Jiong Wang says:
====================
v9:
- Split patch 5 in v8.
make bpf uapi header file sync a separate patch. (Alexei)
v8:
- For stack slot read, mark them as REG_LIVE_READ64. (Alexei)
- Change DEF_NOT_SUBREG from -1 to 0. (Alexei)
- Rebased on top of latest bpf-next.
v7:
- Drop the first patch in v6, the one adding 32-bit return value and
argument type. (Alexei)
- Rename bpf_jit_hardware_zext to bpf_jit_needs_zext. (Alexei)
- Use mov32 with imm == 1 to indicate it is zext. (Alexei)
- JIT back-ends peephole next insn to optimize out unnecessary zext
inserted by verifier. (Alexei)
- Testing:
+ patch set tested (bpf selftest) on x64 host with llvm 9.0
no regression observed no both JIT and interpreter modes.
+ patch set tested (bpf selftest) on x32 host.
By Yanqing Wang, thanks!
no regression observed on both JIT and interpreter modes.
+ patch set tested (bpf selftest) on RV64 host with llvm 9.0,
By Björn Töpel, thanks!
no regression observed before and after this set with JIT_ALWAYS_ON.
test_progs_32 also enabled as LLVM 9.0 is used by Björn.
+ cross compiled the other affected targets, arm, PowerPC, SPARC, S390.
v6:
- Fixed s390 kbuild test robot error. (kbuild)
- Make comment style in backends patches more consistent.
v5:
- Adjusted several test_verifier helpers to make them works on hosts
w and w/o hardware zext. (Naveen)
- Make sure zext flag not set when verifier by-passed, for example,
libtest_bpf.ko. (Naveen)
- Conservatively mark bpf main return value as 64-bit. (Alexei)
- Make sure read flag is either READ64 or READ32, not the mix of both.
(Alexei)
- Merged patch 1 and 2 in v4. (Alexei)
- Fixed kbuild test robot warning on NFP. (kbuild)
- Proposed new BPF_ZEXT insn to have optimal code-gen for various JIT
back-ends.
- Conservately set zext flags for patched-insn.
- Fixed return value zext for helper function calls.
- Also adjusted test_verifier scalability unit test to avoid triggerring
too many insn patch which will hang computer.
- re-tested on x86 host with llvm 9.0, no regression on test_verifier,
test_progs, test_progs_32.
- re-tested offload target (nfp), no regression on local testsuite.
v4:
- added the two missing fixes which addresses two Jakub's reviewes in v3.
- rebase on top of bpf-next.
v3:
- remove redundant check in "propagate_liveness_reg". (Jakub)
- add extra check in "mark_reg_read" to prune more search. (Jakub)
- re-implemented "prog_flags" passing mechanism, removed use of
global switch inside libbpf.
- enabled high 32-bit randomization beyond "test_verifier" and
"test_progs". Now it should have been enabled for all possible
tests. Re-run all tests, haven't noticed regression.
- remove RFC tag.
v2:
- rebased on top of bpf-next master.
- added comments for what is sub-register def index. (Edward, Alexei)
- removed patch 1 which turns bit mask from enum to macro. (Alexei)
- removed sysctl/bpf_jit_32bit_opt. (Alexei)
- merged sub-register def insn index into reg state. (Alexei)
- change test methodology (Alexei):
+ instead of simple unit tests on x86_64 for which this optimization
doesn't enabled due to there is hardware support, poison high
32-bit for whose def identified as safe to do so. this could let
the correctness of this patch set checked when daily bpf selftest
ran which delivers very stressful test on host machine like x86_64.
+ hi32 poisoning is gated by a new BPF_F_TEST_RND_HI32 prog flags.
+ BPF_F_TEST_RND_HI32 is enabled for all tests of "test_progs" and
"test_verifier", the latter needs minor tweak on two unit tests,
please see the patch for the change.
+ introduced a new global variable "libbpf_test_mode" into libbpf.
once it is set to true, it will set BPF_F_TEST_RND_HI32 for all the
later PROG_LOAD syscall, the goal is to easy the enable of hi32
poison on exsiting testsuite.
we could also introduce new APIs, for example "bpf_prog_test_load",
then use -Dbpf_prog_load=bpf_prog_test_load to migrate tests under
test_progs, but there are several load APIs, and such new API need
some change on struture like "struct bpf_prog_load_attr".
