Linux Audio

Check our new training course

Loading...
v5.4
  1// SPDX-License-Identifier: GPL-2.0-only
  2/* bpf_jit_comp.c: BPF JIT compiler
 
  3 *
  4 * Copyright 2011 Matt Evans <matt@ozlabs.org>, IBM Corporation
 
  5 *
  6 * Based on the x86 BPF compiler, by Eric Dumazet (eric.dumazet@gmail.com)
  7 * Ported to ppc32 by Denis Kirjanov <kda@linux-powerpc.org>
  8 */
  9#include <linux/moduleloader.h>
 10#include <asm/cacheflush.h>
 11#include <asm/asm-compat.h>
 12#include <linux/netdevice.h>
 13#include <linux/filter.h>
 14#include <linux/if_vlan.h>
 
 
 
 15
 16#include "bpf_jit32.h"
 
 17
 18static inline void bpf_flush_icache(void *start, void *end)
 
 
 
 
 
 19{
 20	smp_wmb();
 21	flush_icache_range((unsigned long)start, (unsigned long)end);
 22}
 23
 24static void bpf_jit_build_prologue(struct bpf_prog *fp, u32 *image,
 25				   struct codegen_context *ctx)
 26{
 27	int i;
 28	const struct sock_filter *filter = fp->insns;
 29
 30	if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
 31		/* Make stackframe */
 32		if (ctx->seen & SEEN_DATAREF) {
 33			/* If we call any helpers (for loads), save LR */
 34			EMIT(PPC_INST_MFLR | __PPC_RT(R0));
 35			PPC_BPF_STL(0, 1, PPC_LR_STKOFF);
 36
 37			/* Back up non-volatile regs. */
 38			PPC_BPF_STL(r_D, 1, -(REG_SZ*(32-r_D)));
 39			PPC_BPF_STL(r_HL, 1, -(REG_SZ*(32-r_HL)));
 40		}
 41		if (ctx->seen & SEEN_MEM) {
 42			/*
 43			 * Conditionally save regs r15-r31 as some will be used
 44			 * for M[] data.
 45			 */
 46			for (i = r_M; i < (r_M+16); i++) {
 47				if (ctx->seen & (1 << (i-r_M)))
 48					PPC_BPF_STL(i, 1, -(REG_SZ*(32-i)));
 49			}
 50		}
 51		PPC_BPF_STLU(1, 1, -BPF_PPC_STACKFRAME);
 52	}
 53
 54	if (ctx->seen & SEEN_DATAREF) {
 55		/*
 56		 * If this filter needs to access skb data,
 57		 * prepare r_D and r_HL:
 58		 *  r_HL = skb->len - skb->data_len
 59		 *  r_D	 = skb->data
 60		 */
 61		PPC_LWZ_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
 62							 data_len));
 63		PPC_LWZ_OFFS(r_HL, r_skb, offsetof(struct sk_buff, len));
 64		PPC_SUB(r_HL, r_HL, r_scratch1);
 65		PPC_LL_OFFS(r_D, r_skb, offsetof(struct sk_buff, data));
 66	}
 67
 68	if (ctx->seen & SEEN_XREG) {
 69		/*
 70		 * TODO: Could also detect whether first instr. sets X and
 71		 * avoid this (as below, with A).
 72		 */
 73		PPC_LI(r_X, 0);
 74	}
 
 
 
 
 
 
 
