CONFIG_IA64_NEW_UNWIND
I'm running 2.4.0 with a couple of patches on 2-CPU Bigsur.
I have a problem where a cpu apparently hangs when a process dumps core. The "hang" is caused when do_copy_regs() in arch/ia64/kernel/process.c picks up random values for RSE backing store addresses, and sits in this loop:
for (addr = pt->ar_bspstore; addr < ar_bsp; addr += 8)
if (ia64_peek(pt, current, addr, &val) = 0)
access_process_vm(current, addr, &val, sizeof(val),1);
for a really long time (last time I caught it "addr" was 0x000003e2_47fbf540 and "ar_bsp" was 0x00006000_00000000 ... every call to ia64_peek() failed with EIO ... but I calculated that it would have taken over six weeks to complete the loop).
Looking back at where do_copy_regs() digs these values out of the stack, I think that the problem lies in these lines:
unw_get_sp(info, &sp);
pt = (struct pt_regs *) (sp + 16);We pick up a perfectly reasonable "sp" in the first line, but I can't see why the code believes that that there would be a pt_regs structure 16 bytes further up. I think that the frame looks like this at this point
[high addresses]
frame for ia64_elf_core_copy_regs() <- "unw_get_sp(info, &sp)" switch_stack structure (pushed by unw_init_running) unw_frame_info structure (pushed by unw_init_running) [low addresses]
The CONFIG_IA64_NEW_UNWIND=n version seems to make more sense (it just uses the pt_user structure that was passed to ia64_elf_core_copy_regs()).
A quick scan of 2.4.2 shows no apparent changes here.
Any clues?
-Tony Luck
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line · Tony
Tony> I'm running 2.4.0 with a couple of patches on 2-CPU Bigsur. I Tony> have a problem where a cpu apparently hangs when a process Tony> dumps core. The "hang" is caused when do_copy_regs() in Tony> arch/ia64/kernel/process.c picks up random values for RSE Tony> backing store addresses, and sits in this loop:
Tony> for (addr = pt->ar_bspstore; addr < ar_bsp; addr += 8) Tony> if (ia64_peek(pt, current, addr, &val) = 0) Tony> access_process_vm(current, addr, &val, sizeof(val), 1);
Tony> for a really long time (last time I caught it "addr" was Tony> 0x000003e2_47fbf540 and "ar_bsp" was 0x00006000_00000000 Tony> ... every call to ia64_peek() failed with EIO ... but I Tony> calculated that it would have taken over six weeks to complete Tony> the loop).
That's a sign that the unwind info is incorrect. Are you by chance trying to use the 3.0 branch of gcc? There is a known bug in that compiler that's causing this problem.
Tony> Looking back at where do_copy_regs() digs these values out of Tony> the stack, I think that the problem lies in these lines:
Tony> unw_get_sp(info, &sp); pt = (struct pt_regs *) (sp +
Tony> 16);
Tony> We pick up a perfectly reasonable "sp" in the first line, but Tony> I can't see why the code believes that that there would be a Tony> pt_regs structure 16 bytes further up. I think that the frame Tony> looks like this at this point
You missed the call to unw_unwind_to_user(). It rewinds the stack to the pt-regs structure created on entry to the kernel.
--david