This change adds support for tail growing and shrinking for XDP frags.
When called on a non-linear packet with a grow request, it will work
on the last fragment of the packet. So the maximum grow size is the
last fragments tailroom, i.e. no new buffer will be allocated.
A XDP frags capable driver is expected to set frag_size in xdp_rxq_info
data structure to notify the XDP core the fragment size.
frag_size set to 0 is interpreted by the XDP core as tail growing is
not allowed.
Introduce __xdp_rxq_info_reg utility routine to initialize frag_size field.
When shrinking, it will work from the last fragment, all the way down to
the base buffer depending on the shrinking size. It's important to mention
that once you shrink down the fragment(s) are freed, so you can not grow
again to the original size.
Acked-by: Toke Hoiland-Jorgensen <toke@redhat.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Co-developed-by: Lorenzo Bianconi <lorenzo@kernel.org>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Signed-off-by: Eelco Chaudron <echaudro@redhat.com>
Link: https://lore.kernel.org/r/eabda3485dda4f2f158b477729337327e609461d.1642758637.git.lorenzo@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Introduce xdp_update_skb_shared_info routine to update frags array
metadata in skb_shared_info data structure converting to a skb from
a xdp_buff or xdp_frame.
According to the current skb_shared_info architecture in
xdp_frame/xdp_buff and to the xdp frags support, there is
no need to run skb_add_rx_frag() and reset frags array converting the buffer
to a skb since the frag array will be in the same position for xdp_buff/xdp_frame
and for the skb, we just need to update memory metadata.
Introduce XDP_FLAGS_PF_MEMALLOC flag in xdp_buff_flags in order to mark
the xdp_buff or xdp_frame as under memory-pressure if pages of the frags array
are under memory pressure. Doing so we can avoid looping over all fragments in
xdp_update_skb_shared_info routine. The driver is expected to set the
flag constructing the xdp_buffer using xdp_buff_set_frag_pfmemalloc
utility routine.
Rely on xdp_update_skb_shared_info in __xdp_build_skb_from_frame routine
converting the non-linear xdp_frame to a skb after performing a XDP_REDIRECT.
Acked-by: Toke Hoiland-Jorgensen <toke@redhat.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Link: https://lore.kernel.org/r/bfd23fb8a8d7438724f7819c567cdf99ffd6226f.1642758637.git.lorenzo@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Introduce flags field in xdp_frame and xdp_buffer data structures
to define additional buffer features. At the moment the only
supported buffer feature is frags bit (XDP_FLAGS_HAS_FRAGS).
frags bit is used to specify if this is a linear buffer
(XDP_FLAGS_HAS_FRAGS not set) or a frags frame (XDP_FLAGS_HAS_FRAGS
set). In the latter case the driver is expected to initialize the
skb_shared_info structure at the end of the first buffer to link together
subsequent buffers belonging to the same frame.
Acked-by: Toke Hoiland-Jorgensen <toke@redhat.com>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Link: https://lore.kernel.org/r/e389f14f3a162c0a5bc6a2e1aa8dd01a90be117d.1642758637.git.lorenzo@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Andrii Nakryiko says:
====================
Revamp existing low-level XDP APIs provided by libbpf to follow more
consistent naming (new APIs follow bpf_tc_xxx() approach where it makes
sense) and be extensible without ABI breakages (OPTS-based). See patch #1 for
details, remaining patches switch bpftool, selftests/bpf and samples/bpf to
new APIs.
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Introduce 4 new netlink-based XDP APIs for attaching, detaching, and
querying XDP programs:
- bpf_xdp_attach;
- bpf_xdp_detach;
- bpf_xdp_query;
- bpf_xdp_query_id.
These APIs replace bpf_set_link_xdp_fd, bpf_set_link_xdp_fd_opts,
bpf_get_link_xdp_id, and bpf_get_link_xdp_info APIs ([0]). The latter
don't follow a consistent naming pattern and some of them use
non-extensible approaches (e.g., struct xdp_link_info which can't be
modified without breaking libbpf ABI).
The approach I took with these low-level XDP APIs is similar to what we
did with low-level TC APIs. There is a nice duality of bpf_tc_attach vs
bpf_xdp_attach, and so on. I left bpf_xdp_attach() to support detaching
when -1 is specified for prog_fd for generality and convenience, but
bpf_xdp_detach() is preferred due to clearer naming and associated
semantics. Both bpf_xdp_attach() and bpf_xdp_detach() accept the same
opts struct allowing to specify expected old_prog_fd.
While doing the refactoring, I noticed that old APIs require users to
specify opts with old_fd == -1 to declare "don't care about already
attached XDP prog fd" condition. Otherwise, FD 0 is assumed, which is
essentially never an intended behavior. So I made this behavior
consistent with other kernel and libbpf APIs, in which zero FD means "no
FD". This seems to be more in line with the latest thinking in BPF land
and should cause less user confusion, hopefully.
