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 * A combined memcpy/__copy_user
187 * __copy_user sets len to 0 for success; else to an upper bound of
188 * the number of uncopied bytes.
189 * memcpy sets v0 to dst.
190 */
191	.align	5
192LEAF(memcpy)					/* a0=dst a1=src a2=len */
193	move	v0, dst				/* return value */
194.L__memcpy:
195FEXPORT(__copy_user)
196	/*
197	 * Note: dst & src may be unaligned, len may be 0
198	 * Temps
199	 */
200#define rem t8
201
202	R10KCBARRIER(0(ra))
203	/*
204	 * The "issue break"s below are very approximate.
205	 * Issue delays for dcache fills will perturb the schedule, as will
206	 * load queue full replay traps, etc.
207	 *
208	 * If len < NBYTES use byte operations.
209	 */
210	PREF(	0, 0(src) )
211	PREF(	1, 0(dst) )
212	sltu	t2, len, NBYTES
213	and	t1, dst, ADDRMASK
214	PREF(	0, 1*32(src) )
215	PREF(	1, 1*32(dst) )
216	bnez	t2, .Lcopy_bytes_checklen
217	 and	t0, src, ADDRMASK
218	PREF(	0, 2*32(src) )
219	PREF(	1, 2*32(dst) )
220	bnez	t1, .Ldst_unaligned
221	 nop
222	bnez	t0, .Lsrc_unaligned_dst_aligned
223	/*
224	 * use delay slot for fall-through
225	 * src and dst are aligned; need to compute rem
226	 */
227.Lboth_aligned:
228	 SRL	t0, len, LOG_NBYTES+3    # +3 for 8 units/iter
229	beqz	t0, .Lcleanup_both_aligned # len < 8*NBYTES
230	 and	rem, len, (8*NBYTES-1)	 # rem = len % (8*NBYTES)
231	PREF(	0, 3*32(src) )
232	PREF(	1, 3*32(dst) )
233	.align	4
2341:
235	R10KCBARRIER(0(ra))
236EXC(	LOAD	t0, UNIT(0)(src),	.Ll_exc)
237EXC(	LOAD	t1, UNIT(1)(src),	.Ll_exc_copy)
238EXC(	LOAD	t2, UNIT(2)(src),	.Ll_exc_copy)
239EXC(	LOAD	t3, UNIT(3)(src),	.Ll_exc_copy)
240	SUB	len, len, 8*NBYTES
241EXC(	LOAD	t4, UNIT(4)(src),	.Ll_exc_copy)
242EXC(	LOAD	t7, UNIT(5)(src),	.Ll_exc_copy)
243EXC(	STORE	t0, UNIT(0)(dst),	.Ls_exc_p8u)
244EXC(	STORE	t1, UNIT(1)(dst),	.Ls_exc_p7u)
245EXC(	LOAD	t0, UNIT(6)(src),	.Ll_exc_copy)
246EXC(	LOAD	t1, UNIT(7)(src),	.Ll_exc_copy)
247	ADD	src, src, 8*NBYTES
248	ADD	dst, dst, 8*NBYTES
249EXC(	STORE	t2, UNIT(-6)(dst),	.Ls_exc_p6u)
250EXC(	STORE	t3, UNIT(-5)(dst),	.Ls_exc_p5u)
251EXC(	STORE	t4, UNIT(-4)(dst),	.Ls_exc_p4u)
252EXC(	STORE	t7, UNIT(-3)(dst),	.Ls_exc_p3u)
253EXC(	STORE	t0, UNIT(-2)(dst),	.Ls_exc_p2u)
254EXC(	STORE	t1, UNIT(-1)(dst),	.Ls_exc_p1u)
255	PREF(	0, 8*32(src) )
256	PREF(	1, 8*32(dst) )
257	bne	len, rem, 1b
258	 nop
259
260	/*
261	 * len == rem == the number of bytes left to copy < 8*NBYTES
262	 */
263.Lcleanup_both_aligned:
264	beqz	len, .Ldone
265	 sltu	t0, len, 4*NBYTES
266	bnez	t0, .Lless_than_4units
267	 and	rem, len, (NBYTES-1)	# rem = len % NBYTES
268	/*
269	 * len >= 4*NBYTES
270	 */
271EXC(	LOAD	t0, UNIT(0)(src),	.Ll_exc)
272EXC(	LOAD	t1, UNIT(1)(src),	.Ll_exc_copy)
