Re: [PATCH HID 03/13] HID: bpf: implement HID-BPF through bpf_struct_ops
From: Alexei Starovoitov
Date: Wed May 29 2024 - 00:03:01 EST
On Tue, May 28, 2024 at 6:15 AM Benjamin Tissoires <bentiss@xxxxxxxxxx> wrote:
>
> We do this implementation in several steps to not have the CI failing:
> - first (this patch), we add struct_ops while keeping the existing infra
> available
> - then we change the selftests, the examples and the existing in-tree
> HID-BPF programs
> - then we remove the existing trace points making old HID-BPF obsolete
>
> There are a few advantages of struct_ops over tracing:
> - compatibility with sleepable programs (for hid_hw_raw_request() in
> a later patch)
> - a lot simpler in the kernel: it's a simple rcu protected list
> - we can add more parameters to the function called without much trouble
> - the "attach" is now generic through BPF-core: the caller just needs to
> set hid_id and flags before calling __load().
> - all the BPF tough part is not handled in BPF-core through generic
> processing
> - hid_bpf_ctx is now only writable where it needs be
>
> Signed-off-by: Benjamin Tissoires <bentiss@xxxxxxxxxx>
> ---
> drivers/hid/bpf/Makefile | 2 +-
> drivers/hid/bpf/hid_bpf_dispatch.c | 52 +++++++-
> drivers/hid/bpf/hid_bpf_dispatch.h | 4 +
> drivers/hid/bpf/hid_bpf_jmp_table.c | 3 +
> drivers/hid/bpf/hid_bpf_struct_ops.c | 246 +++++++++++++++++++++++++++++++++++
> include/linux/hid_bpf.h | 64 ++++++++-
> 6 files changed, 362 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/hid/bpf/Makefile b/drivers/hid/bpf/Makefile
> index cf55120cf7d6..1cb3f31e9335 100644
> --- a/drivers/hid/bpf/Makefile
> +++ b/drivers/hid/bpf/Makefile
> @@ -8,4 +8,4 @@ LIBBPF_INCLUDE = $(srctree)/tools/lib
> obj-$(CONFIG_HID_BPF) += hid_bpf.o
> CFLAGS_hid_bpf_dispatch.o += -I$(LIBBPF_INCLUDE)
> CFLAGS_hid_bpf_jmp_table.o += -I$(LIBBPF_INCLUDE)
> -hid_bpf-objs += hid_bpf_dispatch.o hid_bpf_jmp_table.o
> +hid_bpf-objs += hid_bpf_dispatch.o hid_bpf_jmp_table.o hid_bpf_struct_ops.o
> diff --git a/drivers/hid/bpf/hid_bpf_dispatch.c b/drivers/hid/bpf/hid_bpf_dispatch.c
> index c8bb79ce2354..7216c3c7713d 100644
> --- a/drivers/hid/bpf/hid_bpf_dispatch.c
> +++ b/drivers/hid/bpf/hid_bpf_dispatch.c
> @@ -58,6 +58,7 @@ dispatch_hid_bpf_device_event(struct hid_device *hdev, enum hid_report_type type
> },
> .data = hdev->bpf.device_data,
> };
> + struct hid_bpf_ops *e;
> int ret;
>
> if (type >= HID_REPORT_TYPES)
> @@ -70,9 +71,25 @@ dispatch_hid_bpf_device_event(struct hid_device *hdev, enum hid_report_type type
> memset(ctx_kern.data, 0, hdev->bpf.allocated_data);
> memcpy(ctx_kern.data, data, *size);
>
> + rcu_read_lock();
> + list_for_each_entry_rcu(e, &hdev->bpf.prog_list, list) {
In the typical case there will be only one prog per device, right?
The for_each is future proofing?
