[PATCH] linux: arch_prctl() system call

[Benoit Canet]

Needed by libraries produced by the Go compiler.

With this patch and the golang app the runtime
init crash later because of a vdso issue.

It can be examinated by insertion an abort()
before osv send SIGSEGV to the program.

(gdb) bt
at /usr/lib/golang/src/runtime/mheap.go:503
argc=<error reading variable: Cannot access memory at address 0xffffffffffffffcc>, argv=<optimized out>) at core/app.cc:338
Backtrace stopped: previous frame inner to this frame (corrupt stack?)
(gdb) frame 9
188 CALL AX
(gdb) lsit
Undefined command: "lsit". Try "help".
(gdb) list
183 RET
184 fallback:
185 LEAQ 0(SP), DI
186 MOVQ $0, SI
187 MOVQ runtime·__vdso_gettimeofday_sym(SB), AX
188 CALL AX
189 MOVQ 0(SP), AX // sec
190 MOVL 8(SP), DX // usec
191 IMULQ $1000, DX
192 // sec is in AX, nsec in DX
(gdb) p $ax
$1 = 0
(gdb)

Signed-off-by: Pekka Enberg <pen...@scylladb.com>
Signed-off-by: Benoît Canet <ben...@scylladb.com>
---
Makefile | 1 +
arch/x64/msr.hh | 1 +
arch/x64/prctl.cc | 39 +++++++++++++++++++++++++++++++++++++++
linux.cc | 3 +++
4 files changed, 44 insertions(+)
create mode 100644 arch/x64/prctl.cc

diff --git a/Makefile b/Makefile
index e294a03..4518da8 100644
--- a/Makefile
+++ b/Makefile
@@ -882,6 +882,7 @@ objects += arch/x64/ioapic.o
objects += arch/x64/apic.o
objects += arch/x64/apic-clock.o
objects += arch/x64/entry-xen.o
+objects += arch/x64/prctl.o
objects += arch/x64/xen.o
objects += arch/x64/xen_intr.o
objects += core/sampler.o
diff --git a/arch/x64/msr.hh b/arch/x64/msr.hh
index d77c75c..9aa8314 100644
--- a/arch/x64/msr.hh
+++ b/arch/x64/msr.hh
@@ -62,6 +62,7 @@ enum class msr : uint32_t {
IA32_LSTAR = 0xc0000082,
IA32_FMASK = 0xc0000084,
IA32_FS_BASE = 0xc0000100,
+ IA32_GS_BASE = 0xc0000101,

KVM_WALL_CLOCK = 0x11,
KVM_SYSTEM_TIME = 0x12,
diff --git a/arch/x64/prctl.cc b/arch/x64/prctl.cc
new file mode 100644
index 0000000..73ba7c7
--- /dev/null
+++ b/arch/x64/prctl.cc
@@ -0,0 +1,39 @@
+/*
+ * Copyright (C) 2014 Cloudius Systems, Ltd.
+ *
+ * This work is open source software, licensed under the terms of the
+ * BSD license as described in the LICENSE file in the top-level directory.
+ */
+
+#include "arch.hh"
+#include "msr.hh"
+#include "libc/libc.hh"
+
+#include <assert.h>
+#include <stdio.h>
+
+enum {
+ ARCH_SET_GS = 0x1001,
+ ARCH_SET_FS = 0x1002,
+ ARCH_GET_FS = 0x1003,
+ ARCH_GET_GS = 0x1004,
+};
+
+long arch_prctl(int code, unsigned long addr)
+{
+ switch (code) {
+ case ARCH_SET_GS:
+ processor::wrmsr(msr::IA32_GS_BASE, addr);
+ asm volatile ("movq %0, %%gs" :: "r"(addr));
+ break;
+ case ARCH_SET_FS:
+ processor::wrmsr(msr::IA32_FS_BASE, addr);
+ asm volatile ("movq %0, %%fs" :: "r"(addr));
+ break;
+ case ARCH_GET_FS:
+ return processor::rdmsr(msr::IA32_FS_BASE);
+ case ARCH_GET_GS:
+ return processor::rdmsr(msr::IA32_GS_BASE);
+ }
+ return libc_error(EINVAL);
+}
diff --git a/linux.cc b/linux.cc
index e78bb0c..03be002 100644
--- a/linux.cc
+++ b/linux.cc
@@ -297,6 +297,8 @@ int rt_sigprocmask(int how, sigset_t * nset, sigset_t * oset, size_t sigsetsize)
return sigprocmask(how, nset, oset);
}

+extern long arch_prctl(int code, unsigned long addr);
+
long syscall(long number, ...)
{
switch (number) {
@@ -319,6 +321,7 @@ long syscall(long number, ...)
SYSCALL2(munmap, void *, size_t);
SYSCALL4(rt_sigaction, int, const struct k_sigaction *, struct k_sigaction *, size_t);
SYSCALL4(rt_sigprocmask, int, sigset_t *, sigset_t *, size_t);
+ SYSCALL2(arch_prctl, int, unsigned long);
}

debug_always("syscall(): unimplemented system call %d\n", number);
--
2.7.4

[Nadav Har'El]

In an earlier review in https://groups.google.com/forum/#!msg/osv-dev/PW3bkaVCuMg/340vWUeimdMJ

and also today in another thread, I pointed out the problems with this arch_prctl thing which makes this patch either incomplete or even counter-productive:

1. After a thread does ARCH_SET_FS, if it get switched out and then back in, thread::switch_to() will overwrite this setting with a fixed pointer "thread->_tcb". So to support ARCH_SET_FS we need to save and restore this value on context switch (or more efficiently, save it only on ARCH_SET_FS, not on every switch out, and restore on switch in).

2. After a thread does ARCH_SET_FS, it cannot call any OSv code (or any library function in general). We need to modify the system-call code to restore the standard fs_base (the _tcb) when entering the system call, and restore it back to the thread's chosen address when exiting the system call.

3. It's even worse if after a thread does ARCH_SET_FS is calls a function directly, not through a system call. e.g., malloc(). It is possible (???) that Go is aware of this problem and restores the FS_BASE before calling C functions, but if it doesn't, we're in big problems.

--
Nadav Har'El

n...@scylladb.com

--
2.7.4

--
You received this message because you are subscribed to the Google Groups "OSv Development" group.
To unsubscribe from this group and stop receiving emails from it, send an email to osv-dev+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Nadav Har'El]

On Wed, Aug 31, 2016 at 4:04 PM, Benoit Canet <benoit.can...@gmail.com> wrote:

By the way, I think these "movq" things are wrong - I don't know what they are supposed to be doing?

 

--
2.7.4

[Waldek Kozaczuk]

Anybody has an idea if this patch is still relevant? None of the golang tests I have conducted exposed any issues possibly related to lack of arch_prctl SYSCALL.

To unsubscribe from this group and stop receiving emails from it, send an email to osv-dev+u...@googlegroups.com.

[Nadav Har'El]

On Mon, Jan 15, 2018 at 11:20 PM, Waldek Kozaczuk <jwkoz...@gmail.com> wrote:

Anybody has an idea if this patch is still relevant? None of the golang tests I have conducted exposed any issues possibly related to lack of arch_prctl SYSCALL.

I'm not sure it was ever relevant. Perhaps it was relevant to one Go version or one runtime mode (e.g., static compilation but not c-shared), but no longer is relevant to current Go and c-shared.

If you don't see any messages about this system call being missing, we can drop it.

 

To unsubscribe from this group and stop receiving emails from it, send an email to osv-dev+unsubscribe@googlegroups.com.