Consider this code snippet:

  struct node {
    long key;
    bpf_list_node l;
    bpf_rb_node r;
    bpf_refcount ref;
  }

  int some_bpf_prog(void *ctx)
  {
    struct node *n = bpf_obj_new(/*...*/), *m;

    bpf_spin_lock(&glock);

    bpf_rbtree_add(&some_tree, &n->r, /* ... */);
    m = bpf_refcount_acquire(n);
    bpf_rbtree_add(&other_tree, &m->r, /* ... */);

    bpf_spin_unlock(&glock);

    /* ... */
  }

After bpf_refcount_acquire, n and m point to the same underlying memory,
and that node's bpf_rb_node field is being used by the some_tree insert,
so overwriting it as a result of the second insert is an error. In order
to properly support refcounted nodes, the rbtree and list insert
functions must be allowed to fail. This patch adds such support.

The kfuncs bpf_rbtree_add, bpf_list_push_{front,back} are modified to
return an int indicating success/failure, with 0 -> success, nonzero ->
failure.

bpf_obj_drop on failure
=======================

Currently the only reason an insert can fail is the example above: the
bpf_{list,rb}_node is already in use. When such a failure occurs, the
insert kfuncs will bpf_obj_drop the input node. This allows the insert
operations to logically fail without changing their verifier owning ref
behavior, namely the unconditional release_reference of the input
owning ref.

With insert that always succeeds, ownership of the node is always passed
to the collection, since the node always ends up in the collection.

With a possibly-failed insert w/ bpf_obj_drop, ownership of the node
is always passed either to the collection (success), or to bpf_obj_drop
(failure). Regardless, it's correct to continue unconditionally
releasing the input owning ref, as something is always taking ownership
from the calling program on insert.

Keeping owning ref behavior unchanged results in a nice default UX for
insert functions that can fail. If the program's reaction to a failed
insert is "fine, just get rid of this owning ref for me and let me go
on with my business", then there's no reason to check for failure since
that's default behavior. e.g.:

  long important_failures = 0;

  int some_bpf_prog(void *ctx)
  {
    struct node *n, *m, *o; /* all bpf_obj_new'd */

    bpf_spin_lock(&glock);
    bpf_rbtree_add(&some_tree, &n->node, /* ... */);
    bpf_rbtree_add(&some_tree, &m->node, /* ... */);
    if (bpf_rbtree_add(&some_tree, &o->node, /* ... */)) {
      important_failures++;
    }
    bpf_spin_unlock(&glock);
  }

If we instead chose to pass ownership back to the program on failed
insert - by returning NULL on success or an owning ref on failure -
programs would always have to do something with the returned ref on
failure. The most likely action is probably "I'll just get rid of this
owning ref and go about my business", which ideally would look like:

  if (n = bpf_rbtree_add(&some_tree, &n->node, /* ... */))
    bpf_obj_drop(n);

But bpf_obj_drop isn't allowed in a critical section and inserts must
occur within one, so in reality error handling would become a
hard-to-parse mess.

For refcounted nodes, we can replicate the "pass ownership back to
program on failure" logic with this patch's semantics, albeit in an ugly
way:

  struct node *n = bpf_obj_new(/* ... */), *m;

  bpf_spin_lock(&glock);

  m = bpf_refcount_acquire(n);
  if (bpf_rbtree_add(&some_tree, &n->node, /* ... */)) {
    /* Do something with m */
  }

  bpf_spin_unlock(&glock);
  bpf_obj_drop(m);

bpf_refcount_acquire is used to simulate "return owning ref on failure".
This should be an uncommon occurrence, though.

Addition of two verifier-fixup'd args to collection inserts
===========================================================

The actual bpf_obj_drop kfunc is
bpf_obj_drop_impl(void *, struct btf_struct_meta *), with bpf_obj_drop
macro populating the second arg with 0 and the verifier later filling in
the arg during insn fixup.

Because bpf_rbtree_add and bpf_list_push_{front,back} now might do
bpf_obj_drop, these kfuncs need a btf_struct_meta parameter that can be
passed to bpf_obj_drop_impl.

Similarly, because the 'node' param to those insert functions is the
bpf_{list,rb}_node within the node type, and bpf_obj_drop expects a
pointer to the beginning of the node, the insert functions need to be
able to find the beginning of the node struct. A second
verifier-populated param is necessary: the offset of {list,rb}_node within the
node type.

These two new params allow the insert kfuncs to correctly call
__bpf_obj_drop_impl:

  beginning_of_node = bpf_rb_node_ptr - offset
  if (already_inserted)
    __bpf_obj_drop_impl(beginning_of_node, btf_struct_meta->record);

Similarly to other kfuncs with "hidden" verifier-populated params, the
insert functions are renamed with _impl prefix and a macro is provided
for common usage. For example, bpf_rbtree_add kfunc is now
bpf_rbtree_add_impl and bpf_rbtree_add is now a macro which sets
"hidden" args to 0.

Due to the two new args BPF progs will need to be recompiled to work
with the new _impl kfuncs.

This patch also rewrites the "hidden argument" explanation to more
directly say why the BPF program writer doesn't need to populate the
arguments with anything meaningful.

How does this new logic affect non-owning references?
=====================================================

Currently, non-owning refs are valid until the end of the critical
section in which they're created. We can make this guarantee because, if
a non-owning ref exists, the referent was added to some collection. The
collection will drop() its nodes when it goes away, but it can't go away
while our program is accessing it, so that's not a problem. If the
referent is removed from the collection in the same CS that it was added
in, it can't be bpf_obj_drop'd until after CS end. Those are the only
two ways to free the referent's memory and neither can happen until
after the non-owning ref's lifetime ends.

On first glance, having these collection insert functions potentially
bpf_obj_drop their input seems like it breaks the "can't be
bpf_obj_drop'd until after CS end" line of reasoning. But we care about
the memory not being _freed_ until end of CS end, and a previous patch
in the series modified bpf_obj_drop such that it doesn't free refcounted
nodes until refcount == 0. So the statement can be more accurately
rewritten as "can't be free'd until after CS end".

We can prove that this rewritten statement holds for any non-owning
reference produced by collection insert functions:

* If the input to the insert function is _not_ refcounted
  * We have an owning reference to the input, and can conclude it isn't
    in any collection
    * Inserting a node in a collection turns owning refs into
      non-owning, and since our input type isn't refcounted, there's no
      way to obtain additional owning refs to the same underlying
      memory
  * Because our node isn't in any collection, the insert operation
    cannot fail, so bpf_obj_drop will not execute
  * If bpf_obj_drop is guaranteed not to execute, there's no risk of
    memory being free'd

* Otherwise, the input to the insert function is refcounted
  * If the insert operation fails due to the node's list_head or rb_root
    already being in some collection, there was some previous successful
    insert which passed refcount to the collection
  * We have an owning reference to the input, it must have been
    acquired via bpf_refcount_acquire, which bumped the refcount
  * refcount must be >= 2 since there's a valid owning reference and the
    node is already in a collection
  * Insert triggering bpf_obj_drop will decr refcount to >= 1, never
    resulting in a free

So although we may do bpf_obj_drop during the critical section, this
will never result in memory being free'd, and no changes to non-owning
ref logic are needed in this patch.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
Link: https://lore.kernel.org/r/20230415201811.343116-6-davemarchevsky@fb.comSigned-off-by: Alexei Starovoitov <ast@kernel.org>