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.
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 * 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
31 #ifdef CONFIG_CPU_MIPSR6
32 #undef CONFIG_CPU_HAS_PREFETCH
36 #include <asm/asm-offsets.h>
37 #include <asm/export.h>
38 #include <asm/regdef.h>
47 * memcpy copies len bytes from src to dst and sets v0 to dst.
49 * - src and dst don't overlap
52 * memcpy uses the standard calling convention
54 * __copy_user copies up to len bytes from src to dst and sets a2 (len) to
55 * the number of uncopied bytes due to an exception caused by a read or write.
56 * __copy_user assumes that src and dst don't overlap, and that the call is
57 * implementing one of the following:
59 * - src is readable (no exceptions when reading src)
61 * - dst is writable (no exceptions when writing dst)
62 * __copy_user uses a non-standard calling convention; see
63 * include/asm-mips/uaccess.h
65 * When an exception happens on a load, the handler must
66 # ensure that all of the destination buffer is overwritten to prevent
67 * leaking information to user mode programs.
75 * The exception handler for loads requires that:
76 * 1- AT contain the address of the byte just past the end of the source
78 * 2- src_entry <= src < AT, and
79 * 3- (dst - src) == (dst_entry - src_entry),
80 * The _entry suffix denotes values when __copy_user was called.
82 * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user
83 * (2) is met by incrementing src by the number of bytes copied
84 * (3) is met by not doing loads between a pair of increments of dst and src
86 * The exception handlers for stores adjust len (if necessary) and return.
87 * These handlers do not need to overwrite any data.
89 * For __rmemcpy and memmove an exception is always a kernel bug, therefore
90 * they're not protected.
93 /* Instruction type */
97 #define SRC_PREFETCH 1
98 #define DST_PREFETCH 2
105 * Wrapper to add an entry in the exception table
106 * in case the insn causes a memory exception.
108 * insn : Load/store instruction
109 * type : Instruction type
112 * handler : Exception handler
115 #define EXC(insn, type, reg, addr, handler) \
116 .if \mode == LEGACY_MODE; \
118 .section __ex_table,"a"; \
121 /* This is assembled in EVA mode */ \
123 /* If loading from user or storing to user */ \
124 .if ((\from == USEROP) && (type == LD_INSN)) || \
125 ((\to == USEROP) && (type == ST_INSN)); \
126 9: __BUILD_EVA_INSN(insn##e, reg, addr); \
127 .section __ex_table,"a"; \
132 * Still in EVA, but no need for \
133 * exception handler or EVA insn \
140 * Only on the 64-bit kernel we can made use of 64-bit registers.
148 #define LOADK ld /* No exception */
149 #define LOAD(reg, addr, handler) EXC(ld, LD_INSN, reg, addr, handler)
150 #define LOADL(reg, addr, handler) EXC(ldl, LD_INSN, reg, addr, handler)
151 #define LOADR(reg, addr, handler) EXC(ldr, LD_INSN, reg, addr, handler)
152 #define STOREL(reg, addr, handler) EXC(sdl, ST_INSN, reg, addr, handler)
153 #define STORER(reg, addr, handler) EXC(sdr, ST_INSN, reg, addr, handler)
154 #define STORE(reg, addr, handler) EXC(sd, ST_INSN, reg, addr, handler)
166 * As we are sharing code base with the mips32 tree (which use the o32 ABI
167 * register definitions). We need to redefine the register definitions from
168 * the n64 ABI register naming to the o32 ABI register naming.
