Merge tag 'net-next-6.10' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next

Pull networking updates from Jakub Kicinski:
 "Core & protocols:

   - Complete rework of garbage collection of AF_UNIX sockets.

     AF_UNIX is prone to forming reference count cycles due to fd
     passing functionality. New method based on Tarjan's Strongly
     Connected Components algorithm should be both faster and remove a
     lot of workarounds we accumulated over the years.

   - Add TCP fraglist GRO support, allowing chaining multiple TCP
     packets and forwarding them together. Useful for small switches /
     routers which lack basic checksum offload in some scenarios (e.g.
     PPPoE).

   - Support using SMP threads for handling packet backlog i.e. packet
     processing from software interfaces and old drivers which don't use
     NAPI. This helps move the processing out of the softirq jumble.

   - Continue work of converting from rtnl lock to RCU protection.

     Don't require rtnl lock when reading: IPv6 routing FIB, IPv6
     address labels, netdev threaded NAPI sysfs files, bonding driver's
     sysfs files, MPLS devconf, IPv4 FIB rules, netns IDs, tcp metrics,
     TC Qdiscs, neighbor entries, ARP entries via ioctl(SIOCGARP), a lot
     of the link information available via rtnetlink.

   - Small optimizations from Eric to UDP wake up handling, memory
     accounting, RPS/RFS implementation, TCP packet sizing etc.

   - Allow direct page recycling in the bulk API used by XDP, for +2%
     PPS.

   - Support peek with an offset on TCP sockets.

   - Add MPTCP APIs for querying last time packets were received/sent/acked
     and whether MPTCP "upgrade" succeeded on a TCP socket.

   - Add intra-node communication shortcut to improve SMC performance.

   - Add IPv6 (and IPv{4,6}-over-IPv{4,6}) support to the GTP protocol
     driver.

   - Add HSR-SAN (RedBOX) mode of operation to the HSR protocol driver.

   - Add reset reasons for tracing what caused a TCP reset to be sent.

   - Introduce direction attribute for xfrm (IPSec) states. State can be
     used either for input or output packet processing.

  Things we sprinkled into general kernel code:

   - Add bitmap_{read,write}(), bitmap_size(), expose BYTES_TO_BITS().

     This required touch-ups and renaming of a few existing users.

   - Add Endian-dependent __counted_by_{le,be} annotations.

   - Make building selftests "quieter" by printing summaries like
     "CC object.o" rather than full commands with all the arguments.

  Netfilter:

   - Use GFP_KERNEL to clone elements, to deal better with OOM
     situations and avoid failures in the .commit step.

  BPF:

   - Add eBPF JIT for ARCv2 CPUs.

   - Support attaching kprobe BPF programs through kprobe_multi link in
     a session mode, meaning, a BPF program is attached to both function
     entry and return, the entry program can decide if the return
     program gets executed and the entry program can share u64 cookie
     value with return program. "Session mode" is a common use-case for
     tetragon and bpftrace.

   - Add the ability to specify and retrieve BPF cookie for raw
     tracepoint programs in order to ease migration from classic to raw
     tracepoints.

   - Add an internal-only BPF per-CPU instruction for resolving per-CPU
     memory addresses and implement support in x86, ARM64 and RISC-V
     JITs. This allows inlining functions which need to access per-CPU
     state.

   - Optimize x86 BPF JIT's emit_mov_imm64, and add support for various
     atomics in bpf_arena which can be JITed as a single x86
     instruction. Support BPF arena on ARM64.

   - Add a new bpf_wq API for deferring events and refactor
     process-context bpf_timer code to keep common code where possible.

   - Harden the BPF verifier's and/or/xor value tracking.

   - Introduce crypto kfuncs to let BPF programs call kernel crypto
     APIs.

   - Support bpf_tail_call_static() helper for BPF programs with GCC 13.

   - Add bpf_preempt_{disable,enable}() kfuncs in order to allow a BPF
     program to have code sections where preemption is disabled.

  Driver API:

   - Skip software TC processing completely if all installed rules are
     marked as HW-only, instead of checking the HW-only flag rule by
     rule.

   - Add support for configuring PoE (Power over Ethernet), similar to
     the already existing support for PoDL (Power over Data Line)
     config.

   - Initial bits of a queue control API, for now allowing a single
     queue to be reset without disturbing packet flow to other queues.

   - Common (ethtool) statistics for hardware timestamping.

  Tests and tooling:

   - Remove the need to create a config file to run the net forwarding
     tests so that a naive "make run_tests" can exercise them.

   - Define a method of writing tests which require an external endpoint
     to communicate with (to send/receive data towards the test
     machine). Add a few such tests.

   - Create a shared code library for writing Python tests. Expose the
     YAML Netlink library from tools/ to the tests for easy Netlink
     access.

   - Move netfilter tests under net/, extend them, separate performance
     tests from correctness tests, and iron out issues found by running
     them "on every commit".

   - Refactor BPF selftests to use common network helpers.

   - Further work filling in YAML definitions of Netlink messages for:
     nftables, team driver, bonding interfaces, vlan interfaces, VF
     info, TC u32 mark, TC police action.

   - Teach Python YAML Netlink to decode attribute policies.

   - Extend the definition of the "indexed array" construct in the specs
     to cover arrays of scalars rather than just nests.

   - Add hyperlinks between definitions in generated Netlink docs.

  Drivers:

   - Make sure unsupported flower control flags are rejected by drivers,
     and make more drivers report errors directly to the application
     rather than dmesg (large number of driver changes from Asbjørn
     Sloth Tønnesen).

   - Ethernet high-speed NICs:
      - Broadcom (bnxt):
         - support multiple RSS contexts and steering traffic to them
         - support XDP metadata
         - make page pool allocations more NUMA aware
      - Intel (100G, ice, idpf):
         - extract datapath code common among Intel drivers into a library
         - use fewer resources in switchdev by sharing queues with the PF
         - add PFCP filter support
         - add Ethernet filter support
         - use a spinlock instead of HW lock in PTP clock ops
         - support 5 layer Tx scheduler topology
      - nVidia/Mellanox:
         - 800G link modes and 100G SerDes speeds
         - per-queue IRQ coalescing configuration
      - Marvell Octeon:
         - support offloading TC packet mark action

   - Ethernet NICs consumer, embedded and virtual:
      - stop lying about skb->truesize in USB Ethernet drivers, it
        messes up TCP memory calculations
      - Google cloud vNIC:
         - support changing ring size via ethtool
         - support ring reset using the queue control API
      - VirtIO net:
         - expose flow hash from RSS to XDP
         - per-queue statistics
         - add selftests
      - Synopsys (stmmac):
         - support controllers which require an RX clock signal from the
           MII bus to perform their hardware initialization
      - TI:
         - icssg_prueth: support ICSSG-based Ethernet on AM65x SR1.0 devices
         - icssg_prueth: add SW TX / RX Coalescing based on hrtimers
         - cpsw: minimal XDP support
      - Renesas (ravb):
         - support describing the MDIO bus
      - Realtek (r8169):
         - add support for RTL8168M
      - Microchip Sparx5:
         - matchall and flower actions mirred and redirect

   - Ethernet switches:
      - nVidia/Mellanox:
         - improve events processing performance
      - Marvell:
         - add support for MV88E6250 family internal PHYs
      - Microchip:
         - add DCB and DSCP mapping support for KSZ switches
         - vsc73xx: convert to PHYLINK
      - Realtek:
         - rtl8226b/rtl8221b: add C45 instances and SerDes switching

   - Many driver changes related to PHYLIB and PHYLINK deprecated API
     cleanup

   - Ethernet PHYs:
      - Add a new driver for Airoha EN8811H 2.5 Gigabit PHY.
      - micrel: lan8814: add support for PPS out and external timestamp trigger

   - WiFi:
      - Disable Wireless Extensions (WEXT) in all Wi-Fi 7 devices
        drivers. Modern devices can only be configured using nl80211.
      - mac80211/cfg80211
         - handle color change per link for WiFi 7 Multi-Link Operation
      - Intel (iwlwifi):
         - don't support puncturing in 5 GHz
         - support monitor mode on passive channels
         - BZ-W device support
         - P2P with HE/EHT support
         - re-add support for firmware API 90
         - provide channel survey information for Automatic Channel Selection
      - MediaTek (mt76):
         - mt7921 LED control
         - mt7925 EHT radiotap support
         - mt7920e PCI support
      - Qualcomm (ath11k):
         - P2P support for QCA6390, WCN6855 and QCA2066
         - support hibernation
         - ieee80211-freq-limit Device Tree property support
      - Qualcomm (ath12k):
         - refactoring in preparation of multi-link support
         - suspend and hibernation support
         - ACPI support
         - debugfs support, including dfs_simulate_radar support
      - RealTek:
         - rtw88: RTL8723CS SDIO device support
         - rtw89: RTL8922AE Wi-Fi 7 PCI device support
         - rtw89: complete features of new WiFi 7 chip 8922AE including
           BT-coexistence and Wake-on-WLAN
         - rtw89: use BIOS ACPI settings to set TX power and channels
         - rtl8xxxu: enable Management Frame Protection (MFP) support

   - Bluetooth:
      - support for Intel BlazarI and Filmore Peak2 (BE201)
      - support for MediaTek MT7921S SDIO
      - initial support for Intel PCIe BT driver
      - remove HCI_AMP support"

* tag 'net-next-6.10' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next: (1827 commits)
  selftests: netfilter: fix packetdrill conntrack testcase
  net: gro: fix napi_gro_cb zeroed alignment
  Bluetooth: btintel_pcie: Refactor and code cleanup
  Bluetooth: btintel_pcie: Fix warning reported by sparse
  Bluetooth: hci_core: Fix not handling hdev->le_num_of_adv_sets=1
  Bluetooth: btintel: Fix compiler warning for multi_v7_defconfig config
  Bluetooth: btintel_pcie: Fix compiler warnings
  Bluetooth: btintel_pcie: Add *setup* function to download firmware
  Bluetooth: btintel_pcie: Add support for PCIe transport
  Bluetooth: btintel: Export few static functions
  Bluetooth: HCI: Remove HCI_AMP support
  Bluetooth: L2CAP: Fix div-by-zero in l2cap_le_flowctl_init()
  Bluetooth: qca: Fix error code in qca_read_fw_build_info()
  Bluetooth: hci_conn: Use __counted_by() and avoid -Wfamnae warning
  Bluetooth: btintel: Add support for Filmore Peak2 (BE201)
  Bluetooth: btintel: Add support for BlazarI
  LE Create Connection command timeout increased to 20 secs
  dt-bindings: net: bluetooth: Add MediaTek MT7921S SDIO Bluetooth
  Bluetooth: compute LE flow credits based on recvbuf space
  Bluetooth: hci_sync: Use cmd->num_cis instead of magic number
  ...

