Merge branch 'Transit between BPF TCP congestion controls.'

Kui-Feng Lee says:

====================

Major changes:

 - Create bpf_links in the kernel for BPF struct_ops to register and
   unregister it.

 - Enables switching between implementations of bpf-tcp-cc under a
   name instantly by replacing the backing struct_ops map of a
   bpf_link.

Previously, BPF struct_ops didn't go off, as even when the user
program creating it was terminated, none of these ever were pinned.
For instance, the TCP congestion control subsystem indirectly
maintains a reference count on the struct_ops of any registered BPF
implemented algorithm. Thus, the algorithm won't be deactivated until
someone deliberately unregisters it.  For compatibility with other BPF
programs, bpf_links have been created to work in coordination with
struct_ops maps. This ensures that the registration and unregistration
of these respective maps is carried out at the start and end of the
bpf_link.

We also faced complications when attempting to replace an existing TCP
congestion control algorithm with a new implementation on the fly. A
struct_ops map was used to register a TCP congestion control algorithm
with a unique name.  We had to either register the alternative
implementation with a new name and move over or unregister the current
one before being able to reregistration with the same name.  To fix
this problem, we can an option to migrate the registration of the
algorithm from struct_ops maps to bpf_links. By modifying the backing
map of a bpf_link, it suddenly becomes possible to replace an existing
TCP congestion control algorithm with ease.
---

The major differences from v11:

 - Fix incorrectly setting both old_prog_fd and old_map_fd.

The major differences from v10:

 - Add old_map_fd as an additional field instead of an union in
   bpf_link_update_opts.

The major differences from v9:

 - Add test case for BPF_F_LINK.  Includes adding old_map_fd to struct
   bpf_link_update_opts in patch 6.

 - Return -EPERM instead of -EINVAL when the old map fd doesn't match
   with BPF_F_LINK.

 - Fix -EBUSY case in bpf_map__attach_struct_ops().

The major differences form v8:

 - Check bpf_struct_ops::{validate,update} in
   bpf_struct_ops_map_alloc()

The major differences from v7:

 - Use synchronize_rcu_mult(call_rcu, call_rcu_tasks) to replace
   synchronize_rcu() and synchronize_rcu_tasks().

 - Call synchronize_rcu() in tcp_update_congestion_control().

 - Handle -EBUSY in bpf_map__attach_struct_ops() to allow a struct_ops
   can be used to create links more than once.  Include a test case.

 - Add old_map_fd to bpf_attr and handle BPF_F_REPLACE in
   bpf_struct_ops_map_link_update().

 - Remove changes in bpf_dummy_struct_ops.c and add a check of .update
   function pointer of bpf_struct_ops.

The major differences from v6:

 - Reword commit logs of the patch 1, 2, and 8.

 - Call synchronize_rcu_tasks() as well in bpf_struct_ops_map_free().

 - Refactor bpf_struct_ops_map_free() so that
   bpf_struct_ops_map_alloc() can free a struct_ops without waiting
   for a RCU grace period.

The major differences from v5:

 - Add a new step to bpf_object__load() to prepare vdata.

 - Accept BPF_F_REPLACE.

 - Check section IDs in find_struct_ops_map_by_offset()

 - Add a test case to check mixing w/ and w/o link struct_ops.

 - Add a test case of using struct_ops w/o link to update a link.

 - Improve bpf_link__detach_struct_ops() to handle the w/ link case.

The major differences from v4:

 - Rebase.

 - Reorder patches and merge part 4 to part 2 of the v4.

The major differences from v3:

 - Remove bpf_struct_ops_map_free_rcu(), and use synchronize_rcu().

 - Improve the commit log of the part 1.

 - Before transitioning to the READY state, we conduct a value check
   to ensure that struct_ops can be successfully utilized and links
   created later.

The major differences from v2:

 - Simplify states

   - Remove TOBEUNREG.

   - Rename UNREG to READY.

 - Stop using the refcnt of the kvalue of a struct_ops. Explicitly
   increase and decrease the refcount of struct_ops.

 - Prepare kernel vdata during the load phase of libbpf.

The major differences from v1:

 - Added bpf_struct_ops_link to replace the previous union-based
   approach.

 - Added UNREG and TOBEUNREG to the state of bpf_struct_ops_map.

   - bpf_struct_ops_transit_state() maintains state transitions.

 - Fixed synchronization issue.

