- 13 Jul, 2018 3 commits
-
-
Russell King authored
Improve the 64-bit store implementation from: ldr r6, [fp, #-8] str r8, [r6] ldr r6, [fp, #-8] mov r7, #4 add r7, r6, r7 str r9, [r7] to: ldr r6, [fp, #-8] str r8, [r6] str r9, [r6, #4] We leave the store as two separate STR instructions rather than using STRD as the store may not be aligned, and STR can handle misalignment. Signed-off-by:
Russell King <rmk+kernel@armlinux.org.uk> Signed-off-by:
Daniel Borkmann <daniel@iogearbox.net>
-
Russell King authored
Improve the 64-bit sign-extended immediate from: mov r6, #1 str r6, [fp, #-52] ; 0xffffffcc mov r6, #0 str r6, [fp, #-48] ; 0xffffffd0 to: mov r6, #1 mov r7, #0 strd r6, [fp, #-52] ; 0xffffffcc Signed-off-by:
-
Russell King authored
Rather than writing each 32-bit half of the 64-bit immediate value separately when the register is on the stack: movw r6, #45056 ; 0xb000 movt r6, #60979 ; 0xee33 str r6, [fp, #-44] ; 0xffffffd4 mov r6, #0 str r6, [fp, #-40] ; 0xffffffd8 arrange to use the double-word store when available instead: movw r6, #45056 ; 0xb000 movt r6, #60979 ; 0xee33 mov r7, #0 strd r6, [fp, #-44] ; 0xffffffd4 Signed-off-by:
-
- 12 Jul, 2018 19 commits
-
-
Daniel Borkmann authored
Russell King says: ==================== This series improves the ARM BPF JIT compiler by: - enumerating the stack layout rather than using constants that happen to be multiples of four - rejig the BPF "register" accesses to use negative numbers instead of positive, which could be confused with register numbers in the bpf2a32 array. - since we maintain the ARM FP register as a pointer to the top of our scratch space (or, with frame pointers enabled, a valid ARM frame pointer register), we can access our scratch space using FP, which is constant across all BPF programs, including tail-called programs. - use immediate forms of ARM instructions where possible, rather than first loading the immediate into an ARM register. - use load-with-shift instruction rather than seperate shift instruction followed by load - avoid reloading index and array in the tail-call code - use double-word load/store instructions where available Version 2: - Fix ARMv5 test pointed out by Olof - Fix build error found by 0-day (adding an additional patch) ==================== Signed-off-by: -
Russell King authored
Use double-word load and stores where support for this instruction is supported by the CPU architecture. Signed-off-by:
-
Russell King authored
Always use an odd/even register pair for our 64-bit registers, so that we're able to use the double-word load/store instructions in the future. Signed-off-by:
-
Russell King authored
Rearranging the order of the initial tail call code a little allows is to avoid reloading the 'array' pointer. Signed-off-by:
-
Russell King authored
Avoid reloading 'index' after we have validated it - it remains in tmp2[1] up to the point that we begin the code to index the pointer array, so with a little rearrangement of the registers, we can use the already loaded value. Signed-off-by:
-
Russell King authored
Rather than pre-shifting the rm register for the ldr in the tail call, shift it in the load instruction. This eliminates one unnecessary instruction. Signed-off-by:
-
Russell King authored
Rather than moving constants to a register and then using them in a subsequent instruction, use them directly in the desired instruction cutting out the "middle" register. This removes two instructions from the tail call code path. Signed-off-by:
-
Russell King authored
Provide a version of the imm8m() function that the compiler can optimise when used with a constant expression. Signed-off-by:
-
Russell King authored
Access the eBPF scratch space using the frame pointer rather than our stack pointer, as the offsets from the ARM frame pointer are constant across all eBPF programs. Since we no longer reference the scratch space registers from the stack pointer, this simplifies emit_push_r64() as it no longer needs to know how many words are pushed onto the stack. Signed-off-by:
-
Russell King authored
Provide a couple of 64-bit register accessors, and use them where appropriate Signed-off-by:
-
Russell King authored
Many of the code paths need to have knowledge about whether a register is stacked or in a CPU register. Move this decision making to a pair of helper functions instead of having it scattered throughout the code. Signed-off-by:
-
Russell King authored
The decision about whether a BPF register is on the stack or in a CPU register is detected at the top BPF insn processing level, and then percolated throughout the remainder of the code. Since we now use negative register values to represent stacked registers, we can detect where a BPF register is stored without restoring to carrying this additional metadata through all code paths. Signed-off-by:
-
Russell King authored
Use negative numbers for eBPF registers that live on the stack. Signed-off-by:
-
Russell King authored
Provide a set of load/store opcode generators that work with negative immediates as well as positive ones. Signed-off-by:
-
Russell King authored
Enumerate the contents of the JIT scratch stack layout used for storing some of the JITs 64-bit registers, tail call counter and AX register. Signed-off-by:
-
Daniel Borkmann authored
Quentin Monnet says: ==================== The three patches in this series are related to the documentation for eBPF helpers. The first patch brings minor formatting edits to the documentation in include/uapi/linux/bpf.h, and the second one updates the related header file under tools/. The third patch adds a Makefile under tools/bpf for generating the documentation (man pages) about eBPF helpers. The targets defined in this file can also be called from the bpftool directory (please refer to relevant commit logs for details). ==================== Signed-off-by: -
Quentin Monnet authored
Provide a new Makefile.helpers in tools/bpf, in order to build and install the man page for eBPF helpers. This Makefile is also included in the one used to build bpftool documentation, so that it can be called either on its own (cd tools/bpf && make -f Makefile.helpers) or from bpftool directory (cd tools/bpf/bpftool && make doc, or cd tools/bpf/bpftool/Documentation && make helpers). Makefile.helpers is not added directly to bpftool to avoid changing its Makefile too much (helpers are not 100% directly related with bpftool). But the possibility to build the page from bpftool directory makes us able to package the helpers man page with bpftool, and to install it along with bpftool documentation, so that the doc for helpers becomes easily available to developers through the "man" program. Cc: linux-man@vger.kernel.org Suggested-by:
Quentin Monnet <quentin.monnet@netronome.com> Reviewed-by:
Jakub Kicinski <jakub.kicinski@netronome.com> Signed-off-by:
-
Quentin Monnet authored
Update with latest changes from include/uapi/linux/bpf.h header. Signed-off-by:
-
Quentin Monnet authored
Minor formatting edits for eBPF helpers documentation, including blank lines removal, fix of item list for return values in bpf_fib_lookup(), and missing prefix on bpf_skb_load_bytes_relative(). Signed-off-by:
-
- 11 Jul, 2018 14 commits
-
-
Daniel Borkmann authored
Jakub Kicinski says: ==================== This series starts with two minor clean ups to test_offload.py selftest script. The next 11 patches extend the abilities of bpftool prog load beyond the simple cgroup use cases. Three new parameters are added: - type - allows specifying program type, independent of how code sections are named; - map - allows reusing existing maps, instead of creating a new map on every program load; - dev - offload/binding to a device. A number of changes to libbpf is required to accomplish the task. The section - program type logic mapping is exposed. We should probably aim to use the libbpf program section naming everywhere. For reuse of maps we need to allow users to set FD for bpf map object in libbpf. Examples Load program my_xdp.o and pin it as /sys/fs/bpf/my_xdp, for xdp program type: $ bpftool prog load my_xdp.o /sys/fs/bpf/my_xdp \ type xdp As above but for offload: $ bpftool prog load my_xdp.o /sys/fs/bpf/my_xdp \ type xdp \ dev netdevsim0 Load program my_maps.o, but for the first map reuse map id 17, and for the map called "other_map" reuse pinned map /sys/fs/bpf/map0: $ bpftool prog load my_maps.o /sys/fs/bpf/prog \ map idx 0 id 17 \ map name other_map pinned /sys/fs/bpf/map0 v3: - fix return codes in patch 5; - rename libbpf_prog_type_by_string() -> libbpf_prog_type_by_name(); - fold file path into xattr in patch 8; - add patch 10; - use dup3() in patch 12; - depend on fd value in patch 12; - close old fd in patch 12. v2: - add compat for reallocarray(). ==================== Signed-off-by: -
