m68k: fix spinlock race in kernel thread creation
RFC PATCH 1/1 m68k: fix spinlock race in kernel thread creation
Context switching does take care to retain the correct lock owner across the switch from 'prev' to 'next' tasks. This does rely on interrupts remaining disabled for the entire duration of the switch.
This condition is guaranteed for normal process creation and context switching between already running processes, because both 'prev' and 'next' already have interrupts disabled in their saved copies of the status register.
The situation is different for newly created kernel threads. The status register is set to PS_S in copy_thread(), which does leave the IPL at 0. Upon restoring the 'next' thread's status register in switch_to() aka resume(), interrupts then become enabled prematurely. resume() then returns via ret_from_kernel_thread() and schedule_tail() where run queue lock is released (see finish_task_switch() and finish_lock_switch()).
A timer interrupt calling scheduler_tick() before the lock is released in finish_task_switch() will find the lock already taken, with the current task as lock owner. This causes a spinlock recursion warning as reported by Guenter Roeck.
As far as I can ascertain, this race has been opened in commit 533e6903bea0 ("m68k: split ret_from_fork(), simplify kernel_thread()") but I haven't done a detailed study of kernel history so it may well predate that commit.
Interrupts cannot be disabled in the saved status register copy for kernel threads (init will complain about interrupts disabled when finally starting user space). Disable interrupts temporarily when switching the tasks' register sets in resume().
Note that a simple oriw 0x700,%sr after restoring sr is not enough here - this leaves enough of a race for the 'spinlock recursion' warning to still be observed.
Tested on ARAnyM and qemu (Quadra 800 emulation).
--- arch/m68k/kernel/entry.S | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
arch/m68k/kernel/entry.S +3 -1
@@ -446,7 +446,9 @@ resume: movec %a0,%dfc /* restore status register */ - movew %a1@(TASK_THREAD+THREAD_SR),%sr + movew %a1@(TASK_THREAD+THREAD_SR),%d0 + oriw #0x0700,%d0 + movew %d0,%sr rts
Signature · 1 line
2.17.1
Quoted 39 lines · Michael Schmitz
Context switching does take care to retain the correct lock owner across the switch from 'prev' to 'next' tasks. This does rely on interrupts remaining disabled for the entire duration of the switch.
This condition is guaranteed for normal process creation and context switching between already running processes, because both 'prev' and 'next' already have interrupts disabled in their saved copies of the status register.
The situation is different for newly created kernel threads. The status register is set to
PS_Sincopy_thread(), which does leave the IPL at 0. Upon restoring the 'next' thread's status register inswitch_to()akaresume(), interrupts then become enabled prematurely.resume()then returns viaret_from_kernel_thread()andschedule_tail()where run queue lock is released (seefinish_task_switch()andfinish_lock_switch()).A timer interrupt calling
scheduler_tick()before the lock is released infinish_task_switch()will find the lock already taken, with the current task as lock owner. This causes a spinlock recursion warning as reported by Guenter Roeck.As far as I can ascertain, this race has been opened in commit
533e6903bea0("m68k: splitret_from_fork(), simplifykernel_thread()") but I haven't done a detailed study of kernel history so it may well predate that commit.Interrupts cannot be disabled in the saved status register copy for kernel threads (init will complain about interrupts disabled when finally starting user space). Disable interrupts temporarily when switching the tasks' register sets in
resume().Note that a simple oriw 0x700,%sr after restoring sr is not enough here - this leaves enough of a race for the 'spinlock recursion' warning to still be observed.
Tested on ARAnyM and qemu (Quadra 800 emulation).
Quoted 9 lines
Cc Guenter Roeck <[email protected]>Cc Geert Uytterhoeven <[email protected]>Cc Finn Thain <[email protected]>Cc Al Viro <[email protected]>Cc [email protected] # 3.6Fixes 533e6903bea0 ("m68k: split ret_from_fork(), simplify kernel_thread()")Signed-off-by Michael Schmitz <[email protected]>
With this patch applied, the problem was not observed over more than 1,500 test runs. Without this patch, the problem is seen roughly once in ~100 test runs. Tested with qemu's q800 emulation.
Guenter
Quoted 22 lines
--- arch/m68k/kernel/entry.S | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)arch/m68k/kernel/entry.S +3
-1@@ -446,7 +446,9 @@ resume: movec %a0,%dfc /* restore status register */ - movew %a1@(TASK_THREAD+THREAD_SR),%sr + movew %a1@(TASK_THREAD+THREAD_SR),%d0 + oriw #0x0700,%d0 + movew %d0,%sr rtsSignature · 1 line
2.17.1
Quoted 37 lines · Michael Schmitz
Context switching does take care to retain the correct lock owner across the switch from 'prev' to 'next' tasks. This does rely on interrupts remaining disabled for the entire duration of the switch.
This condition is guaranteed for normal process creation and context switching between already running processes, because both 'prev' and 'next' already have interrupts disabled in their saved copies of the status register.
The situation is different for newly created kernel threads. The status register is set to
PS_Sincopy_thread(), which does leave the IPL at 0. Upon restoring the 'next' thread's status register inswitch_to()akaresume(), interrupts then become enabled prematurely.resume()then returns viaret_from_kernel_thread()andschedule_tail()where run queue lock is released (seefinish_task_switch()andfinish_lock_switch()).A timer interrupt calling
scheduler_tick()before the lock is released infinish_task_switch()will find the lock already taken, with the current task as lock owner. This causes a spinlock recursion warning as reported by Guenter Roeck.As far as I can ascertain, this race has been opened in commit
533e6903bea0("m68k: splitret_from_fork(), simplifykernel_thread()") but I haven't done a detailed study of kernel history so it may well predate that commit.Interrupts cannot be disabled in the saved status register copy for kernel threads (init will complain about interrupts disabled when finally starting user space). Disable interrupts temporarily when switching the tasks' register sets in
resume().Note that a simple oriw 0x700,%sr after restoring sr is not enough here - this leaves enough of a race for the 'spinlock recursion' warning to still be observed.
Quoted 8 lines
Tested on ARAnyM and qemu (Quadra 800 emulation).
Cc Guenter Roeck <[email protected]>Cc Geert Uytterhoeven <[email protected]>Cc Finn Thain <[email protected]>Cc Al Viro <[email protected]>Cc [email protected] # 3.6
Quoted 1 line
s/Link/Closes/
Quoted 2 lines
Fixes 533e6903bea0 ("m68k: split ret_from_fork(), simplify kernel_thread()")Signed-off-by Michael Schmitz <[email protected]>
i.e. will queue in the m68k tree for v6.10 (with above tags fixed).
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]
In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds