may hide potential issues reported by kernel.
It would also avoid cluttering the tests for limited gain.
Hmm, if consider the more code lines about disabling/enabling console and the dependency of /proc/sys/kernel/printk, I do prefer current change.
It should really only be the single line above.
But I'm also interested in how the other applications developers to treat this warning, from the new kernel version side, we should use the latest non executable flags for security, but to let applications work with old kernels, we must support old flags, checking the kernel versions may be another choice.
I know that systemd does it the same way as you proposed it, with non-negligible code overhead.
But for nolibc-test I really don't see any security issue.
Yes, but at least as a good reference when people want to reuse some codes from nolibc-test.c ;-)
Perhaps it's time for us to add the 'uname()' for nolibc, but the version comparing may be not that easy when we are in c context ;-)
Please no :-)
So, the current method may be still a 'balanced' solution, it tries supported flags from new kernel to old kernel to get a better and working memfd_create() without the version checking, is this cleaner?
int i; /* kernel >= v6.2 require MFD_NOEXEC_SEAL (0x0008U), but older ones not support this flag */
It is not required, only desired. The functionality still works as expected. I don't think the "old" way can ever stop working as it would break userspace ABI.
unsigned int flags[2] = {0x0008U, 0};
for (i = 0; i < 2; i ++) {
Loops like this should use ARRAY_SIZE() to calculate the termination condition.
Yeah, it should be:
unsigned int flags[] = {0x0008U, 0};
for (i = 0; i < sizeof(flags)/sizeof(flags[0]); i ++) ...
Best regards, Zhangjin
/* try supported flags from new kernels to old kernels */ fd = memfd_create("vfprintf", flags[i]); if (fd != -1) break;}
if (fd == -1) { ... }
<snip>
Thomas
Hi, Thomas
Hi Zhangjin,
some general comments for the whole series.
On 2023-06-21 20:52:30+0800, Zhangjin Wu wrote:
Hi, Willy
This patchset mainly allows speed up the nolibc test with a minimal kernel config.
(snip)
selftests/nolibc: fix up kernel parameters support
kernel cmdline allows pass two types of parameters, one is without '=', another is with '=', the first one is passed as init arguments, the sencond one is passed as init environment variables.
Our nolibc-test prefer arguments to environment variables, this not work when users add such parameters in the kernel cmdline:
noapic NOLIBC_TEST=syscall
So, this patch will verify the setting from arguments at first, if it is no valid, will try the environment variables instead.
This would be much simpler as:
test = getenv("NOLIBC_TEST"); if (!test) test = argv[1];
It changes the semantics a bit, but it doesn't seem to be an issue. (Maybe gated behind getpid() == 1).
Cool suggestion, it looks really better:
if (getpid() == 1) { prepare(); /* kernel cmdline may pass: "noapic NOLIBC_TEST=syscall", * to init program: * * "noapic" as arguments, * "NOLIBC_TEST=syscall" as environment variables, * * to avoid getting null test in this case, parsing * environment variables at first. */ test = getenv("NOLIBC_TEST"); if (!test) test = argv[1]; } else { /* for normal nolibc-test program, prefer arguments */ test = argv[1]; if (!test) test = getenv("NOLIBC_TEST"); }
Test shows, when no NOLIBC_TEST environment variable passed to kernel cmdline, it will still branch to this code:
test = argv[1]; /* nopaic ... */
And therefore report the whole test is ignored and no test will be run:
Ignoring unknown test name 'noapic'
So, we may still need to verify it like my originally proposed method, but let's further verify the one from NOLIBC_TEST=, we should tune the code a litle.
Thanks, Zhangjin
linux-kselftest-mirror@lists.linaro.org
-
Thomas Weißschuh -
Zhangjin Wu