xref: /openbmc/linux/arch/mips/lib/memcpy.S (revision b6bec26c)
1/*
2 * This file is subject to the terms and conditions of the GNU General Public
3 * License.  See the file "COPYING" in the main directory of this archive
4 * for more details.
5 *
6 * Unified implementation of memcpy, memmove and the __copy_user backend.
7 *
8 * Copyright (C) 1998, 99, 2000, 01, 2002 Ralf Baechle (ralf@gnu.org)
9 * Copyright (C) 1999, 2000, 01, 2002 Silicon Graphics, Inc.
10 * Copyright (C) 2002 Broadcom, Inc.
11 *   memcpy/copy_user author: Mark Vandevoorde
12 * Copyright (C) 2007  Maciej W. Rozycki
13 *
14 * Mnemonic names for arguments to memcpy/__copy_user
15 */
16
17/*
18 * Hack to resolve longstanding prefetch issue
19 *
20 * Prefetching may be fatal on some systems if we're prefetching beyond the
21 * end of memory on some systems.  It's also a seriously bad idea on non
22 * dma-coherent systems.
23 */
24#ifdef CONFIG_DMA_NONCOHERENT
25#undef CONFIG_CPU_HAS_PREFETCH
26#endif
27#ifdef CONFIG_MIPS_MALTA
28#undef CONFIG_CPU_HAS_PREFETCH
29#endif
30
31#include <asm/asm.h>
32#include <asm/asm-offsets.h>
33#include <asm/regdef.h>
34
35#define dst a0
36#define src a1
37#define len a2
38
39/*
40 * Spec
41 *
42 * memcpy copies len bytes from src to dst and sets v0 to dst.
43 * It assumes that
44 *   - src and dst don't overlap
45 *   - src is readable
46 *   - dst is writable
47 * memcpy uses the standard calling convention
48 *
49 * __copy_user copies up to len bytes from src to dst and sets a2 (len) to
50 * the number of uncopied bytes due to an exception caused by a read or write.
51 * __copy_user assumes that src and dst don't overlap, and that the call is
52 * implementing one of the following:
53 *   copy_to_user
54 *     - src is readable  (no exceptions when reading src)
55 *   copy_from_user
56 *     - dst is writable  (no exceptions when writing dst)
57 * __copy_user uses a non-standard calling convention; see
58 * include/asm-mips/uaccess.h
59 *
60 * When an exception happens on a load, the handler must
61 # ensure that all of the destination buffer is overwritten to prevent
62 * leaking information to user mode programs.
63 */
64
65/*
66 * Implementation
67 */
68
69/*
70 * The exception handler for loads requires that:
71 *  1- AT contain the address of the byte just past the end of the source
72 *     of the copy,
73 *  2- src_entry <= src < AT, and
74 *  3- (dst - src) == (dst_entry - src_entry),
75 * The _entry suffix denotes values when __copy_user was called.
76 *
77 * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user
78 * (2) is met by incrementing src by the number of bytes copied
79 * (3) is met by not doing loads between a pair of increments of dst and src
80 *
81 * The exception handlers for stores adjust len (if necessary) and return.
82 * These handlers do not need to overwrite any data.
83 *
84 * For __rmemcpy and memmove an exception is always a kernel bug, therefore
85 * they're not protected.
86 */
87
88#define EXC(inst_reg,addr,handler)		\
899:	inst_reg, addr;				\
90	.section __ex_table,"a";		\
91	PTR	9b, handler;			\
92	.previous
93
94/*
95 * Only on the 64-bit kernel we can made use of 64-bit registers.
96 */
97#ifdef CONFIG_64BIT
98#define USE_DOUBLE
99#endif
100
101#ifdef USE_DOUBLE
102
103#define LOAD   ld
104#define LOADL  ldl
105#define LOADR  ldr
106#define STOREL sdl
107#define STORER sdr
108#define STORE  sd
109#define ADD    daddu
110#define SUB    dsubu
111#define SRL    dsrl
112#define SRA    dsra
113#define SLL    dsll
114#define SLLV   dsllv
115#define SRLV   dsrlv
116#define NBYTES 8
117#define LOG_NBYTES 3
118
119/*
120 * As we are sharing code base with the mips32 tree (which use the o32 ABI
121 * register definitions). We need to redefine the register definitions from
122 * the n64 ABI register naming to the o32 ABI register naming.
