On 9/30/19 12:29 PM, Richard Henderson wrote: > This is a workaround for a ppc64le host kernel bug. > > For the test case linux-test, we have an instruction trace > > IN: sig_alarm > ... > > IN: > 0x400080ed28: 380000ac li r0, 0xac > 0x400080ed2c: 44000002 sc > > IN: __libc_nanosleep > 0x1003bb4c: 7c0802a6 mflr r0 > 0x1003bb50: f8010010 std r0, 0x10(r1) > > Our signal return trampoline has, rightly, changed the guest > stack page read-only. Which, rightly, faults on the store of > a return address into a stack frame. > > Checking the host /proc/pid/maps, we see the expected state: > > 4000800000-4000810000 r--p 00000000 00:00 0 > > However, the host kernel has supplied si_code == SEGV_MAPERR, > which is obviously incorrect. > > By dropping this check, we may have an extra walk of the page > tables, but this should be inexpensive. > > Signed-off-by: Richard Henderson <richard.hender...@linaro.org> > --- > > FWIW, filed as > > https://bugzilla.redhat.com/show_bug.cgi?id=1757189 > > out of habit and then > > https://bugs.centos.org/view.php?id=16499 > > when I remembered that the system is running Centos not RHEL. > > --- > accel/tcg/user-exec.c | 7 +++++-- > 1 file changed, 5 insertions(+), 2 deletions(-) > > diff --git a/accel/tcg/user-exec.c b/accel/tcg/user-exec.c > index 71c4bf6477..31ef091a70 100644 > --- a/accel/tcg/user-exec.c > +++ b/accel/tcg/user-exec.c > @@ -143,9 +143,12 @@ static inline int handle_cpu_signal(uintptr_t pc, > siginfo_t *info, > * for some other kind of fault that should really be passed to the > * guest, we'd end up in an infinite loop of retrying the faulting > * access. > + * > + * XXX: At least one host kernel, ppc64le w/Centos 7 4.14.0-115.6.1, > + * incorrectly reports SEGV_MAPERR for a STDX write to a read-only page. > + * Therefore, do not test info->si_code. > */ > - if (is_write && info->si_signo == SIGSEGV && info->si_code == > SEGV_ACCERR && > - h2g_valid(address)) { > + if (is_write && info->si_signo == SIGSEGV && h2g_valid(address)) {
Ho hum. This change is in conflict with Peter's long comment; I should have read the context more thoroughly. There is an even longer comment with the patch description: 9c4bbee9e3b83544257e82566342c29e15a88637 The SEGV_ACCERR check here is to prevent a loop by which page_unprotect races with itself and, from Peter's analysis, > * ...but when B gets the mmap lock it finds that the page is already > PAGE_WRITE, and so it exits page_unprotect() via the "not due to > protected translation" code path, and wrongly delivers the signal > to the guest rather than just retrying the access This bug was fixed in the referenced patch. But then continues: > Since this would cause an infinite loop if we ever called > page_unprotect() for some other kind of fault than "write failed due > to bad access permissions", tighten the condition in > handle_cpu_signal() to check the signal number and si_code, and add a > comment so that if somebody does ever find themselves debugging an > infinite loop of faults they have some clue about why. > > (The trick for identifying the correct setting for > current_tb_invalidated for thread B (needed to handle the precise-SMC > case) is due to Richard Henderson. Paolo Bonzini suggested just > relying on si_code rather than trying anything more complicated.) It is disappointing about the kernel bug. But since this affects Centos 7, which is what *all* of the gcc compile farm ppc64 machines use, I think we need to work around it somehow. Should we simply add SEGV_MAPERR to the set of allowed si_code, to directly work around the bug? If we got that code from a kernel without the bug, then page_find should fail to find an entry, and we should then indicate that the signal should be passed to the guest. Thoughts? r~