21 files changed, 5200 insertions, 71 deletions

diff --git a/arch/arc/Kbuild b/arch/arc/Kbuild
index b94102fff68b..20ea7dd482d4 100644
--- a/arch/arc/Kbuild
+++ b/arch/arc/Kbuild
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: GPL-2.0
 obj-y += kernel/
 obj-y += mm/
+obj-y += net/
 
 # for cleaning
 subdir- += boot
diff --git a/arch/arc/Kconfig b/arch/arc/Kconfig
index 4092bec198be..fd0b0a0d4686 100644
--- a/arch/arc/Kconfig
+++ b/arch/arc/Kconfig
@@ -51,6 +51,7 @@ config ARC
 	select PCI_SYSCALL if PCI
 	select HAVE_ARCH_JUMP_LABEL if ISA_ARCV2 && !CPU_ENDIAN_BE32
 	select TRACE_IRQFLAGS_SUPPORT
+	select HAVE_EBPF_JIT if ISA_ARCV2
 
 config LOCKDEP_SUPPORT
 	def_bool y
diff --git a/arch/arc/net/Makefile b/arch/arc/net/Makefile
new file mode 100644
index 000000000000..ea5790952e9a
--- /dev/null
+++ b/arch/arc/net/Makefile
@@ -0,0 +1,6 @@
+# SPDX-License-Identifier: GPL-2.0-only
+
+ifeq ($(CONFIG_ISA_ARCV2),y)
+	obj-$(CONFIG_BPF_JIT) += bpf_jit_core.o
+	obj-$(CONFIG_BPF_JIT) += bpf_jit_arcv2.o
+endif
diff --git a/arch/arc/net/bpf_jit.h b/arch/arc/net/bpf_jit.h
new file mode 100644
index 000000000000..ec44873c42d1
--- /dev/null
+++ b/arch/arc/net/bpf_jit.h
@@ -0,0 +1,164 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * The interface that a back-end should provide to bpf_jit_core.c.
+ *
+ * Copyright (c) 2024 Synopsys Inc.
+ * Author: Shahab Vahedi <shahab@synopsys.com>
+ */
+
+#ifndef _ARC_BPF_JIT_H
+#define _ARC_BPF_JIT_H
+
+#include <linux/bpf.h>
+#include <linux/filter.h>
+
+/* Print debug info and assert. */
+//#define ARC_BPF_JIT_DEBUG
+
+/* Determine the address type of the target. */
+#ifdef CONFIG_ISA_ARCV2
+#define ARC_ADDR u32
+#endif
+
+/*
+ * For the translation of some BPF instructions, a temporary register
+ * might be needed for some interim data.
+ */
+#define JIT_REG_TMP MAX_BPF_JIT_REG
+
+/*
+ * Buffer access: If buffer "b" is not NULL, advance by "n" bytes.
+ *
+ * This macro must be used in any place that potentially requires a
+ * "buf + len". This way, we make sure that the "buf" argument for
+ * the underlying "arc_*(buf, ...)" ends up as NULL instead of something
+ * like "0+4" or "0+8", etc. Those "arc_*()" functions check their "buf"
+ * value to decide if instructions should be emitted or not.
+ */
+#define BUF(b, n) (((b) != NULL) ? ((b) + (n)) : (b))
+
+/************** Functions that the back-end must provide **************/
+/* Extension for 32-bit operations. */
+inline u8 zext(u8 *buf, u8 rd);
+/***** Moves *****/
+u8 mov_r32(u8 *buf, u8 rd, u8 rs, u8 sign_ext);
+u8 mov_r32_i32(u8 *buf, u8 reg, s32 imm);
+u8 mov_r64(u8 *buf, u8 rd, u8 rs, u8 sign_ext);
+u8 mov_r64_i32(u8 *buf, u8 reg, s32 imm);
+u8 mov_r64_i64(u8 *buf, u8 reg, u32 lo, u32 hi);
+/***** Loads and stores *****/
+u8 load_r(u8 *buf, u8 rd, u8 rs, s16 off, u8 size, bool sign_ext);
+u8 store_r(u8 *buf, u8 rd, u8 rs, s16 off, u8 size);
+u8 store_i(u8 *buf, s32 imm, u8 rd, s16 off, u8 size);
+/***** Addition *****/
+u8 add_r32(u8 *buf, u8 rd, u8 rs);
+u8 add_r32_i32(u8 *buf, u8 rd, s32 imm);
+u8 add_r64(u8 *buf, u8 rd, u8 rs);
+u8 add_r64_i32(u8 *buf, u8 rd, s32 imm);
+/***** Subtraction *****/
+u8 sub_r32(u8 *buf, u8 rd, u8 rs);
+u8 sub_r32_i32(u8 *buf, u8 rd, s32 imm);
+u8 sub_r64(u8 *buf, u8 rd, u8 rs);
+u8 sub_r64_i32(u8 *buf, u8 rd, s32 imm);
+/***** Multiplication *****/
+u8 mul_r32(u8 *buf, u8 rd, u8 rs);
+u8 mul_r32_i32(u8 *buf, u8 rd, s32 imm);
+u8 mul_r64(u8 *buf, u8 rd, u8 rs);
+u8 mul_r64_i32(u8 *buf, u8 rd, s32 imm);
+/***** Division *****/
+u8 div_r32(u8 *buf, u8 rd, u8 rs, bool sign_ext);
+u8 div_r32_i32(u8 *buf, u8 rd, s32 imm, bool sign_ext);
+/***** Remainder *****/
