| 1 |
# Copyright (c) Meta Platforms, Inc. and affiliates. |
| 2 |
# SPDX-License-Identifier: LGPL-2.1-or-later |
| 3 |
|
| 4 |
import contextlib |
| 5 |
import errno |
| 6 |
import os |
| 7 |
import resource |
| 8 |
import unittest |
| 9 |
|
| 10 |
from _drgn_util.platform import NORMALIZED_MACHINE_NAME |
| 11 |
from drgn import cast |
| 12 |
from drgn.helpers.linux.bpf import ( |
| 13 |
bpf_btf_for_each, |
| 14 |
bpf_link_for_each, |
| 15 |
bpf_map_by_id, |
| 16 |
bpf_map_for_each, |
| 17 |
bpf_prog_by_id, |
| 18 |
bpf_prog_for_each, |
| 19 |
bpf_prog_used_maps, |
| 20 |
cgroup_bpf_prog_for_each, |
| 21 |
cgroup_bpf_prog_for_each_effective, |
| 22 |
) |
| 23 |
from drgn.helpers.linux.cgroup import cgroup_get_from_path |
| 24 |
from drgn.helpers.linux.fs import fget |
| 25 |
from drgn.helpers.linux.pid import find_task |
| 26 |
from tests.linux_kernel import LinuxKernelTestCase |
| 27 |
from tests.linux_kernel.bpf import ( |
| 28 |
BPF_CGROUP_INET_INGRESS, |
| 29 |
BPF_EXIT_INSN, |
| 30 |
BPF_F_ALLOW_MULTI, |
| 31 |
BPF_LD_MAP_FD, |
| 32 |
BPF_MAP_TYPE_HASH, |
| 33 |
BPF_MOV64_IMM, |
| 34 |
BPF_PROG_TYPE_CGROUP_SKB, |
| 35 |
BPF_PROG_TYPE_SOCKET_FILTER, |
| 36 |
BPF_REG_0, |
| 37 |
_SYS_bpf, |
| 38 |
bpf_btf_ids, |
| 39 |
bpf_link_create, |
| 40 |
bpf_link_ids, |
| 41 |
bpf_map_create, |
| 42 |
bpf_map_get_info_by_fd, |
| 43 |
bpf_map_ids, |
| 44 |
bpf_prog_attach, |
| 45 |
bpf_prog_get_info_by_fd, |
| 46 |
bpf_prog_ids, |
| 47 |
bpf_prog_load, |
| 48 |
) |
| 49 |
from tests.linux_kernel.helpers.test_cgroup import tmp_cgroups |
| 50 |
|
| 51 |
|
| 52 |
class BpfTestCase(LinuxKernelTestCase): |
| 53 |
INSNS = ( |
| 54 |
BPF_MOV64_IMM(BPF_REG_0, 0), |
| 55 |
BPF_EXIT_INSN(), |
| 56 |
) |
| 57 |
|
| 58 |
@classmethod |
| 59 |
def setUpClass(cls): |
| 60 |
super().setUpClass() |
| 61 |
if _SYS_bpf is None: |
| 62 |
raise unittest.SkipTest( |
| 63 |
f"bpf syscall number is not known on {NORMALIZED_MACHINE_NAME}" |
| 64 |
) |
| 65 |
# Before the patch series culminating in Linux kernel commit |
| 66 |
# 3ac1f01b43b6 ("bpf: Eliminate rlimit-based memory accounting for bpf |
| 67 |
# progs") (in v5.11), BPF program and map memory usage was limited by |
| 68 |
# RLIMIT_MEMLOCK. At that time (before Linux kernel commit 9dcc38e2813e |
| 69 |
# ("Increase default MLOCK_LIMIT to 8 MiB") (in v5.16)), the limit was |
| 70 |
# only 64kB. We only allocate a few small objects at a time, but with |
| 71 |
# 64k pages, we can easily blow that limit. |
| 72 |
memlock_limit = 8 * 1024 * 1024 |
| 73 |
old_limit = resource.getrlimit(resource.RLIMIT_MEMLOCK) |
| 74 |
if old_limit[0] < memlock_limit: |
| 75 |
resource.setrlimit( |
| 76 |
resource.RLIMIT_MEMLOCK, |
| 77 |
(memlock_limit, max(memlock_limit, old_limit[1])), |
| 78 |
) |
| 79 |
cls.addClassCleanup(resource.setrlimit, resource.RLIMIT_MEMLOCK, old_limit) |
| 80 |
try: |
| 81 |
