Content-Type: multipart/mixed; boundary="===============4091220628634494826==" MIME-Version: 1.0 From: Christoph Hellwig To: lkp@lists.01.org Subject: Re: [fs] 36e2c7421f: kernel-selftests.splice.short_splice_read.sh.fail Date: Sat, 19 Sep 2020 07:07:39 +0200 Message-ID: <20200919050739.GA7038@lst.de> In-Reply-To: <202009181443.C2179FB@keescook> List-Id: --===============4091220628634494826== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On Fri, Sep 18, 2020 at 02:49:19PM -0700, Kees Cook wrote: > In response to my recent bug fix for splice vs sysfs binary handler[1], > I added splice testing for other pseudo filesystems[2], for which the > test output is seen above. > = > What is the final verdict on the "should splice have a fallback mode?" > question[3]? Right now /proc and /sys reject splice attempts (which, as > I mentioned in the thread, is fine by me, since it would have blocked > the bug I had to fix from ever being exposed in the first place). The verdict is: without a set_fs()-like mechanism that allows uaccess routines to operate on kernel buffers, or even worse a compat_alloc_user_space-like mechanism we can't have an entirely generic fallback. > Should I update the test to _expect_ that splice should fail? I think so. We can updated individual file operations to support splice where actually used applications except it (even when they shouldn't), but I'd rather not do it just for a test case. --===============4091220628634494826==--