+ removed old unit tests. it is based on insn scan and requires quite
a few test_verifier generic code change. given hi32 randomization
could offer good test coverage, the unit tests doesn't add much
extra test value.
- enhanced register width check ("is_reg64") when record sub-register
write, now, it returns more accurate width.
- Re-run all tests under "test_progs" and "test_verifier" on x86_64, no
regression. Fixed a couple of bugs exposed:
1. ctx field size transformation was not taken into account.
2. insn patch could cause lost of original aux data which is
important for ctx field conversion.
3. return value for propagate_liveness was wrong and caused
regression on processed insn number.
4. helper call arg wasn't handled properly that path prune may cause
64-bit read info in pruned path lost.
- Re-run Cilium bpf prog for processed-insn-number benchmarking, no
regression.
v1:
- Fixed the missing handling on callee-saved for bpf-to-bpf call,
sub-register defs therefore moved to frame state. (Jakub Kicinski)
- Removed redundant "cross_reg". (Jakub Kicinski)
- Various coding styles & grammar fixes. (Jakub Kicinski, Quentin Monnet)
eBPF ISA specification requires high 32-bit cleared when low 32-bit
sub-register is written. This applies to destination register of ALU32 etc.
JIT back-ends must guarantee this semantic when doing code-gen. x86_64 and
AArch64 ISA has the same semantics, so the corresponding JIT back-end
doesn't need to do extra work.
However, 32-bit arches (arm, x86, nfp etc.) and some other 64-bit arches
(PowerPC, SPARC etc) need to do explicit zero extension to meet this
requirement, otherwise code like the following will fail.
u64_value = (u64) u32_value
... other uses of u64_value
This is because compiler could exploit the semantic described above and
save those zero extensions for extending u32_value to u64_value, these JIT
back-ends are expected to guarantee this through inserting extra zero
extensions which however could be a significant increase on the code size.
Some benchmarks show there could be ~40% sub-register writes out of total
insns, meaning at least ~40% extra code-gen.
One observation is these extra zero extensions are not always necessary.
Take above code snippet for example, it is possible u32_value will never be
casted into a u64, the value of high 32-bit of u32_value then could be
ignored and extra zero extension could be eliminated.
This patch implements this idea, insns defining sub-registers will be
marked when the high 32-bit of the defined sub-register matters. For
those unmarked insns, it is safe to eliminate high 32-bit clearnace for
them.
Algo
====
We could use insn scan based static analysis to tell whether one
sub-register def doesn't need zero extension. However, using such static
analysis, we must do conservative assumption at branching point where
multiple uses could be introduced. So, for any sub-register def that is
active at branching point, we need to mark it as needing zero extension.
This could introducing quite a few false alarms, for example ~25% on
Cilium bpf_lxc.
It will be far better to use dynamic data-flow tracing which verifier
fortunately already has and could be easily extend to serve the purpose of
this patch set.
- Split read flags into READ32 and READ64.
- Record index of insn that does sub-register write. Keep the index inside
reg state and update it during verifier insn walking.
- A full register read on a sub-register marks its definition insn as
needing zero extension on dst register.
A new sub-register write overrides the old one.
- When propagating read64 during path pruning, also mark any insn defining
a sub-register that is read in the pruned path as full-register.
Benchmark
=========
- I estimate the JITed image could be 10% ~ 30% smaller on these affected
arches (nfp, arm, x32, risv, ppc, sparc, s390), depending on the prog.
- For Cilium bpf_lxc, there is ~11500 insns in the compiled binary (use
latest LLVM snapshot, and with -mcpu=v3 -mattr=+alu32 enabled), 4460 of
them has sub-register writes (~40%). Calculated by:
cat dump | grep -P "\tw" | wc -l (ALU32)
cat dump | grep -P "r.*=.*u32" | wc -l (READ_W)
cat dump | grep -P "r.*=.*u16" | wc -l (READ_H)
cat dump | grep -P "r.*=.*u8" | wc -l (READ_B)
After this patch set enabled, > 25% of those 4460 could be identified as
doesn't needing zero extension on the destination, and the percentage
could go further up to more than 50% with some follow up optimizations
based on the infrastructure offered by this set. This leads to
significant save on JITed image.