 75
 76	/* make sure we dont leak kernel information to user */
 77	if (bpf_needs_clear_a(&filter[0]))
 78		PPC_LI(r_A, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 79}
 80
 81static void bpf_jit_build_epilogue(u32 *image, struct codegen_context *ctx)
 82{
 83	int i;
 84
 85	if (ctx->seen & (SEEN_MEM | SEEN_DATAREF)) {
 86		PPC_ADDI(1, 1, BPF_PPC_STACKFRAME);
 87		if (ctx->seen & SEEN_DATAREF) {
 88			PPC_BPF_LL(0, 1, PPC_LR_STKOFF);
 89			PPC_MTLR(0);
 90			PPC_BPF_LL(r_D, 1, -(REG_SZ*(32-r_D)));
 91			PPC_BPF_LL(r_HL, 1, -(REG_SZ*(32-r_HL)));
 92		}
 93		if (ctx->seen & SEEN_MEM) {
 94			/* Restore any saved non-vol registers */
 95			for (i = r_M; i < (r_M+16); i++) {
 96				if (ctx->seen & (1 << (i-r_M)))
 97					PPC_BPF_LL(i, 1, -(REG_SZ*(32-i)));
 98			}
 99		}
100	}
101	/* The RETs have left a return value in R3. */
102
103	PPC_BLR();
104}
105
106#define CHOOSE_LOAD_FUNC(K, func) \
107	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
108
109/* Assemble the body code between the prologue & epilogue. */
110static int bpf_jit_build_body(struct bpf_prog *fp, u32 *image,
111			      struct codegen_context *ctx,
112			      unsigned int *addrs)
113{
114	const struct sock_filter *filter = fp->insns;
115	int flen = fp->len;
116	u8 *func;
117	unsigned int true_cond;
118	int i;
119
120	/* Start of epilogue code */
121	unsigned int exit_addr = addrs[flen];
122
123	for (i = 0; i < flen; i++) {
124		unsigned int K = filter[i].k;
125		u16 code = bpf_anc_helper(&filter[i]);
126
127		/*
128		 * addrs[] maps a BPF bytecode address into a real offset from
129		 * the start of the body code.
130		 */
131		addrs[i] = ctx->idx * 4;
132
133		switch (code) {
134			/*** ALU ops ***/
135		case BPF_ALU | BPF_ADD | BPF_X: /* A += X; */
136			ctx->seen |= SEEN_XREG;
137			PPC_ADD(r_A, r_A, r_X);
138			break;
139		case BPF_ALU | BPF_ADD | BPF_K: /* A += K; */
140			if (!K)
141				break;
142			PPC_ADDI(r_A, r_A, IMM_L(K));
143			if (K >= 32768)
144				PPC_ADDIS(r_A, r_A, IMM_HA(K));
145			break;
146		case BPF_ALU | BPF_SUB | BPF_X: /* A -= X; */
147			ctx->seen |= SEEN_XREG;
148			PPC_SUB(r_A, r_A, r_X);
149			break;
150		case BPF_ALU | BPF_SUB | BPF_K: /* A -= K */
151			if (!K)
152				break;
153			PPC_ADDI(r_A, r_A, IMM_L(-K));
154			if (K >= 32768)
155				PPC_ADDIS(r_A, r_A, IMM_HA(-K));
156			break;
157		case BPF_ALU | BPF_MUL | BPF_X: /* A *= X; */
158			ctx->seen |= SEEN_XREG;
159			PPC_MULW(r_A, r_A, r_X);
160			break;
161		case BPF_ALU | BPF_MUL | BPF_K: /* A *= K */
162			if (K < 32768)
163				PPC_MULI(r_A, r_A, K);
164			else {
165				PPC_LI32(r_scratch1, K);
166				PPC_MULW(r_A, r_A, r_scratch1);
167			}
168			break;
169		case BPF_ALU | BPF_MOD | BPF_X: /* A %= X; */
170		case BPF_ALU | BPF_DIV | BPF_X: /* A /= X; */
171			ctx->seen |= SEEN_XREG;
172			PPC_CMPWI(r_X, 0);
173			if (ctx->pc_ret0 != -1) {
174				PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
175			} else {
176				PPC_BCC_SHORT(COND_NE, (ctx->idx*4)+12);
177				PPC_LI(r_ret, 0);
178				PPC_JMP(exit_addr);
179			}
180			if (code == (BPF_ALU | BPF_MOD | BPF_X)) {
181				PPC_DIVWU(r_scratch1, r_A, r_X);
182				PPC_MULW(r_scratch1, r_X, r_scratch1);
183				PPC_SUB(r_A, r_A, r_scratch1);
184			} else {
185				PPC_DIVWU(r_A, r_A, r_X);
186			}
187			break;
188		case BPF_ALU | BPF_MOD | BPF_K: /* A %= K; */
189			PPC_LI32(r_scratch2, K);
190			PPC_DIVWU(r_scratch1, r_A, r_scratch2);
191			PPC_MULW(r_scratch1, r_scratch2, r_scratch1);
192			PPC_SUB(r_A, r_A, r_scratch1);
193			break;
194		case BPF_ALU | BPF_DIV | BPF_K: /* A /= K */
195			if (K == 1)
196				break;
197			PPC_LI32(r_scratch1, K);
198			PPC_DIVWU(r_A, r_A, r_scratch1);
199			break;
200		case BPF_ALU | BPF_AND | BPF_X:
201			ctx->seen |= SEEN_XREG;
202			PPC_AND(r_A, r_A, r_X);
203			break;
204		case BPF_ALU | BPF_AND | BPF_K:
205			if (!IMM_H(K))
206				PPC_ANDI(r_A, r_A, K);
207			else {
208				PPC_LI32(r_scratch1, K);
209				PPC_AND(r_A, r_A, r_scratch1);
210			}
211			break;
212		case BPF_ALU | BPF_OR | BPF_X:
213			ctx->seen |= SEEN_XREG;
214			PPC_OR(r_A, r_A, r_X);
215			break;
216		case BPF_ALU | BPF_OR | BPF_K:
217			if (IMM_L(K))
218				PPC_ORI(r_A, r_A, IMM_L(K));
219			if (K >= 65536)
220				PPC_ORIS(r_A, r_A, IMM_H(K));
221			break;
222		case BPF_ANC | SKF_AD_ALU_XOR_X:
223		case BPF_ALU | BPF_XOR | BPF_X: /* A ^= X */
224			ctx->seen |= SEEN_XREG;
225			PPC_XOR(r_A, r_A, r_X);
226			break;
227		case BPF_ALU | BPF_XOR | BPF_K: /* A ^= K */
228			if (IMM_L(K))
229				PPC_XORI(r_A, r_A, IMM_L(K));
230			if (K >= 65536)
231				PPC_XORIS(r_A, r_A, IMM_H(K));
232			break;
233		case BPF_ALU | BPF_LSH | BPF_X: /* A <<= X; */
234			ctx->seen |= SEEN_XREG;
235			PPC_SLW(r_A, r_A, r_X);
236			break;
237		case BPF_ALU | BPF_LSH | BPF_K:
238			if (K == 0)
239				break;
240			else
241				PPC_SLWI(r_A, r_A, K);
242			break;
243		case BPF_ALU | BPF_RSH | BPF_X: /* A >>= X; */
244			ctx->seen |= SEEN_XREG;
245			PPC_SRW(r_A, r_A, r_X);
246			break;
247		case BPF_ALU | BPF_RSH | BPF_K: /* A >>= K; */
248			if (K == 0)
249				break;
250			else
251				PPC_SRWI(r_A, r_A, K);
252			break;
253		case BPF_ALU | BPF_NEG:
254			PPC_NEG(r_A, r_A);
255			break;
256		case BPF_RET | BPF_K:
257			PPC_LI32(r_ret, K);
258			if (!K) {
259				if (ctx->pc_ret0 == -1)
260					ctx->pc_ret0 = i;
261			}
262			/*
263			 * If this isn't the very last instruction, branch to
264			 * the epilogue if we've stuff to clean up.  Otherwise,
265			 * if there's nothing to tidy, just return.  If we /are/
266			 * the last instruction, we're about to fall through to
267			 * the epilogue to return.
268			 */
269			if (i != flen - 1) {
270				/*
271				 * Note: 'seen' is properly valid only on pass
272				 * #2.	Both parts of this conditional are the
273				 * same instruction size though, meaning the
274				 * first pass will still correctly determine the
275				 * code size/addresses.
276				 */
277				if (ctx->seen)
278					PPC_JMP(exit_addr);
279				else
280					PPC_BLR();
281			}
282			break;
283		case BPF_RET | BPF_A:
284			PPC_MR(r_ret, r_A);
285			if (i != flen - 1) {
286				if (ctx->seen)
287					PPC_JMP(exit_addr);
288				else
289					PPC_BLR();
290			}
291			break;
292		case BPF_MISC | BPF_TAX: /* X = A */
293			PPC_MR(r_X, r_A);
294			break;
295		case BPF_MISC | BPF_TXA: /* A = X */
296			ctx->seen |= SEEN_XREG;
297			PPC_MR(r_A, r_X);
298			break;
299
300			/*** Constant loads/M[] access ***/
301		case BPF_LD | BPF_IMM: /* A = K */
302			PPC_LI32(r_A, K);
303			break;
304		case BPF_LDX | BPF_IMM: /* X = K */
305			PPC_LI32(r_X, K);
306			break;
307		case BPF_LD | BPF_MEM: /* A = mem[K] */
308			PPC_MR(r_A, r_M + (K & 0xf));
309			ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
310			break;
311		case BPF_LDX | BPF_MEM: /* X = mem[K] */
312			PPC_MR(r_X, r_M + (K & 0xf));
313			ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
314			break;
315		case BPF_ST: /* mem[K] = A */
316			PPC_MR(r_M + (K & 0xf), r_A);
317			ctx->seen |= SEEN_MEM | (1<<(K & 0xf));
318			break;
319		case BPF_STX: /* mem[K] = X */
320			PPC_MR(r_M + (K & 0xf), r_X);
321			ctx->seen |= SEEN_XREG | SEEN_MEM | (1<<(K & 0xf));
322			break;
323		case BPF_LD | BPF_W | BPF_LEN: /*	A = skb->len; */
324			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
325			PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff, len));
326			break;
327		case BPF_LDX | BPF_W | BPF_ABS: /* A = *((u32 *)(seccomp_data + K)); */
328			PPC_LWZ_OFFS(r_A, r_skb, K);
329			break;
330		case BPF_LDX | BPF_W | BPF_LEN: /* X = skb->len; */
331			PPC_LWZ_OFFS(r_X, r_skb, offsetof(struct sk_buff, len));
332			break;
333
334			/*** Ancillary info loads ***/
335		case BPF_ANC | SKF_AD_PROTOCOL: /* A = ntohs(skb->protocol); */
336			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
337						  protocol) != 2);
338			PPC_NTOHS_OFFS(r_A, r_skb, offsetof(struct sk_buff,
339							    protocol));
340			break;
341		case BPF_ANC | SKF_AD_IFINDEX:
342		case BPF_ANC | SKF_AD_HATYPE:
343			BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
344						ifindex) != 4);
345			BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
346						type) != 2);
347			PPC_LL_OFFS(r_scratch1, r_skb, offsetof(struct sk_buff,
348								dev));
349			PPC_CMPDI(r_scratch1, 0);
350			if (ctx->pc_ret0 != -1) {
351				PPC_BCC(COND_EQ, addrs[ctx->pc_ret0]);
352			} else {
353				/* Exit, returning 0; first pass hits here. */
354				PPC_BCC_SHORT(COND_NE, ctx->idx * 4 + 12);
355				PPC_LI(r_ret, 0);
356				PPC_JMP(exit_addr);
357			}
358			if (code == (BPF_ANC | SKF_AD_IFINDEX)) {
359				PPC_LWZ_OFFS(r_A, r_scratch1,
360				     offsetof(struct net_device, ifindex));
361			} else {
362				PPC_LHZ_OFFS(r_A, r_scratch1,
363				     offsetof(struct net_device, type));
364			}
365
366			break;
367		case BPF_ANC | SKF_AD_MARK:
368			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
369			PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
370							  mark));
371			break;
372		case BPF_ANC | SKF_AD_RXHASH:
373			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
374			PPC_LWZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
375							  hash));
376			break;
377		case BPF_ANC | SKF_AD_VLAN_TAG:
378			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
379
380			PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
381							  vlan_tci));
382			break;
383		case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT:
384			PPC_LBZ_OFFS(r_A, r_skb, PKT_VLAN_PRESENT_OFFSET());
385			if (PKT_VLAN_PRESENT_BIT)
386				PPC_SRWI(r_A, r_A, PKT_VLAN_PRESENT_BIT);
387			if (PKT_VLAN_PRESENT_BIT < 7)
388				PPC_ANDI(r_A, r_A, 1);
389			break;
390		case BPF_ANC | SKF_AD_QUEUE:
391			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
392						  queue_mapping) != 2);
393			PPC_LHZ_OFFS(r_A, r_skb, offsetof(struct sk_buff,
394							  queue_mapping));
395			break;
396		case BPF_ANC | SKF_AD_PKTTYPE:
397			PPC_LBZ_OFFS(r_A, r_skb, PKT_TYPE_OFFSET());
398			PPC_ANDI(r_A, r_A, PKT_TYPE_MAX);
399			PPC_SRWI(r_A, r_A, 5);
400			break;
401		case BPF_ANC | SKF_AD_CPU:
402			PPC_BPF_LOAD_CPU(r_A);
403			break;
404			/*** Absolute loads from packet header/data ***/
405		case BPF_LD | BPF_W | BPF_ABS:
406			func = CHOOSE_LOAD_FUNC(K, sk_load_word);
407			goto common_load;
408		case BPF_LD | BPF_H | BPF_ABS:
409			func = CHOOSE_LOAD_FUNC(K, sk_load_half);
410			goto common_load;
411		case BPF_LD | BPF_B | BPF_ABS:
412			func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
413		common_load:
414			/* Load from [K]. */
415			ctx->seen |= SEEN_DATAREF;
416			PPC_FUNC_ADDR(r_scratch1, func);
417			PPC_MTLR(r_scratch1);
418			PPC_LI32(r_addr, K);
419			PPC_BLRL();
420			/*
421			 * Helper returns 'lt' condition on error, and an
422			 * appropriate return value in r3
423			 */
424			PPC_BCC(COND_LT, exit_addr);
425			break;
426
427			/*** Indirect loads from packet header/data ***/
428		case BPF_LD | BPF_W | BPF_IND:
429			func = sk_load_word;
430			goto common_load_ind;
431		case BPF_LD | BPF_H | BPF_IND:
432			func = sk_load_half;
433			goto common_load_ind;
434		case BPF_LD | BPF_B | BPF_IND:
435			func = sk_load_byte;
436		common_load_ind:
437			/*
438			 * Load from [X + K].  Negative offsets are tested for
439			 * in the helper functions.
440			 */
441			ctx->seen |= SEEN_DATAREF | SEEN_XREG;
442			PPC_FUNC_ADDR(r_scratch1, func);
443			PPC_MTLR(r_scratch1);
444			PPC_ADDI(r_addr, r_X, IMM_L(K));
445			if (K >= 32768)
446				PPC_ADDIS(r_addr, r_addr, IMM_HA(K));
447			PPC_BLRL();
448			/* If error, cr0.LT set */
449			PPC_BCC(COND_LT, exit_addr);
450			break;
451
452		case BPF_LDX | BPF_B | BPF_MSH:
453			func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
454			goto common_load;
455			break;
456
457			/*** Jump and branches ***/
458		case BPF_JMP | BPF_JA:
459			if (K != 0)
460				PPC_JMP(addrs[i + 1 + K]);
461			break;
462
463		case BPF_JMP | BPF_JGT | BPF_K:
464		case BPF_JMP | BPF_JGT | BPF_X:
465			true_cond = COND_GT;
466			goto cond_branch;
467		case BPF_JMP | BPF_JGE | BPF_K:
468		case BPF_JMP | BPF_JGE | BPF_X:
469			true_cond = COND_GE;
470			goto cond_branch;
471		case BPF_JMP | BPF_JEQ | BPF_K:
472		case BPF_JMP | BPF_JEQ | BPF_X:
473			true_cond = COND_EQ;
474			goto cond_branch;
475		case BPF_JMP | BPF_JSET | BPF_K:
476		case BPF_JMP | BPF_JSET | BPF_X:
477			true_cond = COND_NE;
478			/* Fall through */
479		cond_branch:
480			/* same targets, can avoid doing the test :) */
481			if (filter[i].jt == filter[i].jf) {
482				if (filter[i].jt > 0)
483					PPC_JMP(addrs[i + 1 + filter[i].jt]);
484				break;
485			}
486
487			switch (code) {
488			case BPF_JMP | BPF_JGT | BPF_X:
489			case BPF_JMP | BPF_JGE | BPF_X:
490			case BPF_JMP | BPF_JEQ | BPF_X:
491				ctx->seen |= SEEN_XREG;
492				PPC_CMPLW(r_A, r_X);
493				break;
494			case BPF_JMP | BPF_JSET | BPF_X:
495				ctx->seen |= SEEN_XREG;
496				PPC_AND_DOT(r_scratch1, r_A, r_X);
497				break;
498			case BPF_JMP | BPF_JEQ | BPF_K:
499			case BPF_JMP | BPF_JGT | BPF_K:
500			case BPF_JMP | BPF_JGE | BPF_K:
501				if (K < 32768)
502					PPC_CMPLWI(r_A, K);
503				else {
504					PPC_LI32(r_scratch1, K);
505					PPC_CMPLW(r_A, r_scratch1);
506				}
507				break;
508			case BPF_JMP | BPF_JSET | BPF_K:
509				if (K < 32768)
510					/* PPC_ANDI is /only/ dot-form */
511					PPC_ANDI(r_scratch1, r_A, K);
512				else {
513					PPC_LI32(r_scratch1, K);
514					PPC_AND_DOT(r_scratch1, r_A,
515						    r_scratch1);
516				}
517				break;
518			}
519			/* Sometimes branches are constructed "backward", with
520			 * the false path being the branch and true path being
521			 * a fallthrough to the next instruction.
522			 */
523			if (filter[i].jt == 0)
524				/* Swap the sense of the branch */
525				PPC_BCC(true_cond ^ COND_CMP_TRUE,
526					addrs[i + 1 + filter[i].jf]);
527			else {
528				PPC_BCC(true_cond, addrs[i + 1 + filter[i].jt]);
529				if (filter[i].jf != 0)
530					PPC_JMP(addrs[i + 1 + filter[i].jf]);
531			}
532			break;
533		default:
534			/* The filter contains something cruel & unusual.
535			 * We don't handle it, but also there shouldn't be
536			 * anything missing from our list.
537			 */
538			if (printk_ratelimit())
539				pr_err("BPF filter opcode %04x (@%d) unsupported\n",
540				       filter[i].code, i);
541			return -ENOTSUPP;
542		}
543
544	}
545	/* Set end-of-body-code address for exit. */
546	addrs[i] = ctx->idx * 4;
547
548	return 0;
549}
550
551void bpf_jit_compile(struct bpf_prog *fp)
 