273EXC(	LOAD	t2, UNIT(2)(src),	.Ll_exc_copy)
274EXC(	LOAD	t3, UNIT(3)(src),	.Ll_exc_copy)
275	SUB	len, len, 4*NBYTES
276	ADD	src, src, 4*NBYTES
277	R10KCBARRIER(0(ra))
278EXC(	STORE	t0, UNIT(0)(dst),	.Ls_exc_p4u)
279EXC(	STORE	t1, UNIT(1)(dst),	.Ls_exc_p3u)
280EXC(	STORE	t2, UNIT(2)(dst),	.Ls_exc_p2u)
281EXC(	STORE	t3, UNIT(3)(dst),	.Ls_exc_p1u)
282	.set	reorder				/* DADDI_WAR */
283	ADD	dst, dst, 4*NBYTES
284	beqz	len, .Ldone
285	.set	noreorder
286.Lless_than_4units:
287	/*
288	 * rem = len % NBYTES
289	 */
290	beq	rem, len, .Lcopy_bytes
291	 nop
2921:
293	R10KCBARRIER(0(ra))
294EXC(	LOAD	t0, 0(src),		.Ll_exc)
295	ADD	src, src, NBYTES
296	SUB	len, len, NBYTES
297EXC(	STORE	t0, 0(dst),		.Ls_exc_p1u)
298	.set	reorder				/* DADDI_WAR */
299	ADD	dst, dst, NBYTES
300	bne	rem, len, 1b
301	.set	noreorder
302
303	/*
304	 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
305	 * A loop would do only a byte at a time with possible branch
306	 * mispredicts.  Can't do an explicit LOAD dst,mask,or,STORE
307	 * because can't assume read-access to dst.  Instead, use
308	 * STREST dst, which doesn't require read access to dst.
309	 *
310	 * This code should perform better than a simple loop on modern,
311	 * wide-issue mips processors because the code has fewer branches and
312	 * more instruction-level parallelism.
313	 */
314#define bits t2
315	beqz	len, .Ldone
316	 ADD	t1, dst, len	# t1 is just past last byte of dst
317	li	bits, 8*NBYTES
318	SLL	rem, len, 3	# rem = number of bits to keep
319EXC(	LOAD	t0, 0(src),		.Ll_exc)
320	SUB	bits, bits, rem	# bits = number of bits to discard
321	SHIFT_DISCARD t0, t0, bits
322EXC(	STREST	t0, -1(t1),		.Ls_exc)
323	jr	ra
324	 move	len, zero
325.Ldst_unaligned:
326	/*
327	 * dst is unaligned
328	 * t0 = src & ADDRMASK
329	 * t1 = dst & ADDRMASK; T1 > 0
330	 * len >= NBYTES
331	 *
332	 * Copy enough bytes to align dst
333	 * Set match = (src and dst have same alignment)
334	 */
335#define match rem
336EXC(	LDFIRST	t3, FIRST(0)(src),	.Ll_exc)
337	ADD	t2, zero, NBYTES
338EXC(	LDREST	t3, REST(0)(src),	.Ll_exc_copy)
339	SUB	t2, t2, t1	# t2 = number of bytes copied
340	xor	match, t0, t1
341	R10KCBARRIER(0(ra))
342EXC(	STFIRST t3, FIRST(0)(dst),	.Ls_exc)
343	beq	len, t2, .Ldone
344	 SUB	len, len, t2
345	ADD	dst, dst, t2
346	beqz	match, .Lboth_aligned
347	 ADD	src, src, t2
348
349.Lsrc_unaligned_dst_aligned:
350	SRL	t0, len, LOG_NBYTES+2    # +2 for 4 units/iter
351	PREF(	0, 3*32(src) )
352	beqz	t0, .Lcleanup_src_unaligned
353	 and	rem, len, (4*NBYTES-1)   # rem = len % 4*NBYTES
354	PREF(	1, 3*32(dst) )
3551:
356/*
357 * Avoid consecutive LD*'s to the same register since some mips
358 * implementations can't issue them in the same cycle.