> + if (e->hid_device_event) {
> + ret = e->hid_device_event(&ctx_kern.ctx, type);
> + if (ret < 0) {
> + rcu_read_unlock();
> + return ERR_PTR(ret);
> + }
> +
> + if (ret)
> + ctx_kern.ctx.retval = ret;
> + }
> + }
> + rcu_read_unlock();
> +
> ret = hid_bpf_prog_run(hdev, HID_BPF_PROG_TYPE_DEVICE_EVENT, &ctx_kern);
> if (ret < 0)
> return ERR_PTR(ret);
> + ret = ctx_kern.ctx.retval;
>
> if (ret) {
> if (ret > ctx_kern.ctx.allocated_size)
> @@ -122,7 +139,10 @@ u8 *call_hid_bpf_rdesc_fixup(struct hid_device *hdev, u8 *rdesc, unsigned int *s
>
> memcpy(ctx_kern.data, rdesc, min_t(unsigned int, *size, HID_MAX_DESCRIPTOR_SIZE));
>
> - ret = hid_bpf_prog_run(hdev, HID_BPF_PROG_TYPE_RDESC_FIXUP, &ctx_kern);
> + if (hdev->bpf.rdesc_ops)
> + ret = hdev->bpf.rdesc_ops->hid_rdesc_fixup(&ctx_kern.ctx);
> + else
> + ret = hid_bpf_prog_run(hdev, HID_BPF_PROG_TYPE_RDESC_FIXUP, &ctx_kern);
This is for backward compat?
I don't see it's being removed in the later patches.
> if (ret < 0)
> goto ignore_bpf;
>
> @@ -150,7 +170,7 @@ static int device_match_id(struct device *dev, const void *id)
> return hdev->id == *(int *)id;
> }
>
> -static struct hid_device *hid_get_device(unsigned int hid_id)
> +struct hid_device *hid_get_device(unsigned int hid_id)
> {
> struct device *dev;
>
> @@ -164,7 +184,7 @@ static struct hid_device *hid_get_device(unsigned int hid_id)
> return to_hid_device(dev);
> }
>
> -static void hid_put_device(struct hid_device *hid)
> +void hid_put_device(struct hid_device *hid)
> {
> put_device(&hid->dev);
> }
> @@ -205,7 +225,7 @@ static int __hid_bpf_allocate_data(struct hid_device *hdev, u8 **data, u32 *size
> return 0;
> }
>
> -static int hid_bpf_allocate_event_data(struct hid_device *hdev)
> +int hid_bpf_allocate_event_data(struct hid_device *hdev)
> {
> /* hdev->bpf.device_data is already allocated, abort */
> if (hdev->bpf.device_data)
> @@ -592,14 +612,22 @@ static const struct btf_kfunc_id_set hid_bpf_syscall_kfunc_set = {
>
> int hid_bpf_connect_device(struct hid_device *hdev)
> {
> - struct hid_bpf_prog_list *prog_list;
> + bool need_to_allocate = false;
> + struct hid_bpf_ops *e;
>
> rcu_read_lock();
> - prog_list = rcu_dereference(hdev->bpf.progs[HID_BPF_PROG_TYPE_DEVICE_EVENT]);
> + list_for_each_entry_rcu(e, &hdev->bpf.prog_list, list) {
> + if (e->hid_device_event) {
> + need_to_allocate = true;
> + break;
> + }
> + }
> + if (rcu_dereference(hdev->bpf.progs[HID_BPF_PROG_TYPE_DEVICE_EVENT]))
> + need_to_allocate = true;
> rcu_read_unlock();
>
> /* only allocate BPF data if there are programs attached */
> - if (!prog_list)
> + if (!need_to_allocate)
> return 0;
>
> return hid_bpf_allocate_event_data(hdev);
> @@ -623,12 +651,15 @@ void hid_bpf_destroy_device(struct hid_device *hdev)
> hdev->bpf.destroyed = true;
>
> __hid_bpf_destroy_device(hdev);
> + __hid_bpf_ops_destroy_device(hdev);
> }
> EXPORT_SYMBOL_GPL(hid_bpf_destroy_device);
>
> void hid_bpf_device_init(struct hid_device *hdev)
> {
> spin_lock_init(&hdev->bpf.progs_lock);
> + INIT_LIST_HEAD(&hdev->bpf.prog_list);
> + mutex_init(&hdev->bpf.prog_list_lock);
> }
> EXPORT_SYMBOL_GPL(hid_bpf_device_init);
>
> @@ -662,6 +693,13 @@ static int __init hid_bpf_init(void)
> return 0;
> }
>
> + /* register struct_ops kfuncs after we are sure we can load our preloaded bpf program */