os.close(bpf_map_create(BPF_MAP_TYPE_HASH, 8, 8, 8)) |
| 82 |
except OSError as e: |
| 83 |
if e.errno != errno.ENOSYS: |
| 84 |
raise |
| 85 |
raise unittest.SkipTest( |
| 86 |
"kernel does not support bpf syscall (CONFIG_BPF_SYSCALL)" |
| 87 |
) |
| 88 |
|
| 89 |
|
| 90 |
class TestBpf(BpfTestCase): |
| 91 |
def test_bpf_btf_for_each(self): |
| 92 |
# BTF was added in Linux kernel commit 69b693f0aefa ("bpf: btf: |
| 93 |
# Introduce BPF Type Format (BTF)") (in v4.18) and had IDs from the |
| 94 |
# start, but there was no API to get them until commit 1b9ed84ecf26 |
| 95 |
# ("bpf: add new BPF_BTF_GET_NEXT_ID syscall command") (in v5.4). The |
| 96 |
# only kernel version that we support in between is v4.19, which we can |
| 97 |
# live without testing. |
| 98 |
# |
| 99 |
# Note that before Linux kernel commits 5329722057d4 ("bpf: Assign ID |
| 100 |
# to vmlinux BTF and return extra info for BTF in GET_OBJ_INFO") and |
| 101 |
# 36e68442d1af ("bpf: Load and verify kernel module BTFs") (in v5.11), |
| 102 |
# there won't be any BTF IDs unless they were explicitly added outside |
| 103 |
# of the test suite. |
| 104 |
try: |
| 105 |
expected_ids = list(bpf_btf_ids()) |
| 106 |
except OSError as e: |
| 107 |
if e.errno != errno.EINVAL: |
| 108 |
raise |
| 109 |
self.skipTest("kernel does not support BPF_BTF_GET_NEXT_ID") |
| 110 |
self.assertCountEqual( |
| 111 |
[btf.id.value_() for btf in bpf_btf_for_each(self.prog)], expected_ids |
| 112 |
) |
| 113 |
|
| 114 |
def test_bpf_link_for_each(self): |
| 115 |
with tmp_cgroups() as cgroups: |
| 116 |
fds = [] |
| 117 |
try: |
| 118 |
for cgroup in cgroups: |
| 119 |
fds.append(os.open(cgroup, os.O_RDONLY | os.O_DIRECTORY)) |
| 120 |
for i in range(3): |
| 121 |
# Cgroup BPF programs didn't exist before Linux kernel |
| 122 |
# commit 3007098494be ("cgroup: add support for eBPF |
| 123 |
# programs") (in v4.10). |
| 124 |
try: |
| 125 |
prog_fd = bpf_prog_load( |
| 126 |
BPF_PROG_TYPE_CGROUP_SKB, |
| 127 |
self.INSNS, |
| 128 |
b"GPL", |
| 129 |
expected_attach_type=BPF_CGROUP_INET_INGRESS, |
| 130 |
) |
| 131 |
fds.append(prog_fd) |
| 132 |
except OSError as e: |
| 133 |
if e.errno != errno.EINVAL: |
| 134 |
raise |
| 135 |
self.skipTest( |
| 136 |
"kernel does not support BPF_PROG_TYPE_CGROUP_SKB" |
| 137 |
) |
| 138 |
# BPF links didn't exist before Linux kernel commit |
| 139 |
# 70ed506c3bbc ("bpf: Introduce pinnable bpf_link |
| 140 |
# abstraction") (in v5.7). |
| 141 |
try: |
| 142 |
fds.append( |
| 143 |
bpf_link_create( |
| 144 |
prog_fd, fds[i % len(cgroups)], BPF_CGROUP_INET_INGRESS |
| 145 |
) |
| 146 |
) |
| 147 |
except OSError as e: |
| 148 |
if e.errno != errno.EINVAL: |
| 149 |
raise |
| 150 |
self.skipTest("kernel does not support BPF_LINK_CREATE") |
| 151 |
|
| 152 |
# bpf_link_for_each() isn't supported before Linux v5.8, which |
| 153 |
