From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Message-ID: <4CA2CCC3.8010307@fusionio.com> Date: Wed, 29 Sep 2010 14:21:07 +0900 From: Jens Axboe MIME-Version: 1.0 Subject: Re: [PATCH 1/1] um: ubd: Fix data corruption References: <1285710456-4435-1-git-send-email-richard@nod.at> <20100928150000.f007f43e.akpm@linux-foundation.org> <201009290013.11332.richard@nod.at> <20100928225202.GA30352@foursquare.net> <4CA275CE.6060401@fusionio.com> <20100929012945.GA3324@foursquare.net> In-Reply-To: <20100929012945.GA3324@foursquare.net> Content-Type: text/plain; charset="ISO-8859-1" Content-Transfer-Encoding: 7bit To: Chris Frey Cc: Richard Weinberger , Andrew Morton , "linux-kernel@vger.kernel.org" , "jdike@addtoit.com" , "user-mode-linux-devel@lists.sourceforge.net" , "user-mode-linux-user@lists.sourceforge.net" , "janjaap@bos.nl" , "geert@linux-m68k.org" , "martin.petersen@oracle.com" , "adobriyan@gmail.com" , "syzop@vulnscan.org" List-ID: On 2010-09-29 10:29, Chris Frey wrote: > On Wed, Sep 29, 2010 at 08:10:06AM +0900, Jens Axboe wrote: >> It looks like that if we need to restart the requeue, then >> we use the initial position and not the current index. Does >> this help? >> >> diff --git a/arch/um/drivers/ubd_kern.c b/arch/um/drivers/ubd_kern.c >> index 1bcd208..81ee063 100644 >> --- a/arch/um/drivers/ubd_kern.c >> +++ b/arch/um/drivers/ubd_kern.c >> @@ -162,7 +162,7 @@ struct ubd { >> spinlock_t lock; >> struct scatterlist sg[MAX_SG]; >> struct request *request; >> - int start_sg, end_sg; >> + int start_sg, end_sg, rq_off; >> }; >> >> #define DEFAULT_COW { \ >> @@ -187,6 +187,7 @@ struct ubd { >> .request = NULL, \ >> .start_sg = 0, \ >> .end_sg = 0, \ >> + .rq_off = 0, \ >> } >> >> /* Protected by ubd_lock */ >> @@ -1241,10 +1242,11 @@ static void do_ubd_request(struct request_queue *q) >> dev->request = req; >> dev->start_sg = 0; >> dev->end_sg = blk_rq_map_sg(q, req, dev->sg); >> + dev->rq_off = 0; >> } >> >> req = dev->request; >> - sector = blk_rq_pos(req); >> + sector = blk_rq_pos(req) + dev->rq_off; >> while(dev->start_sg < dev->end_sg){ >> struct scatterlist *sg = &dev->sg[dev->start_sg]; >> >> @@ -1273,6 +1275,7 @@ static void do_ubd_request(struct request_queue *q) >> } >> >> dev->start_sg++; >> + dev->rq_off += sg->length >> 9; >> } >> dev->end_sg = 0; >> dev->request = NULL; >> >> -- > > This patch does not fix the corruption issue for me. I applied the patch > to 2.6.35.5, and reproduced the "deleted inode referenced" errors > in both a gentoo and ubuntu guest OS. It does take longer to reproduce > though, with this patch. This seems to imply that the original commit pin pointed is not the only issue we have in that code atm. I think we need to find the real fix here, just disabling merging is not a fix (it's just a nasty work-around for the real bug). -- Jens Axboe