Received-SPF: pass (google.com: domain of linux-kernel+bounces-145918-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
Message-ID: <de942b9c-6e91-4dcc-a386-9d1317e0e0b4@kernel.dk>
Date: Mon, 15 Apr 2024 15:16:12 -0600
Precedence: bulk
MIME-Version: 1.0
User-Agent: Mozilla Thunderbird
Subject: Re: [PATCH 002/437] fs: add generic read/write iterator helpers
Content-Language: en-US
To: Al Viro <viro@zeniv.linux.org.uk>
Cc: linux-kernel@vger.kernel.org
References: <20240411153126.16201-1-axboe@kernel.dk>
 <20240411153126.16201-3-axboe@kernel.dk> <20240415195504.GU2118490@ZenIV>
 <898afc09-428f-4da2-a620-d7ca9f37133c@kernel.dk>
 <20240415210827.GX2118490@ZenIV>
From: Jens Axboe <axboe@kernel.dk>
In-Reply-To: <20240415210827.GX2118490@ZenIV>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit

On 4/15/24 3:08 PM, Al Viro wrote:
> On Mon, Apr 15, 2024 at 02:11:56PM -0600, Jens Axboe wrote:
> 
>> do_loop_{readv,writev} need to look like the one io_uring had, which was
>> just an augmented version of the fs/ version. At least for handling
>> anything that is IOVEC/UBUF/BVEC.
> 
> IOVEC and UBUF: pointer will be __user one; copy_from_user() works.
> KVEC: kernel pointer.  Try copy_from_user() on that on x86 with SMAP (or
> on e.g. sparc64, etc.) and you'll get an oops.
> BVEC: page + offset, page quite possibly not mapped anywhere in kernel page
> tables.  And "just kmap() around the call of your callback" is not a good
> idea either, for even more reason that for KVEC.

The old read/write path only handled user pointers, with the exception
being bvecs mapped on the io_uring side (which the io_uring version
dealt with) which also originally came from user pointers. So just user
memory. Why would we need to expand that now? Who would be doing
->read_iter() or ->write_iter() with something that isn't either UBUF or
IOVEC? Because that would break horrible as it is in the current kernel.

-- 
Jens Axboe