On Tue, 2 Apr 2024 18:09:21 +0300 Mike Rapoport <rppt@xxxxxxxxxxxxx> wrote: > On Tue, Apr 02, 2024 at 04:36:11PM +0800, Herbert Xu wrote: > > On Wed, Mar 20, 2024 at 10:57:53AM +0800, xingwei lee wrote: > > > > > > syscall(__NR_bind, /*fd=*/r[0], /*addr=*/0x20000000ul, /*addrlen=*/0x58ul); > > > res = syscall(__NR_accept, /*fd=*/r[0], /*peer=*/0ul, /*peerlen=*/0ul); > > > if (res != -1) > > > r[1] = res; > > > res = syscall(__NR_memfd_secret, /*flags=*/0ul); > > > if (res != -1) > > > r[2] = res; > > > > So this is the key to the issue. The whole point of memfd_secret is > > to make the pages inaccessible to the kernel. The issue is those > > pages are then gifted to the kernel through sendmsg. Somewhere > > along the line someone is supposed to throw up an error about this, > > or map the pages properly. I guess neither happened which is why > > we end up with a page fault. > > Yeah, there was a bug in folio_is_secretmem() that should have throw an > error about this. > > David Hildenbrand sent a fix, it's in Andrew's tree > > https://lore.kernel.org/all/20240326143210.291116-1-david@xxxxxxxxxx I'll send "mm/secretmem: fix GUP-fast succeeding on secretmem folios" upstream this week.