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
This patch eliminate zero extension code-gen for instructions including
both alu and load/store. The only exception is for ctx load, because
offload target doesn't go through host ctx convert logic so we do
customized load and ignores zext flag set by verifier.
Cc: Jakub Kicinski <jakub.kicinski@netronome.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Cc: Björn Töpel <bjorn.topel@gmail.com>
Acked-by: Björn Töpel <bjorn.topel@gmail.com>
Tested-by: Björn Töpel <bjorn.topel@gmail.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The previous libbpf patch allows user to specify "prog_flags" to bpf
program load APIs. To enable high 32-bit randomization for a test, we need
to set BPF_F_TEST_RND_HI32 in "prog_flags".
To enable such randomization for all tests, we need to make sure all places
are passing BPF_F_TEST_RND_HI32. Changing them one by one is not
convenient, also, it would be better if a test could be switched to
"normal" running mode without code change.
Given the program load APIs used across bpf selftests are mostly:
bpf_prog_load: load from file
bpf_load_program: load from raw insns
A test_stub.c is implemented for bpf seltests, it offers two functions for
testing purpose:
bpf_prog_test_load
bpf_test_load_program
The are the same as "bpf_prog_load" and "bpf_load_program", except they
also set BPF_F_TEST_RND_HI32. Given *_xattr functions are the APIs to
customize any "prog_flags", it makes little sense to put these two
functions into libbpf.
Then, the following CFLAGS are passed to compilations for host programs:
-Dbpf_prog_load=bpf_prog_test_load
-Dbpf_load_program=bpf_test_load_program
They migrate the used load APIs to the test version, hence enable high
32-bit randomization for these tests without changing source code.
Besides all these, there are several testcases are using
"bpf_prog_load_attr" directly, their call sites are updated to pass
BPF_F_TEST_RND_HI32.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
- bpf_fill_ld_abs_vlan_push_pop:
Prevent zext happens inside PUSH_CNT loop. This could happen because
of BPF_LD_ABS (32-bit def) + BPF_JMP (64-bit use), or BPF_LD_ABS +
EXIT (64-bit use of R0). So, change BPF_JMP to BPF_JMP32 and redefine
R0 at exit path to cut off the data-flow from inside the loop.
- bpf_fill_jump_around_ld_abs:
Jump range is limited to 16 bit. every ld_abs is replaced by 6 insns,
but on arches like arm, ppc etc, there will be one BPF_ZEXT inserted
to extend the error value of the inlined ld_abs sequence which then
contains 7 insns. so, set the dividend to 7 so the testcase could
work on all arches.
- bpf_fill_scale1/bpf_fill_scale2:
Both contains ~1M BPF_ALU32_IMM which will trigger ~1M insn patcher
call because of hi32 randomization later when BPF_F_TEST_RND_HI32 is
set for bpf selftests. Insn patcher is not efficient that 1M call to
it will hang computer. So , change to BPF_ALU64_IMM to avoid hi32
randomization.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
libbpf doesn't allow passing "prog_flags" during bpf program load in a
couple of load related APIs, "bpf_load_program_xattr", "load_program" and
"bpf_prog_load_xattr".
It makes sense to allow passing "prog_flags" which is useful for
customizing program loading.
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
This patch randomizes high 32-bit of a definition when BPF_F_TEST_RND_HI32
is set.
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Sync new bpf prog load flag "BPF_F_TEST_RND_HI32" to tools/.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
x86_64 and AArch64 perhaps are two arches that running bpf testsuite
frequently, however the zero extension insertion pass is not enabled for
them because of their hardware support.
It is critical to guarantee the pass correction as it is supposed to be
enabled at default for a couple of other arches, for example PowerPC,
SPARC, arm, NFP etc. Therefore, it would be very useful if there is a way
to test this pass on for example x86_64.
