Re: [PATCH bpf-next 3/5] libbpf: add low level TC-BPF API

[Vlad Buslov <vladbu@nvidia.com>]

On Thu 25 Mar 2021 at 14:00, Kumar Kartikeya Dwivedi <memxor@gmail.com> wrote:
> This adds functions that wrap the netlink API used for adding,
> manipulating, and removing filters and actions. These functions operate
> directly on the loaded prog's fd, and return a handle to the filter and
> action using an out parameter (id for tc_cls, and index for tc_act).
>
> The basic featureset is covered to allow for attaching, manipulation of
> properties, and removal of filters and actions. Some additional features
> like TCA_BPF_POLICE and TCA_RATE for tc_cls have been omitted. These can
> added on top later by extending the bpf_tc_cls_opts struct.
>
> Support for binding actions directly to a classifier by passing them in
> during filter creation has also been omitted for now. These actions
> have an auto clean up property because their lifetime is bound to the
> filter they are attached to. This can be added later, but was omitted
> for now as direct action mode is a better alternative to it.
>
> An API summary:
>
> The BPF TC-CLS API
>
> bpf_tc_act_{attach, change, replace}_{dev, block} may be used to attach,
> change, and replace SCHED_CLS bpf classifiers. Separate set of functions
> are provided for network interfaces and shared filter blocks.
>
> bpf_tc_cls_detach_{dev, block} may be used to detach existing SCHED_CLS
> filter. The bpf_tc_cls_attach_id object filled in during attach,
> change, or replace must be passed in to the detach functions for them to
> remove the filter and its attached classififer correctly.
>
> bpf_tc_cls_get_info is a helper that can be used to obtain attributes
> for the filter and classififer. The opts structure may be used to
> choose the granularity of search, such that info for a specific filter
> corresponding to the same loaded bpf program can be obtained. By
> default, the first match is returned to the user.
>
> Examples:
>
> 	struct bpf_tc_cls_attach_id id = {};
> 	struct bpf_object *obj;
> 	struct bpf_program *p;
> 	int fd, r;
>
> 	obj = bpf_object_open("foo.o");
> 	if (IS_ERR_OR_NULL(obj))
> 		return PTR_ERR(obj);
>
> 	p = bpf_object__find_program_by_title(obj, "classifier");
> 	if (IS_ERR_OR_NULL(p))
> 		return PTR_ERR(p);
>
> 	if (bpf_object__load(obj) < 0)
> 		return -1;
>
> 	fd = bpf_program__fd(p);
>
> 	r = bpf_tc_cls_attach_dev(fd, if_nametoindex("lo"),
> 				  BPF_TC_CLSACT_INGRESS, ETH_P_IP,
> 				  NULL, &id);
> 	if (r < 0)
> 		return r;
>
> ... which is roughly equivalent to (after clsact qdisc setup):
>   # tc filter add dev lo ingress bpf obj /home/kkd/foo.o sec classifier
>
> If a user wishes to modify existing options on an attached filter, the
> bpf_tc_cls_change_{dev, block} API may be used. Parameters like
> chain_index, priority, and handle are ignored in the bpf_tc_cls_opts
> struct as they cannot be modified after attaching a filter.
>
> Example:
>
> 	/* Optional parameters necessary to select the right filter */
> 	DECLARE_LIBBPF_OPTS(bpf_tc_cls_opts, opts,
> 			    .handle = id.handle,
> 			    .priority = id.priority,
> 			    .chain_index = id.chain_index)
> 	/* Turn on direct action mode */
> 	opts.direct_action = true;
> 	r = bpf_tc_cls_change_dev(fd, id.ifindex, id.parent_id,
> 			          id.protocol, &opts, &id);
> 	if (r < 0)
> 		return r;
>
> 	/* Verify that the direct action mode has been set */
> 	struct bpf_tc_cls_info info = {};
> 	r = bpf_tc_cls_get_info_dev(fd, id.ifindex, id.parent_id,
> 			            id.protocol, &opts, &info);
> 	if (r < 0)
> 		return r;
>
> 	assert(info.bpf_flags & TCA_BPF_FLAG_ACT_DIRECT);
>
> This would be roughly equivalent to doing:
>   # tc filter change dev lo egress prio <p> handle <h> bpf obj /home/kkd/foo.o section classifier da
>
> ... except a new bpf program will be loaded and replace existing one.
>
> If a user wishes to either replace an existing filter, or create a new
> one with the same properties, they can use bpf_tc_cls_replace_dev. The
> benefit of bpf_tc_cls_change is that it fails if no matching filter
> exists.
>
> The BPF TC-ACT API
>
> bpf_tc_act_{attach, replace} may be used to attach and replace already
> attached SCHED_ACT actions. Passing an index of 0 has special meaning,
> in that an index will be automatically chosen by the kernel. The index
> chosen by the kernel is the return value of these functions in case of
> success.
>
> bpf_tc_act_detach may be used to detach a SCHED_ACT action prog
> identified by the index parameter. The index 0 again has a special
> meaning, in that passing it will flush all existing SCHED_ACT actions
> loaded using the ACT API.
>
> bpf_tc_act_get_info is a helper to get the required attributes of a
> loaded program to be able to manipulate it futher, by passing them
> into the aforementioned functions.
>
> Example:
>
> 	struct bpf_object *obj;
> 	struct bpf_program *p;
> 	__u32 index;
> 	int fd, r;
>
> 	obj = bpf_object_open("foo.o");
> 	if (IS_ERR_OR_NULL(obj))
> 		return PTR_ERR(obj);
>
> 	p = bpf_object__find_program_by_title(obj, "action");
> 	if (IS_ERR_OR_NULL(p))
> 		return PTR_ERR(p);
>
> 	if (bpf_object__load(obj) < 0)
> 		return -1;
>
> 	fd = bpf_program__fd(p);
>
> 	r = bpf_tc_act_attach(fd, NULL, &index);
> 	if (r < 0)
> 		return r;
>
> 	if (bpf_tc_act_detach(index))
> 		return -1;
>
> ... which is equivalent to the following sequence:
> 	tc action add action bpf obj /home/kkd/foo.o sec action
> 	tc action del action bpf index <idx>

How do you handle the locking here? Please note that while
RTM_{NEW|GET|DEL}FILTER API has been refactored to handle its own
locking internally (and registered with RTNL_FLAG_DOIT_UNLOCKED flag),
RTM_{NEW|GET|DEL}ACTION API still expects to be called with rtnl lock
taken.

[...]