359 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
360 * are to the same unit (unless src is aligned, but it's not).
361 */
362	R10KCBARRIER(0(ra))
363EXC(	LDFIRST	t0, FIRST(0)(src),	.Ll_exc)
364EXC(	LDFIRST	t1, FIRST(1)(src),	.Ll_exc_copy)
365	SUB     len, len, 4*NBYTES
366EXC(	LDREST	t0, REST(0)(src),	.Ll_exc_copy)
367EXC(	LDREST	t1, REST(1)(src),	.Ll_exc_copy)
368EXC(	LDFIRST	t2, FIRST(2)(src),	.Ll_exc_copy)
369EXC(	LDFIRST	t3, FIRST(3)(src),	.Ll_exc_copy)
370EXC(	LDREST	t2, REST(2)(src),	.Ll_exc_copy)
371EXC(	LDREST	t3, REST(3)(src),	.Ll_exc_copy)
372	PREF(	0, 9*32(src) )		# 0 is PREF_LOAD  (not streamed)
373	ADD	src, src, 4*NBYTES
374#ifdef CONFIG_CPU_SB1
375	nop				# improves slotting
376#endif
377EXC(	STORE	t0, UNIT(0)(dst),	.Ls_exc_p4u)
378EXC(	STORE	t1, UNIT(1)(dst),	.Ls_exc_p3u)
379EXC(	STORE	t2, UNIT(2)(dst),	.Ls_exc_p2u)
380EXC(	STORE	t3, UNIT(3)(dst),	.Ls_exc_p1u)
381	PREF(	1, 9*32(dst) )     	# 1 is PREF_STORE (not streamed)
382	.set	reorder				/* DADDI_WAR */
383	ADD	dst, dst, 4*NBYTES
384	bne	len, rem, 1b
385	.set	noreorder
386
387.Lcleanup_src_unaligned:
388	beqz	len, .Ldone
389	 and	rem, len, NBYTES-1  # rem = len % NBYTES
390	beq	rem, len, .Lcopy_bytes
391	 nop
3921:
393	R10KCBARRIER(0(ra))
394EXC(	LDFIRST t0, FIRST(0)(src),	.Ll_exc)
395EXC(	LDREST	t0, REST(0)(src),	.Ll_exc_copy)
396	ADD	src, src, NBYTES
397	SUB	len, len, NBYTES
398EXC(	STORE	t0, 0(dst),		.Ls_exc_p1u)
399	.set	reorder				/* DADDI_WAR */
400	ADD	dst, dst, NBYTES
401	bne	len, rem, 1b
402	.set	noreorder
403
404.Lcopy_bytes_checklen:
405	beqz	len, .Ldone
406	 nop
407.Lcopy_bytes:
408	/* 0 < len < NBYTES  */
409	R10KCBARRIER(0(ra))
410#define COPY_BYTE(N)			\
411EXC(	lb	t0, N(src), .Ll_exc);	\
412	SUB	len, len, 1;		\
413	beqz	len, .Ldone;		\
414EXC(	 sb	t0, N(dst), .Ls_exc_p1)
415
416	COPY_BYTE(0)
417	COPY_BYTE(1)
418#ifdef USE_DOUBLE
419	COPY_BYTE(2)
420	COPY_BYTE(3)
421	COPY_BYTE(4)
422	COPY_BYTE(5)
423#endif
424EXC(	lb	t0, NBYTES-2(src), .Ll_exc)
425	SUB	len, len, 1
426	jr	ra
427EXC(	 sb	t0, NBYTES-2(dst), .Ls_exc_p1)
428.Ldone:
429	jr	ra
430	 nop
431	END(memcpy)
432
433.Ll_exc_copy:
434	/*
435	 * Copy bytes from src until faulting load address (or until a
436	 * lb faults)
437	 *
438	 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
439	 * may be more than a byte beyond the last address.
440	 * Hence, the lb below may get an exception.