 - Prepare kernel vdata of struct_ops during the loading phase of
   bpf_object.

 - Merged previous patch 3 to patch 1.

v11: https://lore.kernel.org/all/20230323010409.2265383-1-kuifeng@meta.com/
v10: https://lore.kernel.org/all/20230321232813.3376064-1-kuifeng@meta.com/
v9: https://lore.kernel.org/all/20230320195644.1953096-1-kuifeng@meta.com/
v8: https://lore.kernel.org/all/20230318053144.1180301-1-kuifeng@meta.com/
v7: https://lore.kernel.org/all/20230316023641.2092778-1-kuifeng@meta.com/
v6: https://lore.kernel.org/all/20230310043812.3087672-1-kuifeng@meta.com/
v5: https://lore.kernel.org/all/20230308005050.255859-1-kuifeng@meta.com/
v4: https://lore.kernel.org/all/20230307232913.576893-1-andrii@kernel.org/
v3: https://lore.kernel.org/all/20230303012122.852654-1-kuifeng@meta.com/
v2: https://lore.kernel.org/bpf/20230223011238.12313-1-kuifeng@meta.com/
v1: https://lore.kernel.org/bpf/20230214221718.503964-1-kuifeng@meta.com/
====================

Signed-off-by: Martin KaFai Lau <martin.lau@kernel.org>

1 files changed, 224 insertions, 30 deletions

diff --git a/kernel/bpf/bpf_struct_ops.c b/kernel/bpf/bpf_struct_ops.c
index ba7a94276e3b..6401deca3b56 100644
--- a/kernel/bpf/bpf_struct_ops.c
+++ b/kernel/bpf/bpf_struct_ops.c
@@ -11,11 +11,13 @@
 #include <linux/refcount.h>
 #include <linux/mutex.h>
 #include <linux/btf_ids.h>
+#include <linux/rcupdate_wait.h>
 
 enum bpf_struct_ops_state {
 	BPF_STRUCT_OPS_STATE_INIT,
 	BPF_STRUCT_OPS_STATE_INUSE,
 	BPF_STRUCT_OPS_STATE_TOBEFREE,
+	BPF_STRUCT_OPS_STATE_READY,
 };
 
 #define BPF_STRUCT_OPS_COMMON_VALUE			\
@@ -58,6 +60,13 @@ struct bpf_struct_ops_map {
 	struct bpf_struct_ops_value kvalue;
 };
 
+struct bpf_struct_ops_link {
+	struct bpf_link link;
+	struct bpf_map __rcu *map;
+};
+
+static DEFINE_MUTEX(update_mutex);
+
 #define VALUE_PREFIX "bpf_struct_ops_"
 #define VALUE_PREFIX_LEN (sizeof(VALUE_PREFIX) - 1)
 
@@ -249,6 +258,7 @@ int bpf_struct_ops_map_sys_lookup_elem(struct bpf_map *map, void *key,
 	struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
 	struct bpf_struct_ops_value *uvalue, *kvalue;
 	enum bpf_struct_ops_state state;
+	s64 refcnt;
 
 	if (unlikely(*(u32 *)key != 0))
 		return -ENOENT;
@@ -267,7 +277,14 @@ int bpf_struct_ops_map_sys_lookup_elem(struct bpf_map *map, void *key,
 	uvalue = value;
 	memcpy(uvalue, st_map->uvalue, map->value_size);
 	uvalue->state = state;
-	refcount_set(&uvalue->refcnt, refcount_read(&kvalue->refcnt));
+
+	/* This value offers the user space a general estimate of how
+	 * many sockets are still utilizing this struct_ops for TCP
+	 * congestion control. The number might not be exact, but it
+	 * should sufficiently meet our present goals.
+	 */
+	refcnt = atomic64_read(&map->refcnt) - atomic64_read(&map->usercnt);
+	refcount_set(&uvalue->refcnt, max_t(s64, refcnt, 0));
 
 	return 0;
 }
@@ -491,12 +508,29 @@ static long bpf_struct_ops_map_update_elem(struct bpf_map *map, void *key,
 		*(unsigned long *)(udata + moff) = prog->aux->id;
 	}
 
