m68k/mm: Port OOM changes to do_page_fault()
Kautuk Consul [Sun, 25 Mar 2012 10:37:48 +0000 (06:37 -0400)]
Commit d065bd810b6deb67d4897a14bfe21f8eb526ba99
(mm: retry page fault when blocking on disk transfer) and
commit 37b23e0525d393d48a7d59f870b3bc061a30ccdb
(x86,mm: make pagefault killable)

The above commits introduced changes into the x86 pagefault handler
for making the page fault handler retryable as well as killable.

These changes reduce the mmap_sem hold time, which is crucial
during OOM killer invocation.

Port these changes to m68k.

Signed-off-by: Kautuk Consul <consul.kautuk@gmail.com>
Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org>

arch/m68k/mm/fault.c

index 6b020a8..aeebbb7 100644 (file)
@@ -72,7 +72,8 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
 {
        struct mm_struct *mm = current->mm;
        struct vm_area_struct * vma;
-       int write, fault;
+       int fault;
+       unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
 
 #ifdef DEBUG
        printk ("do page fault:\nregs->sr=%#x, regs->pc=%#lx, address=%#lx, %ld, %p\n",
@@ -87,6 +88,7 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
        if (in_atomic() || !mm)
                goto no_context;
 
+retry:
        down_read(&mm->mmap_sem);
 
        vma = find_vma(mm, address);
@@ -117,14 +119,13 @@ good_area:
 #ifdef DEBUG
        printk("do_page_fault: good_area\n");
 #endif
-       write = 0;
        switch (error_code & 3) {
                default:        /* 3: write, present */
                        /* fall through */
                case 2:         /* write, not present */
                        if (!(vma->vm_flags & VM_WRITE))
                                goto acc_err;
-                       write++;
+                       flags |= FAULT_FLAG_WRITE;
                        break;
                case 1:         /* read, present */
                        goto acc_err;
@@ -139,10 +140,14 @@ good_area:
         * the fault.
         */
 
-       fault = handle_mm_fault(mm, vma, address, write ? FAULT_FLAG_WRITE : 0);
+       fault = handle_mm_fault(mm, vma, address, flags);
 #ifdef DEBUG
        printk("handle_mm_fault returns %d\n",fault);
 #endif
+
+       if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current))
+               return 0;
+
        if (unlikely(fault & VM_FAULT_ERROR)) {
                if (fault & VM_FAULT_OOM)
                        goto out_of_memory;
@@ -150,10 +155,31 @@ good_area:
                        goto bus_err;
                BUG();
        }
-       if (fault & VM_FAULT_MAJOR)
-               current->maj_flt++;
-       else
-               current->min_flt++;
+
+       /*
+        * Major/minor page fault accounting is only done on the
+        * initial attempt. If we go through a retry, it is extremely
+        * likely that the page will be found in page cache at that point.
+        */
+       if (flags & FAULT_FLAG_ALLOW_RETRY) {
+               if (fault & VM_FAULT_MAJOR)
+                       current->maj_flt++;
+               else
+                       current->min_flt++;
+               if (fault & VM_FAULT_RETRY) {
+                       /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
+                        * of starvation. */
+                       flags &= ~FAULT_FLAG_ALLOW_RETRY;
+
+                       /*
+                        * No need to up_read(&mm->mmap_sem) as we would
+                        * have already released it in __lock_page_or_retry
+                        * in mm/filemap.c.
+                        */
+
+                       goto retry;
+               }
+       }
 
        up_read(&mm->mmap_sem);
        return 0;