> + err = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &hid_bpf_kfunc_set);
> + if (err) {
> + pr_warn("error while setting HID BPF tracing kfuncs: %d", err);
> + return 0;
> + }
> +
> /* register syscalls after we are sure we can load our preloaded bpf program */
> err = register_btf_kfunc_id_set(BPF_PROG_TYPE_SYSCALL, &hid_bpf_syscall_kfunc_set);
> if (err) {
> diff --git a/drivers/hid/bpf/hid_bpf_dispatch.h b/drivers/hid/bpf/hid_bpf_dispatch.h
> index fbe0639d09f2..e52c43d81650 100644
> --- a/drivers/hid/bpf/hid_bpf_dispatch.h
> +++ b/drivers/hid/bpf/hid_bpf_dispatch.h
> @@ -10,12 +10,16 @@ struct hid_bpf_ctx_kern {
> u8 *data;
> };
>
> +struct hid_device *hid_get_device(unsigned int hid_id);
> +void hid_put_device(struct hid_device *hid);
> +int hid_bpf_allocate_event_data(struct hid_device *hdev);
> int hid_bpf_preload_skel(void);
> void hid_bpf_free_links_and_skel(void);
> int hid_bpf_get_prog_attach_type(struct bpf_prog *prog);
> int __hid_bpf_attach_prog(struct hid_device *hdev, enum hid_bpf_prog_type prog_type, int prog_fd,
> struct bpf_prog *prog, __u32 flags);
> void __hid_bpf_destroy_device(struct hid_device *hdev);
> +void __hid_bpf_ops_destroy_device(struct hid_device *hdev);
> int hid_bpf_prog_run(struct hid_device *hdev, enum hid_bpf_prog_type type,
> struct hid_bpf_ctx_kern *ctx_kern);
> int hid_bpf_reconnect(struct hid_device *hdev);
> diff --git a/drivers/hid/bpf/hid_bpf_jmp_table.c b/drivers/hid/bpf/hid_bpf_jmp_table.c
> index aa8e1c79cdf5..8a54ba447718 100644
> --- a/drivers/hid/bpf/hid_bpf_jmp_table.c
> +++ b/drivers/hid/bpf/hid_bpf_jmp_table.c
> @@ -81,6 +81,9 @@ static int hid_bpf_program_count(struct hid_device *hdev,
> if (type >= HID_BPF_PROG_TYPE_MAX)
> return -EINVAL;
>
> + if (type == HID_BPF_PROG_TYPE_RDESC_FIXUP && hdev->bpf.rdesc_ops)
> + n += 1;
> +
> FOR_ENTRIES(i, jmp_table.tail, jmp_table.head) {
> struct hid_bpf_prog_entry *entry = &jmp_table.entries[i];
>
> diff --git a/drivers/hid/bpf/hid_bpf_struct_ops.c b/drivers/hid/bpf/hid_bpf_struct_ops.c
> new file mode 100644
> index 000000000000..be514a98e55b
> --- /dev/null
> +++ b/drivers/hid/bpf/hid_bpf_struct_ops.c
> @@ -0,0 +1,246 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +
> +/*
> + * HID-BPF support for Linux
> + *
> + * Copyright (c) 2024 Benjamin Tissoires
> + */
> +
> +#include <linux/bitops.h>
> +#include <linux/bpf_verifier.h>
> +#include <linux/bpf.h>
> +#include <linux/btf.h>
> +#include <linux/btf_ids.h>
> +#include <linux/filter.h>
> +#include <linux/hid.h>
> +#include <linux/hid_bpf.h>
> +#include <linux/init.h>
> +#include <linux/module.h>
> +#include <linux/workqueue.h>
> +#include "hid_bpf_dispatch.h"
> +
> +static struct btf *hid_bpf_ops_btf;
> +
> +static int hid_bpf_ops_init(struct btf *btf)
> +{
> + hid_bpf_ops_btf = btf;
> + return 0;
> +}
> +
> +static bool hid_bpf_ops_is_valid_access(int off, int size,
> + enum bpf_access_type type,
> + const struct bpf_prog *prog,
> + struct bpf_insn_access_aux *info)
> +{
> + return bpf_tracing_btf_ctx_access(off, size, type, prog, info);
> +}
> +
> +static int hid_bpf_ops_check_member(const struct btf_type *t,
> + const struct btf_member *member,
> + const struct bpf_prog *prog)
> +{
> + u32 moff = __btf_member_bit_offset(t, member) / 8;
> +
this is good...
> + switch (moff) {
> + case offsetof(struct hid_bpf_ops, hid_rdesc_fixup):
> + break;
> + default:
> + if (prog->sleepable)
> + return -EINVAL;
> + }
> +
> + return 0;
> +}
> +
> +static int hid_bpf_ops_btf_struct_access(struct bpf_verifier_log *log,
> + const struct bpf_reg_state *reg,
> + int off, int size)
> +{
> + const struct btf_type *state;
> + const struct btf_type *t;
> + s32 type_id;
> +
> + type_id = btf_find_by_name_kind(reg->btf, "hid_bpf_ctx",
> + BTF_KIND_STRUCT);
> + if (type_id < 0)
> + return -EINVAL;
> +
> + t = btf_type_by_id(reg->btf, reg->btf_id);
> + state = btf_type_by_id(reg->btf, type_id);
> + if (t != state) {
> + bpf_log(log, "only access to hid_bpf_ctx is supported\n");
> + return -EACCES;
> + }
> +
> + /* out-of-bound access in hid_bpf_ctx */
> + if (off + size > sizeof(struct hid_bpf_ctx)) {
> + bpf_log(log, "write access at off %d with size %d\n", off, size);
> + return -EACCES;
> + }
> +
> + if (off < offsetof(struct hid_bpf_ctx, retval)) {
> + bpf_log(log,
> + "write access at off %d with size %d on read-only part of hid_bpf_ctx\n",
> + off, size);
> + return -EACCES;
> + }
> +
> + return NOT_INIT;
> +}
> +
> +static const struct bpf_verifier_ops hid_bpf_verifier_ops = {
> + .is_valid_access = hid_bpf_ops_is_valid_access,
> + .btf_struct_access = hid_bpf_ops_btf_struct_access,
> +};
> +
> +static int hid_bpf_ops_init_member(const struct btf_type *t,
> + const struct btf_member *member,
> + void *kdata, const void *udata)
> +{
> + u32 flags;
> +
> + switch (member->offset) {
> + case offsetof(struct hid_bpf_ops, hid_id) * 8:
but here you're open coding it and adding ugly * 8
Just do:
moff = __btf_member_bit_offset(t, member) / 8;
switch (moff) {
case offsetof(struct tcp_congestion_ops, flags):
the way bpf_tcp_ca.c is doing?
Open code of ->offset is asking for trouble.
> + /* For hid_id and flags fields, this function has to copy it
> + * and return 1 to indicate that the data has been handled by
> + * the struct_ops type, or the verifier will reject the map if
> + * the value of those fields is not zero.
