On Mon, Dec 8, 2014 at 10:46 AM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> On Mon, Dec 08, 2014 at 10:37:51AM -0800, Linus Torvalds wrote:
>
>> How about we make "kernel_read()" just clear O_DIRECT? Does that fix
>> it to just use copies?
>
> Umm... clearing O_DIRECT on struct file that might have other references
> to it isn't nice, to put it mildly...
Yeah.
> Frankly, stopping iov_iter_get_pages() on the first page in vmalloc/module
> space looks like the sanest strategy anyway. We'd get the same behaviour
> as we used to, and as for finit_module(2), well... put "fails if given
> an O_DIRECT descriptor" and be done with that.
If it used to fail, then by all means, just make this a failure in the
new model. I really don't want to make core infrastructure silently
just call vmalloc_to_page() to make things "work".
And if it used to do "get_user_pages_fast()" then the old code really
didn't work on vmalloc ranges anyway, since that one checks for not
just _PAGE_PRESENT but also _PAGE_USER, which won't be set on a
vmalloc() page. In fact, it should have failed on *all* kernel pages.
> Alternatively, we can really go for
> page = is_vmalloc_or_module_addr(addr) ? vmalloc_to_page(addr)
> : virt_to_page(addr)
> *pages++ = get_page(page);
actually, no we cannot. Thinking some more about it, that
"get_page(page)" is wrong in _all_ cases. It actually works better for
vmalloc pages than for normal 1:1 pages, since it's actually seriously
and *horrendously* wrong for the case of random kernel addresses which
may not even be refcounted to begin with.
So the whole "get_page()" thing is broken. Iterating over pages in a
KVEC is simply wrong, wrong, wrong. It needs to fail.
Iterating over a KVEC to *copy* data is ok. But no page lookup stuff
or page reference things.
The old code that apparently used "get_user_pages_fast()" was ok
almost by mistake, because it fails on all kernel pages.
Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
