Martin,
Thank you for the detailed feedback.
> Can a separate global lock/mutex (not the lock_sock) be acquired first before
> using the sock pointer in the kfuncs?
Sure. I will add the mutex around the socket operations. As for the
single global sock pointer, I wanted to keep it simple in this patch
series to fulfill the current use case. I agree it might be overkill
for now to add the bpf map and such.
> Is it better to set sk_sndtimeo in bpf_kfunc_init_sock() ?
> All these new kfunc should have the KF_SLEEPABLE flag.
> bpf_testmod_exit() should probably do this NULL check and sock_release() also.
Ack. I will add this.
> nit. Can "struct sockaddr_storage addr;" be directly used instead of a char array?
When using "struct sockaddr_storage addr;" directly, the BPF program
fails to load with the following error message.
> libbpf: prog 'kernel_connect': BPF program load failed: Invalid argument
> libbpf: prog 'kernel_connect': -- BEGIN PROG LOAD LOG --
> 0: R1=ctx() R10=fp0
> ; return bpf_kfunc_call_kernel_connect(args); @ sock_addr_kern.c:26
> 0: (85) call bpf_kfunc_call_kernel_connect#99994
> arg#0 pointer type STRUCT addr_args must point to scalar, or struct with scalar
> processed 1 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
> -- END PROG LOAD LOG --
> libbpf: prog 'kernel_connect': failed to load: -22
> libbpf: failed to load object 'sock_addr_kern'
> libbpf: failed to load BPF skeleton 'sock_addr_kern': -22
> load_sock_addr_kern:FAIL:skel unexpected error: -22
> test_sock_addr:FAIL:load_sock_addr_kern unexpected error: -1 (errno 22)
> #288 sock_addr:FAIL
-Jordan
On Tue, Apr 16, 2024 at 2:43 AM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
>
> On 4/12/24 9:52 AM, Jordan Rife wrote:
> > This patch adds a set of kfuncs to bpf_testmod that can be used to
> > manipulate a socket from kernel space.
> >
> > Signed-off-by: Jordan Rife <jrife@xxxxxxxxxx>
> > ---
> > .../selftests/bpf/bpf_testmod/bpf_testmod.c | 139 ++++++++++++++++++
> > .../bpf/bpf_testmod/bpf_testmod_kfunc.h | 27 ++++
> > 2 files changed, 166 insertions(+)
> >
> > diff --git a/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod.c b/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod.c
> > index 39ad96a18123f..663df8148097e 100644
> > --- a/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod.c
> > +++ b/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod.c
> > @@ -10,18 +10,29 @@
> > #include <linux/percpu-defs.h>
> > #include <linux/sysfs.h>
> > #include <linux/tracepoint.h>
> > +#include <linux/net.h>
> > +#include <linux/socket.h>
> > +#include <linux/nsproxy.h>
> > +#include <linux/inet.h>
> > +#include <linux/in.h>
> > +#include <linux/in6.h>
> > +#include <linux/un.h>
> > +#include <net/sock.h>
> > #include "bpf_testmod.h"
> > #include "bpf_testmod_kfunc.h"
> >
> > #define CREATE_TRACE_POINTS
> > #include "bpf_testmod-events.h"
> >
> > +#define CONNECT_TIMEOUT_SEC 1
> > +
> > typedef int (*func_proto_typedef)(long);
> > typedef int (*func_proto_typedef_nested1)(func_proto_typedef);
> > typedef int (*func_proto_typedef_nested2)(func_proto_typedef_nested1);
> >
> > DEFINE_PER_CPU(int, bpf_testmod_ksym_percpu) = 123;
> > long bpf_testmod_test_struct_arg_result;
> > +static struct socket *sock;
> >
> > struct bpf_testmod_struct_arg_1 {
> > int a;
> > @@ -494,6 +505,124 @@ __bpf_kfunc static u32 bpf_kfunc_call_test_static_unused_arg(u32 arg, u32 unused
> > return arg;
> > }
> >
> > +__bpf_kfunc int bpf_kfunc_init_sock(struct init_sock_args *args)
> > +{
> > + int proto;
> > +
> > + if (sock)
> > + pr_warn("%s called without releasing old sock", __func__);
>
> hmm...this global sock pointer is quite unease. e.g. what if multiple tasks
> trying to use init/close/connect... in parallel.
