proc: pagemap: Hold mmap_sem during page walk
KAMEZAWA Hiroyuki [Fri, 2 Apr 2010 00:11:29 +0000 (09:11 +0900)]
In initial design, walk_page_range() was designed just for walking page
table and it didn't require mmap_sem.  Now, find_vma() etc..  are used
in walk_page_range() and we need mmap_sem around it.

This patch adds mmap_sem around walk_page_range().

Because /proc/<pid>/pagemap's callback routine use put_user(), we have
to get rid of it to do sane fix.

Changelog: 2010/Apr/2
 - fixed start_vaddr and end overflow
Changelog: 2010/Apr/1
 - fixed start_vaddr calculation
 - removed unnecessary cast.
 - removed unnecessary change in smaps.
 - use GFP_TEMPORARY instead of GFP_KERNEL

Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Matt Mackall <mpm@selenic.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: San Mehat <san@google.com>
Cc: Brian Swetland <swetland@google.com>
Cc: Dave Hansen <haveblue@us.ibm.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
[ Fixed kmalloc failure return code as per Matt ]
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

fs/proc/task_mmu.c

index 183f8ff..0962739 100644 (file)
@@ -406,6 +406,7 @@ static int show_smap(struct seq_file *m, void *v)
 
        memset(&mss, 0, sizeof mss);
        mss.vma = vma;
+       /* mmap_sem is held in m_start */
        if (vma->vm_mm && !is_vm_hugetlb_page(vma))
                walk_page_range(vma->vm_start, vma->vm_end, &smaps_walk);
 
@@ -552,7 +553,8 @@ const struct file_operations proc_clear_refs_operations = {
 };
 
 struct pagemapread {
-       u64 __user *out, *end;
+       int pos, len;
+       u64 *buffer;
 };
 
 #define PM_ENTRY_BYTES      sizeof(u64)
@@ -575,10 +577,8 @@ struct pagemapread {
 static int add_to_pagemap(unsigned long addr, u64 pfn,
                          struct pagemapread *pm)
 {
-       if (put_user(pfn, pm->out))
-               return -EFAULT;
-       pm->out++;
-       if (pm->out >= pm->end)
+       pm->buffer[pm->pos++] = pfn;
+       if (pm->pos >= pm->len)
                return PM_END_OF_BUFFER;
        return 0;
 }
@@ -720,21 +720,20 @@ static int pagemap_hugetlb_range(pte_t *pte, unsigned long addr,
  * determine which areas of memory are actually mapped and llseek to
  * skip over unmapped regions.
  */
+#define PAGEMAP_WALK_SIZE      (PMD_SIZE)
 static ssize_t pagemap_read(struct file *file, char __user *buf,
                            size_t count, loff_t *ppos)
 {
        struct task_struct *task = get_proc_task(file->f_path.dentry->d_inode);
-       struct page **pages, *page;
-       unsigned long uaddr, uend;
        struct mm_struct *mm;
        struct pagemapread pm;
-       int pagecount;
        int ret = -ESRCH;
        struct mm_walk pagemap_walk = {};
        unsigned long src;
        unsigned long svpfn;
        unsigned long start_vaddr;
        unsigned long end_vaddr;
+       int copied = 0;
 
        if (!task)
                goto out;
@@ -757,35 +756,12 @@ static ssize_t pagemap_read(struct file *file, char __user *buf,
        if (!mm)
                goto out_task;
 
-
-       uaddr = (unsigned long)buf & PAGE_MASK;
-       uend = (unsigned long)(buf + count);
-       pagecount = (PAGE_ALIGN(uend) - uaddr) / PAGE_SIZE;
-       ret = 0;
-       if (pagecount == 0)
-               goto out_mm;
-       pages = kcalloc(pagecount, sizeof(struct page *), GFP_KERNEL);
+       pm.len = PM_ENTRY_BYTES * (PAGEMAP_WALK_SIZE >> PAGE_SHIFT);
+       pm.buffer = kmalloc(pm.len, GFP_TEMPORARY);
        ret = -ENOMEM;
-       if (!pages)
+       if (!pm.buffer)
                goto out_mm;
 
-       down_read(&current->mm->mmap_sem);
-       ret = get_user_pages(current, current->mm, uaddr, pagecount,
-                            1, 0, pages, NULL);
-       up_read(&current->mm->mmap_sem);
-
-       if (ret < 0)
-               goto out_free;
-
-       if (ret != pagecount) {
-               pagecount = ret;
-               ret = -EFAULT;
-               goto out_pages;
-       }
-
-       pm.out = (u64 __user *)buf;
-       pm.end = (u64 __user *)(buf + count);
-
        pagemap_walk.pmd_entry = pagemap_pte_range;
        pagemap_walk.pte_hole = pagemap_pte_hole;
        pagemap_walk.hugetlb_entry = pagemap_hugetlb_range;
@@ -807,23 +783,36 @@ static ssize_t pagemap_read(struct file *file, char __user *buf,
         * user buffer is tracked in "pm", and the walk
         * will stop when we hit the end of the buffer.
         */
-       ret = walk_page_range(start_vaddr, end_vaddr, &pagemap_walk);
-       if (ret == PM_END_OF_BUFFER)
-               ret = 0;
-       /* don't need mmap_sem for these, but this looks cleaner */
-       *ppos += (char __user *)pm.out - buf;
-       if (!ret)
-               ret = (char __user *)pm.out - buf;
-
-out_pages:
-       for (; pagecount; pagecount--) {
-               page = pages[pagecount-1];
-               if (!PageReserved(page))
-                       SetPageDirty(page);
-               page_cache_release(page);
+       ret = 0;
+       while (count && (start_vaddr < end_vaddr)) {
+               int len;
+               unsigned long end;
+
+               pm.pos = 0;
+               end = start_vaddr + PAGEMAP_WALK_SIZE;
+               /* overflow ? */
+               if (end < start_vaddr || end > end_vaddr)
+                       end = end_vaddr;
+               down_read(&mm->mmap_sem);
+               ret = walk_page_range(start_vaddr, end, &pagemap_walk);
+               up_read(&mm->mmap_sem);
+               start_vaddr = end;
+
+               len = min(count, PM_ENTRY_BYTES * pm.pos);
+               if (copy_to_user(buf, pm.buffer, len) < 0) {
+                       ret = -EFAULT;
+                       goto out_free;
+               }
+               copied += len;
+               buf += len;
+               count -= len;
        }
+       *ppos += copied;
+       if (!ret || ret == PM_END_OF_BUFFER)
+               ret = copied;
+
 out_free:
-       kfree(pages);
+       kfree(pm.buffer);
 out_mm:
        mmput(mm);
 out_task: