The arm64 JIT has the same issue as ppc64 JIT in that the relative BPF
to BPF call offset can be too far away from core kernel in that relative
encoding into imm is not sufficient and could potentially be truncated,
see also fd045f6cd9 ("arm64: add support for module PLTs") which adds
spill-over space for module_alloc() and therefore bpf_jit_binary_alloc().
Therefore, use the recently added bpf_jit_get_func_addr() helper for
properly fetching the address through prog->aux->func[off]->bpf_func
instead. This also has the benefit to optimize normal helper calls since
their address can use the optimized emission. Tested on Cavium ThunderX
CN8890.
Fixes: db496944fd ("bpf: arm64: add JIT support for multi-function programs")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
We can trivially save 4 bytes in prologue for cBPF since tail calls
can never be used from there. The register push/pop is pairwise,
here, x25 (fp) and x26 (tcc), so no point in changing that, only
reset to zero is not needed.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Improve the JIT to emit 64 and 32 bit immediates, the current
algorithm is not optimal and we often emit more instructions
than actually needed. arm64 has movz, movn, movk variants but
for the current 64 bit immediates we only use movz with a
series of movk when needed.
For example loading ffffffffffffabab emits the following 4
instructions in the JIT today:
* movz: abab, shift: 0, result: 000000000000abab
* movk: ffff, shift: 16, result: 00000000ffffabab
* movk: ffff, shift: 32, result: 0000ffffffffabab
* movk: ffff, shift: 48, result: ffffffffffffabab
Whereas after the patch the same load only needs a single
instruction:
* movn: 5454, shift: 0, result: ffffffffffffabab
Another example where two extra instructions can be saved:
* movz: abab, shift: 0, result: 000000000000abab
* movk: 1f2f, shift: 16, result: 000000001f2fabab
* movk: ffff, shift: 32, result: 0000ffff1f2fabab
* movk: ffff, shift: 48, result: ffffffff1f2fabab
After the patch:
* movn: e0d0, shift: 16, result: ffffffff1f2fffff
* movk: abab, shift: 0, result: ffffffff1f2fabab
Another example with movz, before:
* movz: 0000, shift: 0, result: 0000000000000000
* movk: fea0, shift: 32, result: 0000fea000000000
After:
* movz: fea0, shift: 32, result: 0000fea000000000
Moreover, reuse emit_a64_mov_i() for 32 bit immediates that
are loaded via emit_a64_mov_i64() which is a similar optimization
as done in 6fe8b9c1f4 ("bpf, x64: save several bytes by using
mov over movabsq when possible"). On arm64, the latter allows to
use a single instruction with movn due to zero extension where
otherwise two would be needed. And last but not least add a
missing optimization in emit_a64_mov_i() where movn is used but
the subsequent movk not needed. With some of the Cilium programs
in use, this shrinks the needed instructions by about three
percent. Tested on Cavium ThunderX CN8890.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Follow-up to 816d9ef32a ("bpf, arm64: remove ld_abs/ld_ind") in
that the extra 4 byte JIT scratchpad is not needed anymore since it
was in ld_abs/ld_ind as stack buffer for bpf_load_pointer().
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Since LD_ABS/LD_IND instructions are now removed from the core and
reimplemented through a combination of inlined BPF instructions and
a slow-path helper, we can get rid of the complexity from arm64 JIT.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
I recently noticed a crash on arm64 when feeding a bogus index
into BPF tail call helper. The crash would not occur when the
interpreter is used, but only in case of JIT. Output looks as
follows:
[ 347.007486] Unable to handle kernel paging request at virtual address fffb850e96492510
[...]
[ 347.043065] [fffb850e96492510] address between user and kernel address ranges
[ 347.050205] Internal error: Oops: 96000004 [#1] SMP
[...]
[ 347.190829] x13: 0000000000000000 x12: 0000000000000000
[ 347.196128] x11: fffc047ebe782800 x10: ffff808fd7d0fd10
[ 347.201427] x9 : 0000000000000000 x8 : 0000000000000000
[ 347.206726] x7 : 0000000000000000 x6 : 001c991738000000
[ 347.212025] x5 : 0000000000000018 x4 : 000000000000ba5a
[ 347.217325] x3 : 00000000000329c4 x2 : ffff808fd7cf0500
[ 347.222625] x1 : ffff808fd7d0fc00 x0 : ffff808fd7cf0500
[ 347.227926] Process test_verifier (pid: 4548, stack limit = 0x000000007467fa61)
[ 347.235221] Call trace:
[ 347.237656] 0xffff000002f3a4fc
[ 347.240784] bpf_test_run+0x78/0xf8
[ 347.244260] bpf_prog_test_run_skb+0x148/0x230
[ 347.248694] SyS_bpf+0x77c/0x1110
[ 347.251999] el0_svc_naked+0x30/0x34
[ 347.255564] Code: 9100075a d280220a 8b0a002a d37df04b (f86b694b)
[...]
