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
6 * Unified implementation of memcpy, memmove and the __copy_user backend.
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 * Copyright (C) 2014 Imagination Technologies Ltd.
15 * Mnemonic names for arguments to memcpy/__copy_user
19 * Hack to resolve longstanding prefetch issue
21 * Prefetching may be fatal on some systems if we're prefetching beyond the
22 * end of memory on some systems. It's also a seriously bad idea on non
23 * dma-coherent systems.
25 #ifdef CONFIG_DMA_NONCOHERENT
26 #undef CONFIG_CPU_HAS_PREFETCH
28 #ifdef CONFIG_MIPS_MALTA
29 #undef CONFIG_CPU_HAS_PREFETCH
33 #include <asm/asm-offsets.h>
34 #include <asm/export.h>
35 #include <asm/regdef.h>
44 * memcpy copies len bytes from src to dst and sets v0 to dst.
46 * - src and dst don't overlap
49 * memcpy uses the standard calling convention
51 * __copy_user copies up to len bytes from src to dst and sets a2 (len) to
52 * the number of uncopied bytes due to an exception caused by a read or write.
53 * __copy_user assumes that src and dst don't overlap, and that the call is
54 * implementing one of the following:
56 * - src is readable (no exceptions when reading src)
58 * - dst is writable (no exceptions when writing dst)
59 * __copy_user uses a non-standard calling convention; see
60 * include/asm-mips/uaccess.h
62 * When an exception happens on a load, the handler must
63 # ensure that all of the destination buffer is overwritten to prevent
64 * leaking information to user mode programs.
72 * The exception handler for loads requires that:
73 * 1- AT contain the address of the byte just past the end of the source
75 * 2- src_entry <= src < AT, and
76 * 3- (dst - src) == (dst_entry - src_entry),
77 * The _entry suffix denotes values when __copy_user was called.
79 * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user
80 * (2) is met by incrementing src by the number of bytes copied
81 * (3) is met by not doing loads between a pair of increments of dst and src
83 * The exception handlers for stores adjust len (if necessary) and return.
84 * These handlers do not need to overwrite any data.
86 * For __rmemcpy and memmove an exception is always a kernel bug, therefore
87 * they're not protected.
90 /* Instruction type */
94 #define SRC_PREFETCH 1
95 #define DST_PREFETCH 2
102 * Wrapper to add an entry in the exception table
103 * in case the insn causes a memory exception.
105 * insn : Load/store instruction
106 * type : Instruction type
109 * handler : Exception handler
112 #define EXC(insn, type, reg, addr, handler) \
113 .if \mode == LEGACY_MODE; \
115 .section __ex_table,"a"; \
118 /* This is assembled in EVA mode */ \
120 /* If loading from user or storing to user */ \
121 .if ((\from == USEROP) && (type == LD_INSN)) || \
122 ((\to == USEROP) && (type == ST_INSN)); \
123 9: __BUILD_EVA_INSN(insn##e, reg, addr); \
124 .section __ex_table,"a"; \
129 * Still in EVA, but no need for \
130 * exception handler or EVA insn \
137 * Only on the 64-bit kernel we can made use of 64-bit registers.
145 #define LOADK ld /* No exception */
146 #define LOAD(reg, addr, handler) EXC(ld, LD_INSN, reg, addr, handler)
147 #define LOADL(reg, addr, handler) EXC(ldl, LD_INSN, reg, addr, handler)
148 #define LOADR(reg, addr, handler) EXC(ldr, LD_INSN, reg, addr, handler)
149 #define STOREL(reg, addr, handler) EXC(sdl, ST_INSN, reg, addr, handler)
150 #define STORER(reg, addr, handler) EXC(sdr, ST_INSN, reg, addr, handler)
151 #define STORE(reg, addr, handler) EXC(sd, ST_INSN, reg, addr, handler)
163 * As we are sharing code base with the mips32 tree (which use the o32 ABI
164 * register definitions). We need to redefine the register definitions from
165 * the n64 ABI register naming to the o32 ABI register naming.
