Skip to content

L
linux
Switch branch/tag
  1. 13 Mar, 2023 1 commit
  3. 11 Mar, 2023 4 commits
    • Alexei Starovoitov's avatar
      Merge branch 'Support stashing local kptrs with bpf_kptr_xchg' · 49b5300f
      Alexei Starovoitov authored 
      Dave Marchevsky says:

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

Local kptrs are kptrs allocated via bpf_obj_new with a type specified in program
BTF. A BPF program which creates a local kptr has exclusive control of the
lifetime of the kptr, and, prior to terminating, must:

  * free the kptr via bpf_obj_drop
  * If the kptr is a {list,rbtree} node, add the node to a {list, rbtree},
    thereby passing control of the lifetime to the collection

This series adds a third option:

  * stash the kptr in a map value using bpf_kptr_xchg

As indicated by the use of "stash" to describe this behavior, the intended use
of this feature is temporary storage of local kptrs. For example, a sched_ext
([0]) scheduler may want to create an rbtree node for each new cgroup on cgroup
init, but to add that node to the rbtree as part of a separate program which
runs on enqueue. Stashing the node in a map_value allows its lifetime to outlive
the execution of the cgroup_init program.

Behavior:

There is no semantic difference between adding a kptr to a graph collection and
"stashing" it in a map. In both cases exclusive ownership of the kptr's lifetime
is passed to some containing data structure, which is responsible for
bpf_obj_drop'ing it when the container goes away.

Since graph collections also expect exclusive ownership of the nodes they
contain, graph nodes cannot be both stashed in a map_value and contained by
their corresponding collection.

Implementation:

Two observations simplify the verifier changes for this feature. First, kptrs
("referenced kptrs" until a recent renaming) require registration of a
dtor function as part of their acquire/release semantics, so that a referenced
kptr which is placed in a map_value is properly released when the map goes away.
We want this exact behavior for local kptrs, but with bpf_obj_drop as the dtor
instead of a per-btf_id dtor.

The second observation is that, in terms of identification, "referenced kptr"
and "local kptr" already don't interfere with one another. Consider the
following example:

  struct node_data {
          long key;
          long data;
          struct bpf_rb_node node;
  };

  struct map_value {
          struct node_data __kptr *node;
  };

  struct {
          __uint(type, BPF_MAP_TYPE_ARRAY);
          __type(key, int);
          __type(value, struct map_value);
          __uint(max_entries, 1);
  } some_nodes SEC(".maps");

  struct map_value *mapval;
  struct node_data *res;
  int key = 0;

  res = bpf_obj_new(typeof(*res));
  if (!res) { /* err handling */ }

  mapval = bpf_map_lookup_elem(&some_nodes, &key);
  if (!mapval) { /* err handling */ }

  res = bpf_kptr_xchg(&mapval->node, res);
  if (res)
          bpf_obj_drop(res);

The __kptr tag identifies map_value's node as a referenced kptr, while the
PTR_TO_BTF_ID which bpf_obj_new returns - a type in some non-vmlinux,
non-module BTF - identifies res as a local kptr. Type tag on the pointer
indicates referenced kptr, while the type of the pointee indicates local kptr.
So using existing facilities we can tell the verifier about a "referenced kptr"
pointer to a "local kptr" pointee.

When kptr_xchg'ing a kptr into a map_value, the verifier can recognize local
kptr types and treat them like referenced kptrs with a properly-typed
bpf_obj_drop as a dtor.

Other implementation notes:
  * We don't need to do anything special to enforce "graph nodes cannot be
    both stashed in a map_value and contained by their corresponding collection"
    * bpf_kptr_xchg both returns and takes as input a (possibly-null) owning
      reference. It does not accept non-owning references as input by virtue
      of requiring a ref_obj_id. By definition, if a program has an owning
      ref to a node, the node isn't in a collection, so it's safe to pass
      ownership via bpf_kptr_xchg.

Summary of patches:

  * Patch 1 modifies BTF plumbing to support using bpf_obj_drop as a dtor
  * Patch 2 adds verifier plumbing to support MEM_ALLOC-flagged param for
    bpf_kptr_xchg
  * Patch 3 adds selftests exercising the new behavior

Changelog:

v1 -> v2: https://lore.kernel.org/bpf/20230309180111.1618459-1-davemarchevsky@fb.com/

Patch #s used below refer to the patch's position in v1 unless otherwise
specified.