The test methodology employed by this set is "poisoning" useless bits. High
32-bit of a definition is randomized if it is identified as not used by any
later insn. Such randomization is only enabled under testing mode which is
gated by the new bpf prog load flags "BPF_F_TEST_RND_HI32".
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
After previous patches, verifier will mark a insn if it really needs zero
extension on dst_reg.
It is then for back-ends to decide how to use such information to eliminate
unnecessary zero extension code-gen during JIT compilation.
One approach is verifier insert explicit zero extension for those insns
that need zero extension in a generic way, JIT back-ends then do not
generate zero extension for sub-register write at default.
However, only those back-ends which do not have hardware zero extension
want this optimization. Back-ends like x86_64 and AArch64 have hardware
zero extension support that the insertion should be disabled.
This patch introduces new target hook "bpf_jit_needs_zext" which returns
false at default, meaning verifier zero extension insertion is disabled at
default. A back-end could override this hook to return true if it doesn't
have hardware support and want verifier insert zero extension explicitly.
Offload targets do not use this native target hook, instead, they could
get the optimization results using bpf_prog_offload_ops.finalize.
NOTE: arches could have diversified features, it is possible for one arch
to have hardware zero extension support for some sub-register write insns
but not for all. For example, PowerPC, SPARC have zero extended loads, but
not for alu32. So when verifier zero extension insertion enabled, these JIT
back-ends need to peephole insns to remove those zero extension inserted
for insn that actually has hardware zero extension support. The peephole
could be as simple as looking the next insn, if it is a special zero
extension insn then it is safe to eliminate it if the current insn has
hardware zero extension support.
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The encoding for this new variant is based on BPF_X format. "imm" field was
0 only, now it could be 1 which means doing zero extension unconditionally
.code = BPF_ALU | BPF_MOV | BPF_X
.dst_reg = DST
.src_reg = SRC
.imm = 1
We use this new form for doing zero extension for which verifier will
guarantee SRC == DST.
Implications on JIT back-ends when doing code-gen for
BPF_ALU | BPF_MOV | BPF_X:
1. No change if hardware already does zero extension unconditionally for
sub-register write.
2. Otherwise, when seeing imm == 1, just generate insns to clear high
32-bit. No need to generate insns for the move because when imm == 1,
dst_reg is the same as src_reg at the moment.
Interpreter doesn't need change as well. It is doing unconditionally zero
extension for mov32 already.
One helper macro BPF_ZEXT_REG is added to help creating zero extension
insn using this new mov32 variant.
One helper function insn_is_zext is added for checking one insn is an
zero extension on dst. This will be widely used by a few JIT back-ends in
later patches in this set.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Patched insns do not go through generic verification, therefore doesn't has
zero extension information collected during insn walking.
We don't bother analyze them at the moment, for any sub-register def comes
from them, just conservatively mark it as needing zero extension.
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
eBPF ISA specification requires high 32-bit cleared when low 32-bit
sub-register is written. This applies to destination register of ALU32 etc.
JIT back-ends must guarantee this semantic when doing code-gen. x86_64 and
AArch64 ISA has the same semantics, so the corresponding JIT back-end
doesn't need to do extra work.
However, 32-bit arches (arm, x86, nfp etc.) and some other 64-bit arches
(PowerPC, SPARC etc) need to do explicit zero extension to meet this
requirement, otherwise code like the following will fail.
u64_value = (u64) u32_value
... other uses of u64_value
This is because compiler could exploit the semantic described above and
save those zero extensions for extending u32_value to u64_value, these JIT
back-ends are expected to guarantee this through inserting extra zero
extensions which however could be a significant increase on the code size.
Some benchmarks show there could be ~40% sub-register writes out of total
insns, meaning at least ~40% extra code-gen.
One observation is these extra zero extensions are not always necessary.
Take above code snippet for example, it is possible u32_value will never be
casted into a u64, the value of high 32-bit of u32_value then could be
ignored and extra zero extension could be eliminated.
This patch implements this idea, insns defining sub-registers will be
marked when the high 32-bit of the defined sub-register matters. For
those unmarked insns, it is safe to eliminate high 32-bit clearnace for
them.