 
 
 
 
 
 
 
 
 
 
552{
553	unsigned int proglen;
554	unsigned int alloclen;
555	u32 *image = NULL;
 
 
 
 
 
556	u32 *code_base;
557	unsigned int *addrs;
 
558	struct codegen_context cgctx;
559	int pass;
560	int flen = fp->len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
561
562	if (!bpf_jit_enable)
563		return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
564
565	addrs = kcalloc(flen + 1, sizeof(*addrs), GFP_KERNEL);
566	if (addrs == NULL)
567		return;
 
 
568
569	/*
570	 * There are multiple assembly passes as the generated code will change
571	 * size as it settles down, figuring out the max branch offsets/exit
572	 * paths required.
573	 *
574	 * The range of standard conditional branches is +/- 32Kbytes.	Since
575	 * BPF_MAXINSNS = 4096, we can only jump from (worst case) start to
576	 * finish with 8 bytes/instruction.  Not feasible, so long jumps are
577	 * used, distinct from short branches.
578	 *
579	 * Current:
580	 *
581	 * For now, both branch types assemble to 2 words (short branches padded
582	 * with a NOP); this is less efficient, but assembly will always complete
583	 * after exactly 3 passes:
584	 *
585	 * First pass: No code buffer; Program is "faux-generated" -- no code
586	 * emitted but maximum size of output determined (and addrs[] filled
587	 * in).	 Also, we note whether we use M[], whether we use skb data, etc.
588	 * All generation choices assumed to be 'worst-case', e.g. branches all
589	 * far (2 instructions), return path code reduction not available, etc.
590	 *
591	 * Second pass: Code buffer allocated with size determined previously.
592	 * Prologue generated to support features we have seen used.  Exit paths
593	 * determined and addrs[] is filled in again, as code may be slightly
594	 * smaller as a result.
595	 *
596	 * Third pass: Code generated 'for real', and branch destinations
597	 * determined from now-accurate addrs[] map.
598	 *
599	 * Ideal:
600	 *
601	 * If we optimise this, near branches will be shorter.	On the
602	 * first assembly pass, we should err on the side of caution and
603	 * generate the biggest code.  On subsequent passes, branches will be
604	 * generated short or long and code size will reduce.  With smaller
605	 * code, more branches may fall into the short category, and code will
606	 * reduce more.
607	 *
608	 * Finally, if we see one pass generate code the same size as the
609	 * previous pass we have converged and should now generate code for
610	 * real.  Allocating at the end will also save the memory that would
611	 * otherwise be wasted by the (small) current code shrinkage.
612	 * Preferably, we should do a small number of passes (e.g. 5) and if we
613	 * haven't converged by then, get impatient and force code to generate
614	 * as-is, even if the odd branch would be left long.  The chances of a
615	 * long jump are tiny with all but the most enormous of BPF filter
616	 * inputs, so we should usually converge on the third pass.
617	 */
618
619	cgctx.idx = 0;
620	cgctx.seen = 0;
621	cgctx.pc_ret0 = -1;
622	/* Scouting faux-generate pass 0 */
623	if (bpf_jit_build_body(fp, 0, &cgctx, addrs))
624		/* We hit something illegal or unsupported. */
625		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
626
 
627	/*
628	 * Pretend to build prologue, given the features we've seen.  This will
629	 * update ctgtx.idx as it pretends to output instructions, then we can
630	 * calculate total size from idx.
631	 */
632	bpf_jit_build_prologue(fp, 0, &cgctx);
633	bpf_jit_build_epilogue(0, &cgctx);
 
 
 
 
634
635	proglen = cgctx.idx * 4;
636	alloclen = proglen + FUNCTION_DESCR_SIZE;
637	image = module_alloc(alloclen);
638	if (!image)
639		goto out;
 
 
 
 
 
 
 
640
641	code_base = image + (FUNCTION_DESCR_SIZE/4);
 
 
642
643	/* Code generation passes 1-2 */
644	for (pass = 1; pass < 3; pass++) {
645		/* Now build the prologue, body code & epilogue for real. */
646		cgctx.idx = 0;
647		bpf_jit_build_prologue(fp, code_base, &cgctx);
648		bpf_jit_build_body(fp, code_base, &cgctx, addrs);
 
 
 
 
 
 
 
649		bpf_jit_build_epilogue(code_base, &cgctx);
650
651		if (bpf_jit_enable > 1)
652			pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
653				proglen - (cgctx.idx * 4), cgctx.seen);
654	}
655
656	if (bpf_jit_enable > 1)
657		/* Note that we output the base address of the code_base
 
658		 * rather than image, since opcodes are in code_base.
659		 */
660		bpf_jit_dump(flen, proglen, pass, code_base);
661
662	bpf_flush_icache(code_base, code_base + (proglen/4));
663
664#ifdef CONFIG_PPC64
665	/* Function descriptor nastiness: Address + TOC */
666	((u64 *)image)[0] = (u64)code_base;
667	((u64 *)image)[1] = local_paca->kernel_toc;
668#endif
669
670	fp->bpf_func = (void *)image;
671	fp->jited = 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
672
673out:
674	kfree(addrs);
675	return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
676}
677
678void bpf_jit_free(struct bpf_prog *fp)
679{
680	if (fp->jited)
681		module_memfree(fp->bpf_func);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
682
683	bpf_prog_unlock_free(fp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
684}
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * eBPF JIT compiler
   4 *
   5 * Copyright 2016 Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
   6 *		  IBM Corporation
   7 *
   8 * Based on the powerpc classic BPF JIT compiler by Matt Evans
 
   9 */
  10#include <linux/moduleloader.h>
  11#include <asm/cacheflush.h>
  12#include <asm/asm-compat.h>
  13#include <linux/netdevice.h>
  14#include <linux/filter.h>
  15#include <linux/if_vlan.h>
  16#include <linux/kernel.h>
  17#include <linux/memory.h>
  18#include <linux/bpf.h>
  19
  20#include <asm/kprobes.h>
  21#include <asm/text-patching.h>
  22
  23#include "bpf_jit.h"
  24
  25/* These offsets are from bpf prog end and stay the same across progs */
  26static int bpf_jit_ool_stub, bpf_jit_long_branch_stub;
  27
  28static void bpf_jit_fill_ill_insns(void *area, unsigned int size)
  29{
  30	memset32(area, BREAKPOINT_INSTRUCTION, size / 4);
 
  31}
  32
  33void dummy_tramp(void);
 
 
 
 
  34
  35asm (
  36"	.pushsection .text, \"ax\", @progbits	;"
  37"	.global dummy_tramp			;"
  38"	.type dummy_tramp, @function		;"
  39"dummy_tramp:					;"
  40#ifdef CONFIG_PPC_FTRACE_OUT_OF_LINE
  41"	blr					;"
  42#else
  43/* LR is always in r11, so we don't need a 'mflr r11' here */
  44"	mtctr	11				;"
  45"	mtlr	0				;"
  46"	bctr					;"
  47#endif
  48"	.size dummy_tramp, .-dummy_tramp	;"
  49"	.popsection				;"
  50);
 
 
 
 
 
 
 
  51
  52void bpf_jit_build_fentry_stubs(u32 *image, struct codegen_context *ctx)
  53{
  54	int ool_stub_idx, long_branch_stub_idx;
 
 
 
 
 
 
 
 
 