+u8 mod_r32(u8 *buf, u8 rd, u8 rs, bool sign_ext);
+u8 mod_r32_i32(u8 *buf, u8 rd, s32 imm, bool sign_ext);
+/***** Bitwise AND *****/
+u8 and_r32(u8 *buf, u8 rd, u8 rs);
+u8 and_r32_i32(u8 *buf, u8 rd, s32 imm);
+u8 and_r64(u8 *buf, u8 rd, u8 rs);
+u8 and_r64_i32(u8 *buf, u8 rd, s32 imm);
+/***** Bitwise OR *****/
+u8 or_r32(u8 *buf, u8 rd, u8 rs);
+u8 or_r32_i32(u8 *buf, u8 rd, s32 imm);
+u8 or_r64(u8 *buf, u8 rd, u8 rs);
+u8 or_r64_i32(u8 *buf, u8 rd, s32 imm);
+/***** Bitwise XOR *****/
+u8 xor_r32(u8 *buf, u8 rd, u8 rs);
+u8 xor_r32_i32(u8 *buf, u8 rd, s32 imm);
+u8 xor_r64(u8 *buf, u8 rd, u8 rs);
+u8 xor_r64_i32(u8 *buf, u8 rd, s32 imm);
+/***** Bitwise Negate *****/
+u8 neg_r32(u8 *buf, u8 r);
+u8 neg_r64(u8 *buf, u8 r);
+/***** Bitwise left shift *****/
+u8 lsh_r32(u8 *buf, u8 rd, u8 rs);
+u8 lsh_r32_i32(u8 *buf, u8 rd, u8 imm);
+u8 lsh_r64(u8 *buf, u8 rd, u8 rs);
+u8 lsh_r64_i32(u8 *buf, u8 rd, s32 imm);
+/***** Bitwise right shift (logical) *****/
+u8 rsh_r32(u8 *buf, u8 rd, u8 rs);
+u8 rsh_r32_i32(u8 *buf, u8 rd, u8 imm);
+u8 rsh_r64(u8 *buf, u8 rd, u8 rs);
+u8 rsh_r64_i32(u8 *buf, u8 rd, s32 imm);
+/***** Bitwise right shift (arithmetic) *****/
+u8 arsh_r32(u8 *buf, u8 rd, u8 rs);
+u8 arsh_r32_i32(u8 *buf, u8 rd, u8 imm);
+u8 arsh_r64(u8 *buf, u8 rd, u8 rs);
+u8 arsh_r64_i32(u8 *buf, u8 rd, s32 imm);
+/***** Frame related *****/
+u32 mask_for_used_regs(u8 bpf_reg, bool is_call);
+u8 arc_prologue(u8 *buf, u32 usage, u16 frame_size);
+u8 arc_epilogue(u8 *buf, u32 usage, u16 frame_size);
+/***** Jumps *****/
+/*
+ * Different sorts of conditions (ARC enum as opposed to BPF_*).
+ *
+ * Do not change the order of enums here. ARC_CC_SLE+1 is used
+ * to determine the number of JCCs.
+ */
+enum ARC_CC {
+	ARC_CC_UGT = 0,		/* unsigned >  */
+	ARC_CC_UGE,		/* unsigned >= */
+	ARC_CC_ULT,		/* unsigned <  */
+	ARC_CC_ULE,		/* unsigned <= */
+	ARC_CC_SGT,		/*   signed >  */
+	ARC_CC_SGE,		/*   signed >= */
+	ARC_CC_SLT,		/*   signed <  */
+	ARC_CC_SLE,		/*   signed <= */
+	ARC_CC_AL,		/* always      */
+	ARC_CC_EQ,		/*          == */
+	ARC_CC_NE,		/*          != */
+	ARC_CC_SET,		/* test        */
+	ARC_CC_LAST
+};
+
+/*
+ * A few notes:
+ *
+ * - check_jmp_*() are prerequisites before calling the gen_jmp_*().
+ *   They return "true" if the jump is possible and "false" otherwise.
+ *
+ * - The notion of "*_off" is to emphasize that these parameters are
+ *   merely offsets in the JIT stream and not absolute addresses. One
+ *   can look at them as addresses if the JIT code would start from
+ *   address 0x0000_0000. Nonetheless, since the buffer address for the
+ *   JIT is on a word-aligned address, this works and actually makes
+ *   things simpler (offsets are in the range of u32 which is more than
+ *   enough).
+ */
+bool check_jmp_32(u32 curr_off, u32 targ_off, u8 cond);
+bool check_jmp_64(u32 curr_off, u32 targ_off, u8 cond);
+u8 gen_jmp_32(u8 *buf, u8 rd, u8 rs, u8 cond, u32 c_off, u32 t_off);
+u8 gen_jmp_64(u8 *buf, u8 rd, u8 rs, u8 cond, u32 c_off, u32 t_off);
+/***** Miscellaneous *****/
+u8 gen_func_call(u8 *buf, ARC_ADDR func_addr, bool external_func);
+u8 arc_to_bpf_return(u8 *buf);
+/*
+ * - Perform byte swaps on "rd" based on the "size".
+ * - If "force" is set, do it unconditionally. Otherwise, consider the
+ *   desired "endian"ness and the host endianness.
+ * - For data "size"s up to 32 bits, perform a zero-extension if asked
+ *   by the "do_zext" boolean.
+ */
+u8 gen_swap(u8 *buf, u8 rd, u8 size, u8 endian, bool force, bool do_zext);
+
+#endif /* _ARC_BPF_JIT_H */
diff --git a/arch/arc/net/bpf_jit_arcv2.c b/arch/arc/net/bpf_jit_arcv2.c
new file mode 100644
index 000000000000..31bfb6e9ce00