185 #define LOADK lw /* No exception */
186 #define LOAD(reg, addr, handler) EXC(lw, LD_INSN, reg, addr, handler)
187 #define LOADL(reg, addr, handler) EXC(lwl, LD_INSN, reg, addr, handler)
188 #define LOADR(reg, addr, handler) EXC(lwr, LD_INSN, reg, addr, handler)
189 #define STOREL(reg, addr, handler) EXC(swl, ST_INSN, reg, addr, handler)
190 #define STORER(reg, addr, handler) EXC(swr, ST_INSN, reg, addr, handler)
191 #define STORE(reg, addr, handler) EXC(sw, ST_INSN, reg, addr, handler)
202 #endif /* USE_DOUBLE */
204 #define LOADB(reg, addr, handler) EXC(lb, LD_INSN, reg, addr, handler)
205 #define STOREB(reg, addr, handler) EXC(sb, ST_INSN, reg, addr, handler)
207 #ifdef CONFIG_CPU_HAS_PREFETCH
208 # define _PREF(hint, addr, type) \
209 .if \mode == LEGACY_MODE; \
210 kernel_pref(hint, addr); \
212 .if ((\from == USEROP) && (type == SRC_PREFETCH)) || \
213 ((\to == USEROP) && (type == DST_PREFETCH)); \
215 * PREFE has only 9 bits for the offset \
216 * compared to PREF which has 16, so it may \
217 * need to use the $at register but this \
218 * register should remain intact because it's \
219 * used later on. Therefore use $v1. \
222 user_pref(hint, addr); \
225 kernel_pref(hint, addr); \
229 # define _PREF(hint, addr, type)
232 #define PREFS(hint, addr) _PREF(hint, addr, SRC_PREFETCH)
233 #define PREFD(hint, addr) _PREF(hint, addr, DST_PREFETCH)
235 #ifdef CONFIG_CPU_LITTLE_ENDIAN
236 #define LDFIRST LOADR
238 #define STFIRST STORER
239 #define STREST STOREL
240 #define SHIFT_DISCARD SLLV
242 #define LDFIRST LOADL
244 #define STFIRST STOREL
245 #define STREST STORER
246 #define SHIFT_DISCARD SRLV
249 #define FIRST(unit) ((unit)*NBYTES)
250 #define REST(unit) (FIRST(unit)+NBYTES-1)
251 #define UNIT(unit) FIRST(unit)
253 #define ADDRMASK (NBYTES-1)
257 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
266 * Macro to build the __copy_user common code
268 * mode : LEGACY_MODE or EVA_MODE
269 * from : Source operand. USEROP or KERNELOP
270 * to : Destination operand. USEROP or KERNELOP
272 .macro __BUILD_COPY_USER mode, from, to
274 /* initialize __memcpy if this the first time we execute this macro */
277 .hidden __memcpy /* make sure it does not leak */
281 * Note: dst & src may be unaligned, len may be 0
288 * The "issue break"s below are very approximate.
289 * Issue delays for dcache fills will perturb the schedule, as will
290 * load queue full replay traps, etc.
292 * If len < NBYTES use byte operations.
297 and t1, dst, ADDRMASK
298 PREFS( 0, 1*32(src) )
299 PREFD( 1, 1*32(dst) )
300 bnez t2, .Lcopy_bytes_checklen\@
301 and t0, src, ADDRMASK
302 PREFS( 0, 2*32(src) )
303 PREFD( 1, 2*32(dst) )
304 #ifndef CONFIG_CPU_NO_LOAD_STORE_LR
305 bnez t1, .Ldst_unaligned\@
307 bnez t0, .Lsrc_unaligned_dst_aligned\@
308 #else /* CONFIG_CPU_NO_LOAD_STORE_LR */
310 bnez t0, .Lcopy_unaligned_bytes\@
311 #endif /* CONFIG_CPU_NO_LOAD_STORE_LR */
313 * use delay slot for fall-through
314 * src and dst are aligned; need to compute rem
317 SRL t0, len, LOG_NBYTES+3 # +3 for 8 units/iter
318 beqz t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
319 and rem, len, (8*NBYTES-1) # rem = len % (8*NBYTES)
320 PREFS( 0, 3*32(src) )
321 PREFD( 1, 3*32(dst) )
325 LOAD(t0, UNIT(0)(src), .Ll_exc\@)
326 LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
327 LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
328 LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