 
  55
  56	/*
  57	 * Out-of-line stub:
  58	 *	mflr	r0
  59	 *	[b|bl]	tramp
  60	 *	mtlr	r0 // only with CONFIG_PPC_FTRACE_OUT_OF_LINE
  61	 *	b	bpf_func + 4
  62	 */
  63	ool_stub_idx = ctx->idx;
  64	EMIT(PPC_RAW_MFLR(_R0));
  65	EMIT(PPC_RAW_NOP());
  66	if (IS_ENABLED(CONFIG_PPC_FTRACE_OUT_OF_LINE))
  67		EMIT(PPC_RAW_MTLR(_R0));
  68	WARN_ON_ONCE(!is_offset_in_branch_range(4 - (long)ctx->idx * 4));
  69	EMIT(PPC_RAW_BRANCH(4 - (long)ctx->idx * 4));
  70
  71	/*
  72	 * Long branch stub:
  73	 *	.long	<dummy_tramp_addr>
  74	 *	mflr	r11
  75	 *	bcl	20,31,$+4
  76	 *	mflr	r12
  77	 *	ld	r12, -8-SZL(r12)
  78	 *	mtctr	r12
  79	 *	mtlr	r11 // needed to retain ftrace ABI
  80	 *	bctr
  81	 */
  82	if (image)
  83		*((unsigned long *)&image[ctx->idx]) = (unsigned long)dummy_tramp;
  84	ctx->idx += SZL / 4;
  85	long_branch_stub_idx = ctx->idx;
  86	EMIT(PPC_RAW_MFLR(_R11));
  87	EMIT(PPC_RAW_BCL4());
  88	EMIT(PPC_RAW_MFLR(_R12));
  89	EMIT(PPC_RAW_LL(_R12, _R12, -8-SZL));
  90	EMIT(PPC_RAW_MTCTR(_R12));
  91	EMIT(PPC_RAW_MTLR(_R11));
  92	EMIT(PPC_RAW_BCTR());
  93
  94	if (!bpf_jit_ool_stub) {
  95		bpf_jit_ool_stub = (ctx->idx - ool_stub_idx) * 4;
  96		bpf_jit_long_branch_stub = (ctx->idx - long_branch_stub_idx) * 4;
  97	}
  98}
  99
 100int bpf_jit_emit_exit_insn(u32 *image, struct codegen_context *ctx, int tmp_reg, long exit_addr)
 101{
 102	if (!exit_addr || is_offset_in_branch_range(exit_addr - (ctx->idx * 4))) {
 103		PPC_JMP(exit_addr);
 104	} else if (ctx->alt_exit_addr) {
 105		if (WARN_ON(!is_offset_in_branch_range((long)ctx->alt_exit_addr - (ctx->idx * 4))))
 106			return -1;
 107		PPC_JMP(ctx->alt_exit_addr);
 108	} else {
 109		ctx->alt_exit_addr = ctx->idx * 4;
 110		bpf_jit_build_epilogue(image, ctx);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 111	}
 
 
 112
 113	return 0;
 114}
 115
 116struct powerpc_jit_data {
 117	/* address of rw header */
 118	struct bpf_binary_header *hdr;
 119	/* address of ro final header */
 120	struct bpf_binary_header *fhdr;
 121	u32 *addrs;
 122	u8 *fimage;
 123	u32 proglen;
 124	struct codegen_context ctx;
 125};
 126
 127bool bpf_jit_needs_zext(void)
 128{
 129	return true;
 130}
 131
 132struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *fp)
 133{
 134	u32 proglen;
 135	u32 alloclen;
 136	u8 *image = NULL;
 137	u32 *code_base;
 138	u32 *addrs;
 139	struct powerpc_jit_data *jit_data;
 140	struct codegen_context cgctx;
 141	int pass;
 142	int flen;
 143	struct bpf_binary_header *fhdr = NULL;
 144	struct bpf_binary_header *hdr = NULL;
 145	struct bpf_prog *org_fp = fp;
 146	struct bpf_prog *tmp_fp;
 147	bool bpf_blinded = false;
 148	bool extra_pass = false;
 149	u8 *fimage = NULL;
 150	u32 *fcode_base;
 151	u32 extable_len;
 152	u32 fixup_len;
 153
 154	if (!fp->jit_requested)
 155		return org_fp;
 156
 157	tmp_fp = bpf_jit_blind_constants(org_fp);
 158	if (IS_ERR(tmp_fp))
 159		return org_fp;
 160
 161	if (tmp_fp != org_fp) {
 162		bpf_blinded = true;
 163		fp = tmp_fp;
 164	}
 165
 166	jit_data = fp->aux->jit_data;
 167	if (!jit_data) {
 168		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
 169		if (!jit_data) {
 170			fp = org_fp;
 171			goto out;
 172		}
 173		fp->aux->jit_data = jit_data;
 174	}
 175
 176	flen = fp->len;
 177	addrs = jit_data->addrs;
 178	if (addrs) {
 179		cgctx = jit_data->ctx;
 180		/*
 181		 * JIT compiled to a writable location (image/code_base) first.
 182		 * It is then moved to the readonly final location (fimage/fcode_base)
 183		 * using instruction patching.
 184		 */
 185		fimage = jit_data->fimage;
 186		fhdr = jit_data->fhdr;
 187		proglen = jit_data->proglen;
 188		hdr = jit_data->hdr;
 189		image = (void *)hdr + ((void *)fimage - (void *)fhdr);
 190		extra_pass = true;
 191		/* During extra pass, ensure index is reset before repopulating extable entries */
 192		cgctx.exentry_idx = 0;
 193		goto skip_init_ctx;
 194	}
 195
 196	addrs = kcalloc(flen + 1, sizeof(*addrs), GFP_KERNEL);
 197	if (addrs == NULL) {
 198		fp = org_fp;
 199		goto out_addrs;
 200	}
 201
 202	memset(&cgctx, 0, sizeof(struct codegen_context));
 203	bpf_jit_init_reg_mapping(&cgctx);
 204
 205	/* Make sure that the stack is quadword aligned. */
 206	cgctx.stack_size = round_up(fp->aux->stack_depth, 16);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 207
 
 
 
 208	/* Scouting faux-generate pass 0 */
 209	if (bpf_jit_build_body(fp, NULL, NULL, &cgctx, addrs, 0, false)) {
 210		/* We hit something illegal or unsupported. */
 211		fp = org_fp;
 212		goto out_addrs;
 213	}
 214
 215	/*
 216	 * If we have seen a tail call, we need a second pass.
 217	 * This is because bpf_jit_emit_common_epilogue() is called
 218	 * from bpf_jit_emit_tail_call() with a not yet stable ctx->seen.
 219	 * We also need a second pass if we ended up with too large
 220	 * a program so as to ensure BPF_EXIT branches are in range.
 221	 */
 222	if (cgctx.seen & SEEN_TAILCALL || !is_offset_in_branch_range((long)cgctx.idx * 4)) {
 223		cgctx.idx = 0;
 224		if (bpf_jit_build_body(fp, NULL, NULL, &cgctx, addrs, 0, false)) {
 225			fp = org_fp;
 226			goto out_addrs;
 227		}
 228	}
 229
 230	bpf_jit_realloc_regs(&cgctx);
 231	/*
 232	 * Pretend to build prologue, given the features we've seen.  This will
 233	 * update ctgtx.idx as it pretends to output instructions, then we can
 234	 * calculate total size from idx.
 235	 */
 236	bpf_jit_build_prologue(NULL, &cgctx);
 237	addrs[fp->len] = cgctx.idx * 4;
 238	bpf_jit_build_epilogue(NULL, &cgctx);
 239
 240	fixup_len = fp->aux->num_exentries * BPF_FIXUP_LEN * 4;
 241	extable_len = fp->aux->num_exentries * sizeof(struct exception_table_entry);
 242
 243	proglen = cgctx.idx * 4;
 244	alloclen = proglen + FUNCTION_DESCR_SIZE + fixup_len + extable_len;
 245
 246	fhdr = bpf_jit_binary_pack_alloc(alloclen, &fimage, 4, &hdr, &image,
 247					      bpf_jit_fill_ill_insns);
 248	if (!fhdr) {
 249		fp = org_fp;
 250		goto out_addrs;
 251	}
 252
 253	if (extable_len)
 254		fp->aux->extable = (void *)fimage + FUNCTION_DESCR_SIZE + proglen + fixup_len;
 255
 256skip_init_ctx:
 257	code_base = (u32 *)(image + FUNCTION_DESCR_SIZE);
 258	fcode_base = (u32 *)(fimage + FUNCTION_DESCR_SIZE);
 259
 260	/* Code generation passes 1-2 */
 261	for (pass = 1; pass < 3; pass++) {
 262		/* Now build the prologue, body code & epilogue for real. */
 263		cgctx.idx = 0;
 264		cgctx.alt_exit_addr = 0;
 265		bpf_jit_build_prologue(code_base, &cgctx);
 266		if (bpf_jit_build_body(fp, code_base, fcode_base, &cgctx, addrs, pass,
 267				       extra_pass)) {
 268			bpf_arch_text_copy(&fhdr->size, &hdr->size, sizeof(hdr->size));
 269			bpf_jit_binary_pack_free(fhdr, hdr);
 270			fp = org_fp;
 271			goto out_addrs;
 272		}
 273		bpf_jit_build_epilogue(code_base, &cgctx);
 274
 275		if (bpf_jit_enable > 1)
 276			pr_info("Pass %d: shrink = %d, seen = 0x%x\n", pass,
 277				proglen - (cgctx.idx * 4), cgctx.seen);
 278	}
 279
 280	if (bpf_jit_enable > 1)
 281		/*
 282		 * Note that we output the base address of the code_base
 283		 * rather than image, since opcodes are in code_base.
 284		 */
 285		bpf_jit_dump(flen, proglen, pass, code_base);
 286
 287#ifdef CONFIG_PPC64_ELF_ABI_V1
 