182 #define LOADK lw /* No exception */
183 #define LOAD(reg, addr, handler) EXC(lw, LD_INSN, reg, addr, handler)
184 #define LOADL(reg, addr, handler) EXC(lwl, LD_INSN, reg, addr, handler)
185 #define LOADR(reg, addr, handler) EXC(lwr, LD_INSN, reg, addr, handler)
186 #define STOREL(reg, addr, handler) EXC(swl, ST_INSN, reg, addr, handler)
187 #define STORER(reg, addr, handler) EXC(swr, ST_INSN, reg, addr, handler)
188 #define STORE(reg, addr, handler) EXC(sw, ST_INSN, reg, addr, handler)
199 #endif /* USE_DOUBLE */
201 #define LOADB(reg, addr, handler) EXC(lb, LD_INSN, reg, addr, handler)
202 #define STOREB(reg, addr, handler) EXC(sb, ST_INSN, reg, addr, handler)
204 #define _PREF(hint, addr, type) \
205 .if \mode == LEGACY_MODE; \
208 .if ((\from == USEROP) && (type == SRC_PREFETCH)) || \
209 ((\to == USEROP) && (type == DST_PREFETCH)); \
211 * PREFE has only 9 bits for the offset \
212 * compared to PREF which has 16, so it may \
213 * need to use the $at register but this \
214 * register should remain intact because it's \
215 * used later on. Therefore use $v1. \
225 #define PREFS(hint, addr) _PREF(hint, addr, SRC_PREFETCH)
226 #define PREFD(hint, addr) _PREF(hint, addr, DST_PREFETCH)
228 #ifdef CONFIG_CPU_LITTLE_ENDIAN
229 #define LDFIRST LOADR
231 #define STFIRST STORER
232 #define STREST STOREL
233 #define SHIFT_DISCARD SLLV
235 #define LDFIRST LOADL
237 #define STFIRST STOREL
238 #define STREST STORER
239 #define SHIFT_DISCARD SRLV
242 #define FIRST(unit) ((unit)*NBYTES)
243 #define REST(unit) (FIRST(unit)+NBYTES-1)
244 #define UNIT(unit) FIRST(unit)
246 #define ADDRMASK (NBYTES-1)
250 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
259 * Macro to build the __copy_user common code
261 * mode : LEGACY_MODE or EVA_MODE
262 * from : Source operand. USEROP or KERNELOP
263 * to : Destination operand. USEROP or KERNELOP
265 .macro __BUILD_COPY_USER mode, from, to
267 /* initialize __memcpy if this the first time we execute this macro */
270 .hidden __memcpy /* make sure it does not leak */
274 * Note: dst & src may be unaligned, len may be 0
281 * The "issue break"s below are very approximate.
282 * Issue delays for dcache fills will perturb the schedule, as will
283 * load queue full replay traps, etc.
285 * If len < NBYTES use byte operations.
290 and t1, dst, ADDRMASK
291 PREFS( 0, 1*32(src) )
292 PREFD( 1, 1*32(dst) )
293 bnez t2, .Lcopy_bytes_checklen\@
294 and t0, src, ADDRMASK
295 PREFS( 0, 2*32(src) )
296 PREFD( 1, 2*32(dst) )
297 #ifndef CONFIG_CPU_MIPSR6
298 bnez t1, .Ldst_unaligned\@
300 bnez t0, .Lsrc_unaligned_dst_aligned\@
303 bnez t0, .Lcopy_unaligned_bytes\@
306 * use delay slot for fall-through
307 * src and dst are aligned; need to compute rem
310 SRL t0, len, LOG_NBYTES+3 # +3 for 8 units/iter
311 beqz t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
312 and rem, len, (8*NBYTES-1) # rem = len % (8*NBYTES)
313 PREFS( 0, 3*32(src) )
314 PREFD( 1, 3*32(dst) )
318 LOAD(t0, UNIT(0)(src), .Ll_exc\@)
319 LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
320 LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
321 LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
322 SUB len, len, 8*NBYTES
323 LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
324 LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@)
325 STORE(t0, UNIT(0)(dst), .Ls_exc_p8u\@)
326 STORE(t1, UNIT(1)(dst), .Ls_exc_p7u\@)
327 LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@)
328 LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@)
329 ADD src, src, 8*NBYTES
330 ADD dst, dst, 8*NBYTES
331 STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@)
332 STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@)
333 STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@)
334 STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@)
335 STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@)
336 STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@)
337 PREFS( 0, 8*32(src) )
338 PREFD( 1, 8*32(dst) )
343 * len == rem == the number of bytes left to copy < 8*NBYTES
345 .Lcleanup_both_aligned\@:
347 sltu t0, len, 4*NBYTES
348 bnez t0, .Lless_than_4units\@
349 and rem, len, (NBYTES-1) # rem = len % NBYTES
353 LOAD( t0, UNIT(0)(src), .Ll_exc\@)
