[PATCH] fadvise invalidating range fix

[WU Fengguang <wfg@mail.ustc.edu.cn>]

Hi Andrew

I noticed fadvise(POSIX_FADV_DONTNEED) was invalidating more parts of the file
than I expected. Here is the patch.

Notes for change in fadvise.c/sys_fadvise64_64():

- It should be noticed that the 'end' param of invalidate_mapping_pages()
  is inclusive;

- When 'offset' and/or 'offset+len' do no align to page boundary, we must
  decide whether to abandon the partial page at the beginning/end of the range. 
  My patch assumes that the application is scanning forward,
  which is the most common case.
  So 'end_index' is set to the page just before the ending partial page.

- Manual pages for posix_fadvise() mentioned the case of 'len=0', which
  is not handled by the code. Perhaps the handling of 'len=0' is useless,
  so I leaved that alone.

Notes for change in truncate.c/invalidate_mapping_pages():

- Is there any reason that I din't know to free any page outside of [begin,end]?
  The origin code will abandon useful trailing pages when there's hole
  in the range, which may cause series of unecessary disk I/O in
  streaming applications.

best regards, Wu Fengguang


diff -Naur linux-2.6.4-rc2/mm/fadvise.c linux-2.6.4-rc2_fadvise_fix/mm/fadvise.c
--- linux-2.6.4-rc2/mm/fadvise.c	2004-02-18 03:57:30.000000000 +0000
+++ linux-2.6.4-rc2_fadvise_fix/mm/fadvise.c	2004-03-08 12:20:06.000000000 +0000
@@ -66,8 +66,7 @@
 		if (!bdi_write_congested(mapping->backing_dev_info))
 			filemap_flush(mapping);
 		start_index = offset >> PAGE_CACHE_SHIFT;
-		end_index = (offset + len + PAGE_CACHE_SIZE - 1) >>
-						PAGE_CACHE_SHIFT;
+		end_index = ((offset + len) >> PAGE_CACHE_SHIFT) - 1;
 		invalidate_mapping_pages(mapping, start_index, end_index);
 		break;
 	default:
diff -Naur linux-2.6.4-rc2/mm/truncate.c linux-2.6.4-rc2_fadvise_fix/mm/truncate.c
--- linux-2.6.4-rc2/mm/truncate.c	2004-02-18 03:59:34.000000000 +0000
+++ linux-2.6.4-rc2_fadvise_fix/mm/truncate.c	2004-03-08 12:20:06.000000000 +0000
@@ -219,6 +219,8 @@
 			ret += invalidate_complete_page(mapping, page);
 unlock:
 			unlock_page(page);
+			if (next > end)
+				break;
 		}
 		pagevec_release(&pvec);
 		cond_resched();