329 SUB len, len, 8*NBYTES
330 LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
331 LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@)
332 STORE(t0, UNIT(0)(dst), .Ls_exc_p8u\@)
333 STORE(t1, UNIT(1)(dst), .Ls_exc_p7u\@)
334 LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@)
335 LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@)
336 ADD src, src, 8*NBYTES
337 ADD dst, dst, 8*NBYTES
338 STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@)
339 STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@)
340 STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@)
341 STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@)
342 STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@)
343 STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@)
344 PREFS( 0, 8*32(src) )
345 PREFD( 1, 8*32(dst) )
350 * len == rem == the number of bytes left to copy < 8*NBYTES
352 .Lcleanup_both_aligned\@:
354 sltu t0, len, 4*NBYTES
355 bnez t0, .Lless_than_4units\@
356 and rem, len, (NBYTES-1) # rem = len % NBYTES
360 LOAD( t0, UNIT(0)(src), .Ll_exc\@)
361 LOAD( t1, UNIT(1)(src), .Ll_exc_copy\@)
362 LOAD( t2, UNIT(2)(src), .Ll_exc_copy\@)
363 LOAD( t3, UNIT(3)(src), .Ll_exc_copy\@)
364 SUB len, len, 4*NBYTES
365 ADD src, src, 4*NBYTES
367 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@)
368 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@)
369 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@)
370 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@)
371 .set reorder /* DADDI_WAR */
372 ADD dst, dst, 4*NBYTES
375 .Lless_than_4units\@:
379 beq rem, len, .Lcopy_bytes\@
383 LOAD(t0, 0(src), .Ll_exc\@)
386 STORE(t0, 0(dst), .Ls_exc_p1u\@)
387 .set reorder /* DADDI_WAR */
392 #ifndef CONFIG_CPU_NO_LOAD_STORE_LR
394 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
395 * A loop would do only a byte at a time with possible branch
396 * mispredicts. Can't do an explicit LOAD dst,mask,or,STORE
397 * because can't assume read-access to dst. Instead, use
398 * STREST dst, which doesn't require read access to dst.
400 * This code should perform better than a simple loop on modern,
401 * wide-issue mips processors because the code has fewer branches and
402 * more instruction-level parallelism.
406 ADD t1, dst, len # t1 is just past last byte of dst
408 SLL rem, len, 3 # rem = number of bits to keep
409 LOAD(t0, 0(src), .Ll_exc\@)
410 SUB bits, bits, rem # bits = number of bits to discard
411 SHIFT_DISCARD t0, t0, bits
412 STREST(t0, -1(t1), .Ls_exc\@)
418 * t0 = src & ADDRMASK
419 * t1 = dst & ADDRMASK; T1 > 0
422 * Copy enough bytes to align dst
423 * Set match = (src and dst have same alignment)
426 LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
428 LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
429 SUB t2, t2, t1 # t2 = number of bytes copied
432 STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
433 beq len, t2, .Ldone\@
436 beqz match, .Lboth_aligned\@
439 .Lsrc_unaligned_dst_aligned\@:
440 SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter
441 PREFS( 0, 3*32(src) )
442 beqz t0, .Lcleanup_src_unaligned\@
443 and rem, len, (4*NBYTES-1) # rem = len % 4*NBYTES
444 PREFD( 1, 3*32(dst) )
447 * Avoid consecutive LD*'s to the same register since some mips
448 * implementations can't issue them in the same cycle.
449 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
450 * are to the same unit (unless src is aligned, but it's not).