--- /dev/null
+++ b/arch/arc/net/bpf_jit_arcv2.c
@@ -0,0 +1,3005 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * The ARCv2 backend of Just-In-Time compiler for eBPF bytecode.
+ *
+ * Copyright (c) 2024 Synopsys Inc.
+ * Author: Shahab Vahedi <shahab@synopsys.com>
+ */
+#include <linux/bug.h>
+#include "bpf_jit.h"
+
+/* ARC core registers. */
+enum {
+	ARC_R_0,  ARC_R_1,  ARC_R_2,  ARC_R_3,  ARC_R_4,  ARC_R_5,
+	ARC_R_6,  ARC_R_7,  ARC_R_8,  ARC_R_9,  ARC_R_10, ARC_R_11,
+	ARC_R_12, ARC_R_13, ARC_R_14, ARC_R_15, ARC_R_16, ARC_R_17,
+	ARC_R_18, ARC_R_19, ARC_R_20, ARC_R_21, ARC_R_22, ARC_R_23,
+	ARC_R_24, ARC_R_25, ARC_R_26, ARC_R_FP, ARC_R_SP, ARC_R_ILINK,
+	ARC_R_30, ARC_R_BLINK,
+	/*
+	 * Having ARC_R_IMM encoded as source register means there is an
+	 * immediate that must be interpreted from the next 4 bytes. If
+	 * encoded as the destination register though, it implies that the
+	 * output of the operation is not assigned to any register. The
+	 * latter is helpful if we only care about updating the CPU status
+	 * flags.
+	 */
+	ARC_R_IMM = 62
+};
+
+/*
+ * Remarks about the rationale behind the chosen mapping:
+ *
+ * - BPF_REG_{1,2,3,4} are the argument registers and must be mapped to
+ *   argument registers in ARCv2 ABI: r0-r7. The r7 registers is the last
+ *   argument register in the ABI. Therefore BPF_REG_5, as the fifth
+ *   argument, must be pushed onto the stack. This is a must for calling
+ *   in-kernel functions.
+ *
+ * - In ARCv2 ABI, the return value is in r0 for 32-bit results and (r1,r0)
+ *   for 64-bit results. However, because they're already used for BPF_REG_1,
+ *   the next available scratch registers, r8 and r9, are the best candidates
+ *   for BPF_REG_0. After a "call" to a(n) (in-kernel) function, the result
+ *   is "mov"ed to these registers. At a BPF_EXIT, their value is "mov"ed to
+ *   (r1,r0).
+ *   It is worth mentioning that scratch registers are the best choice for
+ *   BPF_REG_0, because it is very popular in BPF instruction encoding.
+ *
+ * - JIT_REG_TMP is an artifact needed to translate some BPF instructions.
+ *   Its life span is one single BPF instruction. Since during the
+ *   analyze_reg_usage(), it is not known if temporary registers are used,
+ *   it is mapped to ARC's scratch registers: r10 and r11. Therefore, they
+ *   don't matter in analysing phase and don't need saving. This temporary
+ *   register is added as yet another index in the bpf2arc array, so it will
+ *   unfold like the rest of registers during the code generation process.
+ *
+ * - Mapping of callee-saved BPF registers, BPF_REG_{6,7,8,9}, starts from
+ *   (r15,r14) register pair. The (r13,r12) is not a good choice, because
+ *   in ARCv2 ABI, r12 is not a callee-saved register and this can cause
+ *   problem when calling an in-kernel function. Theoretically, the mapping
+ *   could start from (r14,r13), but it is not a conventional ARCv2 register
+ *   pair. To have a future proof design, I opted for this arrangement.
+ *   If/when we decide to add ARCv2 instructions that do use register pairs,
+ *   the mapping, hopefully, doesn't need to be revisited.
+ */
+const u8 bpf2arc[][2] = {
+	/* Return value from in-kernel function, and exit value from eBPF */
+	[BPF_REG_0] = {ARC_R_8, ARC_R_9},
+	/* Arguments from eBPF program to in-kernel function */
+	[BPF_REG_1] = {ARC_R_0, ARC_R_1},
+	[BPF_REG_2] = {ARC_R_2, ARC_R_3},
+	[BPF_REG_3] = {ARC_R_4, ARC_R_5},
+	[BPF_REG_4] = {ARC_R_6, ARC_R_7},
+	/* Remaining arguments, to be passed on the stack per 32-bit ABI */
+	[BPF_REG_5] = {ARC_R_22, ARC_R_23},
+	/* Callee-saved registers that in-kernel function will preserve */
