xref: /linux/drivers/net/ethernet/netronome/nfp/bpf/jit.c (revision c411ed854584a71b0e86ac3019b60e4789d88086)
1 /*
2  * Copyright (C) 2016 Netronome Systems, Inc.
3  *
4  * This software is dual licensed under the GNU General License Version 2,
5  * June 1991 as shown in the file COPYING in the top-level directory of this
6  * source tree or the BSD 2-Clause License provided below.  You have the
7  * option to license this software under the complete terms of either license.
8  *
9  * The BSD 2-Clause License:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      1. Redistributions of source code must retain the above
16  *         copyright notice, this list of conditions and the following
17  *         disclaimer.
18  *
19  *      2. Redistributions in binary form must reproduce the above
20  *         copyright notice, this list of conditions and the following
21  *         disclaimer in the documentation and/or other materials
22  *         provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33 
34 #define pr_fmt(fmt)	"NFP net bpf: " fmt
35 
36 #include <linux/kernel.h>
37 #include <linux/bpf.h>
38 #include <linux/filter.h>
39 #include <linux/pkt_cls.h>
40 #include <linux/unistd.h>
41 
42 #include "main.h"
43 #include "../nfp_asm.h"
44 
45 /* --- NFP prog --- */
46 /* Foreach "multiple" entries macros provide pos and next<n> pointers.
47  * It's safe to modify the next pointers (but not pos).
48  */
49 #define nfp_for_each_insn_walk2(nfp_prog, pos, next)			\
50 	for (pos = list_first_entry(&(nfp_prog)->insns, typeof(*pos), l), \
51 	     next = list_next_entry(pos, l);			\
52 	     &(nfp_prog)->insns != &pos->l &&			\
53 	     &(nfp_prog)->insns != &next->l;			\
54 	     pos = nfp_meta_next(pos),				\
55 	     next = nfp_meta_next(pos))
56 
57 #define nfp_for_each_insn_walk3(nfp_prog, pos, next, next2)		\
58 	for (pos = list_first_entry(&(nfp_prog)->insns, typeof(*pos), l), \
59 	     next = list_next_entry(pos, l),			\
60 	     next2 = list_next_entry(next, l);			\
61 	     &(nfp_prog)->insns != &pos->l &&			\
62 	     &(nfp_prog)->insns != &next->l &&			\
63 	     &(nfp_prog)->insns != &next2->l;			\
64 	     pos = nfp_meta_next(pos),				\
65 	     next = nfp_meta_next(pos),				\
66 	     next2 = nfp_meta_next(next))
67 
68 static bool
69 nfp_meta_has_next(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
70 {
71 	return meta->l.next != &nfp_prog->insns;
72 }
73 
74 static bool
75 nfp_meta_has_prev(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
76 {
77 	return meta->l.prev != &nfp_prog->insns;
78 }
79 
80 static void nfp_prog_free(struct nfp_prog *nfp_prog)
81 {
82 	struct nfp_insn_meta *meta, *tmp;
83 
84 	list_for_each_entry_safe(meta, tmp, &nfp_prog->insns, l) {
85 		list_del(&meta->l);
86 		kfree(meta);
87 	}
88 	kfree(nfp_prog);
89 }
90 
91 static void nfp_prog_push(struct nfp_prog *nfp_prog, u64 insn)
92 {
93 	if (nfp_prog->__prog_alloc_len == nfp_prog->prog_len) {
94 		nfp_prog->error = -ENOSPC;
95 		return;
96 	}
97 
98 	nfp_prog->prog[nfp_prog->prog_len] = insn;
99 	nfp_prog->prog_len++;
100 }
101 
102 static unsigned int nfp_prog_current_offset(struct nfp_prog *nfp_prog)
103 {
104 	return nfp_prog->start_off + nfp_prog->prog_len;
105 }
106 
107 static unsigned int
108 nfp_prog_offset_to_index(struct nfp_prog *nfp_prog, unsigned int offset)
109 {
110 	return offset - nfp_prog->start_off;
111 }
112 
113 /* --- SW reg --- */
114 struct nfp_insn_ur_regs {
115 	enum alu_dst_ab dst_ab;
116 	u16 dst;
117 	u16 areg, breg;
118 	bool swap;
119 	bool wr_both;
120 };
121 
122 struct nfp_insn_re_regs {
123 	enum alu_dst_ab dst_ab;
124 	u8 dst;
125 	u8 areg, breg;
126 	bool swap;
127 	bool wr_both;
128 	bool i8;
129 };
130 
131 static u16 nfp_swreg_to_unreg(u32 swreg, bool is_dst)
132 {
133 	u16 val = FIELD_GET(NN_REG_VAL, swreg);
134 
135 	switch (FIELD_GET(NN_REG_TYPE, swreg)) {
136 	case NN_REG_GPR_A:
137 	case NN_REG_GPR_B:
138 	case NN_REG_GPR_BOTH:
139 		return val;
140 	case NN_REG_NNR:
141 		return UR_REG_NN | val;
142 	case NN_REG_XFER:
143 		return UR_REG_XFR | val;
144 	case NN_REG_IMM:
145 		if (val & ~0xff) {
146 			pr_err("immediate too large\n");
147 			return 0;
148 		}
149 		return UR_REG_IMM_encode(val);
150 	case NN_REG_NONE:
151 		return is_dst ? UR_REG_NO_DST : REG_NONE;
152 	default:
153 		pr_err("unrecognized reg encoding %08x\n", swreg);
154 		return 0;
155 	}
156 }
157 
158 static int
159 swreg_to_unrestricted(u32 dst, u32 lreg, u32 rreg, struct nfp_insn_ur_regs *reg)
160 {
161 	memset(reg, 0, sizeof(*reg));
162 
163 	/* Decode destination */
164 	if (FIELD_GET(NN_REG_TYPE, dst) == NN_REG_IMM)
165 		return -EFAULT;
166 
167 	if (FIELD_GET(NN_REG_TYPE, dst) == NN_REG_GPR_B)
168 		reg->dst_ab = ALU_DST_B;
169 	if (FIELD_GET(NN_REG_TYPE, dst) == NN_REG_GPR_BOTH)
170 		reg->wr_both = true;
171 	reg->dst = nfp_swreg_to_unreg(dst, true);
172 
173 	/* Decode source operands */
174 	if (FIELD_GET(NN_REG_TYPE, lreg) == FIELD_GET(NN_REG_TYPE, rreg))
175 		return -EFAULT;
176 
177 	if (FIELD_GET(NN_REG_TYPE, lreg) == NN_REG_GPR_B ||
178 	    FIELD_GET(NN_REG_TYPE, rreg) == NN_REG_GPR_A) {
179 		reg->areg = nfp_swreg_to_unreg(rreg, false);
180 		reg->breg = nfp_swreg_to_unreg(lreg, false);
181 		reg->swap = true;
182 	} else {
183 		reg->areg = nfp_swreg_to_unreg(lreg, false);
184 		reg->breg = nfp_swreg_to_unreg(rreg, false);
185 	}
186 
187 	return 0;
188 }
189 
190 static u16 nfp_swreg_to_rereg(u32 swreg, bool is_dst, bool has_imm8, bool *i8)
191 {
192 	u16 val = FIELD_GET(NN_REG_VAL, swreg);
193 
194 	switch (FIELD_GET(NN_REG_TYPE, swreg)) {
195 	case NN_REG_GPR_A:
196 	case NN_REG_GPR_B:
197 	case NN_REG_GPR_BOTH:
198 		return val;
199 	case NN_REG_XFER:
200 		return RE_REG_XFR | val;
201 	case NN_REG_IMM:
202 		if (val & ~(0x7f | has_imm8 << 7)) {
203 			pr_err("immediate too large\n");
204 			return 0;
205 		}
206 		*i8 = val & 0x80;
207 		return RE_REG_IMM_encode(val & 0x7f);
208 	case NN_REG_NONE:
209 		return is_dst ? RE_REG_NO_DST : REG_NONE;
210 	default:
211 		pr_err("unrecognized reg encoding\n");
212 		return 0;
213 	}
214 }
215 
216 static int
217 swreg_to_restricted(u32 dst, u32 lreg, u32 rreg, struct nfp_insn_re_regs *reg,
218 		    bool has_imm8)
219 {
220 	memset(reg, 0, sizeof(*reg));
221 
222 	/* Decode destination */
223 	if (FIELD_GET(NN_REG_TYPE, dst) == NN_REG_IMM)
224 		return -EFAULT;
225 
226 	if (FIELD_GET(NN_REG_TYPE, dst) == NN_REG_GPR_B)
227 		reg->dst_ab = ALU_DST_B;
228 	if (FIELD_GET(NN_REG_TYPE, dst) == NN_REG_GPR_BOTH)
229 		reg->wr_both = true;
230 	reg->dst = nfp_swreg_to_rereg(dst, true, false, NULL);
231 
232 	/* Decode source operands */
233 	if (FIELD_GET(NN_REG_TYPE, lreg) == FIELD_GET(NN_REG_TYPE, rreg))
234 		return -EFAULT;
235 
236 	if (FIELD_GET(NN_REG_TYPE, lreg) == NN_REG_GPR_B ||
237 	    FIELD_GET(NN_REG_TYPE, rreg) == NN_REG_GPR_A) {
238 		reg->areg = nfp_swreg_to_rereg(rreg, false, has_imm8, &reg->i8);
239 		reg->breg = nfp_swreg_to_rereg(lreg, false, has_imm8, &reg->i8);
240 		reg->swap = true;
241 	} else {
242 		reg->areg = nfp_swreg_to_rereg(lreg, false, has_imm8, &reg->i8);
243 		reg->breg = nfp_swreg_to_rereg(rreg, false, has_imm8, &reg->i8);
244 	}
245 
246 	return 0;
247 }
248 
249 /* --- Emitters --- */
250 static const struct cmd_tgt_act cmd_tgt_act[__CMD_TGT_MAP_SIZE] = {