123 */
124#undef t0
125#undef t1
126#undef t2
127#undef t3
128#define t0	$8
129#define t1	$9
130#define t2	$10
131#define t3	$11
132#define t4	$12
133#define t5	$13
134#define t6	$14
135#define t7	$15
136
137#else
138
139#define LOAD   lw
140#define LOADL  lwl
141#define LOADR  lwr
142#define STOREL swl
143#define STORER swr
144#define STORE  sw
145#define ADD    addu
146#define SUB    subu
147#define SRL    srl
148#define SLL    sll
149#define SRA    sra
150#define SLLV   sllv
151#define SRLV   srlv
152#define NBYTES 4
153#define LOG_NBYTES 2
154
155#endif /* USE_DOUBLE */
156
157#ifdef CONFIG_CPU_LITTLE_ENDIAN
158#define LDFIRST LOADR
159#define LDREST  LOADL
160#define STFIRST STORER
161#define STREST  STOREL
162#define SHIFT_DISCARD SLLV
163#else
164#define LDFIRST LOADL
165#define LDREST  LOADR
166#define STFIRST STOREL
167#define STREST  STORER
168#define SHIFT_DISCARD SRLV
169#endif
170
171#define FIRST(unit) ((unit)*NBYTES)
172#define REST(unit)  (FIRST(unit)+NBYTES-1)
173#define UNIT(unit)  FIRST(unit)
174
175#define ADDRMASK (NBYTES-1)
176
177	.text
178	.set	noreorder
179#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
180	.set	noat
181#else
182	.set	at=v1
183#endif
184
185/*
186 * t6 is used as a flag to note inatomic mode.
187 */
188LEAF(__copy_user_inatomic)
189	b	__copy_user_common
190	 li	t6, 1
191	END(__copy_user_inatomic)
192
193/*
194 * A combined memcpy/__copy_user
195 * __copy_user sets len to 0 for success; else to an upper bound of
196 * the number of uncopied bytes.
197 * memcpy sets v0 to dst.
198 */
199	.align	5
200LEAF(memcpy)					/* a0=dst a1=src a2=len */
201	move	v0, dst				/* return value */
202.L__memcpy:
203FEXPORT(__copy_user)
204	li	t6, 0	/* not inatomic */
205__copy_user_common:
206	/*
207	 * Note: dst & src may be unaligned, len may be 0
208	 * Temps
209	 */
210#define rem t8
211
212	R10KCBARRIER(0(ra))
213	/*
214	 * The "issue break"s below are very approximate.
215	 * Issue delays for dcache fills will perturb the schedule, as will
216	 * load queue full replay traps, etc.
217	 *
218	 * If len < NBYTES use byte operations.
219	 */
220	PREF(	0, 0(src) )
221	PREF(	1, 0(dst) )
222	sltu	t2, len, NBYTES
223	and	t1, dst, ADDRMASK
224	PREF(	0, 1*32(src) )
225	PREF(	1, 1*32(dst) )
226	bnez	t2, .Lcopy_bytes_checklen
227	 and	t0, src, ADDRMASK
228	PREF(	0, 2*32(src) )
229	PREF(	1, 2*32(dst) )
230	bnez	t1, .Ldst_unaligned
231	 nop
232	bnez	t0, .Lsrc_unaligned_dst_aligned
233	/*
234	 * use delay slot for fall-through
235	 * src and dst are aligned; need to compute rem
236	 */
237.Lboth_aligned:
238	 SRL	t0, len, LOG_NBYTES+3    # +3 for 8 units/iter
239	beqz	t0, .Lcleanup_both_aligned # len < 8*NBYTES
240	 and	rem, len, (8*NBYTES-1)	 # rem = len % (8*NBYTES)
241	PREF(	0, 3*32(src) )
242	PREF(	1, 3*32(dst) )
243	.align	4
2441:
245	R10KCBARRIER(0(ra))
246EXC(	LOAD	t0, UNIT(0)(src),	.Ll_exc)
247EXC(	LOAD	t1, UNIT(1)(src),	.Ll_exc_copy)
248EXC(	LOAD	t2, UNIT(2)(src),	.Ll_exc_copy)
249EXC(	LOAD	t3, UNIT(3)(src),	.Ll_exc_copy)
250	SUB	len, len, 8*NBYTES
251EXC(	LOAD	t4, UNIT(4)(src),	.Ll_exc_copy)
252EXC(	LOAD	t7, UNIT(5)(src),	.Ll_exc_copy)
253EXC(	STORE	t0, UNIT(0)(dst),	.Ls_exc_p8u)
254EXC(	STORE	t1, UNIT(1)(dst),	.Ls_exc_p7u)
255EXC(	LOAD	t0, UNIT(6)(src),	.Ll_exc_copy)
256EXC(	LOAD	t1, UNIT(7)(src),	.Ll_exc_copy)
257	ADD	src, src, 8*NBYTES
258	ADD	dst, dst, 8*NBYTES
259EXC(	STORE	t2, UNIT(-6)(dst),	.Ls_exc_p6u)