+	[BPF_REG_6] = {ARC_R_14, ARC_R_15},
+	[BPF_REG_7] = {ARC_R_16, ARC_R_17},
+	[BPF_REG_8] = {ARC_R_18, ARC_R_19},
+	[BPF_REG_9] = {ARC_R_20, ARC_R_21},
+	/* Read-only frame pointer to access the eBPF stack. 32-bit only. */
+	[BPF_REG_FP] = {ARC_R_FP, },
+	/* Register for blinding constants */
+	[BPF_REG_AX] = {ARC_R_24, ARC_R_25},
+	/* Temporary registers for internal use */
+	[JIT_REG_TMP] = {ARC_R_10, ARC_R_11}
+};
+
+#define ARC_CALLEE_SAVED_REG_FIRST ARC_R_13
+#define ARC_CALLEE_SAVED_REG_LAST  ARC_R_25
+
+#define REG_LO(r) (bpf2arc[(r)][0])
+#define REG_HI(r) (bpf2arc[(r)][1])
+
+/*
+ * To comply with ARCv2 ABI, BPF's arg5 must be put on stack. After which,
+ * the stack needs to be restored by ARG5_SIZE.
+ */
+#define ARG5_SIZE 8
+
+/* Instruction lengths in bytes. */
+enum {
+	INSN_len_normal = 4,	/* Normal instructions length. */
+	INSN_len_imm = 4	/* Length of an extra 32-bit immediate. */
+};
+
+/* ZZ defines the size of operation in encodings that it is used. */
+enum {
+	ZZ_1_byte = 1,
+	ZZ_2_byte = 2,
+	ZZ_4_byte = 0,
+	ZZ_8_byte = 3
+};
+
+/*
+ * AA is mostly about address write back mode. It determines if the
+ * address in question should be updated before usage or after:
+ *   addr += offset; data = *addr;
+ *   data = *addr; addr += offset;
+ *
+ * In "scaling" mode, the effective address will become the sum
+ * of "address" + "index"*"size". The "size" is specified by the
+ * "ZZ" field. There is no write back when AA is set for scaling:
+ *   data = *(addr + offset<<zz)
+ */
+enum {
+	AA_none  = 0,
+	AA_pre   = 1,	/* in assembly known as "a/aw". */
+	AA_post  = 2,	/* in assembly known as "ab". */
+	AA_scale = 3	/* in assembly known as "as". */
+};
+
+/* X flag determines the mode of extension. */
+enum {
+	X_zero = 0,
+	X_sign = 1
+};
+
+/* Condition codes. */
+enum {
+	CC_always     = 0,	/* condition is true all the time */
+	CC_equal      = 1,	/* if status32.z flag is set */
+	CC_unequal    = 2,	/* if status32.z flag is clear */
+	CC_positive   = 3,	/* if status32.n flag is clear */
+	CC_negative   = 4,	/* if status32.n flag is set */
+	CC_less_u     = 5,	/* less than (unsigned) */
+	CC_less_eq_u  = 14,	/* less than or equal (unsigned) */
+	CC_great_eq_u = 6,	/* greater than or equal (unsigned) */
+	CC_great_u    = 13,	/* greater than (unsigned) */
+	CC_less_s     = 11,	/* less than (signed) */
+	CC_less_eq_s  = 12,	/* less than or equal (signed) */
+	CC_great_eq_s = 10,	/* greater than or equal (signed) */
+	CC_great_s    = 9	/* greater than (signed) */
+};
+
+#define IN_U6_RANGE(x)	((x) <= (0x40      - 1) && (x) >= 0)
+#define IN_S9_RANGE(x)	((x) <= (0x100     - 1) && (x) >= -0x100)
+#define IN_S12_RANGE(x)	((x) <= (0x800     - 1) && (x) >= -0x800)
+#define IN_S21_RANGE(x)	((x) <= (0x100000  - 1) && (x) >= -0x100000)
+#define IN_S25_RANGE(x)	((x) <= (0x1000000 - 1) && (x) >= -0x1000000)
+
+/* Operands in most of the encodings. */
+#define OP_A(x)	((x) & 0x03f)
+#define OP_B(x)	((((x) & 0x07) << 24) | (((x) & 0x38) <<  9))
+#define OP_C(x)	(((x) & 0x03f) << 6)
+#define OP_IMM	(OP_C(ARC_R_IMM))
+#define COND(x)	(OP_A((x) & 31))
+#define FLAG(x)	(((x) & 1) << 15)
+
+/*
+ * The 4-byte encoding of "mov b,c":
+ *
+ * 0010_0bbb 0000_1010 0BBB_cccc cc00_0000
+ *
+ * b:  BBBbbb		destination register
+ * c:  cccccc		source register
+ */
+#define OPC_MOV		0x200a0000
+
+/*
+ * The 4-byte encoding of "mov b,s12" (used for moving small immediates):
+ *
+ * 0010_0bbb 1000_1010 0BBB_ssss ssSS_SSSS
+ *
+ * b:  BBBbbb		destination register
+ * s:  SSSSSSssssss	source immediate (signed)
+ */
+#define OPC_MOVI	0x208a0000
+#define MOVI_S12(x)	((((x) & 0xfc0) >> 6) | (((x) & 0x3f) << 6))
