readdir: make user_access_begin() use the real access range

readdir: make user_access_begin() use the real access range

In commit 449246b07d1f ("Convert filldir[64]() from __put_user() to
unsafe_put_user()") I changed filldir to not do individual __put_user()
accesses, but instead use unsafe_put_user() surrounded by the proper
user_access_begin/end() pair.

That make them enormously faster on modern x86, where the STAC/CLAC
games make individual user accesses fairly heavy-weight.

However, the user_access_begin() range was not really the exact right
one, since filldir() has the unfortunate problem that it needs to not
only fill out the new directory entry, it also needs to fix up the
previous one to contain the proper file offset.

It's unfortunate, but the "d_off" field in "struct dirent" is _not_ the
file offset of the directory entry itself - it's the offset of the next
one.  So we end up backfilling the offset in the previous entry as we
walk along.

But since x86 didn't really care about the exact range, and used to be
the only architecture that did anything fancy in user_access_begin() to
begin with, the filldir[64]() changes did something lazy, and even
commented on it:

	/*
	 * Note! This range-checks 'previous' (which may be NULL).
	 * The real range was checked in getdents
	 */
	if (!user_access_begin(dirent, sizeof(*dirent)))
		goto efault;

and it all worked fine.

But now 32-bit ppc is starting to also implement user_access_begin(),
and the fact that we faked the range to only be the (possibly not even
valid) previous directory entry becomes a problem, because ppc32 will
actually be using the range that is passed in for more than just "check
that it's user space".

This is a complete rewrite of Christophe's original patch.

By saving off the record length of the previous entry instead of a
pointer to it in the filldir data structures, we can simplify the range
check and the writing of the previous entry d_off field.  No need for
any conditionals in the user accesses themselves, although we retain the
conditional EINTR checking for the "was this the first directory entry"
signal handling latency logic.

Fixes: 449246b07d1f ("Convert filldir[64]() from __put_user() to unsafe_put_user()")
Link: https://lore.kernel.org/lkml/a02d3426f93f7eb04960a4d9140902d278cab0bb.1579697910.git.christophe.leroy@c-s.fr/
Link: https://lore.kernel.org/lkml/408c90c4068b00ea8f1c41cca45b84ec23d4946b.1579783936.git.christophe.leroy@c-s.fr/
Reported-and-tested-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>