Algo:
- Split read flags into READ32 and READ64.
- Record index of insn that does sub-register write. Keep the index inside
reg state and update it during verifier insn walking.
- A full register read on a sub-register marks its definition insn as
needing zero extension on dst register.
A new sub-register write overrides the old one.
- When propagating read64 during path pruning, also mark any insn defining
a sub-register that is read in the pruned path as full-register.
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Jiong Wang <jiong.wang@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Yonghong Song says:
====================
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces
through kernel function get_perf_callchain()
when certain events happens (e.g., cache miss counter or
cpu clock counter overflows). But such stack traces are
not enough for jitted programs, e.g., hhvm (jited php).
To get real stack trace, jit engine internal data structures
need to be traversed in order to get the real user functions.
bpf program itself may not be the best place to traverse
the jit engine as the traversing logic could be complex and
it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which
it can dump stack traces. When it receives a signal, it will
dump the stack in next such program location.
This patch implements bpf_send_signal() helper to send
a signal to hhvm in real time, resulting in intended stack traces.
Patch #1 implemented the bpf_send_helper() in the kernel.
Patch #2 synced uapi header bpf.h to tools directory.
Patch #3 added a self test which covers tracepoint
and perf_event bpf programs.
Changelogs:
v4 => v5:
. pass the "current" task struct to irq_work as well
since the current task struct may change between
nmi and subsequent irq_work_interrupt.
Discovered by Daniel.
v3 => v4:
. fix one typo and declare "const char *id_path = ..."
to avoid directly use the long string in the func body
in Patch #3.
v2 => v3:
. change the standalone test to be part of prog_tests.
RFC v1 => v2:
. previous version allows to send signal to an arbitrary
pid. This version just sends the signal to current
task to avoid unstable pid and potential races between
sending signals and task state changes for the pid.
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
The bpf uapi header include/uapi/linux/bpf.h is sync'ed
to tools/include/uapi/linux/bpf.h.
Acked-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
This patch tries to solve the following specific use case.
Currently, bpf program can already collect stack traces
through kernel function get_perf_callchain()
when certain events happens (e.g., cache miss counter or
cpu clock counter overflows). But such stack traces are
not enough for jitted programs, e.g., hhvm (jited php).
To get real stack trace, jit engine internal data structures
need to be traversed in order to get the real user functions.
bpf program itself may not be the best place to traverse
the jit engine as the traversing logic could be complex and
it is not a stable interface either.
Instead, hhvm implements a signal handler,
e.g. for SIGALARM, and a set of program locations which
it can dump stack traces. When it receives a signal, it will
dump the stack in next such program location.
Such a mechanism can be implemented in the following way:
. a perf ring buffer is created between bpf program
and tracing app.
. once a particular event happens, bpf program writes
to the ring buffer and the tracing app gets notified.
. the tracing app sends a signal SIGALARM to the hhvm.
But this method could have large delays and causing profiling
results skewed.
This patch implements bpf_send_signal() helper to send
a signal to hhvm in real time, resulting in intended stack traces.
Acked-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Yonghong Song <yhs@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Andrii Nakryiko says:
====================
This patch set adds BTF-to-C dumping APIs to libbpf, allowing to output
a subset of BTF types as a compilable C type definitions. This is useful by
itself, as raw BTF output is not easy to inspect and comprehend. But it's also
a big part of BPF CO-RE (compile once - run everywhere) initiative aimed at
allowing to write relocatable BPF programs, that won't require on-the-host
kernel headers (and would be able to inspect internal kernel structures, not
exposed through kernel headers).
This patch set consists of three groups of patches and one pre-patch, with the
BTF-to-C dumper API depending on the first two groups.
Pre-patch #1 fixes issue with libbpf_internal.h.
btf__parse_elf() API patches:
- patch #2 adds btf__parse_elf() API to libbpf, allowing to load BTF and/or
BTF.ext from ELF file;
- patch #3 utilizies btf__parse_elf() from bpftool for `btf dump file` command;
- patch #4 switches test_btf.c to use btf__parse_elf() to check for presence
of BTF data in object file.
libbpf's internal hashmap patches:
- patch #5 adds resizeable non-thread safe generic hashmap to libbpf;
- patch #6 adds tests for that hashmap;
- patch #7 migrates btf_dedup()'s dedup_table to use hashmap w/ APPEND.