Patches 1-3 were applied and are not included in v2.
Rebase onto latest bpf-next: "libbpf: Revert poisoning of strlcpy"

Patch 4: "bpf: Support __kptr to local kptrs"
  * Remove !btf_is_kernel(btf) check, WARN_ON_ONCE instead (Alexei)

Patch 6: "selftests/bpf: Add local kptr stashing test"
  * Add test which stashes 2 nodes and later unstashes one of them using a
    separate BPF program (Alexei)
  * Fix incorrect runner subtest name for original test (was
    "rbtree_add_nodes")
====================
Signed-off-by: default avatarAlexei Starovoitov <ast@kernel.org>
      49b5300f
    • Dave Marchevsky's avatar
      selftests/bpf: Add local kptr stashing test · 5d8d6634
      Dave Marchevsky authored 
      Add a new selftest, local_kptr_stash, which uses bpf_kptr_xchg to stash
a bpf_obj_new-allocated object in a map. Test the following scenarios:

  * Stash two rb_nodes in an arraymap, don't unstash them, rely on map
    free to destruct them
  * Stash two rb_nodes in an arraymap, unstash the second one in a
    separate program, rely on map free to destruct first
Signed-off-by: default avatarDave Marchevsky <davemarchevsky@fb.com>
Link: https://lore.kernel.org/r/20230310230743.2320707-4-davemarchevsky@fb.comSigned-off-by: default avatarAlexei Starovoitov <ast@kernel.org>
      5d8d6634
    • Dave Marchevsky's avatar
      bpf: Allow local kptrs to be exchanged via bpf_kptr_xchg · 738c96d5
      Dave Marchevsky authored 
      The previous patch added necessary plumbing for verifier and runtime to
know what to do with non-kernel PTR_TO_BTF_IDs in map values, but didn't
provide any way to get such local kptrs into a map value. This patch
modifies verifier handling of bpf_kptr_xchg to allow MEM_ALLOC kptr
types.

check_reg_type is modified accept MEM_ALLOC-flagged input to
bpf_kptr_xchg despite such types not being in btf_ptr_types. This could
have been done with a MAYBE_MEM_ALLOC equivalent to MAYBE_NULL, but
bpf_kptr_xchg is the only helper that I can forsee using
MAYBE_MEM_ALLOC, so keep it special-cased for now.

The verifier tags bpf_kptr_xchg retval MEM_ALLOC if and only if the BTF
associated with the retval is not kernel BTF.
Signed-off-by: default avatarDave Marchevsky <davemarchevsky@fb.com>
Link: https://lore.kernel.org/r/20230310230743.2320707-3-davemarchevsky@fb.comSigned-off-by: default avatarAlexei Starovoitov <ast@kernel.org>
      738c96d5
    • Dave Marchevsky's avatar
      bpf: Support __kptr to local kptrs · c8e18754
      Dave Marchevsky authored 
      If a PTR_TO_BTF_ID type comes from program BTF - not vmlinux or module
BTF - it must have been allocated by bpf_obj_new and therefore must be
free'd with bpf_obj_drop. Such a PTR_TO_BTF_ID is considered a "local
kptr" and is tagged with MEM_ALLOC type tag by bpf_obj_new.

This patch adds support for treating __kptr-tagged pointers to "local
kptrs" as having an implicit bpf_obj_drop destructor for referenced kptr
acquire / release semantics. Consider the following example:

  struct node_data {
          long key;
          long data;
          struct bpf_rb_node node;
  };

  struct map_value {
          struct node_data __kptr *node;
  };

  struct {
          __uint(type, BPF_MAP_TYPE_ARRAY);
          __type(key, int);
          __type(value, struct map_value);
          __uint(max_entries, 1);
  } some_nodes SEC(".maps");

If struct node_data had a matching definition in kernel BTF, the verifier would
expect a destructor for the type to be registered. Since struct node_data does
not match any type in kernel BTF, the verifier knows that there is no kfunc
that provides a PTR_TO_BTF_ID to this type, and that such a PTR_TO_BTF_ID can
only come from bpf_obj_new. So instead of searching for a registered dtor,
a bpf_obj_drop dtor can be assumed.

This allows the runtime to properly destruct such kptrs in
bpf_obj_free_fields, which enables maps to clean up map_vals w/ such
kptrs when going away.

Implementation notes:
  * "kernel_btf" variable is renamed to "kptr_btf" in btf_parse_kptr.
    Before this patch, the variable would only ever point to vmlinux or
    module BTFs, but now it can point to some program BTF for local kptr
    type. It's later used to populate the (btf, btf_id) pair in kptr btf
    field.
  * It's necessary to btf_get the program BTF when populating btf_field
    for local kptr. btf_record_free later does a btf_put.
  * Behavior for non-local referenced kptrs is not modified, as
    bpf_find_btf_id helper only searches vmlinux and module BTFs for
    matching BTF type. If such a type is found, btf_field_kptr's btf will
    pass btf_is_kernel check, and the associated release function is
    some one-argument dtor. If btf_is_kernel check fails, associated
    release function is two-arg bpf_obj_drop_impl. Before this patch
    only btf_field_kptr's w/ kernel or module BTFs were created.
Signed-off-by: default avatarDave Marchevsky <davemarchevsky@fb.com>
Link: https://lore.kernel.org/r/20230310230743.2320707-2-davemarchevsky@fb.comSigned-off-by: default avatarAlexei Starovoitov <ast@kernel.org>
      c8e18754
  5. 10 Mar, 2023 27 commits
  7. 09 Mar, 2023 8 commits
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7