In this case the index used in BPF r3 is the same as in r1
at the time of the call, meaning we fed a pointer as index;
here, it had the value 0xffff808fd7cf0500 which sits in x2.
While I found tail calls to be working in general (also for
hitting the error cases), I noticed the following in the code
emission:
# bpftool p d j i 988
[...]
38: ldr w10, [x1,x10]
3c: cmp w2, w10
40: b.ge 0x000000000000007c <-- signed cmp
44: mov x10, #0x20 // #32
48: cmp x26, x10
4c: b.gt 0x000000000000007c
50: add x26, x26, #0x1
54: mov x10, #0x110 // #272
58: add x10, x1, x10
5c: lsl x11, x2, #3
60: ldr x11, [x10,x11] <-- faulting insn (f86b694b)
64: cbz x11, 0x000000000000007c
[...]
Meaning, the tests passed because commit ddb55992b0 ("arm64:
bpf: implement bpf_tail_call() helper") was using signed compares
instead of unsigned which as a result had the test wrongly passing.
Change this but also the tail call count test both into unsigned
and cap the index as u32. Latter we did as well in 90caccdd8c
("bpf: fix bpf_tail_call() x64 JIT") and is needed in addition here,
too. Tested on HiSilicon Hi1616.
Result after patch:
# bpftool p d j i 268
[...]
38: ldr w10, [x1,x10]
3c: add w2, w2, #0x0
40: cmp w2, w10
44: b.cs 0x0000000000000080
48: mov x10, #0x20 // #32
4c: cmp x26, x10
50: b.hi 0x0000000000000080
54: add x26, x26, #0x1
58: mov x10, #0x110 // #272
5c: add x10, x1, x10
60: lsl x11, x2, #3
64: ldr x11, [x10,x11]
68: cbz x11, 0x0000000000000080
[...]
Fixes: ddb55992b0 ("arm64: bpf: implement bpf_tail_call() helper")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Since we've changed div/mod exception handling for src_reg in
eBPF verifier itself, remove the leftovers from arm64 JIT.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Alexei Starovoitov says:
====================
pull-request: bpf-next 2018-01-19
The following pull-request contains BPF updates for your *net-next* tree.
The main changes are:
1) bpf array map HW offload, from Jakub.
2) support for bpf_get_next_key() for LPM map, from Yonghong.
3) test_verifier now runs loaded programs, from Alexei.
4) xdp cpumap monitoring, from Jesper.
5) variety of tests, cleanups and small x64 JIT optimization, from Daniel.
6) user space can now retrieve HW JITed program, from Jiong.
Note there is a minor conflict between Russell's arm32 JIT fixes
and removal of bpf_jit_enable variable by Daniel which should
be resolved by keeping Russell's comment and removing that variable.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
The BPF verifier conflict was some minor contextual issue.
The TUN conflict was less trivial. Cong Wang fixed a memory leak of
tfile->tx_array in 'net'. This is an skb_array. But meanwhile in
net-next tun changed tfile->tx_arry into tfile->tx_ring which is a
ptr_ring.
Signed-off-by: David S. Miller <davem@davemloft.net>
Having a pure_initcall() callback just to permanently enable BPF
JITs under CONFIG_BPF_JIT_ALWAYS_ON is unnecessary and could leave
a small race window in future where JIT is still disabled on boot.
Since we know about the setting at compilation time anyway, just
initialize it properly there. Also consolidate all the individual
bpf_jit_enable variables into a single one and move them under one
location. Moreover, don't allow for setting unspecified garbage
values on them.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Using dynamic stack_depth tracking in arm64 JIT is currently broken in
combination with tail calls. In prologue, we cache ctx->stack_size and
adjust SP reg for setting up function call stack, and tearing it down
again in epilogue. Problem is that when doing a tail call, the cached
ctx->stack_size might not be the same.
One way to fix the problem with minimal overhead is to re-adjust SP in
emit_bpf_tail_call() and properly adjust it to the current program's
ctx->stack_size. Tested on Cavium ThunderX ARMv8.
Fixes: f1c9eed7f4 ("bpf, arm64: take advantage of stack_depth tracking")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
fix the following issue:
arch/arm64/net/bpf_jit_comp.c: In function 'bpf_int_jit_compile':
arch/arm64/net/bpf_jit_comp.c:982:18: error: 'image_size' may be used
uninitialized in this function [-Werror=maybe-uninitialized]
Fixes: db496944fd ("bpf: arm64: add JIT support for multi-function programs")
Reported-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
similar to x64 add support for bpf-to-bpf calls.
When program has calls to in-kernel helpers the target call offset
is known at JIT time and arm64 architecture needs 2 passes.