251 	[CMD_TGT_WRITE8] =		{ 0x00, 0x42 },
252 	[CMD_TGT_READ8] =		{ 0x01, 0x43 },
253 	[CMD_TGT_READ_LE] =		{ 0x01, 0x40 },
254 	[CMD_TGT_READ_SWAP_LE] =	{ 0x03, 0x40 },
255 };
256 
257 static void
258 __emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op,
259 	   u8 mode, u8 xfer, u8 areg, u8 breg, u8 size, bool sync)
260 {
261 	enum cmd_ctx_swap ctx;
262 	u64 insn;
263 
264 	if (sync)
265 		ctx = CMD_CTX_SWAP;
266 	else
267 		ctx = CMD_CTX_NO_SWAP;
268 
269 	insn =	FIELD_PREP(OP_CMD_A_SRC, areg) |
270 		FIELD_PREP(OP_CMD_CTX, ctx) |
271 		FIELD_PREP(OP_CMD_B_SRC, breg) |
272 		FIELD_PREP(OP_CMD_TOKEN, cmd_tgt_act[op].token) |
273 		FIELD_PREP(OP_CMD_XFER, xfer) |
274 		FIELD_PREP(OP_CMD_CNT, size) |
275 		FIELD_PREP(OP_CMD_SIG, sync) |
276 		FIELD_PREP(OP_CMD_TGT_CMD, cmd_tgt_act[op].tgt_cmd) |
277 		FIELD_PREP(OP_CMD_MODE, mode);
278 
279 	nfp_prog_push(nfp_prog, insn);
280 }
281 
282 static void
283 emit_cmd(struct nfp_prog *nfp_prog, enum cmd_tgt_map op,
284 	 u8 mode, u8 xfer, u32 lreg, u32 rreg, u8 size, bool sync)
285 {
286 	struct nfp_insn_re_regs reg;
287 	int err;
288 
289 	err = swreg_to_restricted(reg_none(), lreg, rreg, &reg, false);
290 	if (err) {
291 		nfp_prog->error = err;
292 		return;
293 	}
294 	if (reg.swap) {
295 		pr_err("cmd can't swap arguments\n");
296 		nfp_prog->error = -EFAULT;
297 		return;
298 	}
299 
300 	__emit_cmd(nfp_prog, op, mode, xfer, reg.areg, reg.breg, size, sync);
301 }
302 
303 static void
304 __emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, enum br_ev_pip ev_pip,
305 	  enum br_ctx_signal_state css, u16 addr, u8 defer)
306 {
307 	u16 addr_lo, addr_hi;
308 	u64 insn;
309 
310 	addr_lo = addr & (OP_BR_ADDR_LO >> __bf_shf(OP_BR_ADDR_LO));
311 	addr_hi = addr != addr_lo;
312 
313 	insn = OP_BR_BASE |
314 		FIELD_PREP(OP_BR_MASK, mask) |
315 		FIELD_PREP(OP_BR_EV_PIP, ev_pip) |
316 		FIELD_PREP(OP_BR_CSS, css) |
317 		FIELD_PREP(OP_BR_DEFBR, defer) |
318 		FIELD_PREP(OP_BR_ADDR_LO, addr_lo) |
319 		FIELD_PREP(OP_BR_ADDR_HI, addr_hi);
320 
321 	nfp_prog_push(nfp_prog, insn);
322 }
323 
324 static void emit_br_def(struct nfp_prog *nfp_prog, u16 addr, u8 defer)
325 {
326 	if (defer > 2) {
327 		pr_err("BUG: branch defer out of bounds %d\n", defer);
328 		nfp_prog->error = -EFAULT;
329 		return;
330 	}
331 	__emit_br(nfp_prog, BR_UNC, BR_EV_PIP_UNCOND, BR_CSS_NONE, addr, defer);
332 }
333 
334 static void
335 emit_br(struct nfp_prog *nfp_prog, enum br_mask mask, u16 addr, u8 defer)
336 {
337 	__emit_br(nfp_prog, mask,
338 		  mask != BR_UNC ? BR_EV_PIP_COND : BR_EV_PIP_UNCOND,
339 		  BR_CSS_NONE, addr, defer);
340 }
341 
342 static void
343 __emit_br_byte(struct nfp_prog *nfp_prog, u8 areg, u8 breg, bool imm8,
344 	       u8 byte, bool equal, u16 addr, u8 defer)
345 {
346 	u16 addr_lo, addr_hi;
347 	u64 insn;
348 
349 	addr_lo = addr & (OP_BB_ADDR_LO >> __bf_shf(OP_BB_ADDR_LO));
350 	addr_hi = addr != addr_lo;
351 
352 	insn = OP_BBYTE_BASE |
353 		FIELD_PREP(OP_BB_A_SRC, areg) |
354 		FIELD_PREP(OP_BB_BYTE, byte) |
355 		FIELD_PREP(OP_BB_B_SRC, breg) |
356 		FIELD_PREP(OP_BB_I8, imm8) |
357 		FIELD_PREP(OP_BB_EQ, equal) |
358 		FIELD_PREP(OP_BB_DEFBR, defer) |
359 		FIELD_PREP(OP_BB_ADDR_LO, addr_lo) |
360 		FIELD_PREP(OP_BB_ADDR_HI, addr_hi);
361 
362 	nfp_prog_push(nfp_prog, insn);
363 }
364 
365 static void
366 emit_br_byte_neq(struct nfp_prog *nfp_prog,
367 		 u32 dst, u8 imm, u8 byte, u16 addr, u8 defer)
368 {
369 	struct nfp_insn_re_regs reg;
370 	int err;
371 
372 	err = swreg_to_restricted(reg_none(), dst, reg_imm(imm), &reg, true);
373 	if (err) {
374 		nfp_prog->error = err;
375 		return;
376 	}
377 
378 	__emit_br_byte(nfp_prog, reg.areg, reg.breg, reg.i8, byte, false, addr,
379 		       defer);
380 }
381 
382 static void
383 __emit_immed(struct nfp_prog *nfp_prog, u16 areg, u16 breg, u16 imm_hi,
384 	     enum immed_width width, bool invert,
385 	     enum immed_shift shift, bool wr_both)
386 {
387 	u64 insn;
388 
389 	insn = OP_IMMED_BASE |
390 		FIELD_PREP(OP_IMMED_A_SRC, areg) |
391 		FIELD_PREP(OP_IMMED_B_SRC, breg) |
392 		FIELD_PREP(OP_IMMED_IMM, imm_hi) |
393 		FIELD_PREP(OP_IMMED_WIDTH, width) |
394 		FIELD_PREP(OP_IMMED_INV, invert) |
395 		FIELD_PREP(OP_IMMED_SHIFT, shift) |
396 		FIELD_PREP(OP_IMMED_WR_AB, wr_both);
397 
398 	nfp_prog_push(nfp_prog, insn);
399 }
400 
401 static void
402 emit_immed(struct nfp_prog *nfp_prog, u32 dst, u16 imm,
403 	   enum immed_width width, bool invert, enum immed_shift shift)
404 {
405 	struct nfp_insn_ur_regs reg;
406 	int err;
407 
408 	if (FIELD_GET(NN_REG_TYPE, dst) == NN_REG_IMM) {
409 		nfp_prog->error = -EFAULT;
410 		return;
411 	}
412 
413 	err = swreg_to_unrestricted(dst, dst, reg_imm(imm & 0xff), &reg);
414 	if (err) {
415 		nfp_prog->error = err;
416 		return;
417 	}
418 
419 	__emit_immed(nfp_prog, reg.areg, reg.breg, imm >> 8, width,
420 		     invert, shift, reg.wr_both);
421 }
422 
423 static void
424 __emit_shf(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab,
425 	   enum shf_sc sc, u8 shift,
426 	   u16 areg, enum shf_op op, u16 breg, bool i8, bool sw, bool wr_both)
427 {
428 	u64 insn;
429 
430 	if (!FIELD_FIT(OP_SHF_SHIFT, shift)) {
431 		nfp_prog->error = -EFAULT;
432 		return;
433 	}
434 
435 	if (sc == SHF_SC_L_SHF)
436 		shift = 32 - shift;
437 
438 	insn = OP_SHF_BASE |
439 		FIELD_PREP(OP_SHF_A_SRC, areg) |
440 		FIELD_PREP(OP_SHF_SC, sc) |
441 		FIELD_PREP(OP_SHF_B_SRC, breg) |
442 		FIELD_PREP(OP_SHF_I8, i8) |
443 		FIELD_PREP(OP_SHF_SW, sw) |
444 		FIELD_PREP(OP_SHF_DST, dst) |
445 		FIELD_PREP(OP_SHF_SHIFT, shift) |
446 		FIELD_PREP(OP_SHF_OP, op) |
447 		FIELD_PREP(OP_SHF_DST_AB, dst_ab) |
448 		FIELD_PREP(OP_SHF_WR_AB, wr_both);
449 
450 	nfp_prog_push(nfp_prog, insn);
451 }
452 
453 static void
454 emit_shf(struct nfp_prog *nfp_prog, u32 dst, u32 lreg, enum shf_op op, u32 rreg,
455 	 enum shf_sc sc, u8 shift)
456 {
457 	struct nfp_insn_re_regs reg;
458 	int err;
459 
460 	err = swreg_to_restricted(dst, lreg, rreg, &reg, true);
461 	if (err) {
462 		nfp_prog->error = err;
463 		return;
464 	}
465 
466 	__emit_shf(nfp_prog, reg.dst, reg.dst_ab, sc, shift,
467 		   reg.areg, op, reg.breg, reg.i8, reg.swap, reg.wr_both);
468 }
469 
470 static void
471 __emit_alu(struct nfp_prog *nfp_prog, u16 dst, enum alu_dst_ab dst_ab,
472 	   u16 areg, enum alu_op op, u16 breg, bool swap, bool wr_both)
473 {
474 	u64 insn;
475 
476 	insn = OP_ALU_BASE |
477 		FIELD_PREP(OP_ALU_A_SRC, areg) |
478 		FIELD_PREP(OP_ALU_B_SRC, breg) |
479 		FIELD_PREP(OP_ALU_DST, dst) |
480 		FIELD_PREP(OP_ALU_SW, swap) |
481 		FIELD_PREP(OP_ALU_OP, op) |
482 		FIELD_PREP(OP_ALU_DST_AB, dst_ab) |
483 		FIELD_PREP(OP_ALU_WR_AB, wr_both);
484 
485 	nfp_prog_push(nfp_prog, insn);
486 }
487 
488 static void
489 emit_alu(struct nfp_prog *nfp_prog, u32 dst, u32 lreg, enum alu_op op, u32 rreg)
490 {
491 	struct nfp_insn_ur_regs reg;
492 	int err;
493 
494 	err = swreg_to_unrestricted(dst, lreg, rreg, &reg);
495 	if (err) {
496 		nfp_prog->error = err;
497 		return;
498 	}
499 
500 	__emit_alu(nfp_prog, reg.dst, reg.dst_ab,
501 		   reg.areg, op, reg.breg, reg.swap, reg.wr_both);
502 }
503 
504 static void
505 __emit_ld_field(struct nfp_prog *nfp_prog, enum shf_sc sc,
506 		u8 areg, u8 bmask, u8 breg, u8 shift, bool imm8,