260EXC(	STORE	t3, UNIT(-5)(dst),	.Ls_exc_p5u)
261EXC(	STORE	t4, UNIT(-4)(dst),	.Ls_exc_p4u)
262EXC(	STORE	t7, UNIT(-3)(dst),	.Ls_exc_p3u)
263EXC(	STORE	t0, UNIT(-2)(dst),	.Ls_exc_p2u)
264EXC(	STORE	t1, UNIT(-1)(dst),	.Ls_exc_p1u)
265	PREF(	0, 8*32(src) )
266	PREF(	1, 8*32(dst) )
267	bne	len, rem, 1b
268	 nop
269
270	/*
271	 * len == rem == the number of bytes left to copy < 8*NBYTES
272	 */
273.Lcleanup_both_aligned:
274	beqz	len, .Ldone
275	 sltu	t0, len, 4*NBYTES
276	bnez	t0, .Lless_than_4units
277	 and	rem, len, (NBYTES-1)	# rem = len % NBYTES
278	/*
279	 * len >= 4*NBYTES
280	 */
281EXC(	LOAD	t0, UNIT(0)(src),	.Ll_exc)
282EXC(	LOAD	t1, UNIT(1)(src),	.Ll_exc_copy)
283EXC(	LOAD	t2, UNIT(2)(src),	.Ll_exc_copy)
284EXC(	LOAD	t3, UNIT(3)(src),	.Ll_exc_copy)
285	SUB	len, len, 4*NBYTES
286	ADD	src, src, 4*NBYTES
287	R10KCBARRIER(0(ra))
288EXC(	STORE	t0, UNIT(0)(dst),	.Ls_exc_p4u)
289EXC(	STORE	t1, UNIT(1)(dst),	.Ls_exc_p3u)
290EXC(	STORE	t2, UNIT(2)(dst),	.Ls_exc_p2u)
291EXC(	STORE	t3, UNIT(3)(dst),	.Ls_exc_p1u)
292	.set	reorder				/* DADDI_WAR */
293	ADD	dst, dst, 4*NBYTES
294	beqz	len, .Ldone
295	.set	noreorder
296.Lless_than_4units:
297	/*
298	 * rem = len % NBYTES
299	 */
300	beq	rem, len, .Lcopy_bytes
301	 nop
3021:
303	R10KCBARRIER(0(ra))
304EXC(	LOAD	t0, 0(src),		.Ll_exc)
305	ADD	src, src, NBYTES
306	SUB	len, len, NBYTES
307EXC(	STORE	t0, 0(dst),		.Ls_exc_p1u)
308	.set	reorder				/* DADDI_WAR */
309	ADD	dst, dst, NBYTES
310	bne	rem, len, 1b
311	.set	noreorder
312
313	/*
314	 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
315	 * A loop would do only a byte at a time with possible branch
316	 * mispredicts.  Can't do an explicit LOAD dst,mask,or,STORE
317	 * because can't assume read-access to dst.  Instead, use
318	 * STREST dst, which doesn't require read access to dst.
319	 *
320	 * This code should perform better than a simple loop on modern,
321	 * wide-issue mips processors because the code has fewer branches and
322	 * more instruction-level parallelism.
323	 */
324#define bits t2
325	beqz	len, .Ldone
326	 ADD	t1, dst, len	# t1 is just past last byte of dst
327	li	bits, 8*NBYTES
328	SLL	rem, len, 3	# rem = number of bits to keep
329EXC(	LOAD	t0, 0(src),		.Ll_exc)
330	SUB	bits, bits, rem	# bits = number of bits to discard
331	SHIFT_DISCARD t0, t0, bits
332EXC(	STREST	t0, -1(t1),		.Ls_exc)
333	jr	ra
334	 move	len, zero
335.Ldst_unaligned:
336	/*
337	 * dst is unaligned
338	 * t0 = src & ADDRMASK
339	 * t1 = dst & ADDRMASK; T1 > 0
340	 * len >= NBYTES
341	 *
342	 * Copy enough bytes to align dst
343	 * Set match = (src and dst have same alignment)
344	 */
345#define match rem
346EXC(	LDFIRST	t3, FIRST(0)(src),	.Ll_exc)
347	ADD	t2, zero, NBYTES
348EXC(	LDREST	t3, REST(0)(src),	.Ll_exc_copy)
349	SUB	t2, t2, t1	# t2 = number of bytes copied
350	xor	match, t0, t1
351	R10KCBARRIER(0(ra))
352EXC(	STFIRST t3, FIRST(0)(dst),	.Ls_exc)
353	beq	len, t2, .Ldone
354	 SUB	len, len, t2
355	ADD	dst, dst, t2
356	beqz	match, .Lboth_aligned
357	 ADD	src, src, t2
358
359.Lsrc_unaligned_dst_aligned:
360	SRL	t0, len, LOG_NBYTES+2    # +2 for 4 units/iter
361	PREF(	0, 3*32(src) )
362	beqz	t0, .Lcleanup_src_unaligned
363	 and	rem, len, (4*NBYTES-1)   # rem = len % 4*NBYTES
364	PREF(	1, 3*32(dst) )
3651:
366/*
367 * Avoid consecutive LD*'s to the same register since some mips
368 * implementations can't issue them in the same cycle.