With bpf-to-bpf calls the dynamically allocated function start
is unknown until all functions of the program are JITed.
Therefore (just like x64) arm64 JIT needs one extra pass over
the program to emit correct call offsets.
Implementation detail:
Avoid being too clever in 64-bit immediate moves and
always use 4 instructions (instead of 3-4 depending on the address)
to make sure only one extra pass is needed.
If some future optimization would make it worth while to optimize
'call 64-bit imm' further, the JIT would need to do 4 passes
over the program instead of 3 as in this patch.
For typical bpf program address the mov needs 3 or 4 insns,
so unconditional 4 insns to save extra pass is a worthy trade off
at this state of JIT.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
global bpf_jit_enable variable is tested multiple times in JITs,
blinding and verifier core. The malicious root can try to toggle
it while loading the programs. This race condition was accounted
for and there should be no issues, but it's safer to avoid
this race condition.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
This work implements jiting of BPF_J{LT,LE,SLT,SLE} instructions
with BPF_X/BPF_K variants for the arm64 eBPF JIT.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
struct jit_ctx::image is used the store a pointer to the jitted
intructions, which are always little-endian. These instructions
are thus correctly converted from native order to little-endian
before being stored but the pointer 'image' is declared as for
native order values.
Fix this by declaring the field as __le32* instead of u32*.
Same for the pointer used in jit_fill_hole() to initialize
the image.
Signed-off-by: Luc Van Oostenryck <luc.vanoostenryck@gmail.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Make use of recently implemented stack_depth tracking for arm64 JIT,
so that stack usage can be reduced heavily for programs not using
tail calls at least.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Will reported that in BPF_XADD we must use a different register in stxr
instruction for the status flag due to otherwise CONSTRAINED UNPREDICTABLE
behavior per architecture. Reference manual says [1]:
If s == t, then one of the following behaviors must occur:
* The instruction is UNDEFINED.
* The instruction executes as a NOP.
* The instruction performs the store to the specified address, but
the value stored is UNKNOWN.
Thus, use a different temporary register for the status flag to fix it.
Disassembly extract from test 226/STX_XADD_DW from test_bpf.ko:
[...]
0000003c: c85f7d4b ldxr x11, [x10]
00000040: 8b07016b add x11, x11, x7
00000044: c80c7d4b stxr w12, x11, [x10]
00000048: 35ffffac cbnz w12, 0x0000003c
[...]
[1] https://static.docs.arm.com/ddi0487/b/DDI0487B_a_armv8_arm.pdf, p.6132
Fixes: 85f68fe898 ("bpf, arm64: implement jiting of BPF_XADD")
Reported-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add jited_len to struct bpf_prog. It will be
useful for the struct bpf_prog_info which will
be added in the later patch.
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Acked-by: Alexei Starovoitov <ast@fb.com>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
free up BPF_JMP | BPF_CALL | BPF_X opcode to be used by actual
indirect call by register and use kernel internal opcode to
mark call instruction into bpf_tail_call() helper.
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Shubham was recently asking on netdev why in arm64 JIT we don't multiply
the index for accessing the tail call map by 8. That led me into testing
out arm64 JIT wrt tail calls and it turned out I got a NULL pointer
dereference on the tail call.
The buggy access is at:
prog = array->ptrs[index];
if (prog == NULL)
goto out;
[...]
00000060: d2800e0a mov x10, #0x70 // #112
00000064: f86a682a ldr x10, [x1,x10]
00000068: f862694b ldr x11, [x10,x2]
0000006c: b40000ab cbz x11, 0x00000080
[...]
The code triggering the crash is f862694b. x1 at the time contains the
address of the bpf array, x10 offsetof(struct bpf_array, ptrs). Meaning,
above we load the pointer to the program at map slot 0 into x10. x10
can then be NULL if the slot is not occupied, which we later on try to
access with a user given offset in x2 that is the map index.
Fix this by emitting the following instead:
[...]
00000060: d2800e0a mov x10, #0x70 // #112
00000064: 8b0a002a add x10, x1, x10
00000068: d37df04b lsl x11, x2, #3
0000006c: f86b694b ldr x11, [x10,x11]
00000070: b40000ab cbz x11, 0x00000084
[...]
This basically adds the offset to ptrs to the base address of the bpf
array we got and we later on access the map with an index * 8 offset
relative to that. The tail call map itself is basically one large area
with meta data at the head followed by the array of prog pointers.
This makes tail calls working again, tested on Cavium ThunderX ARMv8.
Fixes: ddb55992b0 ("arm64: bpf: implement bpf_tail_call() helper")
Reported-by: Shubham Bansal <illusionist.neo@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
Daniel Borkmann
David S. Miller
Alexei Starovoitov
Linus Torvalds
Luc Van Oostenryck
Martin KaFai Lau
Laura Abbott