507 		bool zero, bool swap, bool wr_both)
508 {
509 	u64 insn;
510 
511 	insn = OP_LDF_BASE |
512 		FIELD_PREP(OP_LDF_A_SRC, areg) |
513 		FIELD_PREP(OP_LDF_SC, sc) |
514 		FIELD_PREP(OP_LDF_B_SRC, breg) |
515 		FIELD_PREP(OP_LDF_I8, imm8) |
516 		FIELD_PREP(OP_LDF_SW, swap) |
517 		FIELD_PREP(OP_LDF_ZF, zero) |
518 		FIELD_PREP(OP_LDF_BMASK, bmask) |
519 		FIELD_PREP(OP_LDF_SHF, shift) |
520 		FIELD_PREP(OP_LDF_WR_AB, wr_both);
521 
522 	nfp_prog_push(nfp_prog, insn);
523 }
524 
525 static void
526 emit_ld_field_any(struct nfp_prog *nfp_prog, enum shf_sc sc, u8 shift,
527 		  u32 dst, u8 bmask, u32 src, bool zero)
528 {
529 	struct nfp_insn_re_regs reg;
530 	int err;
531 
532 	err = swreg_to_restricted(reg_none(), dst, src, &reg, true);
533 	if (err) {
534 		nfp_prog->error = err;
535 		return;
536 	}
537 
538 	__emit_ld_field(nfp_prog, sc, reg.areg, bmask, reg.breg, shift,
539 			reg.i8, zero, reg.swap, reg.wr_both);
540 }
541 
542 static void
543 emit_ld_field(struct nfp_prog *nfp_prog, u32 dst, u8 bmask, u32 src,
544 	      enum shf_sc sc, u8 shift)
545 {
546 	emit_ld_field_any(nfp_prog, sc, shift, dst, bmask, src, false);
547 }
548 
549 /* --- Wrappers --- */
550 static bool pack_immed(u32 imm, u16 *val, enum immed_shift *shift)
551 {
552 	if (!(imm & 0xffff0000)) {
553 		*val = imm;
554 		*shift = IMMED_SHIFT_0B;
555 	} else if (!(imm & 0xff0000ff)) {
556 		*val = imm >> 8;
557 		*shift = IMMED_SHIFT_1B;
558 	} else if (!(imm & 0x0000ffff)) {
559 		*val = imm >> 16;
560 		*shift = IMMED_SHIFT_2B;
561 	} else {
562 		return false;
563 	}
564 
565 	return true;
566 }
567 
568 static void wrp_immed(struct nfp_prog *nfp_prog, u32 dst, u32 imm)
569 {
570 	enum immed_shift shift;
571 	u16 val;
572 
573 	if (pack_immed(imm, &val, &shift)) {
574 		emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, false, shift);
575 	} else if (pack_immed(~imm, &val, &shift)) {
576 		emit_immed(nfp_prog, dst, val, IMMED_WIDTH_ALL, true, shift);
577 	} else {
578 		emit_immed(nfp_prog, dst, imm & 0xffff, IMMED_WIDTH_ALL,
579 			   false, IMMED_SHIFT_0B);
580 		emit_immed(nfp_prog, dst, imm >> 16, IMMED_WIDTH_WORD,
581 			   false, IMMED_SHIFT_2B);
582 	}
583 }
584 
585 /* ur_load_imm_any() - encode immediate or use tmp register (unrestricted)
586  * If the @imm is small enough encode it directly in operand and return
587  * otherwise load @imm to a spare register and return its encoding.
588  */
589 static u32 ur_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, u32 tmp_reg)
590 {
591 	if (FIELD_FIT(UR_REG_IMM_MAX, imm))
592 		return reg_imm(imm);
593 
594 	wrp_immed(nfp_prog, tmp_reg, imm);
595 	return tmp_reg;
596 }
597 
598 /* re_load_imm_any() - encode immediate or use tmp register (restricted)
599  * If the @imm is small enough encode it directly in operand and return
600  * otherwise load @imm to a spare register and return its encoding.
601  */
602 static u32 re_load_imm_any(struct nfp_prog *nfp_prog, u32 imm, u32 tmp_reg)
603 {
604 	if (FIELD_FIT(RE_REG_IMM_MAX, imm))
605 		return reg_imm(imm);
606 
607 	wrp_immed(nfp_prog, tmp_reg, imm);
608 	return tmp_reg;
609 }
610 
611 static void
612 wrp_br_special(struct nfp_prog *nfp_prog, enum br_mask mask,
613 	       enum br_special special)
614 {
615 	emit_br(nfp_prog, mask, 0, 0);
616 
617 	nfp_prog->prog[nfp_prog->prog_len - 1] |=
618 		FIELD_PREP(OP_BR_SPECIAL, special);
619 }
620 
621 static void wrp_reg_mov(struct nfp_prog *nfp_prog, u16 dst, u16 src)
622 {
623 	emit_alu(nfp_prog, reg_both(dst), reg_none(), ALU_OP_NONE, reg_b(src));
624 }
625 
626 static int
627 construct_data_ind_ld(struct nfp_prog *nfp_prog, u16 offset,
628 		      u16 src, bool src_valid, u8 size)
629 {
630 	unsigned int i;
631 	u16 shift, sz;
632 	u32 tmp_reg;
633 
634 	/* We load the value from the address indicated in @offset and then
635 	 * shift out the data we don't need.  Note: this is big endian!
636 	 */
637 	sz = size < 4 ? 4 : size;
638 	shift = size < 4 ? 4 - size : 0;
639 
640 	if (src_valid) {
641 		/* Calculate the true offset (src_reg + imm) */
642 		tmp_reg = ur_load_imm_any(nfp_prog, offset, imm_b(nfp_prog));
643 		emit_alu(nfp_prog, imm_both(nfp_prog),
644 			 reg_a(src), ALU_OP_ADD, tmp_reg);
645 		/* Check packet length (size guaranteed to fit b/c it's u8) */
646 		emit_alu(nfp_prog, imm_a(nfp_prog),
647 			 imm_a(nfp_prog), ALU_OP_ADD, reg_imm(size));
648 		emit_alu(nfp_prog, reg_none(),
649 			 NFP_BPF_ABI_LEN, ALU_OP_SUB, imm_a(nfp_prog));
650 		wrp_br_special(nfp_prog, BR_BLO, OP_BR_GO_ABORT);
651 		/* Load data */
652 		emit_cmd(nfp_prog, CMD_TGT_READ8, CMD_MODE_32b, 0,
653 			 pkt_reg(nfp_prog), imm_b(nfp_prog), sz - 1, true);
654 	} else {
655 		/* Check packet length */
656 		tmp_reg = ur_load_imm_any(nfp_prog, offset + size,
657 					  imm_a(nfp_prog));
658 		emit_alu(nfp_prog, reg_none(),
659 			 NFP_BPF_ABI_LEN, ALU_OP_SUB, tmp_reg);
660 		wrp_br_special(nfp_prog, BR_BLO, OP_BR_GO_ABORT);
661 		/* Load data */
662 		tmp_reg = re_load_imm_any(nfp_prog, offset, imm_b(nfp_prog));
663 		emit_cmd(nfp_prog, CMD_TGT_READ8, CMD_MODE_32b, 0,
664 			 pkt_reg(nfp_prog), tmp_reg, sz - 1, true);
665 	}
666 
667 	i = 0;
668 	if (shift)
669 		emit_shf(nfp_prog, reg_both(0), reg_none(), SHF_OP_NONE,
670 			 reg_xfer(0), SHF_SC_R_SHF, shift * 8);
671 	else
672 		for (; i * 4 < size; i++)
673 			emit_alu(nfp_prog, reg_both(i),
674 				 reg_none(), ALU_OP_NONE, reg_xfer(i));
675 
676 	if (i < 2)
677 		wrp_immed(nfp_prog, reg_both(1), 0);
678 
679 	return 0;
680 }
681 
682 static int construct_data_ld(struct nfp_prog *nfp_prog, u16 offset, u8 size)
683 {
684 	return construct_data_ind_ld(nfp_prog, offset, 0, false, size);
685 }
686 
687 static int wrp_set_mark(struct nfp_prog *nfp_prog, u8 src)
688 {
689 	emit_alu(nfp_prog, NFP_BPF_ABI_MARK,
690 		 reg_none(), ALU_OP_NONE, reg_b(src));
691 	emit_alu(nfp_prog, NFP_BPF_ABI_FLAGS,
692 		 NFP_BPF_ABI_FLAGS, ALU_OP_OR, reg_imm(NFP_BPF_ABI_FLAG_MARK));
693 
694 	return 0;
695 }
696 
697 static void
698 wrp_alu_imm(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u32 imm)
699 {
700 	u32 tmp_reg;
701 
702 	if (alu_op == ALU_OP_AND) {
703 		if (!imm)
704 			wrp_immed(nfp_prog, reg_both(dst), 0);
705 		if (!imm || !~imm)
706 			return;
707 	}
708 	if (alu_op == ALU_OP_OR) {
709 		if (!~imm)
710 			wrp_immed(nfp_prog, reg_both(dst), ~0U);
711 		if (!imm || !~imm)
712 			return;
713 	}
714 	if (alu_op == ALU_OP_XOR) {
715 		if (!~imm)
716 			emit_alu(nfp_prog, reg_both(dst), reg_none(),
717 				 ALU_OP_NEG, reg_b(dst));
718 		if (!imm || !~imm)
719 			return;
720 	}
721 
722 	tmp_reg = ur_load_imm_any(nfp_prog, imm, imm_b(nfp_prog));
723 	emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, tmp_reg);
724 }
725 
726 static int
727 wrp_alu64_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
728 	      enum alu_op alu_op, bool skip)
729 {
730 	const struct bpf_insn *insn = &meta->insn;
731 	u64 imm = insn->imm; /* sign extend */
732 
733 	if (skip) {
734 		meta->skip = true;
735 		return 0;
736 	}
737 
738 	wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, imm & ~0U);
739 	wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, alu_op, imm >> 32);
740 
741 	return 0;
742 }
743 
744 static int
745 wrp_alu64_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
746 	      enum alu_op alu_op)