369 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
370 * are to the same unit (unless src is aligned, but it's not).
371 */
372	R10KCBARRIER(0(ra))
373EXC(	LDFIRST	t0, FIRST(0)(src),	.Ll_exc)
374EXC(	LDFIRST	t1, FIRST(1)(src),	.Ll_exc_copy)
375	SUB     len, len, 4*NBYTES
376EXC(	LDREST	t0, REST(0)(src),	.Ll_exc_copy)
377EXC(	LDREST	t1, REST(1)(src),	.Ll_exc_copy)
378EXC(	LDFIRST	t2, FIRST(2)(src),	.Ll_exc_copy)
379EXC(	LDFIRST	t3, FIRST(3)(src),	.Ll_exc_copy)
380EXC(	LDREST	t2, REST(2)(src),	.Ll_exc_copy)
381EXC(	LDREST	t3, REST(3)(src),	.Ll_exc_copy)
382	PREF(	0, 9*32(src) )		# 0 is PREF_LOAD  (not streamed)
383	ADD	src, src, 4*NBYTES
384#ifdef CONFIG_CPU_SB1
385	nop				# improves slotting
386#endif
387EXC(	STORE	t0, UNIT(0)(dst),	.Ls_exc_p4u)
388EXC(	STORE	t1, UNIT(1)(dst),	.Ls_exc_p3u)
389EXC(	STORE	t2, UNIT(2)(dst),	.Ls_exc_p2u)
390EXC(	STORE	t3, UNIT(3)(dst),	.Ls_exc_p1u)
391	PREF(	1, 9*32(dst) )     	# 1 is PREF_STORE (not streamed)
392	.set	reorder				/* DADDI_WAR */
393	ADD	dst, dst, 4*NBYTES
394	bne	len, rem, 1b
395	.set	noreorder
396
397.Lcleanup_src_unaligned:
398	beqz	len, .Ldone
399	 and	rem, len, NBYTES-1  # rem = len % NBYTES
400	beq	rem, len, .Lcopy_bytes
401	 nop
4021:
403	R10KCBARRIER(0(ra))
404EXC(	LDFIRST t0, FIRST(0)(src),	.Ll_exc)
405EXC(	LDREST	t0, REST(0)(src),	.Ll_exc_copy)
406	ADD	src, src, NBYTES
407	SUB	len, len, NBYTES
408EXC(	STORE	t0, 0(dst),		.Ls_exc_p1u)
409	.set	reorder				/* DADDI_WAR */
410	ADD	dst, dst, NBYTES
411	bne	len, rem, 1b
412	.set	noreorder
413
414.Lcopy_bytes_checklen:
415	beqz	len, .Ldone
416	 nop
417.Lcopy_bytes:
418	/* 0 < len < NBYTES  */
419	R10KCBARRIER(0(ra))
420#define COPY_BYTE(N)			\
421EXC(	lb	t0, N(src), .Ll_exc);	\
422	SUB	len, len, 1;		\
423	beqz	len, .Ldone;		\
424EXC(	 sb	t0, N(dst), .Ls_exc_p1)
425
426	COPY_BYTE(0)
427	COPY_BYTE(1)
428#ifdef USE_DOUBLE
429	COPY_BYTE(2)
430	COPY_BYTE(3)
431	COPY_BYTE(4)
432	COPY_BYTE(5)
433#endif
434EXC(	lb	t0, NBYTES-2(src), .Ll_exc)
435	SUB	len, len, 1
436	jr	ra
437EXC(	 sb	t0, NBYTES-2(dst), .Ls_exc_p1)
438.Ldone:
439	jr	ra
440	 nop
441	END(memcpy)
442
443.Ll_exc_copy:
444	/*
445	 * Copy bytes from src until faulting load address (or until a
446	 * lb faults)
447	 *
448	 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
449	 * may be more than a byte beyond the last address.