# added IDs for BPF links in commit a3b80e107894 ("bpf: |
| 154 |
# Allocate ID for bpf_link") and an API to get them in commit |
| 155 |
# 2d602c8cf40d ("bpf: Support GET_FD_BY_ID and GET_NEXT_ID for |
| 156 |
# bpf_link"). |
| 157 |
try: |
| 158 |
expected_ids = list(bpf_link_ids()) |
| 159 |
except OSError as e: |
| 160 |
if e.errno != errno.EINVAL: |
| 161 |
raise |
| 162 |
self.skipTest("kernel does not support BPF_LINK_GET_NEXT_ID") |
| 163 |
self.assertCountEqual( |
| 164 |
[link.id.value_() for link in bpf_link_for_each(self.prog)], |
| 165 |
expected_ids, |
| 166 |
) |
| 167 |
finally: |
| 168 |
for fd in fds: |
| 169 |
os.close(fd) |
| 170 |
|
| 171 |
def test_bpf_map_for_each(self): |
| 172 |
fds = [] |
| 173 |
try: |
| 174 |
for i in range(3): |
| 175 |
fds.append(bpf_map_create(BPF_MAP_TYPE_HASH, 8, 8, 8)) |
| 176 |
|
| 177 |
# bpf_map_for_each() isn't supported before Linux v4.13, which |
| 178 |
# added IDs for BPF maps in commit f3f1c054c288 ("bpf: Introduce |
| 179 |
# bpf_map ID") and a API to get them in commit 34ad5580f8f9 ("bpf: |
| 180 |
# Add BPF_(PROG|MAP)_GET_NEXT_ID command"). |
| 181 |
try: |
| 182 |
expected_ids = list(bpf_map_ids()) |
| 183 |
except OSError as e: |
| 184 |
if e.errno != errno.EINVAL: |
| 185 |
raise |
| 186 |
self.skipTest("kernel does not support BPF_MAP_GET_NEXT_ID") |
| 187 |
|
| 188 |
self.assertCountEqual( |
| 189 |
[map.id.value_() for map in bpf_map_for_each(self.prog)], expected_ids |
| 190 |
) |
| 191 |
finally: |
| 192 |
for fd in fds: |
| 193 |
os.close(fd) |
| 194 |
|
| 195 |
def test_bpf_prog_for_each(self): |
| 196 |
fds = [] |
| 197 |
try: |
| 198 |
for i in range(3): |
| 199 |
fds.append( |
| 200 |
bpf_prog_load(BPF_PROG_TYPE_SOCKET_FILTER, self.INSNS, b"GPL") |
| 201 |
) |
| 202 |
|
| 203 |
# bpf_prog_for_each() isn't supported before Linux v4.13, which |
| 204 |
# added IDs for BPF programs in commit dc4bb0e23561 ("bpf: |
| 205 |
# Introduce bpf_prog ID") and an API to get them in commit |
| 206 |
# 34ad5580f8f9 ("bpf: Add BPF_(PROG|MAP)_GET_NEXT_ID command"). |
| 207 |
try: |
| 208 |
expected_ids = list(bpf_prog_ids()) |
| 209 |
except OSError as e: |
| 210 |
if e.errno != errno.EINVAL: |
| 211 |
raise |
| 212 |
self.skipTest("kernel does not support BPF_PROG_GET_NEXT_ID") |
| 213 |
|
| 214 |
self.assertCountEqual( |
| 215 |
[prog.aux.id.value_() for prog in bpf_prog_for_each(self.prog)], |
| 216 |
expected_ids, |
| 217 |
) |
| 218 |
finally: |
| 219 |
for fd in fds: |
| 220 |
os.close(fd) |
| 221 |
|
| 222 |
def test_bpf_prog_used_maps(self): |
| 223 |
with contextlib.ExitStack() as exit_stack: |
| 224 |
map_fd = bpf_map_create(BPF_MAP_TYPE_HASH, 8, 8, 8) |
| 225 |
exit_stack.callback(os.close, map_fd) |
| 226 |
|
| 227 |
prog_fd = bpf_prog_load( |
| 228 |
BPF_PROG_TYPE_SOCKET_FILTER, |
| 229 |
BPF_LD_MAP_FD(BPF_REG_0, map_fd) + self.INSNS, |
| 230 |
b"GPL", |
| 231 |
) |
| 232 |