-	refcount_set(&kvalue->refcnt, 1);
-	bpf_map_inc(map);
+	if (st_map->map.map_flags & BPF_F_LINK) {
+		err = st_ops->validate(kdata);
+		if (err)
+			goto reset_unlock;
+		set_memory_rox((long)st_map->image, 1);
+		/* Let bpf_link handle registration & unregistration.
+		 *
+		 * Pair with smp_load_acquire() during lookup_elem().
+		 */
+		smp_store_release(&kvalue->state, BPF_STRUCT_OPS_STATE_READY);
+		goto unlock;
+	}
 
 	set_memory_rox((long)st_map->image, 1);
 	err = st_ops->reg(kdata);
 	if (likely(!err)) {
+		/* This refcnt increment on the map here after
+		 * 'st_ops->reg()' is secure since the state of the
+		 * map must be set to INIT at this moment, and thus
+		 * bpf_struct_ops_map_delete_elem() can't unregister
+		 * or transition it to TOBEFREE concurrently.
+		 */
+		bpf_map_inc(map);
 		/* Pair with smp_load_acquire() during lookup_elem().
 		 * It ensures the above udata updates (e.g. prog->aux->id)
 		 * can be seen once BPF_STRUCT_OPS_STATE_INUSE is set.
@@ -512,7 +546,6 @@ static long bpf_struct_ops_map_update_elem(struct bpf_map *map, void *key,
 	 */
 	set_memory_nx((long)st_map->image, 1);
 	set_memory_rw((long)st_map->image, 1);
-	bpf_map_put(map);
 
 reset_unlock:
 	bpf_struct_ops_map_put_progs(st_map);
@@ -530,14 +563,16 @@ static long bpf_struct_ops_map_delete_elem(struct bpf_map *map, void *key)
 	struct bpf_struct_ops_map *st_map;
 
 	st_map = (struct bpf_struct_ops_map *)map;
+	if (st_map->map.map_flags & BPF_F_LINK)
+		return -EOPNOTSUPP;
+
 	prev_state = cmpxchg(&st_map->kvalue.state,
 			     BPF_STRUCT_OPS_STATE_INUSE,
 			     BPF_STRUCT_OPS_STATE_TOBEFREE);
 	switch (prev_state) {
 	case BPF_STRUCT_OPS_STATE_INUSE:
 		st_map->st_ops->unreg(&st_map->kvalue.data);
-		if (refcount_dec_and_test(&st_map->kvalue.refcnt))
-			bpf_map_put(map);
+		bpf_map_put(map);
 		return 0;
 	case BPF_STRUCT_OPS_STATE_TOBEFREE:
 		return -EINPROGRESS;
@@ -570,7 +605,7 @@ static void bpf_struct_ops_map_seq_show_elem(struct bpf_map *map, void *key,
 	kfree(value);
 }
 
-static void bpf_struct_ops_map_free(struct bpf_map *map)
+static void __bpf_struct_ops_map_free(struct bpf_map *map)
 {
 	struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
 
@@ -582,10 +617,32 @@ static void bpf_struct_ops_map_free(struct bpf_map *map)
 	bpf_map_area_free(st_map);
 }
 
+static void bpf_struct_ops_map_free(struct bpf_map *map)
+{
+	/* The struct_ops's function may switch to another struct_ops.
+	 *
+	 * For example, bpf_tcp_cc_x->init() may switch to
+	 * another tcp_cc_y by calling
+	 * setsockopt(TCP_CONGESTION, "tcp_cc_y").
+	 * During the switch,  bpf_struct_ops_put(tcp_cc_x) is called
+	 * and its refcount may reach 0 which then free its
+	 * trampoline image while tcp_cc_x is still running.
+	 *
+	 * A vanilla rcu gp is to wait for all bpf-tcp-cc prog
+	 * to finish. bpf-tcp-cc prog is non sleepable.
+	 * A rcu_tasks gp is to wait for the last few insn
+	 * in the tramopline image to finish before releasing
+	 * the trampoline image.
+	 */
+	synchronize_rcu_mult(call_rcu, call_rcu_tasks);
+
+	__bpf_struct_ops_map_free(map);
+}
+
 static int bpf_struct_ops_map_alloc_check(union bpf_attr *attr)
 {
 	if (attr->key_size != sizeof(unsigned int) || attr->max_entries != 1 ||
-	    attr->map_flags || !attr->btf_vmlinux_value_type_id)
+	    (attr->map_flags & ~BPF_F_LINK) || !attr->btf_vmlinux_value_type_id)
 		return -EINVAL;
 	return 0;
 }
@@ -609,6 +666,9 @@ static struct bpf_map *bpf_struct_ops_map_alloc(union bpf_attr *attr)
 	if (attr->value_size != vt->size)
 		return ERR_PTR(-EINVAL);
 