 
 288	/* Function descriptor nastiness: Address + TOC */
 289	((u64 *)image)[0] = (u64)fcode_base;
 290	((u64 *)image)[1] = local_paca->kernel_toc;
 291#endif
 292
 293	fp->bpf_func = (void *)fimage;
 294	fp->jited = 1;
 295	fp->jited_len = cgctx.idx * 4 + FUNCTION_DESCR_SIZE;
 296
 297	if (!fp->is_func || extra_pass) {
 298		if (bpf_jit_binary_pack_finalize(fhdr, hdr)) {
 299			fp = org_fp;
 300			goto out_addrs;
 301		}
 302		bpf_prog_fill_jited_linfo(fp, addrs);
 303out_addrs:
 304		kfree(addrs);
 305		kfree(jit_data);
 306		fp->aux->jit_data = NULL;
 307	} else {
 308		jit_data->addrs = addrs;
 309		jit_data->ctx = cgctx;
 310		jit_data->proglen = proglen;
 311		jit_data->fimage = fimage;
 312		jit_data->fhdr = fhdr;
 313		jit_data->hdr = hdr;
 314	}
 315
 316out:
 317	if (bpf_blinded)
 318		bpf_jit_prog_release_other(fp, fp == org_fp ? tmp_fp : org_fp);
 319
 320	return fp;
 321}
 322
 323/*
 324 * The caller should check for (BPF_MODE(code) == BPF_PROBE_MEM) before calling
 325 * this function, as this only applies to BPF_PROBE_MEM, for now.
 326 */
 327int bpf_add_extable_entry(struct bpf_prog *fp, u32 *image, u32 *fimage, int pass,
 328			  struct codegen_context *ctx, int insn_idx, int jmp_off,
 329			  int dst_reg)
 330{
 331	off_t offset;
 332	unsigned long pc;
 333	struct exception_table_entry *ex, *ex_entry;
 334	u32 *fixup;
 335
 336	/* Populate extable entries only in the last pass */
 337	if (pass != 2)
 338		return 0;
 339
 340	if (!fp->aux->extable ||
 341	    WARN_ON_ONCE(ctx->exentry_idx >= fp->aux->num_exentries))
 342		return -EINVAL;
 343
 344	/*
 345	 * Program is first written to image before copying to the
 346	 * final location (fimage). Accordingly, update in the image first.
 347	 * As all offsets used are relative, copying as is to the
 348	 * final location should be alright.
 349	 */
 350	pc = (unsigned long)&image[insn_idx];
 351	ex = (void *)fp->aux->extable - (void *)fimage + (void *)image;
 352
 353	fixup = (void *)ex -
 354		(fp->aux->num_exentries * BPF_FIXUP_LEN * 4) +
 355		(ctx->exentry_idx * BPF_FIXUP_LEN * 4);
 356
 357	fixup[0] = PPC_RAW_LI(dst_reg, 0);
 358	if (IS_ENABLED(CONFIG_PPC32))
 359		fixup[1] = PPC_RAW_LI(dst_reg - 1, 0); /* clear higher 32-bit register too */
 360
 361	fixup[BPF_FIXUP_LEN - 1] =
 362		PPC_RAW_BRANCH((long)(pc + jmp_off) - (long)&fixup[BPF_FIXUP_LEN - 1]);
 363
 364	ex_entry = &ex[ctx->exentry_idx];
 365
 366	offset = pc - (long)&ex_entry->insn;
 367	if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
 368		return -ERANGE;
 369	ex_entry->insn = offset;
 370
 371	offset = (long)fixup - (long)&ex_entry->fixup;
 372	if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
 373		return -ERANGE;
 374	ex_entry->fixup = offset;
 375
 376	ctx->exentry_idx++;
 377	return 0;
 378}
 379
 380void *bpf_arch_text_copy(void *dst, void *src, size_t len)
 381{
 382	int err;
 383
 384	if (WARN_ON_ONCE(core_kernel_text((unsigned long)dst)))
 385		return ERR_PTR(-EINVAL);
 386
 387	mutex_lock(&text_mutex);
 388	err = patch_instructions(dst, src, len, false);
 389	mutex_unlock(&text_mutex);
 390
 391	return err ? ERR_PTR(err) : dst;
 392}
 393
 394int bpf_arch_text_invalidate(void *dst, size_t len)
 395{
 396	u32 insn = BREAKPOINT_INSTRUCTION;
 397	int ret;
 398
 399	if (WARN_ON_ONCE(core_kernel_text((unsigned long)dst)))
 400		return -EINVAL;
 401
 402	mutex_lock(&text_mutex);
 403	ret = patch_instructions(dst, &insn, len, true);
 404	mutex_unlock(&text_mutex);
 405
 406	return ret;
 407}
 408
 409void bpf_jit_free(struct bpf_prog *fp)
 410{
 411	if (fp->jited) {
 412		struct powerpc_jit_data *jit_data = fp->aux->jit_data;
 413		struct bpf_binary_header *hdr;
 414
 415		/*
 416		 * If we fail the final pass of JIT (from jit_subprogs),
 417		 * the program may not be finalized yet. Call finalize here
 418		 * before freeing it.
 419		 */
 420		if (jit_data) {
 421			bpf_jit_binary_pack_finalize(jit_data->fhdr, jit_data->hdr);
 422			kvfree(jit_data->addrs);
 423			kfree(jit_data);
 424		}
 425		hdr = bpf_jit_binary_pack_hdr(fp);
 426		bpf_jit_binary_pack_free(hdr, NULL);
 427		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
 428	}
 429
 430	bpf_prog_unlock_free(fp);
 431}
 432
 433bool bpf_jit_supports_kfunc_call(void)
 434{
 435	return true;
 436}
 437
 438bool bpf_jit_supports_far_kfunc_call(void)
 439{
 440	return IS_ENABLED(CONFIG_PPC64);
 441}
 442
 443void *arch_alloc_bpf_trampoline(unsigned int size)
 444{
 445	return bpf_prog_pack_alloc(size, bpf_jit_fill_ill_insns);
 446}
 447
 448void arch_free_bpf_trampoline(void *image, unsigned int size)
 449{
 450	bpf_prog_pack_free(image, size);
 451}
 452
 453int arch_protect_bpf_trampoline(void *image, unsigned int size)
 454{
 455	return 0;
 456}
 457
 458static int invoke_bpf_prog(u32 *image, u32 *ro_image, struct codegen_context *ctx,
 459			   struct bpf_tramp_link *l, int regs_off, int retval_off,
 460			   int run_ctx_off, bool save_ret)
 461{
 462	struct bpf_prog *p = l->link.prog;
 463	ppc_inst_t branch_insn;
 464	u32 jmp_idx;
 465	int ret = 0;
 466
 467	/* Save cookie */
 468	if (IS_ENABLED(CONFIG_PPC64)) {
 469		PPC_LI64(_R3, l->cookie);
 470		EMIT(PPC_RAW_STD(_R3, _R1, run_ctx_off + offsetof(struct bpf_tramp_run_ctx,
 471				 bpf_cookie)));
 472	} else {
 473		PPC_LI32(_R3, l->cookie >> 32);
 474		PPC_LI32(_R4, l->cookie);
 475		EMIT(PPC_RAW_STW(_R3, _R1,
 476				 run_ctx_off + offsetof(struct bpf_tramp_run_ctx, bpf_cookie)));
 477		EMIT(PPC_RAW_STW(_R4, _R1,
 478				 run_ctx_off + offsetof(struct bpf_tramp_run_ctx, bpf_cookie) + 4));
 479	}
 480
 481	/* __bpf_prog_enter(p, &bpf_tramp_run_ctx) */
 482	PPC_LI_ADDR(_R3, p);
 483	EMIT(PPC_RAW_MR(_R25, _R3));
 484	EMIT(PPC_RAW_ADDI(_R4, _R1, run_ctx_off));
 485	ret = bpf_jit_emit_func_call_rel(image, ro_image, ctx,
 486					 (unsigned long)bpf_trampoline_enter(p));
 487	if (ret)
 488		return ret;
 489
 490	/* Remember prog start time returned by __bpf_prog_enter */
 491	EMIT(PPC_RAW_MR(_R26, _R3));
 492
 493	/*
 494	 * if (__bpf_prog_enter(p) == 0)
 495	 *	goto skip_exec_of_prog;
 496	 *
 497	 * Emit a nop to be later patched with conditional branch, once offset is known
 498	 */
 499	EMIT(PPC_RAW_CMPLI(_R3, 0));
 500	jmp_idx = ctx->idx;
 501	EMIT(PPC_RAW_NOP());
 502
 503	/* p->bpf_func(ctx) */
 504	EMIT(PPC_RAW_ADDI(_R3, _R1, regs_off));
 505	if (!p->jited)
 506		PPC_LI_ADDR(_R4, (unsigned long)p->insnsi);
 507	if (!create_branch(&branch_insn, (u32 *)&ro_image[ctx->idx], (unsigned long)p->bpf_func,
 508			   BRANCH_SET_LINK)) {
 509		if (image)
 510			image[ctx->idx] = ppc_inst_val(branch_insn);
 511		ctx->idx++;
 512	} else {
 513		EMIT(PPC_RAW_LL(_R12, _R25, offsetof(struct bpf_prog, bpf_func)));
 514		EMIT(PPC_RAW_MTCTR(_R12));
 515		EMIT(PPC_RAW_BCTRL());
 516	}
 517
 518	if (save_ret)
 519		EMIT(PPC_RAW_STL(_R3, _R1, retval_off));
 520
 521	/* Fix up branch */
 522	if (image) {
 523		if (create_cond_branch(&branch_insn, &image[jmp_idx],
 524				       (unsigned long)&image[ctx->idx], COND_EQ << 16))
 525			return -EINVAL;
 526		image[jmp_idx] = ppc_inst_val(branch_insn);
 527	}
 528
 529	/* __bpf_prog_exit(p, start_time, &bpf_tramp_run_ctx) */
 530	EMIT(PPC_RAW_MR(_R3, _R25));
 531	EMIT(PPC_RAW_MR(_R4, _R26));
 532	EMIT(PPC_RAW_ADDI(_R5, _R1, run_ctx_off));
 533	ret = bpf_jit_emit_func_call_rel(image, ro_image, ctx,
 534					 (unsigned long)bpf_trampoline_exit(p));
 535
 536	return ret;
 537}
 538
 539static int invoke_bpf_mod_ret(u32 *image, u32 *ro_image, struct codegen_context *ctx,
 540			      struct bpf_tramp_links *tl, int regs_off, int retval_off,
 541			      int run_ctx_off, u32 *branches)
 542{
 543	int i;
 544
 545	/*
 546	 * The first fmod_ret program will receive a garbage return value.
 547	 * Set this to 0 to avoid confusing the program.
 548	 */
 549	EMIT(PPC_RAW_LI(_R3, 0));
 550	EMIT(PPC_RAW_STL(_R3, _R1, retval_off));
 551	for (i = 0; i < tl->nr_links; i++) {