747 {
748 	u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2;
749 
750 	emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src));
751 	emit_alu(nfp_prog, reg_both(dst + 1),
752 		 reg_a(dst + 1), alu_op, reg_b(src + 1));
753 
754 	return 0;
755 }
756 
757 static int
758 wrp_alu32_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
759 	      enum alu_op alu_op, bool skip)
760 {
761 	const struct bpf_insn *insn = &meta->insn;
762 
763 	if (skip) {
764 		meta->skip = true;
765 		return 0;
766 	}
767 
768 	wrp_alu_imm(nfp_prog, insn->dst_reg * 2, alu_op, insn->imm);
769 	wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0);
770 
771 	return 0;
772 }
773 
774 static int
775 wrp_alu32_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
776 	      enum alu_op alu_op)
777 {
778 	u8 dst = meta->insn.dst_reg * 2, src = meta->insn.src_reg * 2;
779 
780 	emit_alu(nfp_prog, reg_both(dst), reg_a(dst), alu_op, reg_b(src));
781 	wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0);
782 
783 	return 0;
784 }
785 
786 static void
787 wrp_test_reg_one(struct nfp_prog *nfp_prog, u8 dst, enum alu_op alu_op, u8 src,
788 		 enum br_mask br_mask, u16 off)
789 {
790 	emit_alu(nfp_prog, reg_none(), reg_a(dst), alu_op, reg_b(src));
791 	emit_br(nfp_prog, br_mask, off, 0);
792 }
793 
794 static int
795 wrp_test_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
796 	     enum alu_op alu_op, enum br_mask br_mask)
797 {
798 	const struct bpf_insn *insn = &meta->insn;
799 
800 	if (insn->off < 0) /* TODO */
801 		return -EOPNOTSUPP;
802 
803 	wrp_test_reg_one(nfp_prog, insn->dst_reg * 2, alu_op,
804 			 insn->src_reg * 2, br_mask, insn->off);
805 	wrp_test_reg_one(nfp_prog, insn->dst_reg * 2 + 1, alu_op,
806 			 insn->src_reg * 2 + 1, br_mask, insn->off);
807 
808 	return 0;
809 }
810 
811 static int
812 wrp_cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
813 	    enum br_mask br_mask, bool swap)
814 {
815 	const struct bpf_insn *insn = &meta->insn;
816 	u64 imm = insn->imm; /* sign extend */
817 	u8 reg = insn->dst_reg * 2;
818 	u32 tmp_reg;
819 
820 	if (insn->off < 0) /* TODO */
821 		return -EOPNOTSUPP;
822 
823 	tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
824 	if (!swap)
825 		emit_alu(nfp_prog, reg_none(), reg_a(reg), ALU_OP_SUB, tmp_reg);
826 	else
827 		emit_alu(nfp_prog, reg_none(), tmp_reg, ALU_OP_SUB, reg_a(reg));
828 
829 	tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog));
830 	if (!swap)
831 		emit_alu(nfp_prog, reg_none(),
832 			 reg_a(reg + 1), ALU_OP_SUB_C, tmp_reg);
833 	else
834 		emit_alu(nfp_prog, reg_none(),
835 			 tmp_reg, ALU_OP_SUB_C, reg_a(reg + 1));
836 
837 	emit_br(nfp_prog, br_mask, insn->off, 0);
838 
839 	return 0;
840 }
841 
842 static int
843 wrp_cmp_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
844 	    enum br_mask br_mask, bool swap)
845 {
846 	const struct bpf_insn *insn = &meta->insn;
847 	u8 areg = insn->src_reg * 2, breg = insn->dst_reg * 2;
848 
849 	if (insn->off < 0) /* TODO */
850 		return -EOPNOTSUPP;
851 
852 	if (swap) {
853 		areg ^= breg;
854 		breg ^= areg;
855 		areg ^= breg;
856 	}
857 
858 	emit_alu(nfp_prog, reg_none(), reg_a(areg), ALU_OP_SUB, reg_b(breg));
859 	emit_alu(nfp_prog, reg_none(),
860 		 reg_a(areg + 1), ALU_OP_SUB_C, reg_b(breg + 1));
861 	emit_br(nfp_prog, br_mask, insn->off, 0);
862 
863 	return 0;
864 }
865 
866 /* --- Callbacks --- */
867 static int mov_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
868 {
869 	const struct bpf_insn *insn = &meta->insn;
870 
871 	wrp_reg_mov(nfp_prog, insn->dst_reg * 2, insn->src_reg * 2);
872 	wrp_reg_mov(nfp_prog, insn->dst_reg * 2 + 1, insn->src_reg * 2 + 1);
873 
874 	return 0;
875 }
876 
877 static int mov_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
878 {
879 	u64 imm = meta->insn.imm; /* sign extend */
880 
881 	wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2), imm & ~0U);
882 	wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), imm >> 32);
883 
884 	return 0;
885 }
886 
887 static int xor_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
888 {
889 	return wrp_alu64_reg(nfp_prog, meta, ALU_OP_XOR);
890 }
891 
892 static int xor_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
893 {
894 	return wrp_alu64_imm(nfp_prog, meta, ALU_OP_XOR, !meta->insn.imm);
895 }
896 
897 static int and_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
898 {
899 	return wrp_alu64_reg(nfp_prog, meta, ALU_OP_AND);
900 }
901 
902 static int and_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
903 {
904 	return wrp_alu64_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm);
905 }
906 
907 static int or_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
908 {
909 	return wrp_alu64_reg(nfp_prog, meta, ALU_OP_OR);
910 }
911 
912 static int or_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
913 {
914 	return wrp_alu64_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm);
915 }
916 
917 static int add_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
918 {
919 	const struct bpf_insn *insn = &meta->insn;
920 
921 	emit_alu(nfp_prog, reg_both(insn->dst_reg * 2),
922 		 reg_a(insn->dst_reg * 2), ALU_OP_ADD,
923 		 reg_b(insn->src_reg * 2));
924 	emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1),
925 		 reg_a(insn->dst_reg * 2 + 1), ALU_OP_ADD_C,
926 		 reg_b(insn->src_reg * 2 + 1));
927 
928 	return 0;
929 }
930 
931 static int add_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
932 {
933 	const struct bpf_insn *insn = &meta->insn;
934 	u64 imm = insn->imm; /* sign extend */
935 
936 	wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_ADD, imm & ~0U);
937 	wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_ADD_C, imm >> 32);
938 
939 	return 0;
940 }
941 
942 static int sub_reg64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
943 {
944 	const struct bpf_insn *insn = &meta->insn;
945 
946 	emit_alu(nfp_prog, reg_both(insn->dst_reg * 2),
947 		 reg_a(insn->dst_reg * 2), ALU_OP_SUB,
948 		 reg_b(insn->src_reg * 2));
949 	emit_alu(nfp_prog, reg_both(insn->dst_reg * 2 + 1),
950 		 reg_a(insn->dst_reg * 2 + 1), ALU_OP_SUB_C,
951 		 reg_b(insn->src_reg * 2 + 1));
952 
953 	return 0;
954 }
955 
956 static int sub_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
957 {
958 	const struct bpf_insn *insn = &meta->insn;
959 	u64 imm = insn->imm; /* sign extend */
960 
961 	wrp_alu_imm(nfp_prog, insn->dst_reg * 2, ALU_OP_SUB, imm & ~0U);
962 	wrp_alu_imm(nfp_prog, insn->dst_reg * 2 + 1, ALU_OP_SUB_C, imm >> 32);
963 
964 	return 0;
965 }
966 
967 static int shl_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
968 {
969 	const struct bpf_insn *insn = &meta->insn;
970 
971 	if (insn->imm != 32)
972 		return 1; /* TODO */
973 
974 	wrp_reg_mov(nfp_prog, insn->dst_reg * 2 + 1, insn->dst_reg * 2);
975 	wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2), 0);
976 
977 	return 0;
978 }
979 
980 static int shr_imm64(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
981 {
982 	const struct bpf_insn *insn = &meta->insn;
983 
984 	if (insn->imm != 32)
985 		return 1; /* TODO */
986 
987 	wrp_reg_mov(nfp_prog, insn->dst_reg * 2, insn->dst_reg * 2 + 1);
988 	wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0);
989 
990 	return 0;
991 }
992 