+	if (attr->map_flags & BPF_F_LINK && (!st_ops->validate || !st_ops->update))
+		return ERR_PTR(-EOPNOTSUPP);
+
 	t = st_ops->type;
 
 	st_map_size = sizeof(*st_map) +
@@ -630,7 +690,7 @@ static struct bpf_map *bpf_struct_ops_map_alloc(union bpf_attr *attr)
 				   NUMA_NO_NODE);
 	st_map->image = bpf_jit_alloc_exec(PAGE_SIZE);
 	if (!st_map->uvalue || !st_map->links || !st_map->image) {
-		bpf_struct_ops_map_free(map);
+		__bpf_struct_ops_map_free(map);
 		return ERR_PTR(-ENOMEM);
 	}
 
@@ -676,41 +736,175 @@ const struct bpf_map_ops bpf_struct_ops_map_ops = {
 bool bpf_struct_ops_get(const void *kdata)
 {
 	struct bpf_struct_ops_value *kvalue;
+	struct bpf_struct_ops_map *st_map;
+	struct bpf_map *map;
 
 	kvalue = container_of(kdata, struct bpf_struct_ops_value, data);
+	st_map = container_of(kvalue, struct bpf_struct_ops_map, kvalue);
 
-	return refcount_inc_not_zero(&kvalue->refcnt);
+	map = __bpf_map_inc_not_zero(&st_map->map, false);
+	return !IS_ERR(map);
 }
 
-static void bpf_struct_ops_put_rcu(struct rcu_head *head)
+void bpf_struct_ops_put(const void *kdata)
 {
+	struct bpf_struct_ops_value *kvalue;
 	struct bpf_struct_ops_map *st_map;
 
-	st_map = container_of(head, struct bpf_struct_ops_map, rcu);
+	kvalue = container_of(kdata, struct bpf_struct_ops_value, data);
+	st_map = container_of(kvalue, struct bpf_struct_ops_map, kvalue);
+
 	bpf_map_put(&st_map->map);
 }
 
-void bpf_struct_ops_put(const void *kdata)
+static bool bpf_struct_ops_valid_to_reg(struct bpf_map *map)
 {
-	struct bpf_struct_ops_value *kvalue;
+	struct bpf_struct_ops_map *st_map = (struct bpf_struct_ops_map *)map;
 
-	kvalue = container_of(kdata, struct bpf_struct_ops_value, data);
-	if (refcount_dec_and_test(&kvalue->refcnt)) {
-		struct bpf_struct_ops_map *st_map;
+	return map->map_type == BPF_MAP_TYPE_STRUCT_OPS &&
+		map->map_flags & BPF_F_LINK &&
+		/* Pair with smp_store_release() during map_update */
+		smp_load_acquire(&st_map->kvalue.state) == BPF_STRUCT_OPS_STATE_READY;
+}
 
