On Wed, 4 Jan 2006, Ryan Richter wrote: > On Tue, Dec 27, 2005 at 06:21:39PM +0200, Kai Makisara wrote: ... > > I don't have any new ideas but I will tell you what I have tried. In order > > to get more information about what is happening, I inserted the patch at > > the end of this message to my kernel. The purpose of the patch was to > > print something about the page mappings (prt_pages) and to catch possible > > double freeing from st earlier (clear page pointers). > > > > Running dump gave me the following output: > > Here's what I got: > > > st: page attributes before page_release 8 > 0: flags:0x060000000000006c mapping:ffff810102bba238 mapcount:2 count:4 pfn:1392956 > 1: flags:0x060000000000006c mapping:ffff810102bba238 mapcount:2 count:4 pfn:1397945 > 2: flags:0x060000000000006c mapping:ffff810102bba238 mapcount:2 count:4 pfn:1537473 > 3: flags:0x060000000000006c mapping:ffff810102bba238 mapcount:2 count:4 pfn:1398161 > 4: flags:0x060000000000006c mapping:ffff810102bba238 mapcount:2 count:4 pfn:1398261 > 5: flags:0x060000000000006c mapping:ffff810102bba238 mapcount:2 count:4 pfn:1392778 > 6: flags:0x060000000000006c mapping:ffff810102bba238 mapcount:2 count:4 pfn:1402858 > 7: flags:0x060000000000006c mapping:ffff810102bba238 mapcount:2 count:4 pfn:1396799 This looks similar to my output except that the page flags are 0x060000000000006c in your case whereas I have 0x010000000000006c. I have run amanda with a patch modified so that it prints results for 32 consecutive writes each 10000 writes (if ((ctr++ % 10000) < 32)). This showed that taper (at least in my case) is actually using a buffer of 20 x 32 kB = 640 kB. In all writes the mapping, counts, and flags were the same. Instead of remapping a single 32 kB buffer, st is cyclically mapping 20 32 kB buffers. > Bad page state at free_hot_cold_page (in process 'taper', page ffff81000427ff60) > flags:0x010000000000000c mapping:ffff810102bbaaf0 mapcount:2 count:0 The mapping and flags have changed. The flags have lost one bit at the high end and the PG_lru and PG_active flags. The patch sets the page pointer to NULL after page_cache_release() is called in st. This means that st is not calling page_cache_release() more than once for each page returned by get_user_pages(). The page state seems to get corrupted somewhere. I have to continue thinking. -- Kai - : send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