+
+/*
+ * The 4-byte encoding of "mov[.qq] b,u6", used for conditional
+ * moving of even smaller immediates:
+ *
+ * 0010_0bbb 1100_1010 0BBB_cccc cciq_qqqq
+ *
+ * qq: qqqqq		condition code
+ * i:			If set, c is considered a 6-bit immediate, else a reg.
+ *
+ * b:  BBBbbb		destination register
+ * c:  cccccc		source
+ */
+#define OPC_MOV_CC	0x20ca0000
+#define MOV_CC_I	BIT(5)
+#define OPC_MOVU_CC	(OPC_MOV_CC | MOV_CC_I)
+
+/*
+ * The 4-byte encoding of "sexb b,c" (8-bit sign extension):
+ *
+ * 0010_0bbb 0010_1111 0BBB_cccc cc00_0101
+ *
+ * b:  BBBbbb		destination register
+ * c:  cccccc		source register
+ */
+#define OPC_SEXB	0x202f0005
+
+/*
+ * The 4-byte encoding of "sexh b,c" (16-bit sign extension):
+ *
+ * 0010_0bbb 0010_1111 0BBB_cccc cc00_0110
+ *
+ * b:  BBBbbb		destination register
+ * c:  cccccc		source register
+ */
+#define OPC_SEXH	0x202f0006
+
+/*
+ * The 4-byte encoding of "ld[zz][.x][.aa] c,[b,s9]":
+ *
+ * 0001_0bbb ssss_ssss SBBB_0aaz zxcc_cccc
+ *
+ * zz:			size mode
+ * aa:			address write back mode
+ * x:			extension mode
+ *
+ * s9: S_ssss_ssss	9-bit signed number
+ * b:  BBBbbb		source reg for address
+ * c:  cccccc		destination register
+ */
+#define OPC_LOAD	0x10000000
+#define LOAD_X(x)	((x) << 6)
+#define LOAD_ZZ(x)	((x) << 7)
+#define LOAD_AA(x)	((x) << 9)
+#define LOAD_S9(x)	((((x) & 0x0ff) << 16) | (((x) & 0x100) <<  7))
+#define LOAD_C(x)	((x) & 0x03f)
+/* Unsigned and signed loads. */
+#define OPC_LDU		(OPC_LOAD | LOAD_X(X_zero))
+#define OPC_LDS		(OPC_LOAD | LOAD_X(X_sign))
+/* 32-bit load. */
+#define OPC_LD32	(OPC_LDU | LOAD_ZZ(ZZ_4_byte))
+/* "pop reg" is merely a "ld.ab reg,[sp,4]". */
+#define OPC_POP		\
+	(OPC_LD32 | LOAD_AA(AA_post) | LOAD_S9(4) | OP_B(ARC_R_SP))
+
+/*
+ * The 4-byte encoding of "st[zz][.aa] c,[b,s9]":
+ *
+ * 0001_1bbb ssss_ssss SBBB_cccc cc0a_azz0
+ *
+ * zz: zz		size mode
+ * aa: aa		address write back mode
+ *
+ * s9: S_ssss_ssss	9-bit signed number
+ * b:  BBBbbb		source reg for address
+ * c:  cccccc		source reg to be stored
+ */
+#define OPC_STORE	0x18000000
+#define STORE_ZZ(x)	((x) << 1)
+#define STORE_AA(x)	((x) << 3)
+#define STORE_S9(x)	((((x) & 0x0ff) << 16) | (((x) & 0x100) <<  7))
+/* 32-bit store. */
+#define OPC_ST32	(OPC_STORE | STORE_ZZ(ZZ_4_byte))
+/* "push reg" is merely a "st.aw reg,[sp,-4]". */
+#define OPC_PUSH	\
+	(OPC_ST32 | STORE_AA(AA_pre) | STORE_S9(-4) | OP_B(ARC_R_SP))
+
+/*
+ * The 4-byte encoding of "add a,b,c":
+ *
+ * 0010_0bbb 0i00_0000 fBBB_cccc ccaa_aaaa
+ *
+ * f:                   indicates if flags (carry, etc.) should be updated
+ * i:			If set, c is considered a 6-bit immediate, else a reg.
+ *
+ * a:  aaaaaa		result
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand
+ */
+#define OPC_ADD		0x20000000
+/* Addition with updating the pertinent flags in "status32" register. */
+#define OPC_ADDF	(OPC_ADD | FLAG(1))
+#define ADDI		BIT(22)
+#define ADDI_U6(x)	OP_C(x)
+#define OPC_ADDI	(OPC_ADD | ADDI)
+#define OPC_ADDIF	(OPC_ADDI | FLAG(1))
+#define OPC_ADD_I	(OPC_ADD | OP_IMM)
+
+/*
+ * The 4-byte encoding of "adc a,b,c" (addition with carry):
+ *
+ * 0010_0bbb 0i00_0001 0BBB_cccc ccaa_aaaa
+ *
+ * i:			if set, c is considered a 6-bit immediate, else a reg.
+ *
+ * a:  aaaaaa		result
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand
+ */
+#define OPC_ADC		0x20010000
+#define ADCI		BIT(22)
+#define ADCI_U6(x)	OP_C(x)
+#define OPC_ADCI	(OPC_ADC | ADCI)
+
+/*
+ * The 4-byte encoding of "sub a,b,c":
+ *
+ * 0010_0bbb 0i00_0010 fBBB_cccc ccaa_aaaa
+ *
+ * f:                   indicates if flags (carry, etc.) should be updated
+ * i:			if set, c is considered a 6-bit immediate, else a reg.
