Received-SPF: pass (google.com: best guess record for domain of linux-ext4-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Mon, 24 Feb 2020 15:36:25 -0800
From:   Christoph Hellwig <hch@infradead.org>
To:     Eric Biggers <ebiggers@kernel.org>
Cc:     Christoph Hellwig <hch@infradead.org>,
        Satya Tangirala <satyat@google.com>,
        linux-block@vger.kernel.org, linux-scsi@vger.kernel.org,
        linux-fscrypt@vger.kernel.org, linux-fsdevel@vger.kernel.org,
        linux-f2fs-devel@lists.sourceforge.net, linux-ext4@vger.kernel.org,
        Barani Muthukumaran <bmuthuku@qti.qualcomm.com>,
        Kuohong Wang <kuohong.wang@mediatek.com>,
        Kim Boojin <boojin.kim@samsung.com>
Subject: Re: [PATCH v7 3/9] block: blk-crypto-fallback for Inline Encryption
Message-ID: <20200224233625.GB30288@infradead.org>
References: <20200221115050.238976-1-satyat@google.com>
 <20200221115050.238976-4-satyat@google.com>
 <20200221173539.GA6525@infradead.org>
 <20200221183437.GC925@sol.localdomain>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200221183437.GC925@sol.localdomain>
Sender: linux-ext4-owner@vger.kernel.org
Precedence: bulk

On Fri, Feb 21, 2020 at 10:34:37AM -0800, Eric Biggers wrote:
> On Fri, Feb 21, 2020 at 09:35:39AM -0800, Christoph Hellwig wrote:
> > High-level question:  Does the whole keyslot manager concept even make
> > sense for the fallback?  With the work-queue we have item that exectutes
> > at a time per cpu.  So just allocatea per-cpu crypto_skcipher for
> > each encryption mode and there should never be a slot limitation.  Or
> > do I miss something?
> 
> It does make sense because if blk-crypto-fallback didn't use a keyslot manager,
> it would have to call crypto_skcipher_setkey() on the I/O path for every bio to
> ensure that the CPU's crypto_skcipher has the correct key.  That's undesirable,
> because setting a new key can be expensive with some encryption algorithms, and
> also it can require a memory allocation which can fail.  For example, with the
> Adiantum algorithm, setting a key requires encrypting ~1100 bytes of data in
> order to generate subkeys.  It's better to set a key once and use it many times.

I didn't think of such expensive operations when setting the key.
Note that you would not have to do it on every I/O, as chances are high
you'll get I/O from the same submitter and thus the same key, and we
can optimize for that case pretty easily.

But if you think the keyslot manager is better I accept that, this was
just a throught when looking over the code.