On 01/31/2017 01:55 PM, Bart Van Assche wrote: > On Tue, 2017-01-31 at 13:34 -0800, Bart Van Assche wrote: >> On Mon, 2017-01-30 at 17:38 -0800, Jens Axboe wrote: >>> That's a known bug in mainline. Pull it into 4.10-rc6, >>> or use my for-next where everything is already merged. >> >> Hello Jens, >> >> With your for-next branch (commit c2e60b3a2602) I haven't hit any block >> layer crashes so far. The only issue I encountered that is new is a >> memory leak triggered by the SG-IO code. These memory leak reports >> started to appear after I started testing the mq-deadline scheduler. >> kmemleak reported the following call stack multiple times after my tests >> had finished: >> >> unreferenced object 0xffff88041119e528 (size 192): >> comm "multipathd", pid 2353, jiffies 4295128020 (age 1332.440s) >> hex dump (first 32 bytes): >> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ >> 00 00 00 00 00 00 00 00 12 01 00 00 00 00 00 00 ................ >> backtrace: >> [<ffffffff8165e3b5>] kmemleak_alloc+0x45/0xa0 >> [<ffffffff811cc23d>] __kmalloc+0x15d/0x2f0 >> [<ffffffff81310e35>] bio_alloc_bioset+0x185/0x1f0 >> [<ffffffff813117f4>] bio_map_user_iov+0x124/0x400 >> [<ffffffff81320b7a>] blk_rq_map_user_iov+0x11a/0x210 >> [<ffffffff81320cbd>] blk_rq_map_user+0x4d/0x60 >> [<ffffffff81336694>] sg_io+0x3d4/0x410 >> [<ffffffff813369d0>] scsi_cmd_ioctl+0x300/0x490 >> [<ffffffff81336b9d>] scsi_cmd_blk_ioctl+0x3d/0x50 >> [<ffffffff814b4360>] sd_ioctl+0x80/0x100 >> [<ffffffff8132ddde>] blkdev_ioctl+0x51e/0x9f0 >> [<ffffffff8122f388>] block_ioctl+0x38/0x40 >> [<ffffffff8120097f>] do_vfs_ioctl+0x8f/0x700 >> [<ffffffff8120102c>] SyS_ioctl+0x3c/0x70 >> [<ffffffff8166c4aa>] entry_SYSCALL_64_fastpath+0x18/0xad > > After I repeated my test the above findings were confirmed: no memory leaks > were reported by kmemleak after a test with I/O scheduler "none" and the > above call stack was reported 44 times by kmemleak after a test with I/O > scheduler "mq-deadline". Interesting, I'll check this. Doesn't make any sense why the scheduler would be implicated in that, given how we run completions now. But if it complains, then something must be up. -- Jens Axboe