For querying, I left two APIs, both more generic bpf_xdp_query()
allowing to query multiple IDs and attach mode, but also
a specialization of it, bpf_xdp_query_id(), which returns only requested
prog_id. Uses of prog_id returning bpf_get_link_xdp_id() were so
prevalent across selftests and samples, that it seemed a very common use
case and using bpf_xdp_query() for doing it felt very cumbersome with
a highly branches if/else chain based on flags and attach mode.
Old APIs are scheduled for deprecation in libbpf 0.8 release.
[0] Closes: https://github.com/libbpf/libbpf/issues/309
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>
Link: https://lore.kernel.org/r/20220120061422.2710637-2-andrii@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Andrii Nakryiko says:
====================
Officially deprecate legacy BPF map definitions in libbpf. They've been slated
for deprecation for a while in favor of more powerful BTF-defined map
definitions and this patch set adds warnings and a way to enforce this in
libbpf through LIBBPF_STRICT_MAP_DEFINITIONS strict mode flag.
Selftests are fixed up and updated, BPF documentation is updated, bpftool's
strict mode usage is adjusted to avoid breaking users unnecessarily.
v1->v2:
- replace missed bpf_map_def case in Documentation/bpf/btf.rst (Alexei).
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Enact deprecation of legacy BPF map definition in SEC("maps") ([0]). For
the definitions themselves introduce LIBBPF_STRICT_MAP_DEFINITIONS flag
for libbpf strict mode. If it is set, error out on any struct
bpf_map_def-based map definition. If not set, libbpf will print out
a warning for each legacy BPF map to raise awareness that it goes away.
For any use of BPF_ANNOTATE_KV_PAIR() macro providing a legacy way to
associate BTF key/value type information with legacy BPF map definition,
warn through libbpf's pr_warn() error message (but don't fail BPF object
open).
BPF-side struct bpf_map_def is marked as deprecated. User-space struct
bpf_map_def has to be used internally in libbpf, so it is left
untouched. It should be enough for bpf_map__def() to be marked
deprecated to raise awareness that it goes away.
bpftool is an interesting case that utilizes libbpf to open BPF ELF
object to generate skeleton. As such, even though bpftool itself uses
full on strict libbpf mode (LIBBPF_STRICT_ALL), it has to relax it a bit
for BPF map definition handling to minimize unnecessary disruptions. So
opt-out of LIBBPF_STRICT_MAP_DEFINITIONS for bpftool. User's code that
will later use generated skeleton will make its own decision whether to
enforce LIBBPF_STRICT_MAP_DEFINITIONS or not.
There are few tests in selftests/bpf that are consciously using legacy
BPF map definitions to test libbpf functionality. For those, temporary
opt out of LIBBPF_STRICT_MAP_DEFINITIONS mode for the duration of those
tests.
[0] Closes: https://github.com/libbpf/libbpf/issues/272
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/r/20220120060529.1890907-4-andrii@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Converted few remaining legacy BPF map definition to BTF-defined ones.
For the remaining two bpf_map_def-based legacy definitions that we want
to keep for testing purposes until libbpf 1.0 release, guard them in
pragma to suppres deprecation warnings which will be added in libbpf in
the next commit.
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/r/20220120060529.1890907-3-andrii@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
It's very easy to miss compilation warnings without -Werror, which is
not set for selftests. libbpf and bpftool are already strict about this,
so make selftests/bpf also treat compilation warnings as errors to catch
such regressions early.
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/r/20220120060529.1890907-2-andrii@kernel.org
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Some users have complained that selftests fail to build when
CONFIG_NF_CONNTRACK=m. It would be useful to allow building as long as
it is set to module or built-in, even though in case of building as
module, user would need to load it before running the selftest. Note
that this also allows building selftest when CONFIG_NF_CONNTRACK is
disabled.
Signed-off-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20220120164932.2798544-1-memxor@gmail.com
The bind_perm BPF selftest failed when port 111/tcp was already in use
during the test. To fix this, the test now runs in its own network name
space.
To use unshare, it is necessary to reorder the includes. The style of
the includes is adapted to be consistent with the other prog_tests.
v2: Replace deprecated CHECK macro with ASSERT_OK
Fixes: 8259fdeb30 ("selftests/bpf: Verify that rebinding to port < 1024 from BPF works")
Signed-off-by: Felix Maurer <fmaurer@redhat.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Reviewed-by: Jakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/bpf/551ee65533bb987a43f93d88eaf2368b416ccd32.1642518457.git.fmaurer@redhat.com
Lorenzo Bianconi says:
====================
Rely on ASSERT* macros and get rid of deprecated CHECK ones in xdp_bpf2bpf and
xdp_adjust_tail bpf selftests.