354 LOAD( t1, UNIT(1)(src), .Ll_exc_copy\@)
355 LOAD( t2, UNIT(2)(src), .Ll_exc_copy\@)
356 LOAD( t3, UNIT(3)(src), .Ll_exc_copy\@)
357 SUB len, len, 4*NBYTES
358 ADD src, src, 4*NBYTES
360 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@)
361 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@)
362 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@)
363 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@)
364 .set reorder /* DADDI_WAR */
365 ADD dst, dst, 4*NBYTES
368 .Lless_than_4units\@:
372 beq rem, len, .Lcopy_bytes\@
376 LOAD(t0, 0(src), .Ll_exc\@)
379 STORE(t0, 0(dst), .Ls_exc_p1u\@)
380 .set reorder /* DADDI_WAR */
385 #ifndef CONFIG_CPU_MIPSR6
387 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
388 * A loop would do only a byte at a time with possible branch
389 * mispredicts. Can't do an explicit LOAD dst,mask,or,STORE
390 * because can't assume read-access to dst. Instead, use
391 * STREST dst, which doesn't require read access to dst.
393 * This code should perform better than a simple loop on modern,
394 * wide-issue mips processors because the code has fewer branches and
395 * more instruction-level parallelism.
399 ADD t1, dst, len # t1 is just past last byte of dst
401 SLL rem, len, 3 # rem = number of bits to keep
402 LOAD(t0, 0(src), .Ll_exc\@)
403 SUB bits, bits, rem # bits = number of bits to discard
404 SHIFT_DISCARD t0, t0, bits
405 STREST(t0, -1(t1), .Ls_exc\@)
411 * t0 = src & ADDRMASK
412 * t1 = dst & ADDRMASK; T1 > 0
415 * Copy enough bytes to align dst
416 * Set match = (src and dst have same alignment)
419 LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
421 LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
422 SUB t2, t2, t1 # t2 = number of bytes copied
425 STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
426 beq len, t2, .Ldone\@
429 beqz match, .Lboth_aligned\@
432 .Lsrc_unaligned_dst_aligned\@:
433 SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter
434 PREFS( 0, 3*32(src) )
435 beqz t0, .Lcleanup_src_unaligned\@
436 and rem, len, (4*NBYTES-1) # rem = len % 4*NBYTES
437 PREFD( 1, 3*32(dst) )
440 * Avoid consecutive LD*'s to the same register since some mips
441 * implementations can't issue them in the same cycle.
442 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
443 * are to the same unit (unless src is aligned, but it's not).
446 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
447 LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
448 SUB len, len, 4*NBYTES
449 LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
450 LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
451 LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
452 LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
453 LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
454 LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
455 PREFS( 0, 9*32(src) ) # 0 is PREF_LOAD (not streamed)
456 ADD src, src, 4*NBYTES
457 #ifdef CONFIG_CPU_SB1
458 nop # improves slotting
460 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@)
461 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@)
462 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@)
463 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@)
464 PREFD( 1, 9*32(dst) ) # 1 is PREF_STORE (not streamed)
465 .set reorder /* DADDI_WAR */
466 ADD dst, dst, 4*NBYTES
470 .Lcleanup_src_unaligned\@:
472 and rem, len, NBYTES-1 # rem = len % NBYTES
473 beq rem, len, .Lcopy_bytes\@
477 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
478 LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
481 STORE(t0, 0(dst), .Ls_exc_p1u\@)
482 .set reorder /* DADDI_WAR */
487 #endif /* !CONFIG_CPU_MIPSR6 */
488 .Lcopy_bytes_checklen\@:
492 /* 0 < len < NBYTES */
494 #define COPY_BYTE(N) \
495 LOADB(t0, N(src), .Ll_exc\@); \
497 beqz len, .Ldone\@; \
498 STOREB(t0, N(dst), .Ls_exc_p1\@)
508 LOADB(t0, NBYTES-2(src), .Ll_exc\@)
511 STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@)
516 #ifdef CONFIG_CPU_MIPSR6
517 .Lcopy_unaligned_bytes\@:
530 #endif /* CONFIG_CPU_MIPSR6 */
539 * Copy bytes from src until faulting load address (or until a
542 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
543 * may be more than a byte beyond the last address.
