install_page() lockup
Subsystems (2)
I'm getting lockups in install_page() with shared pagetables enabled. I haven't really delved into it. It happens under heavy memory pressure on SMP.
Ingo's new patch is using install_page much more than we used to (I don't think I've ever run it before), so we're running fairly untested codepaths here.
I tried this:
mm/fremap.c | 2 ++ 1 files changed, 2 insertions(+)
mm/fremap.c renamed from mm/fremap.c~a +2 -0
@@ -72,7 +72,9 @@ int install_page(struct mm_struct *mm, spte_page_lock (ptepage);if (page_count (ptepage) >1 ) { pte =pte_unshare (mm, pmd, addr); +pte_page_unlock (ptepage); ptepage =pmd_page (*pmd); +pte_page_lock (ptepage); }else pte =pte_offset_map (pmd, addr); }else {
.
Because doing a pte_page_lock(ptepage) and then losing track of the page we just locked looks fishy. Didn't help though.
Dave could you please review the code in there? It's probably something simple.
Thanks.
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/
Quoted 3 lines · Andrew Morton
Ingo's new patch is using
install_pagemuch more than we used to (I don't think I've ever run it before), so we're running fairly untested codepaths here.
i added install_page() for fremap()'s purposes so i'd be surprised if anything else used it. I have shared-pte turned off in my tests, will try with it on as well.
Ingo
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/
--On Wednesday, October 23, 2002 08:49:11 +0200 Ingo Molnar <[email protected]>
Quoted 3 lines · wrote:
i added
install_page()for fremap()'s purposes so i'd be surprised if anything else used it. I have shared-pte turned off in my tests, will try with it on as well.
As I sent in an earlier mail, I found the bug. do_file_page needs to unlock the pte_page_lock, not the page_table_lock.
If I understand the use of install_page correctly, I don't see a reason why it should be calling pte_unshare. If it's only installing pages from existing shared regions it should leave the pte page shared. The only reason to call pte_unshare is if the vma for that mm has changed, making the sharing decision invalid. Am I missing how this is being used?
Dave McCracken
======================================================================
Dave McCracken IBM Linux Base Kernel Team 1-512-838-3059
[email protected] T/L 678-3059
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to [email protected]. For more info on Linux MM,
see: http://www.linux-mm.org/
--On Tuesday, October 22, 2002 22:37:10 -0700 Andrew Morton
Quoted 8 lines · <[email protected]> wrote:
I'm getting lockups in
install_page()with shared pagetables enabled. I haven't really delved into it. It happens under heavy memory pressure on SMP.Ingo's new patch is using
install_pagemuch more than we used to (I don't think I've ever run it before), so we're running fairly untested codepaths here.
As Ingo said, he added install_page.
Quoted 7 lines
I tried this:
(snip)
Because doing a
pte_page_lock(ptepage)and then losing track of the page we just locked looks fishy. Didn't help though.
The code was correct. pte_unshare moves the lock to the new pte page if it installs one. I know that's not real clean, but it eliminates multiple unlock/relock sequences.
Quoted 2 lines
Dave could you please review the code in there? It's probably something simple.
I found the problem. In memory.c:do_file_page it unlocks the page_table_lock. In the new locking scheme, it's actually the pte_page_lock that's held instead.
The patch to fix this is attached.
Dave McCracken
======================================================================
Dave McCracken IBM Linux Base Kernel Team 1-512-838-3059
[email protected] T/L 678-3059
mm/memory.c +2 -1
@@ -1823,6 +1823,7 @@static int do_file_page (struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,int write_access,pte_t *pte,pmd_t *pmd) { +struct page *ptepage =pmd_page (*pmd);unsigned long pgoff;int err;
@@ -1840,7 +1841,7 @@ pgoff =pte_to_pgoff (*pte);pte_unmap (pte); -spin_unlock (&mm->page_table_lock); +pte_page_unlock (ptepage); err = vma->vm_ops->populate (vma, address &PAGE_MASK ,PAGE_SIZE , vma->vm_page_prot, pgoff,0 );if (err == -ENOMEM )