>
> Storing sock in a bpf map will be better but that may be overkill for testing.
> Can a separate global lock/mutex (not the lock_sock) be acquired first before
> using the sock pointer in the kfuncs?
>
> > +
> > + switch (args->af) {
> > + case AF_INET:
> > + case AF_INET6:
> > + proto = args->type == SOCK_STREAM ? IPPROTO_TCP : IPPROTO_UDP;
> > + break;
> > + case AF_UNIX:
> > + proto = PF_UNIX;
> > + break;
> > + default:
> > + pr_err("invalid address family %d\n", args->af);
> > + return -EINVAL;
> > + }
> > +
> > + return sock_create_kern(&init_net, args->af, args->type, proto, &sock);
> > +}
> > +
> > +__bpf_kfunc void bpf_kfunc_close_sock(void)
> > +{
> > + if (sock) {
> > + sock_release(sock);
>
> bpf_testmod_exit() should probably do this NULL check and sock_release() also.
>
> > + sock = NULL;
> > + }
> > +}
> > +
> > +__bpf_kfunc int bpf_kfunc_call_kernel_connect(struct addr_args *args)
> > +{
> > + /* Set timeout for call to kernel_connect() to prevent it from hanging,
> > + * and consider the connection attempt failed if it returns
> > + * -EINPROGRESS.
> > + */
> > + sock->sk->sk_sndtimeo = CONNECT_TIMEOUT_SEC * HZ;
>
> Is it better to set sk_sndtimeo in bpf_kfunc_init_sock() ?
>
> > +
> > + return kernel_connect(sock, (struct sockaddr *)&args->addr,
> > + args->addrlen, 0);
> > +}
> > +
> > +__bpf_kfunc int bpf_kfunc_call_kernel_bind(struct addr_args *args)
> > +{
> > + return kernel_bind(sock, (struct sockaddr *)&args->addr, args->addrlen);
> > +}
> > +
> > +__bpf_kfunc int bpf_kfunc_call_kernel_listen(void)
> > +{
> > + return kernel_listen(sock, 128);
> > +}
> > +
> > +__bpf_kfunc int bpf_kfunc_call_kernel_sendmsg(struct sendmsg_args *args)
> > +{
> > + struct msghdr msg = {
> > + .msg_name = &args->addr.addr,
> > + .msg_namelen = args->addr.addrlen,
> > + };
> > + struct kvec iov;
> > + int err;
> > +
> > + iov.iov_base = args->msg;
> > + iov.iov_len = args->msglen;
> > +
> > + err = kernel_sendmsg(sock, &msg, &iov, 1, args->msglen);
> > + args->addr.addrlen = msg.msg_namelen;
> > +
> > + return err;
> > +}
> > +
> > +__bpf_kfunc int bpf_kfunc_call_sock_sendmsg(struct sendmsg_args *args)
> > +{
> > + struct msghdr msg = {
> > + .msg_name = &args->addr.addr,
> > + .msg_namelen = args->addr.addrlen,
> > + };
> > + struct kvec iov;
> > + int err;
> > +
> > + iov.iov_base = args->msg;
> > + iov.iov_len = args->msglen;
> > +
> > + iov_iter_kvec(&msg.msg_iter, ITER_SOURCE, &iov, 1, args->msglen);
> > + err = sock_sendmsg(sock, &msg);
> > + args->addr.addrlen = msg.msg_namelen;
> > +
> > + return err;
> > +}
> > +
> > +__bpf_kfunc int bpf_kfunc_call_kernel_getsockname(struct addr_args *args)
> > +{
> > + int err;
> > +
> > + err = kernel_getsockname(sock, (struct sockaddr *)&args->addr);
> > + if (err < 0)
> > + goto out;
> > +
> > + args->addrlen = err;
> > + err = 0;
> > +out:
> > + return err;
> > +}
> > +
> > +__bpf_kfunc int bpf_kfunc_call_kernel_getpeername(struct addr_args *args)
> > +{
> > + int err;
> > +
> > + err = kernel_getpeername(sock, (struct sockaddr *)&args->addr);
> > + if (err < 0)
> > + goto out;
> > +
> > + args->addrlen = err;
> > + err = 0;
> > +out:
> > + return err;
> > +}
> > +
> > BTF_KFUNCS_START(bpf_testmod_check_kfunc_ids)
> > BTF_ID_FLAGS(func, bpf_testmod_test_mod_kfunc)
> > BTF_ID_FLAGS(func, bpf_kfunc_call_test1)
> > @@ -520,6 +649,15 @@ BTF_ID_FLAGS(func, bpf_kfunc_call_test_ref, KF_TRUSTED_ARGS | KF_RCU)
> > BTF_ID_FLAGS(func, bpf_kfunc_call_test_destructive, KF_DESTRUCTIVE)
> > BTF_ID_FLAGS(func, bpf_kfunc_call_test_static_unused_arg)
> > BTF_ID_FLAGS(func, bpf_kfunc_call_test_offset)
> > +BTF_ID_FLAGS(func, bpf_kfunc_init_sock)
> > +BTF_ID_FLAGS(func, bpf_kfunc_close_sock)
> > +BTF_ID_FLAGS(func, bpf_kfunc_call_kernel_connect)
> > +BTF_ID_FLAGS(func, bpf_kfunc_call_kernel_bind)
> > +BTF_ID_FLAGS(func, bpf_kfunc_call_kernel_listen)
> > +BTF_ID_FLAGS(func, bpf_kfunc_call_kernel_sendmsg)
> > +BTF_ID_FLAGS(func, bpf_kfunc_call_sock_sendmsg)
> > +BTF_ID_FLAGS(func, bpf_kfunc_call_kernel_getsockname)
> > +BTF_ID_FLAGS(func, bpf_kfunc_call_kernel_getpeername)
>
> All these new kfunc should have the KF_SLEEPABLE flag.
>
> > BTF_KFUNCS_END(bpf_testmod_check_kfunc_ids)
> >
> > static int bpf_testmod_ops_init(struct btf *btf)
> > @@ -650,6 +788,7 @@ static int bpf_testmod_init(void)
> > return ret;
> > if (bpf_fentry_test1(0) < 0)
> > return -EINVAL;
> > + sock = NULL;
> > return sysfs_create_bin_file(kernel_kobj, &bin_attr_bpf_testmod_file);
> > }
> >
> > diff --git a/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod_kfunc.h b/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod_kfunc.h
> > index 7c664dd610597..cdf7769a7d8ca 100644
> > --- a/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod_kfunc.h
> > +++ b/tools/testing/selftests/bpf/bpf_testmod/bpf_testmod_kfunc.h
> > @@ -64,6 +64,22 @@ struct prog_test_fail3 {
> > char arr2[];
> > };
> >
> > +struct init_sock_args {
> > + int af;
> > + int type;
> > +};
> > +
> > +struct addr_args {
> > + char addr[sizeof(struct __kernel_sockaddr_storage)];
>
> nit. Can "struct sockaddr_storage addr;" be directly used instead of a char array?
>
> > + int addrlen;
> > +};
> > +
> > +struct sendmsg_args {
> > + struct addr_args addr;
> > + char msg[10];
> > + int msglen;
> > +};
> > +
> > struct prog_test_ref_kfunc *
> > bpf_kfunc_call_test_acquire(unsigned long *scalar_ptr) __ksym;
> > void bpf_kfunc_call_test_release(struct prog_test_ref_kfunc *p) __ksym;
> > @@ -106,4 +122,15 @@ void bpf_kfunc_call_test_fail3(struct prog_test_fail3 *p);
> > void bpf_kfunc_call_test_mem_len_fail1(void *mem, int len);
> >
> > void bpf_kfunc_common_test(void) __ksym;
> > +
> > +int bpf_kfunc_init_sock(struct init_sock_args *args) __ksym;
> > +void bpf_kfunc_close_sock(void) __ksym;
> > +int bpf_kfunc_call_kernel_connect(struct addr_args *args) __ksym;
> > +int bpf_kfunc_call_kernel_bind(struct addr_args *args) __ksym;
> > +int bpf_kfunc_call_kernel_listen(void) __ksym;
> > +int bpf_kfunc_call_kernel_sendmsg(struct sendmsg_args *args) __ksym;
> > +int bpf_kfunc_call_sock_sendmsg(struct sendmsg_args *args) __ksym;
> > +int bpf_kfunc_call_kernel_getsockname(struct addr_args *args) __ksym;
> > +int bpf_kfunc_call_kernel_getpeername(struct addr_args *args) __ksym;
> > +
> > #endif /* _BPF_TESTMOD_KFUNC_H */
>