993 static int mov_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
994 {
995 	const struct bpf_insn *insn = &meta->insn;
996 
997 	wrp_reg_mov(nfp_prog, insn->dst_reg * 2,  insn->src_reg * 2);
998 	wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0);
999 
1000 	return 0;
1001 }
1002 
1003 static int mov_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1004 {
1005 	const struct bpf_insn *insn = &meta->insn;
1006 
1007 	wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2), insn->imm);
1008 	wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0);
1009 
1010 	return 0;
1011 }
1012 
1013 static int xor_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1014 {
1015 	return wrp_alu32_reg(nfp_prog, meta, ALU_OP_XOR);
1016 }
1017 
1018 static int xor_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1019 {
1020 	return wrp_alu32_imm(nfp_prog, meta, ALU_OP_XOR, !~meta->insn.imm);
1021 }
1022 
1023 static int and_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1024 {
1025 	return wrp_alu32_reg(nfp_prog, meta, ALU_OP_AND);
1026 }
1027 
1028 static int and_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1029 {
1030 	return wrp_alu32_imm(nfp_prog, meta, ALU_OP_AND, !~meta->insn.imm);
1031 }
1032 
1033 static int or_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1034 {
1035 	return wrp_alu32_reg(nfp_prog, meta, ALU_OP_OR);
1036 }
1037 
1038 static int or_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1039 {
1040 	return wrp_alu32_imm(nfp_prog, meta, ALU_OP_OR, !meta->insn.imm);
1041 }
1042 
1043 static int add_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1044 {
1045 	return wrp_alu32_reg(nfp_prog, meta, ALU_OP_ADD);
1046 }
1047 
1048 static int add_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1049 {
1050 	return wrp_alu32_imm(nfp_prog, meta, ALU_OP_ADD, !meta->insn.imm);
1051 }
1052 
1053 static int sub_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1054 {
1055 	return wrp_alu32_reg(nfp_prog, meta, ALU_OP_SUB);
1056 }
1057 
1058 static int sub_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1059 {
1060 	return wrp_alu32_imm(nfp_prog, meta, ALU_OP_SUB, !meta->insn.imm);
1061 }
1062 
1063 static int shl_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1064 {
1065 	const struct bpf_insn *insn = &meta->insn;
1066 
1067 	if (!insn->imm)
1068 		return 1; /* TODO: zero shift means indirect */
1069 
1070 	emit_shf(nfp_prog, reg_both(insn->dst_reg * 2),
1071 		 reg_none(), SHF_OP_NONE, reg_b(insn->dst_reg * 2),
1072 		 SHF_SC_L_SHF, insn->imm);
1073 	wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2 + 1), 0);
1074 
1075 	return 0;
1076 }
1077 
1078 static int imm_ld8_part2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1079 {
1080 	wrp_immed(nfp_prog, reg_both(nfp_meta_prev(meta)->insn.dst_reg * 2 + 1),
1081 		  meta->insn.imm);
1082 
1083 	return 0;
1084 }
1085 
1086 static int imm_ld8(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1087 {
1088 	const struct bpf_insn *insn = &meta->insn;
1089 
1090 	meta->double_cb = imm_ld8_part2;
1091 	wrp_immed(nfp_prog, reg_both(insn->dst_reg * 2), insn->imm);
1092 
1093 	return 0;
1094 }
1095 
1096 static int data_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1097 {
1098 	return construct_data_ld(nfp_prog, meta->insn.imm, 1);
1099 }
1100 
1101 static int data_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1102 {
1103 	return construct_data_ld(nfp_prog, meta->insn.imm, 2);
1104 }
1105 
1106 static int data_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1107 {
1108 	return construct_data_ld(nfp_prog, meta->insn.imm, 4);
1109 }
1110 
1111 static int data_ind_ld1(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1112 {
1113 	return construct_data_ind_ld(nfp_prog, meta->insn.imm,
1114 				     meta->insn.src_reg * 2, true, 1);
1115 }
1116 
1117 static int data_ind_ld2(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1118 {
1119 	return construct_data_ind_ld(nfp_prog, meta->insn.imm,
1120 				     meta->insn.src_reg * 2, true, 2);
1121 }
1122 
1123 static int data_ind_ld4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1124 {
1125 	return construct_data_ind_ld(nfp_prog, meta->insn.imm,
1126 				     meta->insn.src_reg * 2, true, 4);
1127 }
1128 
1129 static int mem_ldx4_skb(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1130 {
1131 	if (meta->insn.off == offsetof(struct sk_buff, len))
1132 		emit_alu(nfp_prog, reg_both(meta->insn.dst_reg * 2),
1133 			 reg_none(), ALU_OP_NONE, NFP_BPF_ABI_LEN);
1134 	else
1135 		return -EOPNOTSUPP;
1136 
1137 	return 0;
1138 }
1139 
1140 static int mem_ldx4_xdp(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1141 {
1142 	u32 dst = reg_both(meta->insn.dst_reg * 2);
1143 
1144 	if (meta->insn.off != offsetof(struct xdp_md, data) &&
1145 	    meta->insn.off != offsetof(struct xdp_md, data_end))
1146 		return -EOPNOTSUPP;
1147 
1148 	emit_alu(nfp_prog, dst, reg_none(), ALU_OP_NONE, NFP_BPF_ABI_PKT);
1149 
1150 	if (meta->insn.off == offsetof(struct xdp_md, data))
1151 		return 0;
1152 
1153 	emit_alu(nfp_prog, dst,	dst, ALU_OP_ADD, NFP_BPF_ABI_LEN);
1154 
1155 	return 0;
1156 }
1157 
1158 static int mem_ldx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1159 {
1160 	int ret;
1161 
1162 	if (nfp_prog->act == NN_ACT_XDP)
1163 		ret = mem_ldx4_xdp(nfp_prog, meta);
1164 	else
1165 		ret = mem_ldx4_skb(nfp_prog, meta);
1166 
1167 	wrp_immed(nfp_prog, reg_both(meta->insn.dst_reg * 2 + 1), 0);
1168 
1169 	return ret;
1170 }
1171 
1172 static int mem_stx4_skb(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1173 {
1174 	if (meta->insn.off == offsetof(struct sk_buff, mark))
1175 		return wrp_set_mark(nfp_prog, meta->insn.src_reg * 2);
1176 
1177 	return -EOPNOTSUPP;
1178 }
1179 
1180 static int mem_stx4_xdp(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1181 {
1182 	return -EOPNOTSUPP;
1183 }
1184 
1185 static int mem_stx4(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1186 {
1187 	if (nfp_prog->act == NN_ACT_XDP)
1188 		return mem_stx4_xdp(nfp_prog, meta);
1189 	return mem_stx4_skb(nfp_prog, meta);
1190 }
1191 
1192 static int jump(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1193 {
1194 	if (meta->insn.off < 0) /* TODO */
1195 		return -EOPNOTSUPP;
1196 	emit_br(nfp_prog, BR_UNC, meta->insn.off, 0);
1197 
1198 	return 0;
1199 }
1200 
1201 static int jeq_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1202 {
1203 	const struct bpf_insn *insn = &meta->insn;
1204 	u64 imm = insn->imm; /* sign extend */
1205 	u32 or1 = reg_a(insn->dst_reg * 2), or2 = reg_b(insn->dst_reg * 2 + 1);
1206 	u32 tmp_reg;
1207 
1208 	if (insn->off < 0) /* TODO */
1209 		return -EOPNOTSUPP;
1210 
1211 	if (imm & ~0U) {
1212 		tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
1213 		emit_alu(nfp_prog, imm_a(nfp_prog),
1214 			 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg);
1215 		or1 = imm_a(nfp_prog);
1216 	}
1217 
1218 	if (imm >> 32) {
1219 		tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog));
1220 		emit_alu(nfp_prog, imm_b(nfp_prog),
1221 			 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg);
1222 		or2 = imm_b(nfp_prog);
1223 	}
1224 
1225 	emit_alu(nfp_prog, reg_none(), or1, ALU_OP_OR, or2);
1226 	emit_br(nfp_prog, BR_BEQ, insn->off, 0);
1227 
1228 	return 0;
1229 }
