tls-slot allocation not thread-safe

[James Y Knight]

It seems as though the ALLOC-TLS-INDEX-IN-* code is not thread-safe. I cannot figure out how it's not thread-safe, because it looks correct to me, so far as I can see.

But this test case shows that it's not.

sbcl --eval '(load (compile-file "thread-test.lisp"))' --eval "(run-test)" --eval "(quit)"

It will fail with something like:
   The assertion (BOUNDP '*VAR50*) failed.
The variable number will vary randomly, of course.

SBCL 1.0.37.44. Debian Linux 2.6.32-amd64, x86-64.

[James Y Knight (foom)]

Okay, turns out to be an utterly trivial problem. This is the disassembly of SB-VM::ALLOC-TLS-INDEX-IN-RAX. See instructions 0x20000f42 and 0x20000f47. You can't use %rax as the new value and the compare-to value for cmpxchg! So, the lock was never actually taken. The variants for allocating in everything else worked fine. Just RAX is broken.

0x20000f38: mov %rbp,0xb8(%r12)
0x20000f40: push %rcx
0x20000f41: push %rax
0x20000f42: mov $0x1,%eax
0x20000f47: xor %eax,%eax
0x20000f49: lock cmpxchg %rax,0x20100b88
0x20000f53: jne 0x20000f42
0x20000f55: pop %rcx
0x20000f56: mov 0x21(%rcx),%rax
0x20000f5a: or %rax,%rax
0x20000f5d: jne 0x20000f8d
0x20000f5f: mov 0x20100b48,%rax
0x20000f67: cmp $0x8000,%rax
0x20000f6d: jl 0x20000f80
0x20000f6f: movq $0x0,0xb8(%r12)
0x20000f7b: jmpq 0x20001874
0x20000f80: addq $0x8,0x20100b48
0x20000f89: mov %rax,0x21(%rcx)
0x20000f8d: xor %ecx,%ecx
0x20000f8f: xchg %rcx,0x20100b88
0x20000f97: pop %rcx
0x20000f98: xor %rbp,0xb8(%r12)
0x20000fa0: je 0x20000fa4
0x20000fa2: int3
0x20000fa3: 0x09
0x20000fa4: ret

[James Y Knight (foom)]

Here's a fix.

[Nikodemus Siivola (nikodemus)]

In SBCL 1.0.39.16