450	 * Hence, the lb below may get an exception.
451	 *
452	 * Assumes src < THREAD_BUADDR($28)
453	 */
454	LOAD	t0, TI_TASK($28)
455	 nop
456	LOAD	t0, THREAD_BUADDR(t0)
4571:
458EXC(	lb	t1, 0(src),	.Ll_exc)
459	ADD	src, src, 1
460	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
461	.set	reorder				/* DADDI_WAR */
462	ADD	dst, dst, 1
463	bne	src, t0, 1b
464	.set	noreorder
465.Ll_exc:
466	LOAD	t0, TI_TASK($28)
467	 nop
468	LOAD	t0, THREAD_BUADDR(t0)	# t0 is just past last good address
469	 nop
470	SUB	len, AT, t0		# len number of uncopied bytes
471	bnez	t6, .Ldone	/* Skip the zeroing part if inatomic */
472	/*
473	 * Here's where we rely on src and dst being incremented in tandem,
474	 *   See (3) above.
475	 * dst += (fault addr - src) to put dst at first byte to clear
476	 */
477	ADD	dst, t0			# compute start address in a1
478	SUB	dst, src
479	/*
480	 * Clear len bytes starting at dst.  Can't call __bzero because it
481	 * might modify len.  An inefficient loop for these rare times...
482	 */
483	.set	reorder				/* DADDI_WAR */
484	SUB	src, len, 1
485	beqz	len, .Ldone
486	.set	noreorder
4871:	sb	zero, 0(dst)
488	ADD	dst, dst, 1
489#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
490	bnez	src, 1b
491	 SUB	src, src, 1
492#else
493	.set	push
494	.set	noat
495	li	v1, 1
496	bnez	src, 1b
497	 SUB	src, src, v1
498	.set	pop
499#endif
500	jr	ra
501	 nop
502
503
504#define SEXC(n)							\
505	.set	reorder;			/* DADDI_WAR */	\
506.Ls_exc_p ## n ## u:						\
507	ADD	len, len, n*NBYTES;				\
508	jr	ra;						\
509	.set	noreorder
510
511SEXC(8)
512SEXC(7)
513SEXC(6)
514SEXC(5)
515SEXC(4)
516SEXC(3)
517SEXC(2)
518SEXC(1)
519
520.Ls_exc_p1:
521	.set	reorder				/* DADDI_WAR */
522	ADD	len, len, 1
523	jr	ra
524	.set	noreorder
525.Ls_exc:
526	jr	ra
527	 nop
528
529	.align	5
530LEAF(memmove)
531	ADD	t0, a0, a2
532	ADD	t1, a1, a2
533	sltu	t0, a1, t0			# dst + len <= src -> memcpy
534	sltu	t1, a0, t1			# dst >= src + len -> memcpy
535	and	t0, t1
536	beqz	t0, .L__memcpy
537	 move	v0, a0				/* return value */
538	beqz	a2, .Lr_out
539	END(memmove)
540
541	/* fall through to __rmemcpy */
542LEAF(__rmemcpy)					/* a0=dst a1=src a2=len */
543	 sltu	t0, a1, a0
544	beqz	t0, .Lr_end_bytes_up		# src >= dst
545	 nop
546	ADD	a0, a2				# dst = dst + len
547	ADD	a1, a2				# src = src + len
548
549.Lr_end_bytes:
550	R10KCBARRIER(0(ra))
551	lb	t0, -1(a1)
552	SUB	a2, a2, 0x1
553	sb	t0, -1(a0)
554	SUB	a1, a1, 0x1
555	.set	reorder				/* DADDI_WAR */
556	SUB	a0, a0, 0x1
557	bnez	a2, .Lr_end_bytes
558	.set	noreorder
559
560.Lr_out:
561	jr	ra
562	 move	a2, zero
563
564.Lr_end_bytes_up:
565	R10KCBARRIER(0(ra))
566	lb	t0, (a1)
567	SUB	a2, a2, 0x1
568	sb	t0, (a0)
569	ADD	a1, a1, 0x1
570	.set	reorder				/* DADDI_WAR */
571	ADD	a0, a0, 0x1
572	bnez	a2, .Lr_end_bytes_up
573	.set	noreorder
574
575	jr	ra
576	 move	a2, zero
577	END(__rmemcpy)
578