1230 
1231 static int jgt_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1232 {
1233 	return wrp_cmp_imm(nfp_prog, meta, BR_BLO, false);
1234 }
1235 
1236 static int jge_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1237 {
1238 	return wrp_cmp_imm(nfp_prog, meta, BR_BHS, true);
1239 }
1240 
1241 static int jlt_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1242 {
1243 	return wrp_cmp_imm(nfp_prog, meta, BR_BHS, false);
1244 }
1245 
1246 static int jle_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1247 {
1248 	return wrp_cmp_imm(nfp_prog, meta, BR_BLO, true);
1249 }
1250 
1251 static int jset_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1252 {
1253 	const struct bpf_insn *insn = &meta->insn;
1254 	u64 imm = insn->imm; /* sign extend */
1255 	u32 tmp_reg;
1256 
1257 	if (insn->off < 0) /* TODO */
1258 		return -EOPNOTSUPP;
1259 
1260 	if (!imm) {
1261 		meta->skip = true;
1262 		return 0;
1263 	}
1264 
1265 	if (imm & ~0U) {
1266 		tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
1267 		emit_alu(nfp_prog, reg_none(),
1268 			 reg_a(insn->dst_reg * 2), ALU_OP_AND, tmp_reg);
1269 		emit_br(nfp_prog, BR_BNE, insn->off, 0);
1270 	}
1271 
1272 	if (imm >> 32) {
1273 		tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog));
1274 		emit_alu(nfp_prog, reg_none(),
1275 			 reg_a(insn->dst_reg * 2 + 1), ALU_OP_AND, tmp_reg);
1276 		emit_br(nfp_prog, BR_BNE, insn->off, 0);
1277 	}
1278 
1279 	return 0;
1280 }
1281 
1282 static int jne_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1283 {
1284 	const struct bpf_insn *insn = &meta->insn;
1285 	u64 imm = insn->imm; /* sign extend */
1286 	u32 tmp_reg;
1287 
1288 	if (insn->off < 0) /* TODO */
1289 		return -EOPNOTSUPP;
1290 
1291 	if (!imm) {
1292 		emit_alu(nfp_prog, reg_none(), reg_a(insn->dst_reg * 2),
1293 			 ALU_OP_OR, reg_b(insn->dst_reg * 2 + 1));
1294 		emit_br(nfp_prog, BR_BNE, insn->off, 0);
1295 	}
1296 
1297 	tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
1298 	emit_alu(nfp_prog, reg_none(),
1299 		 reg_a(insn->dst_reg * 2), ALU_OP_XOR, tmp_reg);
1300 	emit_br(nfp_prog, BR_BNE, insn->off, 0);
1301 
1302 	tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog));
1303 	emit_alu(nfp_prog, reg_none(),
1304 		 reg_a(insn->dst_reg * 2 + 1), ALU_OP_XOR, tmp_reg);
1305 	emit_br(nfp_prog, BR_BNE, insn->off, 0);
1306 
1307 	return 0;
1308 }
1309 
1310 static int jeq_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1311 {
1312 	const struct bpf_insn *insn = &meta->insn;
1313 
1314 	if (insn->off < 0) /* TODO */
1315 		return -EOPNOTSUPP;
1316 
1317 	emit_alu(nfp_prog, imm_a(nfp_prog), reg_a(insn->dst_reg * 2),
1318 		 ALU_OP_XOR, reg_b(insn->src_reg * 2));
1319 	emit_alu(nfp_prog, imm_b(nfp_prog), reg_a(insn->dst_reg * 2 + 1),
1320 		 ALU_OP_XOR, reg_b(insn->src_reg * 2 + 1));
1321 	emit_alu(nfp_prog, reg_none(),
1322 		 imm_a(nfp_prog), ALU_OP_OR, imm_b(nfp_prog));
1323 	emit_br(nfp_prog, BR_BEQ, insn->off, 0);
1324 
1325 	return 0;
1326 }
1327 
1328 static int jgt_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1329 {
1330 	return wrp_cmp_reg(nfp_prog, meta, BR_BLO, false);
1331 }
1332 
1333 static int jge_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1334 {
1335 	return wrp_cmp_reg(nfp_prog, meta, BR_BHS, true);
1336 }
1337 
1338 static int jlt_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1339 {
1340 	return wrp_cmp_reg(nfp_prog, meta, BR_BHS, false);
1341 }
1342 
1343 static int jle_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1344 {
1345 	return wrp_cmp_reg(nfp_prog, meta, BR_BLO, true);
1346 }
1347 
1348 static int jset_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1349 {
1350 	return wrp_test_reg(nfp_prog, meta, ALU_OP_AND, BR_BNE);
1351 }
1352 
1353 static int jne_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1354 {
1355 	return wrp_test_reg(nfp_prog, meta, ALU_OP_XOR, BR_BNE);
1356 }
1357 
1358 static int goto_out(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
1359 {
1360 	wrp_br_special(nfp_prog, BR_UNC, OP_BR_GO_OUT);
1361 
1362 	return 0;
1363 }
1364 
1365 static const instr_cb_t instr_cb[256] = {
1366 	[BPF_ALU64 | BPF_MOV | BPF_X] =	mov_reg64,
1367 	[BPF_ALU64 | BPF_MOV | BPF_K] =	mov_imm64,
1368 	[BPF_ALU64 | BPF_XOR | BPF_X] =	xor_reg64,
1369 	[BPF_ALU64 | BPF_XOR | BPF_K] =	xor_imm64,
1370 	[BPF_ALU64 | BPF_AND | BPF_X] =	and_reg64,
1371 	[BPF_ALU64 | BPF_AND | BPF_K] =	and_imm64,
1372 	[BPF_ALU64 | BPF_OR | BPF_X] =	or_reg64,
1373 	[BPF_ALU64 | BPF_OR | BPF_K] =	or_imm64,
1374 	[BPF_ALU64 | BPF_ADD | BPF_X] =	add_reg64,
1375 	[BPF_ALU64 | BPF_ADD | BPF_K] =	add_imm64,
1376 	[BPF_ALU64 | BPF_SUB | BPF_X] =	sub_reg64,
1377 	[BPF_ALU64 | BPF_SUB | BPF_K] =	sub_imm64,
1378 	[BPF_ALU64 | BPF_LSH | BPF_K] =	shl_imm64,
1379 	[BPF_ALU64 | BPF_RSH | BPF_K] =	shr_imm64,
1380 	[BPF_ALU | BPF_MOV | BPF_X] =	mov_reg,
1381 	[BPF_ALU | BPF_MOV | BPF_K] =	mov_imm,
1382 	[BPF_ALU | BPF_XOR | BPF_X] =	xor_reg,
1383 	[BPF_ALU | BPF_XOR | BPF_K] =	xor_imm,
1384 	[BPF_ALU | BPF_AND | BPF_X] =	and_reg,
1385 	[BPF_ALU | BPF_AND | BPF_K] =	and_imm,
1386 	[BPF_ALU | BPF_OR | BPF_X] =	or_reg,
1387 	[BPF_ALU | BPF_OR | BPF_K] =	or_imm,
1388 	[BPF_ALU | BPF_ADD | BPF_X] =	add_reg,
1389 	[BPF_ALU | BPF_ADD | BPF_K] =	add_imm,
1390 	[BPF_ALU | BPF_SUB | BPF_X] =	sub_reg,
1391 	[BPF_ALU | BPF_SUB | BPF_K] =	sub_imm,
1392 	[BPF_ALU | BPF_LSH | BPF_K] =	shl_imm,
1393 	[BPF_LD | BPF_IMM | BPF_DW] =	imm_ld8,
1394 	[BPF_LD | BPF_ABS | BPF_B] =	data_ld1,
1395 	[BPF_LD | BPF_ABS | BPF_H] =	data_ld2,
1396 	[BPF_LD | BPF_ABS | BPF_W] =	data_ld4,
1397 	[BPF_LD | BPF_IND | BPF_B] =	data_ind_ld1,
1398 	[BPF_LD | BPF_IND | BPF_H] =	data_ind_ld2,
1399 	[BPF_LD | BPF_IND | BPF_W] =	data_ind_ld4,
1400 	[BPF_LDX | BPF_MEM | BPF_W] =	mem_ldx4,
1401 	[BPF_STX | BPF_MEM | BPF_W] =	mem_stx4,
1402 	[BPF_JMP | BPF_JA | BPF_K] =	jump,
1403 	[BPF_JMP | BPF_JEQ | BPF_K] =	jeq_imm,
1404 	[BPF_JMP | BPF_JGT | BPF_K] =	jgt_imm,
1405 	[BPF_JMP | BPF_JGE | BPF_K] =	jge_imm,
1406 	[BPF_JMP | BPF_JLT | BPF_K] =	jlt_imm,
1407 	[BPF_JMP | BPF_JLE | BPF_K] =	jle_imm,
1408 	[BPF_JMP | BPF_JSET | BPF_K] =	jset_imm,
1409 	[BPF_JMP | BPF_JNE | BPF_K] =	jne_imm,
1410 	[BPF_JMP | BPF_JEQ | BPF_X] =	jeq_reg,
1411 	[BPF_JMP | BPF_JGT | BPF_X] =	jgt_reg,
1412 	[BPF_JMP | BPF_JGE | BPF_X] =	jge_reg,
1413 	[BPF_JMP | BPF_JLT | BPF_X] =	jlt_reg,
1414 	[BPF_JMP | BPF_JLE | BPF_X] =	jle_reg,
1415 	[BPF_JMP | BPF_JSET | BPF_X] =	jset_reg,
1416 	[BPF_JMP | BPF_JNE | BPF_X] =	jne_reg,
1417 	[BPF_JMP | BPF_EXIT] =		goto_out,
1418 };
1419 
1420 /* --- Misc code --- */
1421 static void br_set_offset(u64 *instr, u16 offset)
1422 {
1423 	u16 addr_lo, addr_hi;
1424 
1425 	addr_lo = offset & (OP_BR_ADDR_LO >> __bf_shf(OP_BR_ADDR_LO));
1426 	addr_hi = offset != addr_lo;
1427 	*instr &= ~(OP_BR_ADDR_HI | OP_BR_ADDR_LO);
1428 	*instr |= FIELD_PREP(OP_BR_ADDR_HI, addr_hi);
1429 	*instr |= FIELD_PREP(OP_BR_ADDR_LO, addr_lo);
1430 }
1431 
1432 /* --- Assembler logic --- */
1433 static int nfp_fixup_branches(struct nfp_prog *nfp_prog)
1434 {
1435 	struct nfp_insn_meta *meta, *next;
1436 	u32 off, br_idx;
1437 	u32 idx;
1438 
1439 	nfp_for_each_insn_walk2(nfp_prog, meta, next) {
1440 		if (meta->skip)
1441 			continue;
1442 		if (BPF_CLASS(meta->insn.code) != BPF_JMP)
1443 			continue;
1444 
