On 4/29/20 10:31 AM, Christian Borntraeger wrote: > On 29.04.20 18:07, Dave Hansen wrote: >> On 4/28/20 3:50 PM, Claudio Imbrenda wrote: >>> If a page is inaccesible and it is used for things like sendfile, then >>> the content of the page is not always touched, and can be passed >>> directly to a driver, causing issues. >>> >>> This patch fixes the issue by adding a call to arch_make_page_accessible >>> in page_cache_pipe_buf_confirm; this fixes the issue. >> I spent about 5 minutes putting together a patch: >> >> https://sr71.net/~dave/intel/accessible.patch >> >> It adds a page flag ("daccess") which starts out set. It clears the >> flag it when the page is added to the page cache or mapped as anonymous. > And that of course does not work. Pages are not made unaccessible at a random > point in time. We do check for several page flags and page count before doing > so and we also do this while with paqe_ref_freeze to avoid several races. > I guess you just hit one of those. Actually, that's the problem. You've gone through all these careful checks and made the page inaccessible. *After* that process, how do you keep the page from being hit by an I/O device before it's made accessible again? My patch just assumes that *all* pages have gone through that process and passed those checks. I'm pretty sure if I lifted all the checks in arch/s390/kernel/uv.c::make_secure_pte() and duplicated them at the sites where I'm doing ClearPageAccessible(), they'd happily pass. Freezing page refs is a transient thing you do *during* the conversion, but it doesn't stop future access to the page. That's what these incomplete hooks are trying to do. Anyway, I look forward to seeing the patch for the FOLL_PIN issue I pointed out, and I hope to see another copy of the fs/splice changes with a proper changelog and the maintainer on cc. It's starting to get late in the rc's.