This is a preliminary series for XDP multi-frags support.
Changes since v1:
- run each ASSERT test separately
- drop unnecessary return statements
- drop unnecessary if condition in test_xdp_bpf2bpf()
====================
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
YiFei Zhu says:
====================
Right now, most cgroup hooks are best used for permission checks. They
can only reject a syscall with -EPERM, so a cause of a rejection, if
the rejected by eBPF cgroup hooks, is ambiguous to userspace.
Additionally, if the syscalls are implemented in eBPF, all permission
checks and the implementation has to happen within the same filter,
as programs executed later in the series of progs are unaware of the
return values return by the previous progs.
This patch series adds two helpers, bpf_get_retval and bpf_set_retval,
that allows hooks to get/set the return value of syscall to userspace.
This also allows later progs to retrieve retval set by previous progs.
For legacy programs that rejects a syscall without setting the retval,
for backwards compatibility, if a prog rejects without itself or a
prior prog setting retval to an -err, the retval is set by the kernel
to -EPERM.
For getsockopt hooks that has ctx->retval, this variable mirrors that
that accessed by the helpers.
Additionally, the following user-visible behavior for getsockopt
hooks has changed:
- If a prior filter rejected the syscall, it will be visible
in ctx->retval.
- Attempting to change the retval arbitrarily is now allowed and
will not cause an -EFAULT.
- If kernel rejects a getsockopt syscall before running the hooks,
the error will be visible in ctx->retval. Returning 0 from the
prog will not overwrite the error to -EPERM unless there is an
explicit call of bpf_set_retval(-EPERM)
Tests have been added in this series to test the behavior of the helper
with cgroup setsockopt getsockopt hooks.
Patch 1 changes the API of macros to prepare for the next patch and
should be a no-op.
Patch 2 moves ctx->retval to a struct pointed to by current
task_struct.
Patch 3 implements the helpers.
Patch 4 tests the behaviors of the helpers.
Patch 5 updates a test after the test broke due to the visible changes.
v1 -> v2:
- errno -> retval
- split one helper to get & set helpers
- allow retval to be set arbitrarily in the general case
- made the helper retval and context retval mirror each other
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The tests would break without this patch, because at one point it calls
getsockopt(fd, SOL_TCP, TCP_ZEROCOPY_RECEIVE, &buf, &optlen)
This getsockopt receives the kernel-set -EINVAL. Prior to this patch
series, the eBPF getsockopt hook's -EPERM would override kernel's
-EINVAL, however, after this patch series, return 0's automatic
-EPERM will not; the eBPF prog has to explicitly bpf_set_retval(-EPERM)
if that is wanted.
I also removed the explicit mentions of EPERM in the comments in the
prog.
Signed-off-by: YiFei Zhu <zhuyifei@google.com>
Reviewed-by: Stanislav Fomichev <sdf@google.com>
Link: https://lore.kernel.org/r/4f20b77cb46812dbc2bdcd7e3fa87c7573bde55e.1639619851.git.zhuyifei@google.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The tests checks how different ways of interacting with the helpers
(getting retval, setting EUNATCH, EISCONN, and legacy reject
returning 0 without setting retval), produce different results in
both the setsockopt syscall and the retval returned by the helper.
A few more tests verify the interaction between the retval of the
helper and the retval in getsockopt context.
Signed-off-by: YiFei Zhu <zhuyifei@google.com>
Reviewed-by: Stanislav Fomichev <sdf@google.com>
Link: https://lore.kernel.org/r/43ec60d679ae3f4f6fd2460559c28b63cb93cd12.1639619851.git.zhuyifei@google.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The helpers continue to use int for retval because all the hooks
are int-returning rather than long-returning. The return value of
bpf_set_retval is int for future-proofing, in case in the future
there may be errors trying to set the retval.
After the previous patch, if a program rejects a syscall by
returning 0, an -EPERM will be generated no matter if the retval
is already set to -err. This patch change it being forced only if
retval is not -err. This is because we want to support, for
example, invoking bpf_set_retval(-EINVAL) and return 0, and have
the syscall return value be -EINVAL not -EPERM.
For BPF_PROG_CGROUP_INET_EGRESS_RUN_ARRAY, the prior behavior is
that, if the return value is NET_XMIT_DROP, the packet is silently
dropped. We preserve this behavior for backward compatibility
reasons, so even if an errno is set, the errno does not return to
caller. However, setting a non-err to retval cannot propagate so
this is not allowed and we return a -EFAULT in that case.
Signed-off-by: YiFei Zhu <zhuyifei@google.com>
Reviewed-by: Stanislav Fomichev <sdf@google.com>
Link: https://lore.kernel.org/r/b4013fd5d16bed0b01977c1fafdeae12e1de61fb.1639619851.git.zhuyifei@google.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The retval value is moved to struct bpf_cg_run_ctx for ease of access
in different prog types with different context structs layouts. The
helper implementation (to be added in a later patch in the series) can
simply perform a container_of from current->bpf_ctx to retrieve
bpf_cg_run_ctx.