 552		if (invoke_bpf_prog(image, ro_image, ctx, tl->links[i], regs_off, retval_off,
 553				    run_ctx_off, true))
 554			return -EINVAL;
 555
 556		/*
 557		 * mod_ret prog stored return value after prog ctx. Emit:
 558		 * if (*(u64 *)(ret_val) !=  0)
 559		 *	goto do_fexit;
 560		 */
 561		EMIT(PPC_RAW_LL(_R3, _R1, retval_off));
 562		EMIT(PPC_RAW_CMPLI(_R3, 0));
 563
 564		/*
 565		 * Save the location of the branch and generate a nop, which is
 566		 * replaced with a conditional jump once do_fexit (i.e. the
 567		 * start of the fexit invocation) is finalized.
 568		 */
 569		branches[i] = ctx->idx;
 570		EMIT(PPC_RAW_NOP());
 571	}
 572
 573	return 0;
 574}
 575
 576static void bpf_trampoline_setup_tail_call_cnt(u32 *image, struct codegen_context *ctx,
 577					       int func_frame_offset, int r4_off)
 578{
 579	if (IS_ENABLED(CONFIG_PPC64)) {
 580		/* See bpf_jit_stack_tailcallcnt() */
 581		int tailcallcnt_offset = 6 * 8;
 582
 583		EMIT(PPC_RAW_LL(_R3, _R1, func_frame_offset - tailcallcnt_offset));
 584		EMIT(PPC_RAW_STL(_R3, _R1, -tailcallcnt_offset));
 585	} else {
 586		/* See bpf_jit_stack_offsetof() and BPF_PPC_TC */
 587		EMIT(PPC_RAW_LL(_R4, _R1, r4_off));
 588	}
 589}
 590
 591static void bpf_trampoline_restore_tail_call_cnt(u32 *image, struct codegen_context *ctx,
 592						 int func_frame_offset, int r4_off)
 593{
 594	if (IS_ENABLED(CONFIG_PPC64)) {
 595		/* See bpf_jit_stack_tailcallcnt() */
 596		int tailcallcnt_offset = 6 * 8;
 597
 598		EMIT(PPC_RAW_LL(_R3, _R1, -tailcallcnt_offset));
 599		EMIT(PPC_RAW_STL(_R3, _R1, func_frame_offset - tailcallcnt_offset));
 600	} else {
 601		/* See bpf_jit_stack_offsetof() and BPF_PPC_TC */
 602		EMIT(PPC_RAW_STL(_R4, _R1, r4_off));
 603	}
 604}
 605
 606static void bpf_trampoline_save_args(u32 *image, struct codegen_context *ctx, int func_frame_offset,
 607				     int nr_regs, int regs_off)
 608{
 609	int param_save_area_offset;
 610
 611	param_save_area_offset = func_frame_offset; /* the two frames we alloted */
 612	param_save_area_offset += STACK_FRAME_MIN_SIZE; /* param save area is past frame header */
 613
 614	for (int i = 0; i < nr_regs; i++) {
 615		if (i < 8) {
 616			EMIT(PPC_RAW_STL(_R3 + i, _R1, regs_off + i * SZL));
 617		} else {
 618			EMIT(PPC_RAW_LL(_R3, _R1, param_save_area_offset + i * SZL));
 619			EMIT(PPC_RAW_STL(_R3, _R1, regs_off + i * SZL));
 620		}
 621	}
 622}
 623
 624/* Used when restoring just the register parameters when returning back */
 625static void bpf_trampoline_restore_args_regs(u32 *image, struct codegen_context *ctx,
 626					     int nr_regs, int regs_off)
 627{
 628	for (int i = 0; i < nr_regs && i < 8; i++)
 629		EMIT(PPC_RAW_LL(_R3 + i, _R1, regs_off + i * SZL));
 630}
 631
 632/* Used when we call into the traced function. Replicate parameter save area */
 633static void bpf_trampoline_restore_args_stack(u32 *image, struct codegen_context *ctx,
 634					      int func_frame_offset, int nr_regs, int regs_off)
 635{
 636	int param_save_area_offset;
 637
 638	param_save_area_offset = func_frame_offset; /* the two frames we alloted */
 639	param_save_area_offset += STACK_FRAME_MIN_SIZE; /* param save area is past frame header */
 640
 641	for (int i = 8; i < nr_regs; i++) {
 642		EMIT(PPC_RAW_LL(_R3, _R1, param_save_area_offset + i * SZL));
 643		EMIT(PPC_RAW_STL(_R3, _R1, STACK_FRAME_MIN_SIZE + i * SZL));
 644	}
 645	bpf_trampoline_restore_args_regs(image, ctx, nr_regs, regs_off);
 646}
 647
 648static int __arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *rw_image,
 649					 void *rw_image_end, void *ro_image,
 650					 const struct btf_func_model *m, u32 flags,
 651					 struct bpf_tramp_links *tlinks,
 652					 void *func_addr)
 653{
 654	int regs_off, nregs_off, ip_off, run_ctx_off, retval_off, nvr_off, alt_lr_off, r4_off = 0;
 655	int i, ret, nr_regs, bpf_frame_size = 0, bpf_dummy_frame_size = 0, func_frame_offset;
 656	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
 657	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
 658	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
 659	struct codegen_context codegen_ctx, *ctx;
 660	u32 *image = (u32 *)rw_image;
 661	ppc_inst_t branch_insn;
 662	u32 *branches = NULL;
 663	bool save_ret;
 664
 665	if (IS_ENABLED(CONFIG_PPC32))
 666		return -EOPNOTSUPP;
 667
 668	nr_regs = m->nr_args;
 669	/* Extra registers for struct arguments */
 670	for (i = 0; i < m->nr_args; i++)
 671		if (m->arg_size[i] > SZL)
 672			nr_regs += round_up(m->arg_size[i], SZL) / SZL - 1;
 673
 674	if (nr_regs > MAX_BPF_FUNC_ARGS)
 675		return -EOPNOTSUPP;
 676
 677	ctx = &codegen_ctx;
 678	memset(ctx, 0, sizeof(*ctx));
 679
 680	/*
 681	 * Generated stack layout:
 682	 *
 683	 * func prev back chain         [ back chain        ]
 684	 *                              [                   ]
 685	 * bpf prog redzone/tailcallcnt [ ...               ] 64 bytes (64-bit powerpc)
 686	 *                              [                   ] --
 687	 * LR save area                 [ r0 save (64-bit)  ]   | header
 688	 *                              [ r0 save (32-bit)  ]   |
 689	 * dummy frame for unwind       [ back chain 1      ] --
 690	 *                              [ padding           ] align stack frame
 691	 *       r4_off                 [ r4 (tailcallcnt)  ] optional - 32-bit powerpc
 692	 *       alt_lr_off             [ real lr (ool stub)] optional - actual lr
 693	 *                              [ r26               ]
 694	 *       nvr_off                [ r25               ] nvr save area
 695	 *       retval_off             [ return value      ]
 696	 *                              [ reg argN          ]
 697	 *                              [ ...               ]
 698	 *       regs_off               [ reg_arg1          ] prog ctx context
 699	 *       nregs_off              [ args count        ]
 700	 *       ip_off                 [ traced function   ]
 701	 *                              [ ...               ]
 702	 *       run_ctx_off            [ bpf_tramp_run_ctx ]
 703	 *                              [ reg argN          ]
 704	 *                              [ ...               ]
 705	 *       param_save_area        [ reg_arg1          ] min 8 doublewords, per ABI
 706	 *                              [ TOC save (64-bit) ] --
 707	 *                              [ LR save (64-bit)  ]   | header
 708	 *                              [ LR save (32-bit)  ]   |
 709	 * bpf trampoline frame	        [ back chain 2      ] --
 710	 *
 711	 */
 712
 713	/* Minimum stack frame header */
 714	bpf_frame_size = STACK_FRAME_MIN_SIZE;
 715
 716	/*
 717	 * Room for parameter save area.
 718	 *
 719	 * As per the ABI, this is required if we call into the traced
 720	 * function (BPF_TRAMP_F_CALL_ORIG):
 721	 * - if the function takes more than 8 arguments for the rest to spill onto the stack
 722	 * - or, if the function has variadic arguments
 723	 * - or, if this functions's prototype was not available to the caller
 724	 *
 725	 * Reserve space for at least 8 registers for now. This can be optimized later.
 726	 */
 727	bpf_frame_size += (nr_regs > 8 ? nr_regs : 8) * SZL;
 728
 729	/* Room for struct bpf_tramp_run_ctx */
 730	run_ctx_off = bpf_frame_size;
 731	bpf_frame_size += round_up(sizeof(struct bpf_tramp_run_ctx), SZL);
 732
 733	/* Room for IP address argument */
 734	ip_off = bpf_frame_size;
 735	if (flags & BPF_TRAMP_F_IP_ARG)
 736		bpf_frame_size += SZL;
 737
 738	/* Room for args count */
 739	nregs_off = bpf_frame_size;
 740	bpf_frame_size += SZL;
 741
 742	/* Room for args */
 743	regs_off = bpf_frame_size;
 744	bpf_frame_size += nr_regs * SZL;
 745
 746	/* Room for return value of func_addr or fentry prog */
 747	retval_off = bpf_frame_size;
 748	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
 749	if (save_ret)
 750		bpf_frame_size += SZL;
 751
 752	/* Room for nvr save area */
 753	nvr_off = bpf_frame_size;
 754	bpf_frame_size += 2 * SZL;
