fixes for 33-rt4 found while testing on powerpc
cover fixes for 33-rt4 found while testing on powerpc
These two commits were originally done by Kevin to solve problems on a 2.6.31 based RT tree. They carry forward onto 33-rt4 as-is, and I've done a quick sanity test on an sbc8641d board (powerpc SMP dual core).
Paul.
PATCH 1/2 powerpc: replace kmap_atomic with kmap in pte_offset_map
From: Kevin Hao <[email protected]>
The pte_offset_map/pte_offset_map_nested use kmap_atomic to get the virtual address for the pte table, but kmap_atomic will disable preempt. Hence there will be call trace if we acquire a spin lock after invoking pte_offset_map/pte_offset_map_nested in preempt-rt. To fix it, I've replaced kmap_atomic with kmap in these macros.
--- arch/powerpc/include/asm/pgtable-ppc32.h | 12 ++++++++++++ 1 files changed, 12 insertions(+), 0 deletions(-)
arch/powerpc/include/asm/pgtable-ppc32.h +12 -0
@@ -307,6 +307,17 @@ static inline void __ptep_set_access_flags(pte_t *ptep, pte_t entry) (((address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1))#define pte_offset_kernel (dir, addr) \ ((pte_t *) pmd_page_vaddr(*(dir)) + pte_index(addr)) +#ifdef CONFIG_PREEMPT_RT +#define pte_offset_map (dir, addr) \ + ((pte_t *) kmap(pmd_page(*(dir))) + pte_index(addr)) +#define pte_offset_map_nested (dir, addr) \ + ((pte_t *) kmap(pmd_page(*(dir))) + pte_index(addr)) + +#define pte_unmap (pte) \ + kunmap((struct page *)_ALIGN_DOWN((unsigned int)pte, PAGE_SIZE)) +#define pte_unmap_nested (pte) \ + kunmap((struct page *)_ALIGN_DOWN((unsigned int)pte, PAGE_SIZE)) +#else #define pte_offset_map (dir, addr) \ ((pte_t *) kmap_atomic(pmd_page(*(dir)), KM_PTE0) + pte_index(addr))#define pte_offset_map_nested (dir, addr) \
@@ -314,6 +325,7 @@ static inline void __ptep_set_access_flags(pte_t *ptep, pte_t entry)#define pte_unmap (pte) kunmap_atomic(pte, KM_PTE0)#define pte_unmap_nested (pte) kunmap_atomic(pte, KM_PTE1) +#endif /* * Encode and decode a swap entry.
Signature · 1 line
1.6.5.2
PATCH 2/2 rt: reserve TASK_STOPPED state when blocking on a spin lock
From: Kevin Hao <[email protected]>
When a process handles a SIGSTOP signal, it will set the state to TASK_STOPPED, acquire tasklist_lock and notifiy the parent of the status change. But in the rt kernel the process state will change to TASK_UNINTERRUPTIBLE if it blocks on the tasklist_lock. So if we send a SIGCONT signal to this process at this time, the SIGCONT signal just does nothing because this process is not in TASK_STOPPED state. Of course this is not what we wanted. Preserving the TASK_STOPPED state when blocking on a spin lock can fix this bug.
--- kernel/rtmutex.c | 5 +++-- 1 files changed, 3 insertions(+), 2 deletions(-)
kernel/rtmutex.c +3 -2
@@ -757,8 +757,9 @@ rt_set_current_blocked_state(unsigned long saved_state)* saved_state. Now we can ignore further wakeups as we will * return in state running from our "spin" sleep. */ -if (saved_state ==TASK_INTERRUPTIBLE ) - block_state =TASK_INTERRUPTIBLE ; +if (saved_state ==TASK_INTERRUPTIBLE || + saved_state ==TASK_STOPPED ) + block_state = saved_state;else block_state =TASK_UNINTERRUPTIBLE ;
Signature · 1 line
1.6.5.2