544 * Hence, the lb below may get an exception.
546 * Assumes src < THREAD_BUADDR($28)
548 LOADK t0, TI_TASK($28)
550 LOADK t0, THREAD_BUADDR(t0)
552 LOADB(t1, 0(src), .Ll_exc\@)
554 sb t1, 0(dst) # can't fault -- we're copy_from_user
555 .set reorder /* DADDI_WAR */
560 LOADK t0, TI_TASK($28)
562 LOADK t0, THREAD_BUADDR(t0) # t0 is just past last good address
564 SUB len, AT, t0 # len number of uncopied bytes
565 bnez t6, .Ldone\@ /* Skip the zeroing part if inatomic */
567 * Here's where we rely on src and dst being incremented in tandem,
569 * dst += (fault addr - src) to put dst at first byte to clear
571 ADD dst, t0 # compute start address in a1
574 * Clear len bytes starting at dst. Can't call __bzero because it
575 * might modify len. An inefficient loop for these rare times...
577 .set reorder /* DADDI_WAR */
583 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
599 .set reorder; /* DADDI_WAR */ \
600 .Ls_exc_p ## n ## u\@: \
601 ADD len, len, n*NBYTES; \
615 .set reorder /* DADDI_WAR */
626 EXPORT_SYMBOL(memmove)
629 sltu t0, a1, t0 # dst + len <= src -> memcpy
630 sltu t1, a0, t1 # dst >= src + len -> memcpy
633 move v0, a0 /* return value */
637 /* fall through to __rmemcpy */
638 LEAF(__rmemcpy) /* a0=dst a1=src a2=len */
640 beqz t0, .Lr_end_bytes_up # src >= dst
642 ADD a0, a2 # dst = dst + len
643 ADD a1, a2 # src = src + len
651 .set reorder /* DADDI_WAR */
653 bnez a2, .Lr_end_bytes
666 .set reorder /* DADDI_WAR */
668 bnez a2, .Lr_end_bytes_up
676 * t6 is used as a flag to note inatomic mode.
678 LEAF(__copy_user_inatomic)
679 EXPORT_SYMBOL(__copy_user_inatomic)
682 END(__copy_user_inatomic)
685 * A combined memcpy/__copy_user
686 * __copy_user sets len to 0 for success; else to an upper bound of
687 * the number of uncopied bytes.
688 * memcpy sets v0 to dst.
691 LEAF(memcpy) /* a0=dst a1=src a2=len */
692 EXPORT_SYMBOL(memcpy)
693 move v0, dst /* return value */
696 EXPORT_SYMBOL(__copy_user)
697 li t6, 0 /* not inatomic */
699 /* Legacy Mode, user <-> user */
700 __BUILD_COPY_USER LEGACY_MODE USEROP USEROP
705 * For EVA we need distinct symbols for reading and writing to user space.
706 * This is because we need to use specific EVA instructions to perform the
707 * virtual <-> physical translation when a virtual address is actually in user
711 LEAF(__copy_user_inatomic_eva)
712 EXPORT_SYMBOL(__copy_user_inatomic_eva)
713 b __copy_from_user_common
715 END(__copy_user_inatomic_eva)
718 * __copy_from_user (EVA)
721 LEAF(__copy_from_user_eva)
722 EXPORT_SYMBOL(__copy_from_user_eva)
723 li t6, 0 /* not inatomic */
724 __copy_from_user_common:
725 __BUILD_COPY_USER EVA_MODE USEROP KERNELOP
726 END(__copy_from_user_eva)
731 * __copy_to_user (EVA)
734 LEAF(__copy_to_user_eva)
735 EXPORT_SYMBOL(__copy_to_user_eva)
736 __BUILD_COPY_USER EVA_MODE KERNELOP USEROP
737 END(__copy_to_user_eva)
740 * __copy_in_user (EVA)
743 LEAF(__copy_in_user_eva)
744 EXPORT_SYMBOL(__copy_in_user_eva)
745 __BUILD_COPY_USER EVA_MODE USEROP USEROP
746 END(__copy_in_user_eva)