exit_stack.callback(os.close, prog_fd) |
| 233 |
|
| 234 |
bpf_prog = cast( |
| 235 |
"struct bpf_prog *", |
| 236 |
fget(find_task(self.prog, os.getpid()), prog_fd).private_data, |
| 237 |
) |
| 238 |
|
| 239 |
bpf_map = cast( |
| 240 |
"struct bpf_map *", |
| 241 |
fget(find_task(self.prog, os.getpid()), map_fd).private_data, |
| 242 |
) |
| 243 |
|
| 244 |
self.assertEqual(list(bpf_prog_used_maps(bpf_prog)), [bpf_map]) |
| 245 |
|
| 246 |
def test_cgroup_bpf_prog_for_each(self): |
| 247 |
with tmp_cgroups() as (parent_cgroup, child_cgroup): |
| 248 |
fds = [] |
| 249 |
try: |
| 250 |
parent_cgroup_fd = os.open(parent_cgroup, os.O_RDONLY | os.O_DIRECTORY) |
| 251 |
fds.append(parent_cgroup_fd) |
| 252 |
child_cgroup_fd = os.open(child_cgroup, os.O_RDONLY | os.O_DIRECTORY) |
| 253 |
fds.append(child_cgroup_fd) |
| 254 |
try: |
| 255 |
parent_prog_fd = bpf_prog_load( |
| 256 |
BPF_PROG_TYPE_CGROUP_SKB, |
| 257 |
self.INSNS, |
| 258 |
b"GPL", |
| 259 |
expected_attach_type=BPF_CGROUP_INET_INGRESS, |
| 260 |
) |
| 261 |
fds.append(parent_prog_fd) |
| 262 |
except OSError as e: |
| 263 |
if e.errno != errno.EINVAL: |
| 264 |
raise |
| 265 |
# If the kernel doesn't support cgroup BPF programs, the |
| 266 |
# helpers should return empty lists. |
| 267 |
parent_ids = child_ids = child_effective_ids = [] |
| 268 |
else: |
| 269 |
parent_prog_id = bpf_prog_get_info_by_fd(parent_prog_fd).id |
| 270 |
parent_ids = [parent_prog_id] |
| 271 |
|
| 272 |
# If the kernel supports BPF_F_ALLOW_MULTI, test with |
| 273 |
# multiple programs. |
| 274 |
try: |
| 275 |
bpf_prog_attach( |
| 276 |
parent_cgroup_fd, |
| 277 |
parent_prog_fd, |
| 278 |
BPF_CGROUP_INET_INGRESS, |
| 279 |
attach_flags=BPF_F_ALLOW_MULTI, |
| 280 |
) |
| 281 |
except OSError as e: |
| 282 |
if e.errno != errno.EINVAL: |
| 283 |
raise |
| 284 |
bpf_prog_attach( |
| 285 |
parent_cgroup_fd, parent_prog_fd, BPF_CGROUP_INET_INGRESS |
| 286 |
) |
| 287 |
child_ids = [] |
| 288 |
child_effective_ids = [parent_prog_id] |
| 289 |
else: |
| 290 |
child_prog_fd = bpf_prog_load( |
| 291 |
BPF_PROG_TYPE_CGROUP_SKB, |
| 292 |
self.INSNS, |
| 293 |
b"GPL", |
| 294 |
expected_attach_type=BPF_CGROUP_INET_INGRESS, |
| 295 |
) |
| 296 |
fds.append(child_prog_fd) |
| 297 |
child_prog_id = bpf_prog_get_info_by_fd(child_prog_fd).id |
| 298 |
bpf_prog_attach( |
| 299 |
child_cgroup_fd, child_prog_fd, BPF_CGROUP_INET_INGRESS |
| 300 |
) |
| 301 |
child_ids = [child_prog_id] |
| 302 |
child_effective_ids = [parent_prog_id, child_prog_id] |
| 303 |
|
| 304 |
parent_cgrp = cgroup_get_from_path(self.prog, parent_cgroup.name) |
| 305 |
child_cgrp = cgroup_get_from_path( |
| 306 |
self.prog, parent_cgroup.name + "/" + child_cgroup.name |
| 307 |
) |
| 308 |
|
| 309 |
self.assertCountEqual( |
| 310 |
[ |
| 311 |
prog.aux.id.value_() |
| 312 |
for prog in cgroup_bpf_prog_for_each( |
| 313 |