Jakub Kicinski authored
Add map parameter to prog load which will allow reuse of existing maps instead of creating new ones. We need feature detection and compat code for reallocarray, since it's not available in many libc versions. Signed-off-by:
Alexei Starovoitov <ast@kernel.org> Signed-off-by:
-
Jakub Kicinski authored
More advanced applications may want to only replace programs without destroying associated maps. Allow libbpf users to achieve that. Instead of always creating all of the maps at load time, expose to users an API to reconstruct the map object from already existing map. The map parameters are read from the kernel and replace the parameters of the ELF map. libbpf does not restrict the map replacement, i.e. the reused map does not have to be compatible with the ELF map definition. We relay on the verifier for checking the compatibility between maps and programs. The ELF map definition is completely overwritten by the information read from the kernel, to make sure libbpf's view of map object corresponds to the actual map. Signed-off-by:
Andrey Ignatov <rdna@fb.com> Signed-off-by:
-
Jakub Kicinski authored
reallocarray() is a safer variant of realloc which checks for multiplication overflow in case of array allocation. Since it's not available in Glibc < 2.26 import kernel's overflow.h and add a static inline implementation when needed. Use feature detection to probe for existence of reallocarray. Signed-off-by:
Jiong Wang <jiong.wang@netronome.com> Signed-off-by:
-
Jakub Kicinski authored
libbpf_strerror() depends on XSI-compliant (POSIX) version of strerror_r(), which prevents us from using GNU-extensions in libbpf.c, like reallocarray() or dup3(). Move error printing code into a separate file to allow it to continue using POSIX strerror_r(). No functional changes. Signed-off-by:
-
Jakub Kicinski authored
bpf_prog_load() is a very useful helper but it doesn't give us full flexibility of modifying the BPF objects before loading. Open code bpf_prog_load() in bpftool so we can add extra logic in following commits. Signed-off-by:
-
Jakub Kicinski authored
Similarly to bpf_prog_load() users of bpf_object__open() may need to specify the expected program type. Program type is needed at open to avoid the kernel version check for program types which don't require it. Signed-off-by:
-
Jakub Kicinski authored
Add helper to libbpf for recognizing maps which should not have ifindex set when program is loaded. These maps only contain host metadata and therefore are not marked for offload, e.g. the perf event map. Use this helper in bpf_prog_load_xattr(). Signed-off-by:
-
Jakub Kicinski authored
Sometimes program section names don't match with libbpf's expectation. In particular XDP's default section names differ between libbpf and iproute2. Allow users to pass program type on command line. Name the types like the libbpf expected section names. Signed-off-by:
-
Jakub Kicinski authored
libbpf can guess program type based on ELF section names. As libbpf becomes more popular its association between section name strings and types becomes more of a standard. Allow libbpf users to use the same logic for matching strings to types, e.g. when the string originates from command line. Signed-off-by:
-
Jakub Kicinski authored
Extend the bpftool prog load command to also accept "dev" parameter, which will allow us to load programs onto devices. Signed-off-by:
-
Jakub Kicinski authored
Add a new macro for printing more informative message than straight usage() when parameters are missing, and use it for prog do_load(). Save the object and pin path argument to variables for clarity. Signed-off-by:
-
Jakub Kicinski authored
Currently the test only checks errors, not warnings, so save typing and prefix the extack messages with "Error:" inside the check helper. Signed-off-by:
-
Jakub Kicinski authored
Trivial removal of duplicated "mode" in error message. Signed-off-by:
-
- 06 Jul, 2018 4 commits
-
-
Daniel Borkmann authored