-		st_map = container_of(kvalue, struct bpf_struct_ops_map,
-				      kvalue);
-		/* The struct_ops's function may switch to another struct_ops.
-		 *
-		 * For example, bpf_tcp_cc_x->init() may switch to
-		 * another tcp_cc_y by calling
-		 * setsockopt(TCP_CONGESTION, "tcp_cc_y").
-		 * During the switch,  bpf_struct_ops_put(tcp_cc_x) is called
-		 * and its map->refcnt may reach 0 which then free its
-		 * trampoline image while tcp_cc_x is still running.
-		 *
-		 * Thus, a rcu grace period is needed here.
+static void bpf_struct_ops_map_link_dealloc(struct bpf_link *link)
+{
+	struct bpf_struct_ops_link *st_link;
+	struct bpf_struct_ops_map *st_map;
+
+	st_link = container_of(link, struct bpf_struct_ops_link, link);
+	st_map = (struct bpf_struct_ops_map *)
+		rcu_dereference_protected(st_link->map, true);
+	if (st_map) {
+		/* st_link->map can be NULL if
+		 * bpf_struct_ops_link_create() fails to register.
 		 */
-		call_rcu(&st_map->rcu, bpf_struct_ops_put_rcu);
+		st_map->st_ops->unreg(&st_map->kvalue.data);
+		bpf_map_put(&st_map->map);
 	}
+	kfree(st_link);
 }
+
+static void bpf_struct_ops_map_link_show_fdinfo(const struct bpf_link *link,
+					    struct seq_file *seq)
+{
+	struct bpf_struct_ops_link *st_link;
+	struct bpf_map *map;
+
+	st_link = container_of(link, struct bpf_struct_ops_link, link);
+	rcu_read_lock();
+	map = rcu_dereference(st_link->map);
+	seq_printf(seq, "map_id:\t%d\n", map->id);
+	rcu_read_unlock();
+}
+
+static int bpf_struct_ops_map_link_fill_link_info(const struct bpf_link *link,
+					       struct bpf_link_info *info)
+{
+	struct bpf_struct_ops_link *st_link;
+	struct bpf_map *map;
+
+	st_link = container_of(link, struct bpf_struct_ops_link, link);
+	rcu_read_lock();
+	map = rcu_dereference(st_link->map);
+	info->struct_ops.map_id = map->id;
+	rcu_read_unlock();
+	return 0;
+}
+
+static int bpf_struct_ops_map_link_update(struct bpf_link *link, struct bpf_map *new_map,
+					  struct bpf_map *expected_old_map)
+{
+	struct bpf_struct_ops_map *st_map, *old_st_map;
+	struct bpf_map *old_map;
+	struct bpf_struct_ops_link *st_link;
+	int err = 0;
+
+	st_link = container_of(link, struct bpf_struct_ops_link, link);
+	st_map = container_of(new_map, struct bpf_struct_ops_map, map);
+
+	if (!bpf_struct_ops_valid_to_reg(new_map))
+		return -EINVAL;
+
+	mutex_lock(&update_mutex);
+
+	old_map = rcu_dereference_protected(st_link->map, lockdep_is_held(&update_mutex));
+	if (expected_old_map && old_map != expected_old_map) {
+		err = -EPERM;
+		goto err_out;
+	}
+
+	old_st_map = container_of(old_map, struct bpf_struct_ops_map, map);
+	/* The new and old struct_ops must be the same type. */
+	if (st_map->st_ops != old_st_map->st_ops) {
+		err = -EINVAL;
+		goto err_out;
+	}
+
+	err = st_map->st_ops->update(st_map->kvalue.data, old_st_map->kvalue.data);
+	if (err)
+		goto err_out;
+
+	bpf_map_inc(new_map);
+	rcu_assign_pointer(st_link->map, new_map);
+	bpf_map_put(old_map);
+
+err_out:
+	mutex_unlock(&update_mutex);
+
+	return err;
+}
+
+static const struct bpf_link_ops bpf_struct_ops_map_lops = {
+	.dealloc = bpf_struct_ops_map_link_dealloc,
+	.show_fdinfo = bpf_struct_ops_map_link_show_fdinfo,
+	.fill_link_info = bpf_struct_ops_map_link_fill_link_info,
+	.update_map = bpf_struct_ops_map_link_update,
+};
+
+int bpf_struct_ops_link_create(union bpf_attr *attr)
+{
+	struct bpf_struct_ops_link *link = NULL;
+	struct bpf_link_primer link_primer;
+	struct bpf_struct_ops_map *st_map;
+	struct bpf_map *map;
+	int err;
+
+	map = bpf_map_get(attr->link_create.map_fd);
+	if (!map)
+		return -EINVAL;
+
+	st_map = (struct bpf_struct_ops_map *)map;
+
+	if (!bpf_struct_ops_valid_to_reg(map)) {
+		err = -EINVAL;
+		goto err_out;
+	}
+
+	link = kzalloc(sizeof(*link), GFP_USER);
+	if (!link) {
+		err = -ENOMEM;
+		goto err_out;
+	}
+	bpf_link_init(&link->link, BPF_LINK_TYPE_STRUCT_OPS, &bpf_struct_ops_map_lops, NULL);
+
+	err = bpf_link_prime(&link->link, &link_primer);
+	if (err)
+		goto err_out;
+
+	err = st_map->st_ops->reg(st_map->kvalue.data);
+	if (err) {
+		bpf_link_cleanup(&link_primer);
+		link = NULL;
+		goto err_out;
+	}
+	RCU_INIT_POINTER(link->map, map);
+
+	return bpf_link_settle(&link_primer);
+
+err_out:
+	bpf_map_put(map);
+	kfree(link);
+	return err;
+}
+