RDS: use pin_user_pages_fast()

pin_user_pages_fast() is used in the upstream version of
rds_pin_pages(), which is better because so long as all the pages are
faulted-in, the operation is lockless. The migration was originally
performed 15 years ago, in 830eb7d ("RDS: use
get_user_pages_fast()"). Unfortunately, several UEK-specific patches
seem to have thad the effect of reverting this change.

In commit 347353a ("rds: Handle unsupported rdma request to fs dax
memory"), we migrated from _fast() to _longterm() as part of UEK5. In
more recent upstream kernels, _longterm() was dropped in favor of the
standard get_user_pages() with FOLL_LONGTERM. Through forward-porting
this patch, the result was that the commit simply switched back from
get_user_pages_fast() to get_user_pages() with FOLL_LONGTERM, adding
back in the mandatory locking.

Finally, d237411 ("net/rds: Switch to the Linux-5.15 API")
converted this call to pin_user_pages(), rather than
pin_user_pages_fast(), which is what is used upstream.

This has caused problems in some cases where RDMA packets are sent
concurrently with very long-running munmap(). To get back the benefit of
lockless page pinning, let's switch back to pin_user_pages_fast(),
matching upstream. I've introduced a bit more noise into this diff in
an effort to make the form of rds_pin_pages exactly match its upstream
variant, given that there is now no functional difference.

Orabug: 37980705

Signed-off-by: Stephen Brennan <[email protected]>
Reviewed-by: Håkon Bugge <[email protected]>
Reviewed-by: Lorenzo Stoakes <[email protected]>
Signed-off-by: Vijayendra Suman <[email protected]>
Signed-off-by: Brian Maly <[email protected]>

Author: brenns10

net/rds/rdma.c

@@ -164,20 +164,17 @@ EXPORT_SYMBOL_GPL(rds_rdma_drop_keys);
 static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
 			struct page **pages, int write)
 {
+	unsigned int gup_flags = FOLL_LONGTERM;
 	int ret;
-	struct mm_struct *mm = current->mm;
-	int gup_flags = FOLL_LONGTERM | (write ? FOLL_WRITE : 0);
 
-	mmgrab(mm);
-	down_read(&mm->mmap_lock);
-	ret = pin_user_pages(user_addr, nr_pages, gup_flags, pages, NULL);
+	if (write)
+		gup_flags |= FOLL_WRITE;
 
-	if (ret >= 0 && (unsigned) ret < nr_pages) {
+	ret = pin_user_pages_fast(user_addr, nr_pages, gup_flags, pages);
+	if (ret >= 0 && ret < nr_pages) {
 		unpin_user_pages(pages, ret);
 		ret = -EFAULT;
 	}
-	up_read(&mm->mmap_lock);
-	mmdrop(mm);
 
 	return ret;
 }