Unfortunately, there is no easy way to access the current task_struct
via the verifier BPF bytecode rewrite, aside from possibly calling a
helper, so a pointer to current task is added to struct bpf_sockopt_kern
so that the rewritten BPF bytecode can access struct bpf_cg_run_ctx with
an indirection.
For backward compatibility, if a getsockopt program rejects a syscall
by returning 0, an -EPERM will be generated, by having the
BPF_PROG_RUN_ARRAY_CG family macros automatically set the retval to
-EPERM. Unlike prior to this patch, this -EPERM will be visible to
ctx->retval for any other hooks down the line in the prog array.
Additionally, the restriction that getsockopt filters can only set
the retval to 0 is removed, considering that certain getsockopt
implementations may return optlen. Filters are now able to set the
value arbitrarily.
Signed-off-by: YiFei Zhu <zhuyifei@google.com>
Reviewed-by: Stanislav Fomichev <sdf@google.com>
Link: https://lore.kernel.org/r/73b0325f5c29912ccea7ea57ec1ed4d388fc1d37.1639619851.git.zhuyifei@google.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Right now BPF_PROG_RUN_ARRAY and related macros return 1 or 0
for whether the prog array allows or rejects whatever is being
hooked. The caller of these macros then return -EPERM or continue
processing based on thw macro's return value. Unforunately this is
inflexible, since -EPERM is the only err that can be returned.
This patch should be a no-op; it prepares for the next patch. The
returning of the -EPERM is moved to inside the macros, so the outer
functions are directly returning what the macros returned if they
are non-zero.
Signed-off-by: YiFei Zhu <zhuyifei@google.com>
Reviewed-by: Stanislav Fomichev <sdf@google.com>
Link: https://lore.kernel.org/r/788abcdca55886d1f43274c918eaa9f792a9f33b.1639619851.git.zhuyifei@google.com
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Add a hashmap to map the string offsets from a source btf to the
string offsets from a target btf to reduce overheads.
btf__add_btf() calls btf__add_str() to add strings from a source to a
target btf. It causes many string comparisons, and it is a major
hotspot when adding a big btf. btf__add_str() uses strcmp() to check
if a hash entry is the right one. The extra hashmap here compares
offsets of strings, that are much cheaper. It remembers the results
of btf__add_str() for later uses to reduce the cost.
We are parallelizing BTF encoding for pahole by creating separated btf
instances for worker threads. These per-thread btf instances will be
added to the btf instance of the main thread by calling btf__add_str()
to deduplicate and write out. With this patch and -j4, the running
time of pahole drops to about 6.0s from 6.6s.
The following lines are the summary of 'perf stat' w/o the change.
6.668126396 seconds time elapsed
13.451054000 seconds user
0.715520000 seconds sys
The following lines are the summary w/ the change.
5.986973919 seconds time elapsed
12.939903000 seconds user
0.724152000 seconds sys
V4 fixes a bug of error checking against the pointer returned by
hashmap__new().
[v3] https://lore.kernel.org/bpf/20220118232053.2113139-1-kuifeng@fb.com/
[v2] https://lore.kernel.org/bpf/20220114193713.461349-1-kuifeng@fb.com/
Signed-off-by: Kui-Feng Lee <kuifeng@fb.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20220119180214.255634-1-kuifeng@fb.com
Currently the syscalls rst and subsequently man page are auto-generated
using function documentation present in bpf.h. If the documentation for the
syscall is missing or doesn't follow a specific format, then that syscall
is not dumped in the auto-generated rst.
This patch checks the number of syscalls documented within the header file
with those present as part of the enum bpf_cmd and raises an Exception if
they don't match. It is not needed with the currently documented upstream
syscalls, but can help in debugging when developing new syscalls when
there might be missing or misformatted documentation.
The function helper_number_check is moved to the Printer parent
class and renamed to elem_number_check as all the most derived children
classes are using this function now.
Signed-off-by: Usama Arif <usama.arif@bytedance.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Reviewed-by: Quentin Monnet <quentin@isovalent.com>
Link: https://lore.kernel.org/bpf/20220119114442.1452088-3-usama.arif@bytedance.com
Both description and returns section will become mandatory
for helpers and syscalls in a later commit to generate man pages.
This commit also adds in the documentation that BPF_PROG_RUN is
an alias for BPF_PROG_TEST_RUN for anyone searching for the
syscall in the generated man pages.
Signed-off-by: Usama Arif <usama.arif@bytedance.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20220119114442.1452088-1-usama.arif@bytedance.com
