Re: copy on write for splice() from file to pipe?

[Jens Axboe]

On 2/10/23 3:17?PM, Linus Torvalds wrote:
> On Fri, Feb 10, 2023 at 2:08 PM Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>>
>>  (a) the first one is to protect from endless loops
> 
> Just to clarify: they're not "endless loops" per se, but we have
> splice sources and destinations that always succeed, like /dev/zero
> and /dev/null.
> 
> So things like "sendfile()" that are happy to just repeat until done
> do need to have some kind of signal handling even for the case when
> we're not actually waiting for data. That's what that whole
> 
>         /*
>          * Check for signal early to make process killable when there are
>          * always buffers available
>          */
> 
> this is all about. See commit c725bfce7968 ("vfs: Make sendfile(2)
> killable even better") for a less obvious example than that
> "zero->null" kind of thing.
> 
> (I actually suspect that /dev/zero no longer works as a splice source,
> since we disabled the whole "fall back to regular IO" that Christoph
> did in 36e2c7421f02 "fs: don't allow splice read/write without
> explicit ops").

Yet another one... Since it has a read_iter, should be fixable with just
adding the generic splice_read.

-- 
Jens Axboe