441	 *
442	 * Assumes src < THREAD_BUADDR($28)
443	 */
444	LOAD	t0, TI_TASK($28)
445	 nop
446	LOAD	t0, THREAD_BUADDR(t0)
4471:
448EXC(	lb	t1, 0(src),	.Ll_exc)
449	ADD	src, src, 1
450	sb	t1, 0(dst)	# can't fault -- we're copy_from_user
451	.set	reorder				/* DADDI_WAR */
452	ADD	dst, dst, 1
453	bne	src, t0, 1b
454	.set	noreorder
455.Ll_exc:
456	LOAD	t0, TI_TASK($28)
457	 nop
458	LOAD	t0, THREAD_BUADDR(t0)	# t0 is just past last good address
459	 nop
460	SUB	len, AT, t0		# len number of uncopied bytes
461	/*
462	 * Here's where we rely on src and dst being incremented in tandem,
463	 *   See (3) above.
464	 * dst += (fault addr - src) to put dst at first byte to clear
465	 */
466	ADD	dst, t0			# compute start address in a1
467	SUB	dst, src
468	/*
469	 * Clear len bytes starting at dst.  Can't call __bzero because it
470	 * might modify len.  An inefficient loop for these rare times...
471	 */
472	.set	reorder				/* DADDI_WAR */
473	SUB	src, len, 1
474	beqz	len, .Ldone
475	.set	noreorder
4761:	sb	zero, 0(dst)
477	ADD	dst, dst, 1
478#ifndef CONFIG_CPU_DADDI_WORKAROUNDS
479	bnez	src, 1b
480	 SUB	src, src, 1
481#else
482	.set	push
483	.set	noat
484	li	v1, 1
485	bnez	src, 1b
486	 SUB	src, src, v1
487	.set	pop
488#endif
489	jr	ra
490	 nop
491
492
493#define SEXC(n)							\
494	.set	reorder;			/* DADDI_WAR */	\
495.Ls_exc_p ## n ## u:						\
496	ADD	len, len, n*NBYTES;				\
497	jr	ra;						\
498	.set	noreorder
499
500SEXC(8)
501SEXC(7)
502SEXC(6)
503SEXC(5)
504SEXC(4)
505SEXC(3)
506SEXC(2)
507SEXC(1)
508
509.Ls_exc_p1:
510	.set	reorder				/* DADDI_WAR */
511	ADD	len, len, 1
512	jr	ra
513	.set	noreorder
514.Ls_exc:
515	jr	ra
516	 nop
517
518	.align	5
519LEAF(memmove)
520	ADD	t0, a0, a2
521	ADD	t1, a1, a2
522	sltu	t0, a1, t0			# dst + len <= src -> memcpy
523	sltu	t1, a0, t1			# dst >= src + len -> memcpy
524	and	t0, t1
525	beqz	t0, .L__memcpy
526	 move	v0, a0				/* return value */
527	beqz	a2, .Lr_out
528	END(memmove)
529
530	/* fall through to __rmemcpy */
531LEAF(__rmemcpy)					/* a0=dst a1=src a2=len */
532	 sltu	t0, a1, a0
533	beqz	t0, .Lr_end_bytes_up		# src >= dst
534	 nop
535	ADD	a0, a2				# dst = dst + len
536	ADD	a1, a2				# src = src + len
537
538.Lr_end_bytes:
539	R10KCBARRIER(0(ra))
540	lb	t0, -1(a1)
541	SUB	a2, a2, 0x1
542	sb	t0, -1(a0)
543	SUB	a1, a1, 0x1
544	.set	reorder				/* DADDI_WAR */
545	SUB	a0, a0, 0x1
546	bnez	a2, .Lr_end_bytes
547	.set	noreorder
548
549.Lr_out:
550	jr	ra
551	 move	a2, zero
552
553.Lr_end_bytes_up:
554	R10KCBARRIER(0(ra))
555	lb	t0, (a1)
556	SUB	a2, a2, 0x1
557	sb	t0, (a0)
558	ADD	a1, a1, 0x1
559	.set	reorder				/* DADDI_WAR */
560	ADD	a0, a0, 0x1
561	bnez	a2, .Lr_end_bytes_up
562	.set	noreorder
563
564	jr	ra
565	 move	a2, zero
566	END(__rmemcpy)
567