Kanoj Sarcar wrote:
> Here's a primitive patch showing the direction I am thinking of. I do not
> have any problem with a spinning lock, but I coded this against 2.2.10,
> where insert_vm_struct could go to sleep, hence I had to use sleeping
> locks to protect the vma chain.
I found a few places where I don't know how to change them.
1) arch/mips/mm/r4xx0.c:
their flush_cache_range() function internally calls find_vma().
flush_cache_range() is called by proc/mem.c, and it seems that this
function cannot get the mmap semaphore.
Currently, every caller of flush_cache_range() either owns the kernel
lock or the mmap_sem.
OTHO, this function contains a race anyway [src_vma can go away if
handle_mm_fault() sleeps, src_vma is used at the end of the function.]
2) arch/sparc/mm/fault.c:
> /* This conditional is 'interesting'. */
> if (pgd_val(*pgdp) && !(write && !(pte_val(*ptep) & _SUN4C_PAGE_WRITE))
> && (pte_val(*ptep) & _SUN4C_PAGE_VALID))
> /* Note: It is safe to not grab the MMAP semaphore here because
> * we know that update_mmu_cache() will not sleep for
> * any reason (at least not in the current implementation)
> * and therefore there is no danger of another thread getting
> * on the CPU and doing a shrink_mmap() on this vma.
> */
> sun4c_update_mmu_cache (find_vma(current->mm, address), address,
> *ptep);
> else
> do_sparc_fault(regs, text_fault, write, address);
> }
could be safe because sun4c is only UP?
3) include/ppc-asm/pgtable.h:
> extern __inline__ pte_t *find_pte(struct mm_struct *mm,unsigned long va)
> {
> pgd_t *dir;
> pmd_t *pmd;
> pte_t *pte;
>
> va &= PAGE_MASK;
>
> dir = pgd_offset( mm, va );
> if (dir)
> {
> pmd = pmd_offset(dir, va & PAGE_MASK);
> if (pmd && pmd_present(*pmd))
> {
> pte = pte_offset(pmd, va);
> if (pte && pte_present(*pte))
> {
> pte_uncache(*pte);
> flush_tlb_page(find_vma(mm,va),va);
> }
> }
> }
> return pte;
> }
Could be safe because only called for "init_mm"?
I've not yet looked at swap_out [mm/swapfile.c and
arch/m68k/atari/stram.c] and proc/array.c
--
Manfred
On Wed, Oct 13, 1999 at 09:32:54AM +0200, Manfred Spraul wrote:
> Kanoj Sarcar wrote:
> > Here's a primitive patch showing the direction I am thinking of. I do not
> > have any problem with a spinning lock, but I coded this against 2.2.10,
> > where insert_vm_struct could go to sleep, hence I had to use sleeping
> > locks to protect the vma chain.
>
> I found a few places where I don't know how to change them.
>
> 1) arch/mips/mm/r4xx0.c:
> their flush_cache_range() function internally calls find_vma().
> flush_cache_range() is called by proc/mem.c, and it seems that this
> function cannot get the mmap semaphore.
> Currently, every caller of flush_cache_range() either owns the kernel
> lock or the mmap_sem.
> OTHO, this function contains a race anyway [src_vma can go away if
> handle_mm_fault() sleeps, src_vma is used at the end of the function.]
The sole reason for fiddling with the VMA is that we try to optimize
icache flushing for non-VM_EXEC vmas. This optimization is broken
as the MIPS hardware doesn't make a difference between read and execute
in page permissions, so the icache might be dirty even though the vma
has no exec permission. So I'll have to re-implement this whole things
anyway. The other problem is an efficience problem. A call like
flush_cache_range(some_mm_ptr, 0, TASK_SIZE) would take a minor eternity
and for MIPS64 a full eternity ...
Ralf
>
> On Wed, Oct 13, 1999 at 09:32:54AM +0200, Manfred Spraul wrote:
>
> > Kanoj Sarcar wrote:
> > > Here's a primitive patch showing the direction I am thinking of. I do not
> > > have any problem with a spinning lock, but I coded this against 2.2.10,
> > > where insert_vm_struct could go to sleep, hence I had to use sleeping
> > > locks to protect the vma chain.
> >
> > I found a few places where I don't know how to change them.
> >
> > 1) arch/mips/mm/r4xx0.c:
> > their flush_cache_range() function internally calls find_vma().
> > flush_cache_range() is called by proc/mem.c, and it seems that this
> > function cannot get the mmap semaphore.
> > Currently, every caller of flush_cache_range() either owns the kernel
> > lock or the mmap_sem.
> > OTHO, this function contains a race anyway [src_vma can go away if
> > handle_mm_fault() sleeps, src_vma is used at the end of the function.]
>
> The sole reason for fiddling with the VMA is that we try to optimize
> icache flushing for non-VM_EXEC vmas. This optimization is broken
> as the MIPS hardware doesn't make a difference between read and execute
> in page permissions, so the icache might be dirty even though the vma
> has no exec permission. So I'll have to re-implement this whole things
> anyway. The other problem is an efficience problem. A call like
> flush_cache_range(some_mm_ptr, 0, TASK_SIZE) would take a minor eternity
> and for MIPS64 a full eternity ...
>
> Ralf
Ralf,
Looking in 2.3.21, all the find_vma's in arch/mips/mm/r4xx0.c are used to
set a flag called "text" which is not used at all. Also, if the find_vma
returns null, the code basically does nothing. So the optimized icache
flushing is probably not implemented yet? Then, the only reason to
do the flush_vma currently is to check whether the lower level flush
routine should be called. Without holding some locks, this is always
tricky to do on a third party mm.
Btw, this probably belongs to linux-mips, but what do you mean by saying
the icache might be dirty? Its been a while since I worked on the
older mips chips, but as far as I remember, the icache can not hold
dirty lines.
Kanoj