ecryptfs: fix fsx data corruption problems
authorEric Sandeen <sandeen@redhat.com>
Tue, 18 Dec 2007 00:20:10 +0000 (16:20 -0800)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Tue, 18 Dec 2007 03:28:17 +0000 (19:28 -0800)
ecryptfs in 2.6.24-rc3 wasn't surviving fsx for me at all, dying after 4
ops.  Generally, encountering problems with stale data and improperly
zeroed pages.  An extending truncate + write for example would expose stale
data.

With the changes below I got to a million ops and beyond with all mmap ops
disabled - mmap still needs work.  (A version of this patch on a RHEL5
kernel ran for over 110 million fsx ops)

I added a few comments as well, to the best of my understanding
as I read through the code.

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Acked-by: Michael Halcrow <mhalcrow@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/ecryptfs/mmap.c
fs/ecryptfs/read_write.c

index 16a7a555f39264a829ce70a9d2e8c9ba954c87c3..32c5711d79a320593e9a536590f3b347905055bd 100644 (file)
@@ -263,14 +263,13 @@ out:
        return 0;
 }
 
+/* This function must zero any hole we create */
 static int ecryptfs_prepare_write(struct file *file, struct page *page,
                                  unsigned from, unsigned to)
 {
        int rc = 0;
+       loff_t prev_page_end_size;
 
-       if (from == 0 && to == PAGE_CACHE_SIZE)
-               goto out;       /* If we are writing a full page, it will be
-                                  up to date. */
        if (!PageUptodate(page)) {
                rc = ecryptfs_read_lower_page_segment(page, page->index, 0,
                                                      PAGE_CACHE_SIZE,
@@ -283,22 +282,32 @@ static int ecryptfs_prepare_write(struct file *file, struct page *page,
                } else
                        SetPageUptodate(page);
        }
-       if (page->index != 0) {
-               loff_t end_of_prev_pg_pos =
-                       (((loff_t)page->index << PAGE_CACHE_SHIFT) - 1);
 
-               if (end_of_prev_pg_pos > i_size_read(page->mapping->host)) {
+       prev_page_end_size = ((loff_t)page->index << PAGE_CACHE_SHIFT);
+
+       /*
+        * If creating a page or more of holes, zero them out via truncate.
+        * Note, this will increase i_size.
+        */
+       if (page->index != 0) {
+               if (prev_page_end_size > i_size_read(page->mapping->host)) {
                        rc = ecryptfs_truncate(file->f_path.dentry,
-                                              end_of_prev_pg_pos);
+                                              prev_page_end_size);
                        if (rc) {
                                printk(KERN_ERR "Error on attempt to "
                                       "truncate to (higher) offset [%lld];"
-                                      " rc = [%d]\n", end_of_prev_pg_pos, rc);
+                                      " rc = [%d]\n", prev_page_end_size, rc);
                                goto out;
                        }
                }
-               if (end_of_prev_pg_pos + 1 > i_size_read(page->mapping->host))
-                       zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
+       }
+       /*
+        * Writing to a new page, and creating a small hole from start of page?
+        * Zero it out.
+        */
+       if ((i_size_read(page->mapping->host) == prev_page_end_size) &&
+           (from != 0)) {
+               zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
        }
 out:
        return rc;
index 6b7474a4336a5844cd004d3b8af923f19d6fc536..948f57624c05dd2bb1750ff05628f7931f44f2b7 100644 (file)
@@ -124,6 +124,10 @@ int ecryptfs_write(struct file *ecryptfs_file, char *data, loff_t offset,
        loff_t pos;
        int rc = 0;
 
+       /*
+        * if we are writing beyond current size, then start pos
+        * at the current size - we'll fill in zeros from there.
+        */
        if (offset > ecryptfs_file_size)
                pos = ecryptfs_file_size;
        else
@@ -137,6 +141,7 @@ int ecryptfs_write(struct file *ecryptfs_file, char *data, loff_t offset,
                if (num_bytes > total_remaining_bytes)
                        num_bytes = total_remaining_bytes;
                if (pos < offset) {
+                       /* remaining zeros to write, up to destination offset */
                        size_t total_remaining_zeros = (offset - pos);
 
                        if (num_bytes > total_remaining_zeros)
@@ -167,17 +172,27 @@ int ecryptfs_write(struct file *ecryptfs_file, char *data, loff_t offset,
                        }
                }
                ecryptfs_page_virt = kmap_atomic(ecryptfs_page, KM_USER0);
+
+               /*
+                * pos: where we're now writing, offset: where the request was
+                * If current pos is before request, we are filling zeros
+                * If we are at or beyond request, we are writing the *data*
+                * If we're in a fresh page beyond eof, zero it in either case
+                */
+               if (pos < offset || !start_offset_in_page) {
+                       /* We are extending past the previous end of the file.
+                        * Fill in zero values to the end of the page */
+                       memset(((char *)ecryptfs_page_virt
+                               + start_offset_in_page), 0,
+                               PAGE_CACHE_SIZE - start_offset_in_page);
+               }
+
+               /* pos >= offset, we are now writing the data request */
                if (pos >= offset) {
                        memcpy(((char *)ecryptfs_page_virt
                                + start_offset_in_page),
                               (data + data_offset), num_bytes);
                        data_offset += num_bytes;
-               } else {
-                       /* We are extending past the previous end of the file.
-                        * Fill in zero values up to the start of where we
-                        * will be writing data. */
-                       memset(((char *)ecryptfs_page_virt
-                               + start_offset_in_page), 0, num_bytes);
                }
                kunmap_atomic(ecryptfs_page_virt, KM_USER0);
                flush_dcache_page(ecryptfs_page);