> + */
> + ((struct hid_bpf_ops *)kdata)->hid_id = ((struct hid_bpf_ops *)udata)->hid_id;
> + return 1;
> + case offsetof(struct hid_bpf_ops, flags) * 8:
> + flags = ((struct hid_bpf_ops *)udata)->flags;
> + if (flags & ~HID_BPF_FLAG_MASK)
> + return -EINVAL;
> + ((struct hid_bpf_ops *)kdata)->flags = flags;
> + return 1;
> + }
> + return 0;
> +}
> +
> +static int hid_bpf_reg(void *kdata)
> +{
> + struct hid_bpf_ops *ops = kdata;
> + struct hid_device *hdev;
> + int count, err = 0;
> +
> + hdev = hid_get_device(ops->hid_id);
> + if (IS_ERR(hdev))
> + return PTR_ERR(hdev);
> +
> + ops->hdev = hdev;
> +
> + mutex_lock(&hdev->bpf.prog_list_lock);
> +
> + count = list_count_nodes(&hdev->bpf.prog_list);
> + if (count >= HID_BPF_MAX_PROGS_PER_DEV) {
> + err = -E2BIG;
> + goto out_unlock;
> + }
> +
> + if (ops->hid_rdesc_fixup) {
> + if (hdev->bpf.rdesc_ops) {
> + err = -EINVAL;
> + goto out_unlock;
> + }
> +
> + hdev->bpf.rdesc_ops = ops;
> + }
> +
> + if (ops->hid_device_event) {
> + err = hid_bpf_allocate_event_data(hdev);
> + if (err)
> + goto out_unlock;
> + }
> +
> + if (ops->flags & HID_BPF_FLAG_INSERT_HEAD)
> + list_add_rcu(&ops->list, &hdev->bpf.prog_list);
> + else
> + list_add_tail_rcu(&ops->list, &hdev->bpf.prog_list);
Looks like future proofing, but I feel it's too little to materialize.
Take a look at include/linux/bpf_mprog.h
I suspect it might be useful here too.
> +
> +out_unlock:
> + mutex_unlock(&hdev->bpf.prog_list_lock);
> +
> + if (err) {
> + if (hdev->bpf.rdesc_ops == ops)
> + hdev->bpf.rdesc_ops = NULL;
> + hid_put_device(hdev);
> + } else if (ops->hid_rdesc_fixup) {
> + hid_bpf_reconnect(hdev);
> + }
> +
> + return err;
> +}
> +
> +static void hid_bpf_unreg(void *kdata)
> +{
> + struct hid_bpf_ops *ops = kdata;
> + struct hid_device *hdev;
> + bool reconnect = false;
> +
> + hdev = ops->hdev;
> +
> + /* check if __hid_bpf_ops_destroy_device() has been called */
> + if (!hdev)
> + return;
> +
> + mutex_lock(&hdev->bpf.prog_list_lock);
> +
> + list_del_rcu(&ops->list);
> +
> + reconnect = hdev->bpf.rdesc_ops == ops;
> + if (reconnect)
> + hdev->bpf.rdesc_ops = NULL;
> +
> + mutex_unlock(&hdev->bpf.prog_list_lock);
> +
> + if (reconnect)
> + hid_bpf_reconnect(hdev);
> +
> + hid_put_device(hdev);
> +}
> +
> +static int __hid_bpf_device_event(struct hid_bpf_ctx *ctx, enum hid_report_type type)
> +{
> + return 0;
> +}
> +
> +static int __hid_bpf_rdesc_fixup(struct hid_bpf_ctx *ctx)
> +{
> + return 0;
> +}
> +
> +static struct hid_bpf_ops __bpf_hid_bpf_ops = {
> + .hid_device_event = __hid_bpf_device_event,
> + .hid_rdesc_fixup = __hid_bpf_rdesc_fixup,
> +};
> +
> +static struct bpf_struct_ops bpf_hid_bpf_ops = {
> + .verifier_ops = &hid_bpf_verifier_ops,
> + .init = hid_bpf_ops_init,
> + .check_member = hid_bpf_ops_check_member,
> + .init_member = hid_bpf_ops_init_member,
> + .reg = hid_bpf_reg,
> + .unreg = hid_bpf_unreg,
> + .name = "hid_bpf_ops",
> + .cfi_stubs = &__bpf_hid_bpf_ops,
> + .owner = THIS_MODULE,
> +};
> +
> +void __hid_bpf_ops_destroy_device(struct hid_device *hdev)
> +{
> + struct hid_bpf_ops *e;
> +
> + rcu_read_lock();
> + list_for_each_entry_rcu(e, &hdev->bpf.prog_list, list) {