parent_cgrp, BPF_CGROUP_INET_INGRESS |
| 314 |
) |
| 315 |
], |
| 316 |
parent_ids, |
| 317 |
) |
| 318 |
self.assertCountEqual( |
| 319 |
[ |
| 320 |
prog.aux.id.value_() |
| 321 |
for prog in cgroup_bpf_prog_for_each( |
| 322 |
child_cgrp, BPF_CGROUP_INET_INGRESS |
| 323 |
) |
| 324 |
], |
| 325 |
child_ids, |
| 326 |
) |
| 327 |
|
| 328 |
self.assertCountEqual( |
| 329 |
[ |
| 330 |
prog.aux.id.value_() |
| 331 |
for prog in cgroup_bpf_prog_for_each_effective( |
| 332 |
parent_cgrp, BPF_CGROUP_INET_INGRESS |
| 333 |
) |
| 334 |
], |
| 335 |
parent_ids, |
| 336 |
) |
| 337 |
self.assertCountEqual( |
| 338 |
[ |
| 339 |
prog.aux.id.value_() |
| 340 |
for prog in cgroup_bpf_prog_for_each_effective( |
| 341 |
child_cgrp, BPF_CGROUP_INET_INGRESS |
| 342 |
) |
| 343 |
], |
| 344 |
child_effective_ids, |
| 345 |
) |
| 346 |
finally: |
| 347 |
for fd in fds: |
| 348 |
os.close(fd) |
| 349 |
|
| 350 |
def test_bpf_prog_by_id(self): |
| 351 |
fd = bpf_prog_load(BPF_PROG_TYPE_SOCKET_FILTER, self.INSNS, b"GPL") |
| 352 |
self.addCleanup(os.close, fd) |
| 353 |
try: |
| 354 |
prog_id = bpf_prog_get_info_by_fd(fd).id |
| 355 |
except OSError as e: |
| 356 |
# bpf_prog_by_id() and BPF program IDs aren't supported before |
| 357 |
# Linux v4.13, which added IDs for BPF programs in commit |
| 358 |
# dc4bb0e23561 ("bpf: Introduce bpf_prog ID") and an API to get |
| 359 |
# them in commit 34ad5580f8f9 ("bpf: Add BPF_(PROG|MAP)_GET_NEXT_ID |
| 360 |
# command"). |
| 361 |
if e.errno != errno.EINVAL: |
| 362 |
raise |
| 363 |
self.skipTest("kernel does not support BPF_PROG_GET_NEXT_ID") |
| 364 |
|
| 365 |
prog = bpf_prog_by_id(self.prog, prog_id) |
| 366 |
self.assertNotEqual(prog.value_(), 0) |
| 367 |
self.assertEqual(prog.aux.id.value_(), prog_id) |
| 368 |
|
| 369 |
non_existent_prog = bpf_prog_by_id(self.prog, 0x7FFFFFFF) |
| 370 |
self.assertEqual(non_existent_prog.value_(), 0) |
| 371 |
|
| 372 |
def test_bpf_map_by_id(self): |
| 373 |
fd = bpf_map_create(BPF_MAP_TYPE_HASH, 8, 8, 8) |
| 374 |
self.addCleanup(os.close, fd) |
| 375 |
try: |
| 376 |
map_id = bpf_map_get_info_by_fd(fd).id |
| 377 |
except OSError as e: |
| 378 |
# bpf_map_by_id() and BPF map IDs aren't supported before |
| 379 |
# Linux v4.13, which added IDs for BPF maps in commit |
| 380 |
# f3f1c054c288 ("bpf: Introduce bpf_map ID") and an API to get |
| 381 |
# them in commit 34ad5580f8f9 ("bpf: Add BPF_(PROG|MAP)_GET_NEXT_ID |
| 382 |
# command"). |
| 383 |
if e.errno != errno.EINVAL: |
| 384 |
raise |
| 385 |
self.skipTest("kernel does not support BPF_MAP_GET_NEXT_ID") |
| 386 |
|
| 387 |
bpf_map = bpf_map_by_id(self.prog, map_id) |
| 388 |
self.assertNotEqual(bpf_map.value_(), 0) |
| 389 |
self.assertEqual(bpf_map.id.value_(), map_id) |
| 390 |
|
| 391 |
non_existent_map = bpf_map_by_id(self.prog, 0x7FFFFFFF) |
| 392 |
self.assertEqual(non_existent_map.value_(), 0) |