[PATCH] mm: split page table lock

Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with
a many-threaded application which concurrently initializes different parts of
a large anonymous area.

This patch corrects that, by using a separate spinlock per page table page, to
guard the page table entries in that page, instead of using the mm's single
page_table_lock.  (But even then, page_table_lock is still used to guard page
table allocation, and anon_vma allocation.)

In this implementation, the spinlock is tucked inside the struct page of the
page table page: with a BUILD_BUG_ON in case it overflows - which it would in
the case of 32-bit PA-RISC with spinlock debugging enabled.

Splitting the lock is not quite for free: another cacheline access.  Ideally,
I suppose we would use split ptlock only for multi-threaded processes on
multi-cpu machines; but deciding that dynamically would have its own costs.
So for now enable it by config, at some number of cpus - since the Kconfig
language doesn't support inequalities, let preprocessor compare that with
NR_CPUS.  But I don't think it's worth being user-configurable: for good
testing of both split and unsplit configs, split now at 4 cpus, and perhaps
change that to 8 later.

There is a benefit even for singly threaded processes: kswapd can be attacking
one part of the mm while another part is busy faulting.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

1 files changed, 10 insertions, 1 deletions

diff --git a/mm/mremap.c b/mm/mremap.c
index 8de77b632a2..b535438c363 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -72,7 +72,7 @@ static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
         struct address_space *mapping = NULL;
         struct mm_struct *mm = vma->vm_mm;
         pte_t *old_pte, *new_pte, pte;
-        spinlock_t *old_ptl;
+        spinlock_t *old_ptl, *new_ptl;
 
         if (vma->vm_file) {
                 /*
@@ -88,8 +88,15 @@ static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
                         new_vma->vm_truncate_count = 0;
         }
 
+        /*
+         * We don't have to worry about the ordering of src and dst
+         * pte locks because exclusive mmap_sem prevents deadlock.
+         */
         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
         new_pte = pte_offset_map_nested(new_pmd, new_addr);
+        new_ptl = pte_lockptr(mm, new_pmd);
+        if (new_ptl != old_ptl)
+                spin_lock(new_ptl);
 
         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
                                    new_pte++, new_addr += PAGE_SIZE) {
@@ -101,6 +108,8 @@ static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
                 set_pte_at(mm, new_addr, new_pte, pte);
         }
 
+        if (new_ptl != old_ptl)
+                spin_unlock(new_ptl);
         pte_unmap_nested(new_pte - 1);
         pte_unmap_unlock(old_pte - 1, old_ptl);
         if (mapping)