[dm-devel] Re: 2.6.23-mm1

Milan Broz mbroz at redhat.com
Sun Oct 14 22:03:27 UTC 2007


Andrew Morton wrote:
> On Sun, 14 Oct 2007 21:12:08 +0200 "Torsten Kaiser" <just.for.lkml at googlemail.com> wrote:
...
>> 354036 Page allocated via order 0, mask 0x11202
>>             1 (PFN/Block always differ) PFN 3072 Block 6 type 0          Flags
>> 354338 [0xffffffff80266373] mempool_alloc+83
>> 354338 [0xffffffff80266373] mempool_alloc+83
>> 354025 [0xffffffff802bb389] bio_alloc_bioset+185
>> 354058 [0xffffffff804d2b40] kcryptd_do_crypt+0
>> 354052 [0xffffffff804d2cc7] kcryptd_do_crypt+391
>> 354058 [0xffffffff804d2b40] kcryptd_do_crypt+0
>> 354052 [0xffffffff80245d3c] run_workqueue+204
>> 354062 [0xffffffff802467b0] worker_thread+0
>>
>> I'm using dm-crypt with CONFIG_CRYPTO_TWOFISH_X86_64
>>
>>> The other info shows a tremendous memory leak, not via slab.  Looks like
>>> someone is running alloc_pages() directly and isnb't giving them back.
>> Blaming it on dm-crypt looks right, as the leak seems to happens, if
>> there is (heavy) disk activity.
>> (updatedb just ate ~500 Mb)
>>
> 
> Yup, it does appear that dm-crypt is leaking.  Let's add some cc's.

More precisely - change below from git-block.patch update
caused that pages are not deallocated at all.
(cc-ing Jens)

-static int crypt_endio(struct bio *clone, unsigned int done, int error)
+static void crypt_endio(struct bio *clone, int error)
...
-	 * free the processed pages, even if
-	 * it's only a partially completed write
+	 * free the processed pages
 	 */
-	if (!read_io)
-		crypt_free_buffer_pages(cc, clone, done);
-
-	/* keep going - not finished yet */
-	if (unlikely(clone->bi_size))
-		return 1;
-
-	if (!read_io)
+	if (!read_io) {
+		crypt_free_buffer_pages(cc, clone, clone->bi_size);
 		goto out;
+	}

clone->bi_size is zero here now, so crypt_free_buffer_pages will not
work correctly (previously there was count of processed bytes).

But because it seems that bio cannot be processed partially now, we can
simplify crypt_free_buffer_pages to always remove all allocated pages.

Milan
--
mbroz at redhat.com




More information about the dm-devel mailing list