bpf: rework how we keep track and attach cgroup bpf programs

So, the kernel's management of cgroup/BPF programs is a bit misdesigned:
if you attach a BPF program to a cgroup and close the fd for it it will
stay pinned to the cgroup with no chance of ever removing it again (or
otherwise getting ahold of it again), because the fd is used for
selecting which BPF program to detach. The only way to get rid of the
program again is to destroy the cgroup itself.

This is particularly bad for root the cgroup (and in fact any other
cgroup that we cannot realistically remove during runtime, such as
/system.slice, /init.scope or /system.slice/dbus.service) as getting rid
of the program only works by rebooting the system.

To counter this let's closely keep track to which cgroup a BPF program
is attached and let's implicitly detach the BPF program when we are
about to close the BPF fd.

This hence changes the bpf_program_cgroup_attach() function to track
where we attached the program and changes bpf_program_cgroup_detach() to
use this information. Moreover bpf_program_unref() will now implicitly
call bpf_program_cgroup_detach().

In order to simplify things, bpf_program_cgroup_attach() will now
implicitly invoke bpf_program_load_kernel() when necessary, simplifying
the caller's side.

Finally, this adds proper reference counting to BPF programs. This
is useful for working with two BPF programs in parallel: the BPF program
we are preparing for installation and the BPF program we so far
installed, shortening the window when we detach the old one and reattach
the new one.

src/basic/bpf-program.c

@@ -28,6 +28,7 @@
 #include "fd-util.h"
 #include "log.h"
 #include "missing.h"
+#include "path-util.h"
 #include "util.h"
 
 int bpf_program_new(uint32_t prog_type, BPFProgram **ret) {
@@ -37,6 +38,7 @@ int bpf_program_new(uint32_t prog_type, BPFProgram **ret) {
         if (!p)
                 return log_oom();
 
+        p->n_ref = 1;
         p->prog_type = prog_type;
         p->kernel_fd = -1;
 
@@ -45,12 +47,39 @@ int bpf_program_new(uint32_t prog_type, BPFProgram **ret) {
         return 0;
 }
 
+BPFProgram *bpf_program_ref(BPFProgram *p) {
+        if (!p)
+                return NULL;
+
+        assert(p->n_ref > 0);
+        p->n_ref++;
+
+        return p;
+}
+
 BPFProgram *bpf_program_unref(BPFProgram *p) {
         if (!p)
                 return NULL;
 
+        assert(p->n_ref > 0);
+        p->n_ref--;
+
+        if (p->n_ref > 0)
+                return NULL;
+
+        /* Unfortunately, the kernel currently doesn't implicitly detach BPF programs from their cgroups when the last
+         * fd to the BPF program is closed. This has nasty side-effects since this means that abnormally terminated
+         * programs that attached one of their BPF programs to a cgroup will leave this programs pinned for good with
+         * zero chance of recovery, until the cgroup is removed. This is particularly problematic if the cgroup in
+         * question is the root cgroup (or any other cgroup belonging to a service that cannot be restarted during
+         * operation, such as dbus), as the memory for the BPF program can only be reclaimed through a reboot. To
+         * counter this, we track closely to which cgroup a program was attached to and will detach it on our own
+         * whenever we close the BPF fd. */
+        (void) bpf_program_cgroup_detach(p);
+
         safe_close(p->kernel_fd);
         free(p->instructions);
+        free(p->attached_path);
 
         return mfree(p);
 }
@@ -99,13 +128,47 @@ int bpf_program_load_kernel(BPFProgram *p, char *log_buf, size_t log_size) {
 }
 
 int bpf_program_cgroup_attach(BPFProgram *p, int type, const char *path, uint32_t flags) {
+        _cleanup_free_ char *copy = NULL;
         _cleanup_close_ int fd = -1;
         union bpf_attr attr;
+        int r;
 
         assert(p);
         assert(type >= 0);
         assert(path);
 
+        if (!IN_SET(flags, 0, BPF_F_ALLOW_OVERRIDE, BPF_F_ALLOW_MULTI))
+                return -EINVAL;
+
+        /* We need to track which cgroup the program is attached to, and we can only track one attachment, hence let's
+        * refuse this early. */
+        if (p->attached_path) {
+                if (!path_equal(p->attached_path, path))
+                        return -EBUSY;
+                if (p->attached_type != type)
+                        return -EBUSY;
+                if (p->attached_flags != flags)
+                        return -EBUSY;
+
+                /* Here's a shortcut: if we previously attached this program already, then we don't have to do so
+                 * again. Well, with one exception: if we are in BPF_F_ALLOW_OVERRIDE mode then someone else might have
+                 * replaced our program since the last time, hence let's reattach it again, just to be safe. In flags
+                 * == 0 mode this is not an issue since nobody else can replace our program in that case, and in flags
+                 * == BPF_F_ALLOW_MULTI mode any other's program would be installed in addition to ours hence ours
+                 * would remain in effect. */
+                if (flags != BPF_F_ALLOW_OVERRIDE)
+                        return 0;
+        }
+
+        /* Ensure we have a kernel object for this. */
+        r = bpf_program_load_kernel(p, NULL, 0);
+        if (r < 0)
+                return r;
+
+        copy = strdup(path);
+        if (!copy)
+                return -ENOMEM;
+
         fd = open(path, O_DIRECTORY|O_RDONLY|O_CLOEXEC);
         if (fd < 0)
                 return -errno;
@@ -120,31 +183,43 @@ int bpf_program_cgroup_attach(BPFProgram *p, int type, const char *path, uint32_
         if (bpf(BPF_PROG_ATTACH, &attr, sizeof(attr)) < 0)
                 return -errno;
 
+        free_and_replace(p->attached_path, copy);
+        p->attached_type = type;
+        p->attached_flags = flags;
+
         return 0;
 }
 
-int bpf_program_cgroup_detach(BPFProgram *p, int type, const char *path) {
+int bpf_program_cgroup_detach(BPFProgram *p) {
         _cleanup_close_ int fd = -1;
-        union bpf_attr attr;
 
-        assert(type >= 0);
-        assert(path);
+        assert(p);
 
-        /* Note that 'p' may be NULL, in which case any program is detached. However, note that if BPF_F_ALLOW_MULTI is
-         * used 'p' is not optional. */
+        if (!p->attached_path)
+                return -EUNATCH;
 
-        fd = open(path, O_DIRECTORY|O_RDONLY|O_CLOEXEC);
-        if (fd < 0)
-                return -errno;
+        fd = open(p->attached_path, O_DIRECTORY|O_RDONLY|O_CLOEXEC);
+        if (fd < 0) {
+                if (errno != ENOENT)
+                        return -errno;
 
-        attr = (union bpf_attr) {
-                .attach_type = type,
-                .target_fd = fd,
-                .attach_bpf_fd = p ? p->kernel_fd : -1,
-        };
+                /* If the cgroup does not exist anymore, then we don't have to explicitly detach, it got detached
+                 * implicitly by the removal, hence don't complain */
 
-        if (bpf(BPF_PROG_DETACH, &attr, sizeof(attr)) < 0)
-                return -errno;
+        } else {
+                union bpf_attr attr;
+
+                attr = (union bpf_attr) {
+                        .attach_type = p->attached_type,
+                        .target_fd = fd,
+                        .attach_bpf_fd = p->kernel_fd,
+                };
+
+                if (bpf(BPF_PROG_DETACH, &attr, sizeof(attr)) < 0)
+                        return -errno;
+        }
+
+        p->attached_path = mfree(p->attached_path);
 
         return 0;
 }