453 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
454 LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
455 SUB len, len, 4*NBYTES
456 LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
457 LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
458 LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
459 LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
460 LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
461 LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
462 PREFS( 0, 9*32(src) ) # 0 is PREF_LOAD (not streamed)
463 ADD src, src, 4*NBYTES
464 #ifdef CONFIG_CPU_SB1
465 nop # improves slotting
467 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@)
468 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@)
469 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@)
470 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@)
471 PREFD( 1, 9*32(dst) ) # 1 is PREF_STORE (not streamed)
472 .set reorder /* DADDI_WAR */
473 ADD dst, dst, 4*NBYTES
477 .Lcleanup_src_unaligned\@:
479 and rem, len, NBYTES-1 # rem = len % NBYTES
480 beq rem, len, .Lcopy_bytes\@
484 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
485 LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
488 STORE(t0, 0(dst), .Ls_exc_p1u\@)
489 .set reorder /* DADDI_WAR */
494 #endif /* !CONFIG_CPU_NO_LOAD_STORE_LR */
495 .Lcopy_bytes_checklen\@:
499 /* 0 < len < NBYTES */
501 #define COPY_BYTE(N) \
502 LOADB(t0, N(src), .Ll_exc\@); \
504 beqz len, .Ldone\@; \
505 STOREB(t0, N(dst), .Ls_exc_p1\@)
515 LOADB(t0, NBYTES-2(src), .Ll_exc\@)
518 STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@)
523 #ifdef CONFIG_CPU_NO_LOAD_STORE_LR
524 .Lcopy_unaligned_bytes\@:
537 #endif /* CONFIG_CPU_NO_LOAD_STORE_LR */
546 * Copy bytes from src until faulting load address (or until a
549 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
550 * may be more than a byte beyond the last address.
551 * Hence, the lb below may get an exception.
553 * Assumes src < THREAD_BUADDR($28)
555 LOADK t0, TI_TASK($28)
557 LOADK t0, THREAD_BUADDR(t0)
559 LOADB(t1, 0(src), .Ll_exc\@)
561 sb t1, 0(dst) # can't fault -- we're copy_from_user
562 .set reorder /* DADDI_WAR */
567 LOADK t0, TI_TASK($28)
569 LOADK t0, THREAD_BUADDR(t0) # t0 is just past last good address
571 SUB len, AT, t0 # len number of uncopied bytes
576 .set reorder; /* DADDI_WAR */ \
577 .Ls_exc_p ## n ## u\@: \
578 ADD len, len, n*NBYTES; \
592 .set reorder /* DADDI_WAR */
603 EXPORT_SYMBOL(memmove)
606 sltu t0, a1, t0 # dst + len <= src -> memcpy
607 sltu t1, a0, t1 # dst >= src + len -> memcpy
610 move v0, a0 /* return value */
614 /* fall through to __rmemcpy */
615 LEAF(__rmemcpy) /* a0=dst a1=src a2=len */
617 beqz t0, .Lr_end_bytes_up # src >= dst
619 ADD a0, a2 # dst = dst + len
620 ADD a1, a2 # src = src + len
628 .set reorder /* DADDI_WAR */
630 bnez a2, .Lr_end_bytes
643 .set reorder /* DADDI_WAR */
645 bnez a2, .Lr_end_bytes_up
653 * A combined memcpy/__copy_user
654 * __copy_user sets len to 0 for success; else to an upper bound of
655 * the number of uncopied bytes.
656 * memcpy sets v0 to dst.
659 LEAF(memcpy) /* a0=dst a1=src a2=len */
660 EXPORT_SYMBOL(memcpy)
661 move v0, dst /* return value */
664 EXPORT_SYMBOL(__copy_user)
665 /* Legacy Mode, user <-> user */
666 __BUILD_COPY_USER LEGACY_MODE USEROP USEROP
671 * For EVA we need distinct symbols for reading and writing to user space.
672 * This is because we need to use specific EVA instructions to perform the
673 * virtual <-> physical translation when a virtual address is actually in user
678 * __copy_from_user (EVA)
681 LEAF(__copy_from_user_eva)
682 EXPORT_SYMBOL(__copy_from_user_eva)
683 __BUILD_COPY_USER EVA_MODE USEROP KERNELOP
684 END(__copy_from_user_eva)
689 * __copy_to_user (EVA)
692 LEAF(__copy_to_user_eva)
693 EXPORT_SYMBOL(__copy_to_user_eva)
694 __BUILD_COPY_USER EVA_MODE KERNELOP USEROP
695 END(__copy_to_user_eva)
698 * __copy_in_user (EVA)
701 LEAF(__copy_in_user_eva)
702 EXPORT_SYMBOL(__copy_in_user_eva)
703 __BUILD_COPY_USER EVA_MODE USEROP USEROP
704 END(__copy_in_user_eva)