lkml.org 
[lkml]   [2023]   [Jan]   [31]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v1] sparc/mm: don't unconditionally set HW writable bit when setting PTE dirty on 64bit
On Tue, Jan 31, 2023 at 09:47:01AM +0100, David Hildenbrand wrote:
> On 12.12.22 14:02, David Hildenbrand wrote:
> > On sparc64, there is no HW modified bit, therefore, SW tracks via a SW
> > bit if the PTE is dirty via pte_mkdirty(). However, pte_mkdirty()
> > currently also unconditionally sets the HW writable bit, which is wrong.
> >
> > pte_mkdirty() is not supposed to make a PTE actually writable, unless the
> > SW writable bit (pte_write()) indicates that the PTE is not
> > write-protected. Fortunately, sparc64 also defines a SW writable bit.
> >
> > For example, this already turned into a problem in the context of
> > THP splitting as documented in commit 624a2c94f5b7 ("Partly revert "mm/thp:
> > carry over dirty bit when thp splits on pmd") and might be an issue during
> > page migration in mm/migrate.c:remove_migration_pte() as well where we:
> > if (folio_test_dirty(folio) && is_migration_entry_dirty(entry))
> > pte = pte_mkdirty(pte);
> >
> > But more general, anything like:
> > maybe_mkwrite(pte_mkdirty(pte), vma)
> > code is broken on sparc64, because it will unconditionally set the HW
> > writable bit even if the SW writable bit is not set.
> >
> > Simple reproducer that will result in a writable PTE after ptrace
> > access, to highlight the problem and as an easy way to verify if it has
> > been fixed:
> >
> > --------------------------------------------------------------------------
> > #include <fcntl.h>
> > #include <signal.h>
> > #include <unistd.h>
> > #include <string.h>
> > #include <errno.h>
> > #include <stdlib.h>
> > #include <sys/mman.h>
> >
> > static void signal_handler(int sig)
> > {
> > if (sig == SIGSEGV)
> > printf("[PASS] SIGSEGV generated\n");
> > else
> > printf("[FAIL] wrong signal generated\n");
> > exit(0);
> > }
> >
> > int main(void)
> > {
> > size_t pagesize = getpagesize();
> > char data = 1;
> > off_t offs;
> > int mem_fd;
> > char *map;
> > int ret;
> >
> > mem_fd = open("/proc/self/mem", O_RDWR);
> > if (mem_fd < 0) {
> > fprintf(stderr, "open(/proc/self/mem) failed: %d\n", errno);
> > return 1;
> > }
> >
> > map = mmap(NULL, pagesize, PROT_READ, MAP_PRIVATE|MAP_ANON, -1 ,0);
> > if (map == MAP_FAILED) {
> > fprintf(stderr, "mmap() failed: %d\n", errno);
> > return 1;
> > }
> >
> > printf("original: %x\n", *map);
> >
> > /* debug access */
> > offs = lseek(mem_fd, (uintptr_t) map, SEEK_SET);
> > ret = write(mem_fd, &data, 1);
> > if (ret != 1) {
> > fprintf(stderr, "pwrite(/proc/self/mem) failed with %d: %d\n", ret, errno);
> > return 1;
> > }
> > if (*map != data) {
> > fprintf(stderr, "pwrite(/proc/self/mem) not visible\n");
> > return 1;
> > }
> >
> > printf("ptrace: %x\n", *map);
> >
> > /* Install signal handler. */
> > if (signal(SIGSEGV, signal_handler) == SIG_ERR) {
> > fprintf(stderr, "signal() failed\n");
> > return 1;
> > }
> >
> > /* Ordinary access. */
> > *map = 2;
> >
> > printf("access: %x\n", *map);
> >
> > printf("[FAIL] SIGSEGV not generated\n");
> >
> > return 0;
> > }
> > --------------------------------------------------------------------------
> >
> > Without this commit (sun4u in QEMU):
> > # ./reproducer
> > original: 0
> > ptrace: 1
> > access: 2
> > [FAIL] SIGSEGV not generated
> >
> > Let's fix this by setting the HW writable bit only if both, the SW dirty
> > bit and the SW writable bit are set. This matches, for example, how
> > s390x handles pte_mkwrite() and pte_mkdirty() -- except, that they have
> > to clear the _PAGE_PROTECT bit.
> >
> > We have to move pte_dirty() and pte_dirty() up. The code patching
> > mechanism and handling constants > 22bit is a bit special on sparc64.
> >
> > With this commit (sun4u in QEMU):
> > # ./reproducer
> > original: 0
> > ptrace: 1
> > [PASS] SIGSEGV generated
> >
> > This handling seems to have been in place forever.
> >
> > Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> > Cc: Andrew Morton <akpm@linux-foundation.org>
> > Cc: "David S. Miller" <davem@davemloft.net>
> > Cc: Peter Xu <peterx@redhat.com>
> > Cc: Hev <r@hev.cc>
> > Cc: Anatoly Pugachev <matorola@gmail.com>
> > Cc: Raghavendra K T <raghavendra.kt@amd.com>
> > Cc: Thorsten Leemhuis <regressions@leemhuis.info>
> > Cc: Mike Kravetz <mike.kravetz@oracle.com>
> > Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
> > Cc: Juergen Gross <jgross@suse.com>
> > Signed-off-by: David Hildenbrand <david@redhat.com>
> > ---
>
> Ping

I agree with David that the current sparc64 impl of pte_mkdirty is
suspecious.

What David mentioned on page migration above is correct and has another
report here from Nick recently:

https://lore.kernel.org/all/CADyTPEzsvdRC15+Z5T3oryofwRYqHmHzwqRmJKJoHB3d7Tdayw@mail.gmail.com/

If this patch is hopefully correct (which I cannot tell as I know little on
sparc64) and can be merged, it'll be the cleanest solution, comparing to
what I provided here:

https://lore.kernel.org/all/Y9bvwz4FIOQ+D8c4@x1n/

And I assume it'll also fix things like the reproducer being attached on
wrongly applying write bit with FOLL_FORCE, so it fixes more than that.

I plan to keep posting that fix I referenced above for the breakage because
that'll still be the safest so far, but that can change if someone from
sparc64 can have a look at this and ack it.

Thanks,

--
Peter Xu

\
 
 \ /
  Last update: 2023-03-27 00:04    [W:0.107 / U:0.328 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site