 755
 756	/* Optional save area for actual LR in case of ool ftrace */
 757	if (IS_ENABLED(CONFIG_PPC_FTRACE_OUT_OF_LINE)) {
 758		alt_lr_off = bpf_frame_size;
 759		bpf_frame_size += SZL;
 760	}
 761
 762	if (IS_ENABLED(CONFIG_PPC32)) {
 763		if (nr_regs < 2) {
 764			r4_off = bpf_frame_size;
 765			bpf_frame_size += SZL;
 766		} else {
 767			r4_off = regs_off + SZL;
 768		}
 769	}
 770
 771	/* Padding to align stack frame, if any */
 772	bpf_frame_size = round_up(bpf_frame_size, SZL * 2);
 773
 774	/* Dummy frame size for proper unwind - includes 64-bytes red zone for 64-bit powerpc */
 775	bpf_dummy_frame_size = STACK_FRAME_MIN_SIZE + 64;
 776
 777	/* Offset to the traced function's stack frame */
 778	func_frame_offset = bpf_dummy_frame_size + bpf_frame_size;
 779
 780	/* Create dummy frame for unwind, store original return value */
 781	EMIT(PPC_RAW_STL(_R0, _R1, PPC_LR_STKOFF));
 782	/* Protect red zone where tail call count goes */
 783	EMIT(PPC_RAW_STLU(_R1, _R1, -bpf_dummy_frame_size));
 784
 785	/* Create our stack frame */
 786	EMIT(PPC_RAW_STLU(_R1, _R1, -bpf_frame_size));
 787
 788	/* 64-bit: Save TOC and load kernel TOC */
 789	if (IS_ENABLED(CONFIG_PPC64_ELF_ABI_V2) && !IS_ENABLED(CONFIG_PPC_KERNEL_PCREL)) {
 790		EMIT(PPC_RAW_STD(_R2, _R1, 24));
 791		PPC64_LOAD_PACA();
 792	}
 793
 794	/* 32-bit: save tail call count in r4 */
 795	if (IS_ENABLED(CONFIG_PPC32) && nr_regs < 2)
 796		EMIT(PPC_RAW_STL(_R4, _R1, r4_off));
 797
 798	bpf_trampoline_save_args(image, ctx, func_frame_offset, nr_regs, regs_off);
 799
 800	/* Save our return address */
 801	EMIT(PPC_RAW_MFLR(_R3));
 802	if (IS_ENABLED(CONFIG_PPC_FTRACE_OUT_OF_LINE))
 803		EMIT(PPC_RAW_STL(_R3, _R1, alt_lr_off));
 804	else
 805		EMIT(PPC_RAW_STL(_R3, _R1, bpf_frame_size + PPC_LR_STKOFF));
 806
 807	/*
 808	 * Save ip address of the traced function.
 809	 * We could recover this from LR, but we will need to address for OOL trampoline,
 810	 * and optional GEP area.
 811	 */
 812	if (IS_ENABLED(CONFIG_PPC_FTRACE_OUT_OF_LINE) || flags & BPF_TRAMP_F_IP_ARG) {
 813		EMIT(PPC_RAW_LWZ(_R4, _R3, 4));
 814		EMIT(PPC_RAW_SLWI(_R4, _R4, 6));
 815		EMIT(PPC_RAW_SRAWI(_R4, _R4, 6));
 816		EMIT(PPC_RAW_ADD(_R3, _R3, _R4));
 817		EMIT(PPC_RAW_ADDI(_R3, _R3, 4));
 818	}
 819
 820	if (flags & BPF_TRAMP_F_IP_ARG)
 821		EMIT(PPC_RAW_STL(_R3, _R1, ip_off));
 822
 823	if (IS_ENABLED(CONFIG_PPC_FTRACE_OUT_OF_LINE))
 824		/* Fake our LR for unwind */
 825		EMIT(PPC_RAW_STL(_R3, _R1, bpf_frame_size + PPC_LR_STKOFF));
 826
 827	/* Save function arg count -- see bpf_get_func_arg_cnt() */
 828	EMIT(PPC_RAW_LI(_R3, nr_regs));
 829	EMIT(PPC_RAW_STL(_R3, _R1, nregs_off));
 830
 831	/* Save nv regs */
 832	EMIT(PPC_RAW_STL(_R25, _R1, nvr_off));
 833	EMIT(PPC_RAW_STL(_R26, _R1, nvr_off + SZL));
 834
 835	if (flags & BPF_TRAMP_F_CALL_ORIG) {
 836		PPC_LI_ADDR(_R3, (unsigned long)im);
 837		ret = bpf_jit_emit_func_call_rel(image, ro_image, ctx,
 838						 (unsigned long)__bpf_tramp_enter);
 839		if (ret)
 840			return ret;
 841	}
 842
 843	for (i = 0; i < fentry->nr_links; i++)
 844		if (invoke_bpf_prog(image, ro_image, ctx, fentry->links[i], regs_off, retval_off,
 845				    run_ctx_off, flags & BPF_TRAMP_F_RET_FENTRY_RET))
 846			return -EINVAL;
 847
 848	if (fmod_ret->nr_links) {
 849		branches = kcalloc(fmod_ret->nr_links, sizeof(u32), GFP_KERNEL);
 850		if (!branches)
 851			return -ENOMEM;
 852
 853		if (invoke_bpf_mod_ret(image, ro_image, ctx, fmod_ret, regs_off, retval_off,
 854				       run_ctx_off, branches)) {
 855			ret = -EINVAL;
 856			goto cleanup;
 857		}
 858	}
 859
 860	/* Call the traced function */
 861	if (flags & BPF_TRAMP_F_CALL_ORIG) {
 862		/*
 863		 * The address in LR save area points to the correct point in the original function
 864		 * with both PPC_FTRACE_OUT_OF_LINE as well as with traditional ftrace instruction
 865		 * sequence
 866		 */
 867		EMIT(PPC_RAW_LL(_R3, _R1, bpf_frame_size + PPC_LR_STKOFF));
 868		EMIT(PPC_RAW_MTCTR(_R3));
 869
 870		/* Replicate tail_call_cnt before calling the original BPF prog */
 871		if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
 872			bpf_trampoline_setup_tail_call_cnt(image, ctx, func_frame_offset, r4_off);
 873
 874		/* Restore args */
 875		bpf_trampoline_restore_args_stack(image, ctx, func_frame_offset, nr_regs, regs_off);
 876
 877		/* Restore TOC for 64-bit */
 878		if (IS_ENABLED(CONFIG_PPC64_ELF_ABI_V2) && !IS_ENABLED(CONFIG_PPC_KERNEL_PCREL))
 879			EMIT(PPC_RAW_LD(_R2, _R1, 24));
 880		EMIT(PPC_RAW_BCTRL());
 881		if (IS_ENABLED(CONFIG_PPC64_ELF_ABI_V2) && !IS_ENABLED(CONFIG_PPC_KERNEL_PCREL))
 882			PPC64_LOAD_PACA();
 883
 884		/* Store return value for bpf prog to access */
 885		EMIT(PPC_RAW_STL(_R3, _R1, retval_off));
 886
 887		/* Restore updated tail_call_cnt */
 888		if (flags & BPF_TRAMP_F_TAIL_CALL_CTX)
 889			bpf_trampoline_restore_tail_call_cnt(image, ctx, func_frame_offset, r4_off);
 890
 891		/* Reserve space to patch branch instruction to skip fexit progs */
 892		im->ip_after_call = &((u32 *)ro_image)[ctx->idx];
 893		EMIT(PPC_RAW_NOP());
 894	}
 895
 896	/* Update branches saved in invoke_bpf_mod_ret with address of do_fexit */
 897	for (i = 0; i < fmod_ret->nr_links && image; i++) {
 898		if (create_cond_branch(&branch_insn, &image[branches[i]],
 899				       (unsigned long)&image[ctx->idx], COND_NE << 16)) {
 900			ret = -EINVAL;
 901			goto cleanup;
 902		}
 903
 904		image[branches[i]] = ppc_inst_val(branch_insn);
 905	}
 906
 907	for (i = 0; i < fexit->nr_links; i++)
 908		if (invoke_bpf_prog(image, ro_image, ctx, fexit->links[i], regs_off, retval_off,
 909				    run_ctx_off, false)) {
 910			ret = -EINVAL;
 911			goto cleanup;
 912		}
 913
 914	if (flags & BPF_TRAMP_F_CALL_ORIG) {
 915		im->ip_epilogue = &((u32 *)ro_image)[ctx->idx];
 916		PPC_LI_ADDR(_R3, im);
 917		ret = bpf_jit_emit_func_call_rel(image, ro_image, ctx,
 918						 (unsigned long)__bpf_tramp_exit);
 919		if (ret)
 920			goto cleanup;
 921	}
 922
 923	if (flags & BPF_TRAMP_F_RESTORE_REGS)
 924		bpf_trampoline_restore_args_regs(image, ctx, nr_regs, regs_off);
 925
 926	/* Restore return value of func_addr or fentry prog */
 927	if (save_ret)
 928		EMIT(PPC_RAW_LL(_R3, _R1, retval_off));
 929
 930	/* Restore nv regs */
 931	EMIT(PPC_RAW_LL(_R26, _R1, nvr_off + SZL));
 932	EMIT(PPC_RAW_LL(_R25, _R1, nvr_off));
 933
 934	/* Epilogue */
 935	if (IS_ENABLED(CONFIG_PPC64_ELF_ABI_V2) && !IS_ENABLED(CONFIG_PPC_KERNEL_PCREL))
 936		EMIT(PPC_RAW_LD(_R2, _R1, 24));
 937	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
 938		/* Skip the traced function and return to parent */
 939		EMIT(PPC_RAW_ADDI(_R1, _R1, func_frame_offset));
 940		EMIT(PPC_RAW_LL(_R0, _R1, PPC_LR_STKOFF));
 941		EMIT(PPC_RAW_MTLR(_R0));
 942		EMIT(PPC_RAW_BLR());
 943	} else {
 944		if (IS_ENABLED(CONFIG_PPC_FTRACE_OUT_OF_LINE)) {
 945			EMIT(PPC_RAW_LL(_R0, _R1, alt_lr_off));
 946			EMIT(PPC_RAW_MTLR(_R0));
 947			EMIT(PPC_RAW_ADDI(_R1, _R1, func_frame_offset));
 948			EMIT(PPC_RAW_LL(_R0, _R1, PPC_LR_STKOFF));
 949			EMIT(PPC_RAW_BLR());
 950		} else {
 951			EMIT(PPC_RAW_LL(_R0, _R1, bpf_frame_size + PPC_LR_STKOFF));
 952			EMIT(PPC_RAW_MTCTR(_R0));
 953			EMIT(PPC_RAW_ADDI(_R1, _R1, func_frame_offset));
 954			EMIT(PPC_RAW_LL(_R0, _R1, PPC_LR_STKOFF));
 955			EMIT(PPC_RAW_MTLR(_R0));
 956			EMIT(PPC_RAW_BCTR());
 957		}
 958	}
 959
 960	/* Make sure the trampoline generation logic doesn't overflow */
 961	if (image && WARN_ON_ONCE(&image[ctx->idx] > (u32 *)rw_image_end - BPF_INSN_SAFETY)) {
 962		ret = -EFAULT;
 963		goto cleanup;
 964	}
 965	ret = ctx->idx * 4 + BPF_INSN_SAFETY * 4;
 966
 967cleanup:
 968	kfree(branches);
 969	return ret;
 970}
 971
 972int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
 973			     struct bpf_tramp_links *tlinks, void *func_addr)
 974{
 975	struct bpf_tramp_image im;
 976	void *image;
 977	int ret;
 978
 979	/*
 980	 * Allocate a temporary buffer for __arch_prepare_bpf_trampoline().