> + hid_put_device(hdev);
> + e->hdev = NULL;
> + }
> + rcu_read_unlock();
> +}
> +
> +static int __init hid_bpf_struct_ops_init(void)
> +{
> + return register_bpf_struct_ops(&bpf_hid_bpf_ops, hid_bpf_ops);
> +}
> +late_initcall(hid_bpf_struct_ops_init);
> diff --git a/include/linux/hid_bpf.h b/include/linux/hid_bpf.h
> index a66103618e6e..96495e977204 100644
> --- a/include/linux/hid_bpf.h
> +++ b/include/linux/hid_bpf.h
> @@ -65,11 +65,12 @@ struct hid_bpf_ctx {
> * @HID_BPF_FLAG_INSERT_HEAD: insert the given program before any other program
> * currently attached to the device. This doesn't
> * guarantee that this program will always be first
> - * @HID_BPF_FLAG_MAX: sentinel value, not to be used by the callers
> */
> enum hid_bpf_attach_flags {
> HID_BPF_FLAG_NONE = 0,
> HID_BPF_FLAG_INSERT_HEAD = _BITUL(0),
> +
> + /* private: internal use only */
> HID_BPF_FLAG_MAX,
> };
>
> @@ -112,6 +113,63 @@ struct hid_ops {
>
> extern struct hid_ops *hid_ops;
>
> +/**
> + * struct hid_bpf_ops - A BPF struct_ops of callbacks allowing to attach HID-BPF
> + * programs to a HID device
> + * @hid_id: the HID uniq ID to attach to. This is writeable before ``load()``, and
> + * cannot be changed after
> + * @flags: &enum hid_bpf_attach_flags to assign flags before ``load()``.
> + * Writeable only before ``load()``
> + */
> +struct hid_bpf_ops {
> + /* hid_id needs to stay first so we can easily change it
> + * from userspace.
> + */
hmm.
patch 5 just does:
+ skel->struct_ops.mouse_invert->hid_id = hid_id;
I don't see a reason why it needs to be first.
> + int hid_id;
> + u32 flags;
> +
> + /* private: internal use only */
> + struct list_head list;
> +
> + /* public: rest is public */
> +
> +/* fast path fields are put first to fill one cache line */
> +
> + /**
> + * @hid_device_event: called whenever an event is coming in from the device
> + *
> + * It has the following arguments:
> + *
> + * ``ctx``: The HID-BPF context as &struct hid_bpf_ctx
> + *
> + * Return: %0 on success and keep processing; a positive
> + * value to change the incoming size buffer; a negative
> + * error code to interrupt the processing of this event
> + *
> + * Context: Interrupt context.
> + */
> + int (*hid_device_event)(struct hid_bpf_ctx *ctx, enum hid_report_type report_type);
> +
> +/* control/slow paths put last */
> +
> + /**
> + * @hid_rdesc_fixup: called when the probe function parses the report descriptor
> + * of the HID device
> + *
> + * It has the following arguments:
> + *
> + * ``ctx``: The HID-BPF context as &struct hid_bpf_ctx
> + *
> + * Return: %0 on success and keep processing; a positive
> + * value to change the incoming size buffer; a negative
> + * error code to interrupt the processing of this device
> + */
> + int (*hid_rdesc_fixup)(struct hid_bpf_ctx *ctx);
It's fine since you want to keep the prog mostly unchanged,
but since the whole api is improving maybe it's time
to do something about hid_bpf_ctx in all callbacks?
Pass into prog what prog needs to see.
In the old api all progs had to have a common hid_bpf_ctx.
With struct-ops it can be fine tuned.
Overall it looks great.
hid_id did the trick :)