Skip to contentSkip to conversation

mm: Remove the access_ok() call from gup_fast_fallback().

PATCH Applied David Laight 4 participants 5 messages Feb 9, 2025 – Feb 10, 2025

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mm: Remove the access_ok() call from gup_fast_fallback().

Historiaclly the code relied on access_ok() to validate the address range. Commit 26f4c328079d7 added an explicit wrap check before access_ok(). Commit c28b1fc70390d then changed the wrap test to use check_add_overflow(). Commit 6014bc27561f2 relaxed the checks in x86-64's access_ok() and added
an explicit check for TASK_SIZE here to make up for it.
That left a pointless access_ok() call with its associated 'lfence' that
can never actually fail.
So just delete the test.

---
 mm/gup.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
mm/gup.c +1 -3
index 3883b307780e..79a3d2228bf9 100644
@@ -2757,7 +2757,7 @@ EXPORT_SYMBOL(get_user_pages_unlocked);
  *
  *  *) ptes can be read atomically by the architecture.
  *
- *  *) access_ok is sufficient to validate userspace address ranges.
+ *  *) valid user addesses are below TASK_MAX_SIZE
  *
  * The last two assumptions can be relaxed by the addition of helper functions.
  *
@@ -3411,8 +3411,6 @@ static int gup_fast_fallback(unsigned long start, unsigned long nr_pages,
 		return -EOVERFLOW;
 	if (end > TASK_SIZE_MAX)
 		return -EFAULT;
-	if (unlikely(!access_ok((void __user *)start, len)))
-		return -EFAULT;
 
 	nr_pinned = gup_fast(start, end, gup_flags, pages);
 	if (nr_pinned == nr_pages || gup_flags & FOLL_FAST_ONLY)
Signature · 1 line
2.39.5
reviewer Reviewed-by on v1 1/1
#1 lore raw
Quoted 9 lines · David Laight

Historiaclly the code relied on access_ok() to validate the address range. Commit 26f4c328079d7 added an explicit wrap check before access_ok(). Commit c28b1fc70390d then changed the wrap test to use check_add_overflow(). Commit 6014bc27561f2 relaxed the checks in x86-64's access_ok() and added
an explicit check for TASK_SIZE here to make up for it.
That left a pointless access_ok() call with its associated 'lfence' that
can never actually fail.
So just delete the test.

Quoted 4 lines
---
 mm/gup.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

I often wonder about about access_ok() calls, if they still do anything..

Jason

#2 lore raw
Quoted 15 lines · On Sun, 9 Feb 2025 14:24:22 -0400 Jason Gunthorpe <[email protected]> wrote:
Quoted 9 lines · David Laight

Historiaclly the code relied on access_ok() to validate the address range. Commit 26f4c328079d7 added an explicit wrap check before access_ok(). Commit c28b1fc70390d then changed the wrap test to use check_add_overflow(). Commit 6014bc27561f2 relaxed the checks in x86-64's access_ok() and added
an explicit check for TASK_SIZE here to make up for it.
That left a pointless access_ok() call with its associated 'lfence' that
can never actually fail.
So just delete the test.

Quoted 4 lines
---
 mm/gup.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)  
Quoted 4 lines

I often wonder about about access_ok() calls, if they still do anything..

They still do 'stuff' and end up containing a slow memory synchronising instruction (to avoid speculative accesses controlled by the application).

But there are better ways to handle bad user pointers. So, mostly access_ok() isn't needed outside the architecture code that handles userspace accesses.

David

#3 lore raw
Quoted 5 lines · David Laight

They still do 'stuff' and end up containing a slow memory synchronising instruction (to avoid speculative accesses controlled by the application).

But there are better ways to handle bad user pointers.

Not always. If you use __get_user() and friends, you basically *have* to use access_ok(), or you have to be playing games (like some tracing code does, which actually wants to use it as a "I want a kernel pointer _or_ a user pointer{".

Now, admittedly probably nobody should be using __get_user() and friends any more. Almost all the reasons for using it are entirely historical and just not true any more.

But also, comparing against TASK_SIZE_MAX isn't actually the same as access_ok() historically. We've moved in that direction, yes, but we very much used to have a distinction between "this is the fixed maximum", and "this is the actual run-time size".

We've moved towards just using TASK_SIZE_MAX mainly because the run-time size check was annoyingly expensive is some pretty critical code.

But historically the TASK_SIZE_MAX thing is the "this is fast but not exact, use it only for special code that knows what it is doing", and "access_ok()" was "this is proper"

Quoted 2 lines

So, mostly access_ok() isn't needed outside the architecture code that handles userspace accesses.

Oh, there is still a fair amount of code that really does need it.

Admittedly most of it should either be converted to just using regular get/put_user(), or into the modern "user_access_begin()" model, but we do have a number of __get/put_user() users that still very much need that access_ok().

And anything that follows page tables had better check that it's proper. But in that case, I do believe checking for TASK_SIZE_MAX tends to be equivalent.

Linus

Acked-by on v1 1/1
#4 lore raw
In reply to David Laight
Quoted 4 lines · David Laight

Historiaclly the code relied on access_ok() to validate the address range. Commit 26f4c328079d7 added an explicit wrap check before access_ok(). Commit c28b1fc70390d then changed the wrap test to use check_add_overflow(). Commit 6014bc27561f2 relaxed the checks in x86-64's access_ok() and added

checkpatch.pl will correctly tell you about the wrongly quoted commit ids.

E.g.,

ERROR: Please use git commit description style 'commit <12+ chars of sha1> ("<title line>")' - ie: 'Commit 26f4c328079d ("mm: simplify gup_fast_permitted")'

Apart from that

Signature · 3 lines
Cheers,

David / dhildenb