Jiong Wang says: ==================== NFP supports u16 and u32 multiplication. Multiplication is done 8-bits per step, therefore we need 2 steps for u16 and 4 steps for u32. We also need one start instruction to initialize the sequence and one or two instructions to fetch the result depending on either you need the high halve of u32 multiplication. For ALU64, if either operand is beyond u32's value range, we reject it. One thing to note, if the source operand is BPF_K, then we need to check "imm" field directly, and we'd reject it if it is negative. Because for ALU64, "imm" (with s32 type) is expected to be sign extended to s64 which NFP mul doesn't support. For ALU32, it is fine for "imm" be negative though, because the result is 32-bits and here is no difference on the low halve of result for signed/unsigned mul, so we will get correct result. NFP doesn't have integer divide instruction, this patch set uses reciprocal algorithm (the basic one, reciprocal_div) to emulate it. For each u32 divide, we would need 11 instructions to finish the operation. 7 (for multiplication) + 4 (various ALUs) = 11 Given NFP only supports multiplication no bigger than u32, we'd require divisor and dividend no bigger than that as well. Also eBPF doesn't support signed divide and has enforced this on C language level by failing compilation. However LLVM assembler hasn't enforced this, so it is possible for negative constant to leak in as a BPF_K operand through assembly code, we reject such cases as well. Meanwhile reciprocal_div.h only implemented the basic version of: "Division by Invariant Integers Using Multiplication" - Torbjörn Granlund and Peter L. Montgomery This patch set further implements the optimized version (Figure 4.2 in the paper) inside existing reciprocal_div.h. When the divider is even and the calculated reciprocal magic number doesn't fit u32, we could reduce the required ALU instructions from 4 to 2 or 1 for some cases. The advanced version requires more complex calculation to get the reciprocal multiplier and other control variables, but then could reduce the required emulation operations. It makes sense to use it for JIT divide code generation (for example eBPF JIT backends) for which we are willing to trade performance of JITed code with that of host. v2: - add warning in l == 32 code path. (Song Liu/Jakub) - jit separate insn sequence for l == 32. (Jakub/Edwin) - should use unrestricted operand for mul. - once place should use "1U << exp" instead of "1 << exp". ==================== Signed-off-by: -
Jiong Wang authored
As we are doing JIT, we would want to use the advanced version of the reciprocal divide (reciprocal_value_adv) to trade performance with host. We could reduce the required ALU instructions from 4 to 2 or 1. Signed-off-by:
-
Jiong Wang authored
NFP doesn't have integer divide instruction, this patch use reciprocal algorithm (the basic one, reciprocal_div) to emulate it. For each u32 divide, we would need 11 instructions to finish the operation. 7 (for multiplication) + 4 (various ALUs) = 11 Given NFP only supports multiplication no bigger than u32, we'd require divisor and dividend no bigger than that as well. Also eBPF doesn't support signed divide and has enforced this on C language level by failing compilation. However LLVM assembler hasn't enforced this, so it is possible for negative constant to leak in as a BPF_K operand through assembly code, we reject such cases as well. Signed-off-by:
Song Liu <songliubraving@fb.com> Signed-off-by:
-
Jiong Wang authored
NFP supports u16 and u32 multiplication. Multiplication is done 8-bits per step, therefore we need 2 steps for u16 and 4 steps for u32. We also need one start instruction to initialize the sequence and one or two instructions to fetch the result depending on either you need the high halve of u32 multiplication. For ALU64, if either operand is beyond u32's value range, we reject it. One thing to note, if the source operand is BPF_K, then we need to check "imm" field directly, and we'd reject it if it is negative. Because for ALU64, "imm" (with s32 type) is expected to be sign extended to s64 which NFP mul doesn't support. For ALU32, it is fine for "imm" be negative though, because the result is 32-bits and here is no difference on the low halve of result for signed/unsigned mul, so we will get correct result. Signed-off-by:
-