+ *
+ * a:  aaaaaa		result
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand
+ */
+#define OPC_SUB		0x20020000
+/* Subtraction with updating the pertinent flags in "status32" register. */
+#define OPC_SUBF	(OPC_SUB | FLAG(1))
+#define SUBI		BIT(22)
+#define SUBI_U6(x)	OP_C(x)
+#define OPC_SUBI	(OPC_SUB | SUBI)
+#define OPC_SUB_I	(OPC_SUB | OP_IMM)
+
+/*
+ * The 4-byte encoding of "sbc a,b,c" (subtraction with carry):
+ *
+ * 0010_0bbb 0000_0011 fBBB_cccc ccaa_aaaa
+ *
+ * f:                   indicates if flags (carry, etc.) should be updated
+ *
+ * a:  aaaaaa		result
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand
+ */
+#define OPC_SBC		0x20030000
+
+/*
+ * The 4-byte encoding of "cmp[.qq] b,c":
+ *
+ * 0010_0bbb 1100_1100 1BBB_cccc cc0q_qqqq
+ *
+ * qq:	qqqqq		condition code
+ *
+ * b:  BBBbbb		the 1st operand
+ * c:  cccccc		the 2nd operand
+ */
+#define OPC_CMP		0x20cc8000
+
+/*
+ * The 4-byte encoding of "neg a,b":
+ *
+ * 0010_0bbb 0100_1110 0BBB_0000 00aa_aaaa
+ *
+ * a:  aaaaaa		result
+ * b:  BBBbbb		input
+ */
+#define OPC_NEG		0x204e0000
+
+/*
+ * The 4-byte encoding of "mpy a,b,c".
+ * mpy is the signed 32-bit multiplication with the lower 32-bit
+ * of the product as the result.
+ *
+ * 0010_0bbb 0001_1010 0BBB_cccc ccaa_aaaa
+ *
+ * a:  aaaaaa		result
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand
+ */
+#define OPC_MPY		0x201a0000
+#define OPC_MPYI	(OPC_MPY | OP_IMM)
+
+/*
+ * The 4-byte encoding of "mpydu a,b,c".
+ * mpydu is the unsigned 32-bit multiplication with the lower 32-bit of
+ * the product in register "a" and the higher 32-bit in register "a+1".
+ *
+ * 0010_1bbb 0001_1001 0BBB_cccc ccaa_aaaa
+ *
+ * a:  aaaaaa		64-bit result in registers (R_a+1,R_a)
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand
+ */
+#define OPC_MPYDU	0x28190000
+#define OPC_MPYDUI	(OPC_MPYDU | OP_IMM)
+
+/*
+ * The 4-byte encoding of "divu a,b,c" (unsigned division):
+ *
+ * 0010_1bbb 0000_0101 0BBB_cccc ccaa_aaaa
+ *
+ * a:  aaaaaa		result (quotient)
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand (divisor)
+ */
+#define OPC_DIVU	0x28050000
+#define OPC_DIVUI	(OPC_DIVU | OP_IMM)
+
+/*
+ * The 4-byte encoding of "div a,b,c" (signed division):
+ *
+ * 0010_1bbb 0000_0100 0BBB_cccc ccaa_aaaa
+ *
+ * a:  aaaaaa		result (quotient)
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand (divisor)
+ */
+#define OPC_DIVS	0x28040000
+#define OPC_DIVSI	(OPC_DIVS | OP_IMM)
+
+/*
+ * The 4-byte encoding of "remu a,b,c" (unsigned remainder):
+ *
+ * 0010_1bbb 0000_1001 0BBB_cccc ccaa_aaaa
+ *
+ * a:  aaaaaa		result (remainder)
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand (divisor)
+ */
+#define OPC_REMU	0x28090000
+#define OPC_REMUI	(OPC_REMU | OP_IMM)
+
+/*
+ * The 4-byte encoding of "rem a,b,c" (signed remainder):
+ *
+ * 0010_1bbb 0000_1000 0BBB_cccc ccaa_aaaa
+ *
+ * a:  aaaaaa		result (remainder)
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand (divisor)
+ */
+#define OPC_REMS	0x28080000
+#define OPC_REMSI	(OPC_REMS | OP_IMM)
+
+/*
+ * The 4-byte encoding of "and a,b,c":
+ *
+ * 0010_0bbb 0000_0100 fBBB_cccc ccaa_aaaa
+ *
+ * f:                   indicates if zero and negative flags should be updated
+ *
+ * a:  aaaaaa		result
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc		the 2nd input operand
+ */
+#define OPC_AND		0x20040000
+#define OPC_ANDI	(OPC_AND | OP_IMM)
+
+/*
+ * The 4-byte encoding of "tst[.qq] b,c".
+ * Checks if the two input operands have any bit set at the same
+ * position.
+ *
+ * 0010_0bbb 1100_1011 1BBB_cccc cc0q_qqqq
+ *
+ * qq:	qqqqq		condition code
+ *
+ * b:  BBBbbb		the 1st input operand
+ * c:  cccccc