BTF-to-C dumper API patches:
- patch #8 adds btf_dump APIs with all the logic for laying out type
definitions in correct order and emitting C syntax for them;
- patch #9 adds lots of tests for common and quirky parts of C type system;
- patch #10 adds support for C-syntax btf dumping to bpftool;
- patch #11 updates bpftool documentation to mention C-syntax dump option;
- patch #12 update bash-completion for btf dump sub-command.
v2->v3:
- fix bpftool-btf.rst formatting (Quentin);
- simplify bash autocompletion script (Quentin);
- better error message in btf dump (Quentin);
v1->v2:
- removed unuseful file header (Jakub);
- removed inlines in .c (Jakub);
- added 'format {c|raw}' keyword/option (Jakub);
- re-use i var for iteration in btf_dump_c() (Jakub);
- bumped libbpf version to 0.0.4;
v0->v1:
- fix bug in hashmap__for_each_bucket_entry() not handling empty hashmap;
- removed `btf dump`-specific libbpf logging hook up (Quentin has more generic
patchset);
- change btf__parse_elf() to always load .BTF and return it as a result, with
.BTF.ext being optional and returned through struct btf_ext** arg (Alexei);
- endianness check to use __BYTE_ORDER__ (Alexei);
- bool:1 to __u8:1 in type_aux_state (Alexei);
- added HASHMAP_APPEND strategy to hashmap, changed
hashmap__for_each_key_entry() to also check for key equality during
iteration (multimap iteration for key);
- added new tests for empty hashmap and hashmap as a multimap;
- tried to clarify weak/strong dependency ordering comments (Alexei)
- btf dump test's expected output - support better commenting aproach (Alexei);
- added bash-completion for a new "c" option (Alexei).
====================
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Utilize new libbpf's btf_dump API to emit BTF as a C definitions.
Acked-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Reviewed-by: Quentin Monnet <quentin.monnet@netronome.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Add new test_btf_dump set of tests, validating BTF-to-C conversion
correctness. Tests rely on clang to generate BTF from provided C test
cases.
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
BTF contains enough type information to allow generating valid
compilable C header w/ correct layout of structs/unions and all the
typedef/enum definitions. This patch adds a new "object" - btf_dump to
facilitate dumping BTF as valid C. btf_dump__dump_type() is the main API
which takes care of dumping out (through user-provided printf-like
callback function) C definitions for given type ID and it's required
dependencies. This allows for not just dumping out entirety of BTF types,
but also selective filtering based on user-provided criterias w/ minimal
set of dependent types.
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
There is a need for fast point lookups inside libbpf for multiple use
cases (e.g., name resolution for BTF-to-C conversion, by-name lookups in
BTF for upcoming BPF CO-RE relocation support, etc). This patch
implements simple resizable non-thread safe hashmap using single linked
list chains.
Four different insert strategies are supported:
- HASHMAP_ADD - only add key/value if key doesn't exist yet;
- HASHMAP_SET - add key/value pair if key doesn't exist yet; otherwise,
update value;
- HASHMAP_UPDATE - update value, if key already exists; otherwise, do
nothing and return -ENOENT;
- HASHMAP_APPEND - always add key/value pair, even if key already exists.
This turns hashmap into a multimap by allowing multiple values to be
associated with the same key. Most useful read API for such hashmap is
hashmap__for_each_key_entry() iteration. If hashmap__find() is still
used, it will return last inserted key/value entry (first in a bucket
chain).
For HASHMAP_SET and HASHMAP_UPDATE, old key/value pair is returned, so
that calling code can handle proper memory management, if necessary.
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Switch test_btf.c to rely on btf__parse_elf to check presence of BTF and
BTF.ext data, instead of implementing its own ELF parsing.
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Loading BTF and BTF.ext from ELF file is a common need. Instead of
requiring every user to re-implement it, let's provide this API from
libbpf itself. It's mostly copy/paste from `bpftool btf dump`
implementation, which will be switched to libbpf's version in next
patch. btf__parse_elf allows to load BTF and optionally BTF.ext.
