From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mga09.intel.com (mga09.intel.com [134.134.136.24]) (using TLSv1 with cipher DHE-RSA-CAMELLIA256-SHA (256/256 bits)) (No client certificate requested) by mail.server123.net (Postfix) with ESMTPS for ; Mon, 26 Sep 2016 15:08:52 +0200 (CEST) From: "Yu, Wenqian" Date: Mon, 26 Sep 2016 13:08:45 +0000 Message-ID: <858B57D94E9DE244922C3CF355D7FFD64EB600BA@SHSMSX104.ccr.corp.intel.com> References: <858B57D94E9DE244922C3CF355D7FFD64EB5FFA0@SHSMSX104.ccr.corp.intel.com> In-Reply-To: Content-Language: en-US Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: quoted-printable MIME-Version: 1.0 Subject: Re: [dm-crypt] Hang problem with dm-crypt List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Milan Broz , "dm-crypt@saout.de" Cc: "Yu, Wenqian" Hi, Milan, Thanks for the detail information. I noticed the comments and the underlyin= g design logic. In dm-crypt existing design, there is an assumption that the acceleration d= river can queue the requests which are not sent to hardware. =20 I think there are at least two scenarios we should consider to make it more= robust. 1. The queue is full even if the driver has the ability to queue a number = of the requests. 2. The acceleration hardware/driver doesn't have the ability to queue the = requests. Should we add other error code to handle this? =20 Thanks, - Wenqian -----Original Message----- From: Milan Broz [mailto:gmazyland@gmail.com]=20 Sent: Monday, September 26, 2016 6:28 PM To: Yu, Wenqian; dm-crypt@saout.de Subject: Re: [dm-crypt] Hang problem with dm-crypt On 09/26/2016 08:50 AM, Yu, Wenqian wrote: > I tried to use dm-crypt for disk encryption with accelerators and=20 > found that it will hang when accelerator returned EBUSY, which means=20 > the driver request queue is full. That is normal state, when request is processed asynchronously later. Please read explicit comments in code we added to understand this logic. added in this commit: http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/drive= rs/md/dm-crypt.c?id=3D54cea3f6681ad9360814e2926d1f723bbd0f74ed =20 > Per the logic in crypt_convert(), the request will be skipped if the=20 > request is not sent to crypto driver when the driver request queue is=20 > full. Is this expected behavior? It is not skipped, it is queued (or it waits if queue is full and then proc= esses as asynchronous branch (EINPROGRESS)) > In crypt_convert_block(), the sector is advanced (bio_advance_iter())=20 > no matter whether crypto_skcipher_encrypt()/crypto_skcipher_decrypt() > send the request to accelerator driver or not. When the driver > request queue is full, EBUSY will be returned from=20 > crypto_skcipher_encrypt()/crypto_skcipher_decrypt(). And in=20 > crypt_convert(), the existing implementation is waiting for a=20 > completion from a request, which is not queued in the driver when=20 > EBUSY is encountered from crypt_convert_block (). In this case, the=20 > sector should not be advanced or should be rolled back as the request=20 > is not sent to accelerator driver. I think it should be queued (IOW the one that returns BUSY should be queued= ). If it is not done, I would say it is bug in acceleration driver. Note this flag: /* * Use REQ_MAY_BACKLOG so a cipher driver internally backlogs * requests if driver request queue is full. */ Anyway, this is more question for crypto API mailing list... I think that dmcrypt processing is correct here. Milan