1445 		br_idx = nfp_prog_offset_to_index(nfp_prog, next->off) - 1;
1446 		if (!nfp_is_br(nfp_prog->prog[br_idx])) {
1447 			pr_err("Fixup found block not ending in branch %d %02x %016llx!!\n",
1448 			       br_idx, meta->insn.code, nfp_prog->prog[br_idx]);
1449 			return -ELOOP;
1450 		}
1451 		/* Leave special branches for later */
1452 		if (FIELD_GET(OP_BR_SPECIAL, nfp_prog->prog[br_idx]))
1453 			continue;
1454 
1455 		/* Find the target offset in assembler realm */
1456 		off = meta->insn.off;
1457 		if (!off) {
1458 			pr_err("Fixup found zero offset!!\n");
1459 			return -ELOOP;
1460 		}
1461 
1462 		while (off && nfp_meta_has_next(nfp_prog, next)) {
1463 			next = nfp_meta_next(next);
1464 			off--;
1465 		}
1466 		if (off) {
1467 			pr_err("Fixup found too large jump!! %d\n", off);
1468 			return -ELOOP;
1469 		}
1470 
1471 		if (next->skip) {
1472 			pr_err("Branch landing on removed instruction!!\n");
1473 			return -ELOOP;
1474 		}
1475 
1476 		for (idx = nfp_prog_offset_to_index(nfp_prog, meta->off);
1477 		     idx <= br_idx; idx++) {
1478 			if (!nfp_is_br(nfp_prog->prog[idx]))
1479 				continue;
1480 			br_set_offset(&nfp_prog->prog[idx], next->off);
1481 		}
1482 	}
1483 
1484 	/* Fixup 'goto out's separately, they can be scattered around */
1485 	for (br_idx = 0; br_idx < nfp_prog->prog_len; br_idx++) {
1486 		enum br_special special;
1487 
1488 		if ((nfp_prog->prog[br_idx] & OP_BR_BASE_MASK) != OP_BR_BASE)
1489 			continue;
1490 
1491 		special = FIELD_GET(OP_BR_SPECIAL, nfp_prog->prog[br_idx]);
1492 		switch (special) {
1493 		case OP_BR_NORMAL:
1494 			break;
1495 		case OP_BR_GO_OUT:
1496 			br_set_offset(&nfp_prog->prog[br_idx],
1497 				      nfp_prog->tgt_out);
1498 			break;
1499 		case OP_BR_GO_ABORT:
1500 			br_set_offset(&nfp_prog->prog[br_idx],
1501 				      nfp_prog->tgt_abort);
1502 			break;
1503 		}
1504 
1505 		nfp_prog->prog[br_idx] &= ~OP_BR_SPECIAL;
1506 	}
1507 
1508 	return 0;
1509 }
1510 
1511 static void nfp_intro(struct nfp_prog *nfp_prog)
1512 {
1513 	emit_alu(nfp_prog, pkt_reg(nfp_prog),
1514 		 reg_none(), ALU_OP_NONE, NFP_BPF_ABI_PKT);
1515 }
1516 
1517 static void nfp_outro_tc_legacy(struct nfp_prog *nfp_prog)
1518 {
1519 	const u8 act2code[] = {
1520 		[NN_ACT_TC_DROP]  = 0x22,
1521 		[NN_ACT_TC_REDIR] = 0x24
1522 	};
1523 	/* Target for aborts */
1524 	nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog);
1525 	wrp_immed(nfp_prog, reg_both(0), 0);
1526 
1527 	/* Target for normal exits */
1528 	nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog);
1529 	/* Legacy TC mode:
1530 	 *   0        0x11 -> pass,  count as stat0
1531 	 *  -1  drop  0x22 -> drop,  count as stat1
1532 	 *     redir  0x24 -> redir, count as stat1
1533 	 *  ife mark  0x21 -> pass,  count as stat1
1534 	 *  ife + tx  0x24 -> redir, count as stat1
1535 	 */
1536 	emit_br_byte_neq(nfp_prog, reg_b(0), 0xff, 0, nfp_prog->tgt_done, 2);
1537 	emit_alu(nfp_prog, reg_a(0),
1538 		 reg_none(), ALU_OP_NONE, NFP_BPF_ABI_FLAGS);
1539 	emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x11), SHF_SC_L_SHF, 16);
1540 
1541 	emit_br(nfp_prog, BR_UNC, nfp_prog->tgt_done, 1);
1542 	emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(act2code[nfp_prog->act]),
1543 		      SHF_SC_L_SHF, 16);
1544 }
1545 
1546 static void nfp_outro_tc_da(struct nfp_prog *nfp_prog)
1547 {
1548 	/* TC direct-action mode:
1549 	 *   0,1   ok        NOT SUPPORTED[1]
1550 	 *   2   drop  0x22 -> drop,  count as stat1
1551 	 *   4,5 nuke  0x02 -> drop
1552 	 *   7  redir  0x44 -> redir, count as stat2
1553 	 *   * unspec  0x11 -> pass,  count as stat0
1554 	 *
1555 	 * [1] We can't support OK and RECLASSIFY because we can't tell TC
1556 	 *     the exact decision made.  We are forced to support UNSPEC
1557 	 *     to handle aborts so that's the only one we handle for passing
1558 	 *     packets up the stack.
1559 	 */
1560 	/* Target for aborts */
1561 	nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog);
1562 
1563 	emit_br_def(nfp_prog, nfp_prog->tgt_done, 2);
1564 
1565 	emit_alu(nfp_prog, reg_a(0),
1566 		 reg_none(), ALU_OP_NONE, NFP_BPF_ABI_FLAGS);
1567 	emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x11), SHF_SC_L_SHF, 16);
1568 
1569 	/* Target for normal exits */
1570 	nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog);
1571 
1572 	/* if R0 > 7 jump to abort */
1573 	emit_alu(nfp_prog, reg_none(), reg_imm(7), ALU_OP_SUB, reg_b(0));
1574 	emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0);
1575 	emit_alu(nfp_prog, reg_a(0),
1576 		 reg_none(), ALU_OP_NONE, NFP_BPF_ABI_FLAGS);
1577 
1578 	wrp_immed(nfp_prog, reg_b(2), 0x41221211);
1579 	wrp_immed(nfp_prog, reg_b(3), 0x41001211);
1580 
1581 	emit_shf(nfp_prog, reg_a(1),
1582 		 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 2);
1583 
1584 	emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0));
1585 	emit_shf(nfp_prog, reg_a(2),
1586 		 reg_imm(0xf), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0);
1587 
1588 	emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0));
1589 	emit_shf(nfp_prog, reg_b(2),
1590 		 reg_imm(0xf), SHF_OP_AND, reg_b(3), SHF_SC_R_SHF, 0);
1591 
1592 	emit_br_def(nfp_prog, nfp_prog->tgt_done, 2);
1593 
1594 	emit_shf(nfp_prog, reg_b(2),
1595 		 reg_a(2), SHF_OP_OR, reg_b(2), SHF_SC_L_SHF, 4);
1596 	emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16);
1597 }
1598 
1599 static void nfp_outro_xdp(struct nfp_prog *nfp_prog)
1600 {
1601 	/* XDP return codes:
1602 	 *   0 aborted  0x82 -> drop,  count as stat3
1603 	 *   1    drop  0x22 -> drop,  count as stat1
1604 	 *   2    pass  0x11 -> pass,  count as stat0
1605 	 *   3      tx  0x44 -> redir, count as stat2
1606 	 *   * unknown  0x82 -> drop,  count as stat3
1607 	 */
1608 	/* Target for aborts */
1609 	nfp_prog->tgt_abort = nfp_prog_current_offset(nfp_prog);
1610 
1611 	emit_br_def(nfp_prog, nfp_prog->tgt_done, 2);
1612 
1613 	emit_alu(nfp_prog, reg_a(0),
1614 		 reg_none(), ALU_OP_NONE, NFP_BPF_ABI_FLAGS);
1615 	emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_imm(0x82), SHF_SC_L_SHF, 16);
1616 
1617 	/* Target for normal exits */
1618 	nfp_prog->tgt_out = nfp_prog_current_offset(nfp_prog);
1619 
1620 	/* if R0 > 3 jump to abort */
1621 	emit_alu(nfp_prog, reg_none(), reg_imm(3), ALU_OP_SUB, reg_b(0));
1622 	emit_br(nfp_prog, BR_BLO, nfp_prog->tgt_abort, 0);
1623 
1624 	wrp_immed(nfp_prog, reg_b(2), 0x44112282);
1625 
1626 	emit_shf(nfp_prog, reg_a(1),
1627 		 reg_none(), SHF_OP_NONE, reg_b(0), SHF_SC_L_SHF, 3);
1628 
1629 	emit_alu(nfp_prog, reg_none(), reg_a(1), ALU_OP_OR, reg_imm(0));
1630 	emit_shf(nfp_prog, reg_b(2),
1631 		 reg_imm(0xff), SHF_OP_AND, reg_b(2), SHF_SC_R_SHF, 0);
1632 
1633 	emit_br_def(nfp_prog, nfp_prog->tgt_done, 2);
1634 
1635 	emit_alu(nfp_prog, reg_a(0),
1636 		 reg_none(), ALU_OP_NONE, NFP_BPF_ABI_FLAGS);
1637 	emit_ld_field(nfp_prog, reg_a(0), 0xc, reg_b(2), SHF_SC_L_SHF, 16);
1638 }
1639 
1640 static void nfp_outro(struct nfp_prog *nfp_prog)
1641 {
1642 	switch (nfp_prog->act) {
1643 	case NN_ACT_DIRECT:
1644 		nfp_outro_tc_da(nfp_prog);
1645 		break;
1646 	case NN_ACT_TC_DROP:
1647 	case NN_ACT_TC_REDIR:
1648 		nfp_outro_tc_legacy(nfp_prog);
1649 		break;
1650 	case NN_ACT_XDP:
1651 		nfp_outro_xdp(nfp_prog);
1652 		break;
1653 	}
1654 }