This is also useful for tests that need to load/work with BTF, loaded
from test ELF files.
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
libbpf_internal.h expects a bunch of stuff defined in libbpf.h to be
defined. This patch makes sure that libbpf.h is always included.
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
The bpf_printk macro was moved to bpf_helpers.h which is included in all
example programs.
Signed-off-by: Michal Rostecki <mrostecki@opensuse.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
bpf_printk is a macro which is commonly used to print out debug messages
in BPF programs and it was copied in many selftests and samples. Since
all of them include bpf_helpers.h, this change moves the macro there.
Signed-off-by: Michal Rostecki <mrostecki@opensuse.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Alexei Starovoitov says:
====================
Convert explored_states array into hash table and use simple hash
to reduce verifier peak memory consumption for programs with bpf2bpf
calls. More details in patch 3.
v1->v2: fixed Jakub's small nit in patch 1
====================
Acked-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
All prune points inside a callee bpf function most likely will have
different callsites. For example, if function foo() is called from
two callsites the half of explored states in all prune points in foo()
will be useless for subsequent walking of one of those callsites.
Fortunately explored_states pruning heuristics keeps the number of states
per prune point small, but walking these states is still a waste of cpu
time when the callsite of the current state is different from the callsite
of the explored state.
To improve pruning logic convert explored_states into hash table and
use simple insn_idx ^ callsite hash to select hash bucket.
This optimization has no effect on programs without bpf2bpf calls
and drastically improves programs with calls.
In the later case it reduces total memory consumption in 1M scale tests
by almost 3 times (peak_states drops from 5752 to 2016).
Care should be taken when comparing the states for equivalency.
Since the same hash bucket can now contain states with different indices
the insn_idx has to be part of verifier_state and compared.
Different hash table sizes and different hash functions were explored,
but the results were not significantly better vs this patch.
They can be improved in the future.
Hit/miss heuristic is not counting index miscompare as a miss.
Otherwise verifier stats become unstable when experimenting
with different hash functions.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
split explored_states into prune_point boolean mark
and link list of explored states.
This removes STATE_LIST_MARK hack and allows marks to be separate from states.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
clean up explored_states to prep for introduction of hashtable
No functional changes.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Alexei Starovoitov says:
====================
Patch 1 - jmp sequence limit
Patch 2 - improve existing tests
Patch 3 - add pyperf-based realistic bpf program that takes
advantage of higher limit and use it as a stress test
v1->v2: fixed nit in patch 3. added Andrii's acks
====================
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Add a snippet of pyperf bpf program used to collect python stack traces
as a scale test for the verifier.
At 189 loop iterations llvm 9.0 starts ignoring '#pragma unroll'
and generates partially unrolled loop instead.
Hence use 50, 100, and 180 loop iterations to stress test.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Adjust scale tests to check for new jmp sequence limit.
BPF_JGT had to be changed to BPF_JEQ because the verifier was
too smart. It tracked the known safe range of R0 values
and pruned the search earlier before hitting exact 8192 limit.
bpf_semi_rand_get() was too (un)?lucky.
k = 0; was missing in bpf_fill_scale2.
It was testing a bit shorter sequence of jumps than intended.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
The limit of 1024 subsequent jumps was causing otherwise valid
programs to be rejected. Bump it to 8192 and make the error more verbose.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
It's easy to have a mismatch of "intended to be public" vs really
exposed API functions. While Makefile does check for this mismatch, if
it actually occurs it's not trivial to determine which functions are
accidentally exposed. This patch dumps out a diff showing what's not
supposed to be exposed facilitating easier fixing.
Signed-off-by: Andrii Nakryiko <andriin@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Currently, the hv_sock send() iterates once over the buffer, puts data into
the VMBUS channel and returns. It doesn't maximize on the case when there
is a simultaneous reader draining data from the channel. In such a case,
the send() can maximize the bandwidth (and consequently minimize the cpu
cycles) by iterating until the channel is found to be full.