 981	 * This will NOT cause fragmentation in direct map, as we do not
 982	 * call set_memory_*() on this buffer.
 983	 *
 984	 * We cannot use kvmalloc here, because we need image to be in
 985	 * module memory range.
 986	 */
 987	image = bpf_jit_alloc_exec(PAGE_SIZE);
 988	if (!image)
 989		return -ENOMEM;
 990
 991	ret = __arch_prepare_bpf_trampoline(&im, image, image + PAGE_SIZE, image,
 992					    m, flags, tlinks, func_addr);
 993	bpf_jit_free_exec(image);
 994
 995	return ret;
 996}
 997
 998int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *image_end,
 999				const struct btf_func_model *m, u32 flags,
1000				struct bpf_tramp_links *tlinks,
1001				void *func_addr)
1002{
1003	u32 size = image_end - image;
1004	void *rw_image, *tmp;
1005	int ret;
1006
1007	/*
1008	 * rw_image doesn't need to be in module memory range, so we can
1009	 * use kvmalloc.
1010	 */
1011	rw_image = kvmalloc(size, GFP_KERNEL);
1012	if (!rw_image)
1013		return -ENOMEM;
1014
1015	ret = __arch_prepare_bpf_trampoline(im, rw_image, rw_image + size, image, m,
1016					    flags, tlinks, func_addr);
1017	if (ret < 0)
1018		goto out;
1019
1020	if (bpf_jit_enable > 1)
1021		bpf_jit_dump(1, ret - BPF_INSN_SAFETY * 4, 1, rw_image);
1022
1023	tmp = bpf_arch_text_copy(image, rw_image, size);
1024	if (IS_ERR(tmp))
1025		ret = PTR_ERR(tmp);
1026
1027out:
1028	kvfree(rw_image);
1029	return ret;
1030}
1031
1032static int bpf_modify_inst(void *ip, ppc_inst_t old_inst, ppc_inst_t new_inst)
1033{
1034	ppc_inst_t org_inst;
1035
1036	if (copy_inst_from_kernel_nofault(&org_inst, ip)) {
1037		pr_err("0x%lx: fetching instruction failed\n", (unsigned long)ip);
1038		return -EFAULT;
1039	}
1040
1041	if (!ppc_inst_equal(org_inst, old_inst)) {
1042		pr_err("0x%lx: expected (%08lx) != found (%08lx)\n",
1043		       (unsigned long)ip, ppc_inst_as_ulong(old_inst), ppc_inst_as_ulong(org_inst));
1044		return -EINVAL;
1045	}
1046
1047	if (ppc_inst_equal(old_inst, new_inst))
1048		return 0;
1049
1050	return patch_instruction(ip, new_inst);
1051}
1052
1053static void do_isync(void *info __maybe_unused)
1054{
1055	isync();
1056}
1057
1058/*
1059 * A 3-step process for bpf prog entry:
1060 * 1. At bpf prog entry, a single nop/b:
1061 * bpf_func:
1062 *	[nop|b]	ool_stub
1063 * 2. Out-of-line stub:
1064 * ool_stub:
1065 *	mflr	r0
1066 *	[b|bl]	<bpf_prog>/<long_branch_stub>
1067 *	mtlr	r0 // CONFIG_PPC_FTRACE_OUT_OF_LINE only
1068 *	b	bpf_func + 4
1069 * 3. Long branch stub:
1070 * long_branch_stub:
1071 *	.long	<branch_addr>/<dummy_tramp>
1072 *	mflr	r11
1073 *	bcl	20,31,$+4
1074 *	mflr	r12
1075 *	ld	r12, -16(r12)
1076 *	mtctr	r12
1077 *	mtlr	r11 // needed to retain ftrace ABI
1078 *	bctr
1079 *
1080 * dummy_tramp is used to reduce synchronization requirements.
1081 *
1082 * When attaching a bpf trampoline to a bpf prog, we do not need any
1083 * synchronization here since we always have a valid branch target regardless
1084 * of the order in which the above stores are seen. dummy_tramp ensures that
1085 * the long_branch stub goes to a valid destination on other cpus, even when
1086 * the branch to the long_branch stub is seen before the updated trampoline
1087 * address.
1088 *
1089 * However, when detaching a bpf trampoline from a bpf prog, or if changing
1090 * the bpf trampoline address, we need synchronization to ensure that other
1091 * cpus can no longer branch into the older trampoline so that it can be
1092 * safely freed. bpf_tramp_image_put() uses rcu_tasks to ensure all cpus
1093 * make forward progress, but we still need to ensure that other cpus
1094 * execute isync (or some CSI) so that they don't go back into the
1095 * trampoline again.
1096 */
1097int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type poke_type,
1098		       void *old_addr, void *new_addr)
1099{
1100	unsigned long bpf_func, bpf_func_end, size, offset;
1101	ppc_inst_t old_inst, new_inst;
1102	int ret = 0, branch_flags;
1103	char name[KSYM_NAME_LEN];
1104
1105	if (IS_ENABLED(CONFIG_PPC32))
1106		return -EOPNOTSUPP;
1107
1108	bpf_func = (unsigned long)ip;
1109	branch_flags = poke_type == BPF_MOD_CALL ? BRANCH_SET_LINK : 0;
1110
1111	/* We currently only support poking bpf programs */
1112	if (!__bpf_address_lookup(bpf_func, &size, &offset, name)) {
1113		pr_err("%s (0x%lx): kernel/modules are not supported\n", __func__, bpf_func);
1114		return -EOPNOTSUPP;
1115	}
1116
1117	/*
1118	 * If we are not poking at bpf prog entry, then we are simply patching in/out
1119	 * an unconditional branch instruction at im->ip_after_call
1120	 */
1121	if (offset) {
1122		if (poke_type != BPF_MOD_JUMP) {
1123			pr_err("%s (0x%lx): calls are not supported in bpf prog body\n", __func__,
1124			       bpf_func);
1125			return -EOPNOTSUPP;
1126		}
1127		old_inst = ppc_inst(PPC_RAW_NOP());
1128		if (old_addr)
1129			if (create_branch(&old_inst, ip, (unsigned long)old_addr, 0))
1130				return -ERANGE;
1131		new_inst = ppc_inst(PPC_RAW_NOP());
1132		if (new_addr)
1133			if (create_branch(&new_inst, ip, (unsigned long)new_addr, 0))
1134				return -ERANGE;
1135		mutex_lock(&text_mutex);
1136		ret = bpf_modify_inst(ip, old_inst, new_inst);
1137		mutex_unlock(&text_mutex);
1138
1139		/* Make sure all cpus see the new instruction */
1140		smp_call_function(do_isync, NULL, 1);
1141		return ret;
1142	}
1143
1144	bpf_func_end = bpf_func + size;
1145
1146	/* Address of the jmp/call instruction in the out-of-line stub */
1147	ip = (void *)(bpf_func_end - bpf_jit_ool_stub + 4);
1148
1149	if (!is_offset_in_branch_range((long)ip - 4 - bpf_func)) {
1150		pr_err("%s (0x%lx): bpf prog too large, ool stub out of branch range\n", __func__,
1151		       bpf_func);
1152		return -ERANGE;
1153	}
1154
1155	old_inst = ppc_inst(PPC_RAW_NOP());
1156	if (old_addr) {
1157		if (is_offset_in_branch_range(ip - old_addr))
1158			create_branch(&old_inst, ip, (unsigned long)old_addr, branch_flags);
1159		else
1160			create_branch(&old_inst, ip, bpf_func_end - bpf_jit_long_branch_stub,
1161				      branch_flags);
1162	}
1163	new_inst = ppc_inst(PPC_RAW_NOP());
1164	if (new_addr) {
1165		if (is_offset_in_branch_range(ip - new_addr))
1166			create_branch(&new_inst, ip, (unsigned long)new_addr, branch_flags);
1167		else
1168			create_branch(&new_inst, ip, bpf_func_end - bpf_jit_long_branch_stub,
1169				      branch_flags);
1170	}
1171
1172	mutex_lock(&text_mutex);
1173
1174	/*
1175	 * 1. Update the address in the long branch stub:
1176	 * If new_addr is out of range, we will have to use the long branch stub, so patch new_addr
1177	 * here. Otherwise, revert to dummy_tramp, but only if we had patched old_addr here.
1178	 */
1179	if ((new_addr && !is_offset_in_branch_range(new_addr - ip)) ||
1180	    (old_addr && !is_offset_in_branch_range(old_addr - ip)))
1181		ret = patch_ulong((void *)(bpf_func_end - bpf_jit_long_branch_stub - SZL),
1182				  (new_addr && !is_offset_in_branch_range(new_addr - ip)) ?
1183				  (unsigned long)new_addr : (unsigned long)dummy_tramp);
1184	if (ret)
1185		goto out;
1186
1187	/* 2. Update the branch/call in the out-of-line stub */
1188	ret = bpf_modify_inst(ip, old_inst, new_inst);
1189	if (ret)
1190		goto out;
1191
1192	/* 3. Update instruction at bpf prog entry */
1193	ip = (void *)bpf_func;
1194	if (!old_addr || !new_addr) {
1195		if (!old_addr) {
1196			old_inst = ppc_inst(PPC_RAW_NOP());
1197			create_branch(&new_inst, ip, bpf_func_end - bpf_jit_ool_stub, 0);
1198		} else {
1199			new_inst = ppc_inst(PPC_RAW_NOP());
1200			create_branch(&old_inst, ip, bpf_func_end - bpf_jit_ool_stub, 0);
1201		}
1202		ret = bpf_modify_inst(ip, old_inst, new_inst);
1203	}
1204
1205out:
1206	mutex_unlock(&text_mutex);
1207
1208	/*
1209	 * Sync only if we are not attaching a trampoline to a bpf prog so the older
1210	 * trampoline can be freed safely.
1211	 */
1212	if (old_addr)
1213		smp_call_function(do_isync, NULL, 1);
1214
1215	return ret;
1216}