1655 
1656 static int nfp_translate(struct nfp_prog *nfp_prog)
1657 {
1658 	struct nfp_insn_meta *meta;
1659 	int err;
1660 
1661 	nfp_intro(nfp_prog);
1662 	if (nfp_prog->error)
1663 		return nfp_prog->error;
1664 
1665 	list_for_each_entry(meta, &nfp_prog->insns, l) {
1666 		instr_cb_t cb = instr_cb[meta->insn.code];
1667 
1668 		meta->off = nfp_prog_current_offset(nfp_prog);
1669 
1670 		if (meta->skip) {
1671 			nfp_prog->n_translated++;
1672 			continue;
1673 		}
1674 
1675 		if (nfp_meta_has_prev(nfp_prog, meta) &&
1676 		    nfp_meta_prev(meta)->double_cb)
1677 			cb = nfp_meta_prev(meta)->double_cb;
1678 		if (!cb)
1679 			return -ENOENT;
1680 		err = cb(nfp_prog, meta);
1681 		if (err)
1682 			return err;
1683 
1684 		nfp_prog->n_translated++;
1685 	}
1686 
1687 	nfp_outro(nfp_prog);
1688 	if (nfp_prog->error)
1689 		return nfp_prog->error;
1690 
1691 	return nfp_fixup_branches(nfp_prog);
1692 }
1693 
1694 static int
1695 nfp_prog_prepare(struct nfp_prog *nfp_prog, const struct bpf_insn *prog,
1696 		 unsigned int cnt)
1697 {
1698 	unsigned int i;
1699 
1700 	for (i = 0; i < cnt; i++) {
1701 		struct nfp_insn_meta *meta;
1702 
1703 		meta = kzalloc(sizeof(*meta), GFP_KERNEL);
1704 		if (!meta)
1705 			return -ENOMEM;
1706 
1707 		meta->insn = prog[i];
1708 		meta->n = i;
1709 
1710 		list_add_tail(&meta->l, &nfp_prog->insns);
1711 	}
1712 
1713 	return 0;
1714 }
1715 
1716 /* --- Optimizations --- */
1717 static void nfp_bpf_opt_reg_init(struct nfp_prog *nfp_prog)
1718 {
1719 	struct nfp_insn_meta *meta;
1720 
1721 	list_for_each_entry(meta, &nfp_prog->insns, l) {
1722 		struct bpf_insn insn = meta->insn;
1723 
1724 		/* Programs converted from cBPF start with register xoring */
1725 		if (insn.code == (BPF_ALU64 | BPF_XOR | BPF_X) &&
1726 		    insn.src_reg == insn.dst_reg)
1727 			continue;
1728 
1729 		/* Programs start with R6 = R1 but we ignore the skb pointer */
1730 		if (insn.code == (BPF_ALU64 | BPF_MOV | BPF_X) &&
1731 		    insn.src_reg == 1 && insn.dst_reg == 6)
1732 			meta->skip = true;
1733 
1734 		/* Return as soon as something doesn't match */
1735 		if (!meta->skip)
1736 			return;
1737 	}
1738 }
1739 
1740 /* Try to rename registers so that program uses only low ones */
1741 static int nfp_bpf_opt_reg_rename(struct nfp_prog *nfp_prog)
1742 {
1743 	bool reg_used[MAX_BPF_REG] = {};
1744 	u8 tgt_reg[MAX_BPF_REG] = {};
1745 	struct nfp_insn_meta *meta;
1746 	unsigned int i, j;
1747 
1748 	list_for_each_entry(meta, &nfp_prog->insns, l) {
1749 		if (meta->skip)
1750 			continue;
1751 
1752 		reg_used[meta->insn.src_reg] = true;
1753 		reg_used[meta->insn.dst_reg] = true;
1754 	}
1755 
1756 	for (i = 0, j = 0; i < ARRAY_SIZE(tgt_reg); i++) {
1757 		if (!reg_used[i])
1758 			continue;
1759 
1760 		tgt_reg[i] = j++;
1761 	}
1762 	nfp_prog->num_regs = j;
1763 
1764 	list_for_each_entry(meta, &nfp_prog->insns, l) {
1765 		meta->insn.src_reg = tgt_reg[meta->insn.src_reg];
1766 		meta->insn.dst_reg = tgt_reg[meta->insn.dst_reg];
1767 	}
1768 
1769 	return 0;
1770 }
1771 
1772 /* Remove masking after load since our load guarantees this is not needed */
1773 static void nfp_bpf_opt_ld_mask(struct nfp_prog *nfp_prog)
1774 {
1775 	struct nfp_insn_meta *meta1, *meta2;
1776 	const s32 exp_mask[] = {
1777 		[BPF_B] = 0x000000ffU,
1778 		[BPF_H] = 0x0000ffffU,
1779 		[BPF_W] = 0xffffffffU,
1780 	};
1781 
1782 	nfp_for_each_insn_walk2(nfp_prog, meta1, meta2) {
1783 		struct bpf_insn insn, next;
1784 
1785 		insn = meta1->insn;
1786 		next = meta2->insn;
1787 
1788 		if (BPF_CLASS(insn.code) != BPF_LD)
1789 			continue;
1790 		if (BPF_MODE(insn.code) != BPF_ABS &&
1791 		    BPF_MODE(insn.code) != BPF_IND)
1792 			continue;
1793 
1794 		if (next.code != (BPF_ALU64 | BPF_AND | BPF_K))
1795 			continue;
1796 
1797 		if (!exp_mask[BPF_SIZE(insn.code)])
1798 			continue;
1799 		if (exp_mask[BPF_SIZE(insn.code)] != next.imm)
1800 			continue;
1801 
1802 		if (next.src_reg || next.dst_reg)
1803 			continue;
1804 
1805 		meta2->skip = true;
1806 	}
1807 }
1808 
1809 static void nfp_bpf_opt_ld_shift(struct nfp_prog *nfp_prog)
1810 {
1811 	struct nfp_insn_meta *meta1, *meta2, *meta3;
1812 
1813 	nfp_for_each_insn_walk3(nfp_prog, meta1, meta2, meta3) {
1814 		struct bpf_insn insn, next1, next2;
1815 
1816 		insn = meta1->insn;
1817 		next1 = meta2->insn;
1818 		next2 = meta3->insn;
1819 
1820 		if (BPF_CLASS(insn.code) != BPF_LD)
1821 			continue;
1822 		if (BPF_MODE(insn.code) != BPF_ABS &&
1823 		    BPF_MODE(insn.code) != BPF_IND)
1824 			continue;
1825 		if (BPF_SIZE(insn.code) != BPF_W)
1826 			continue;
1827 
1828 		if (!(next1.code == (BPF_LSH | BPF_K | BPF_ALU64) &&
1829 		      next2.code == (BPF_RSH | BPF_K | BPF_ALU64)) &&
1830 		    !(next1.code == (BPF_RSH | BPF_K | BPF_ALU64) &&
1831 		      next2.code == (BPF_LSH | BPF_K | BPF_ALU64)))
1832 			continue;
1833 
1834 		if (next1.src_reg || next1.dst_reg ||
1835 		    next2.src_reg || next2.dst_reg)
1836 			continue;
1837 
1838 		if (next1.imm != 0x20 || next2.imm != 0x20)
1839 			continue;
1840 
1841 		meta2->skip = true;
1842 		meta3->skip = true;
1843 	}
1844 }
1845 
1846 static int nfp_bpf_optimize(struct nfp_prog *nfp_prog)
1847 {
1848 	int ret;
1849 
1850 	nfp_bpf_opt_reg_init(nfp_prog);
1851 
1852 	ret = nfp_bpf_opt_reg_rename(nfp_prog);
1853 	if (ret)
1854 		return ret;
1855 
1856 	nfp_bpf_opt_ld_mask(nfp_prog);
1857 	nfp_bpf_opt_ld_shift(nfp_prog);
1858 
1859 	return 0;
1860 }
1861 
1862 /**
1863  * nfp_bpf_jit() - translate BPF code into NFP assembly
1864  * @filter:	kernel BPF filter struct
1865  * @prog_mem:	memory to store assembler instructions
1866  * @act:	action attached to this eBPF program
1867  * @prog_start:	offset of the first instruction when loaded
1868  * @prog_done:	where to jump on exit
1869  * @prog_sz:	size of @prog_mem in instructions
1870  * @res:	achieved parameters of translation results
1871  */
1872 int
1873 nfp_bpf_jit(struct bpf_prog *filter, void *prog_mem,
1874 	    enum nfp_bpf_action_type act,
1875 	    unsigned int prog_start, unsigned int prog_done,
1876 	    unsigned int prog_sz, struct nfp_bpf_result *res)
1877 {
1878 	struct nfp_prog *nfp_prog;
1879 	int ret;
1880 
1881 	nfp_prog = kzalloc(sizeof(*nfp_prog), GFP_KERNEL);
1882 	if (!nfp_prog)
1883 		return -ENOMEM;
1884 
1885 	INIT_LIST_HEAD(&nfp_prog->insns);
1886 	nfp_prog->act = act;
1887 	nfp_prog->start_off = prog_start;
1888 	nfp_prog->tgt_done = prog_done;
1889 
1890 	ret = nfp_prog_prepare(nfp_prog, filter->insnsi, filter->len);
1891 	if (ret)
1892 		goto out;
1893 
1894 	ret = nfp_prog_verify(nfp_prog, filter);
1895 	if (ret)
1896 		goto out;
1897 
1898 	ret = nfp_bpf_optimize(nfp_prog);
1899 	if (ret)
1900 		goto out;
1901 
1902 	if (nfp_prog->num_regs <= 7)
1903 		nfp_prog->regs_per_thread = 16;
1904 	else
1905 		nfp_prog->regs_per_thread = 32;
1906 
1907 	nfp_prog->prog = prog_mem;
1908 	nfp_prog->__prog_alloc_len = prog_sz;
1909 
1910 	ret = nfp_translate(nfp_prog);
1911 	if (ret) {
1912 		pr_err("Translation failed with error %d (translated: %u)\n",
1913 		       ret, nfp_prog->n_translated);
1914 		ret = -EINVAL;
1915 	}
1916 
1917 	res->n_instr = nfp_prog->prog_len;
1918 	res->dense_mode = nfp_prog->num_regs <= 7;
1919 out:
1920 	nfp_prog_free(nfp_prog);
1921 
1922 	return ret;
1923 }
1924