Perf data:
Total Data Transfer: 10GB/iteration
Single threaded reader/writer, Linux hvsocket writer with Windows hvsocket
reader
Packet size: 64KB
CPU sys time was captured using the 'time' command for the writer to send
10GB of data.
'Send Buffer Loop' is with the patch applied.
The values below are over 10 iterations.
|--------------------------------------------------------|
| | Current | Send Buffer Loop |
|--------------------------------------------------------|
| | Throughput | CPU sys | Throughput | CPU sys |
| | (MB/s) | time (s) | (MB/s) | time (s) |
|--------------------------------------------------------|
| Min | 407 | 7.048 | 401 | 5.958 |
|--------------------------------------------------------|
| Max | 455 | 7.563 | 542 | 6.993 |
|--------------------------------------------------------|
| Avg | 440 | 7.411 | 451 | 6.639 |
|--------------------------------------------------------|
| Median | 446 | 7.417 | 447 | 6.761 |
|--------------------------------------------------------|
Observation:
1. The avg throughput doesn't really change much with this change for this
scenario. This is most probably because the bottleneck on throughput is
somewhere else.
2. The average system (or kernel) cpu time goes down by 10%+ with this
change, for the same amount of data transfer.
Signed-off-by: Sunil Muthuswamy <sunilmut@microsoft.com>
Reviewed-by: Dexuan Cui <decui@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the hv_sock buffer size is static and can't scale to the
bandwidth requirements of the application. This change allows the
applications to influence the socket buffer sizes using the SO_SNDBUF and
the SO_RCVBUF socket options.
Few interesting points to note:
1. Since the VMBUS does not allow a resize operation of the ring size, the
socket buffer size option should be set prior to establishing the
connection for it to take effect.
2. Setting the socket option comes with the cost of that much memory being
reserved/allocated by the kernel, for the lifetime of the connection.
Perf data:
Total Data Transfer: 1GB
Single threaded reader/writer
Results below are summarized over 10 iterations.
Linux hvsocket writer + Windows hvsocket reader:
|---------------------------------------------------------------------------------------------|
|Packet size -> | 128B | 1KB | 4KB | 64KB |
|---------------------------------------------------------------------------------------------|
|SO_SNDBUF size | | Throughput in MB/s (min/max/avg/median): |
| v | |
|---------------------------------------------------------------------------------------------|
| Default | 109/118/114/116 | 636/774/701/700 | 435/507/480/476 | 410/491/462/470 |
| 16KB | 110/116/112/111 | 575/705/662/671 | 749/900/854/869 | 592/824/692/676 |
| 32KB | 108/120/115/115 | 703/823/767/772 | 718/878/850/866 | 1593/2124/2000/2085 |
| 64KB | 108/119/114/114 | 592/732/683/688 | 805/934/903/911 | 1784/1943/1862/1843 |
|---------------------------------------------------------------------------------------------|
Windows hvsocket writer + Linux hvsocket reader:
|---------------------------------------------------------------------------------------------|
|Packet size -> | 128B | 1KB | 4KB | 64KB |
|---------------------------------------------------------------------------------------------|
|SO_RCVBUF size | | Throughput in MB/s (min/max/avg/median): |
| v | |
|---------------------------------------------------------------------------------------------|
| Default | 69/82/75/73 | 313/343/333/336 | 418/477/446/445 | 659/701/676/678 |
| 16KB | 69/83/76/77 | 350/401/375/382 | 506/548/517/516 | 602/624/615/615 |
| 32KB | 62/83/73/73 | 471/529/496/494 | 830/1046/935/939 | 944/1180/1070/1100 |
| 64KB | 64/70/68/69 | 467/533/501/497 | 1260/1590/1430/1431 | 1605/1819/1670/1660 |
|---------------------------------------------------------------------------------------------|
Signed-off-by: Sunil Muthuswamy <sunilmut@microsoft.com>
Reviewed-by: Dexuan Cui <decui@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Removing two 4 bytes holes allows to use kmalloc-32
kmem cache instead of kmalloc-64 on 64bit kernels.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add tracepoint to __neigh_create to enable debugging of new entries.
Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Alexei Starovoitov