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
33 #include <asm/asm-offsets.h>
34 #include <asm/regdef.h>
43 * memcpy copies len bytes from src to dst and sets v0 to dst.
45 * - src and dst don't overlap
48 * memcpy uses the standard calling convention
50 * __copy_user copies up to len bytes from src to dst and sets a2 (len) to
51 * the number of uncopied bytes due to an exception caused by a read or write.
52 * __copy_user assumes that src and dst don't overlap, and that the call is
53 * implementing one of the following:
55 * - src is readable (no exceptions when reading src)
57 * - dst is writable (no exceptions when writing dst)
58 * __copy_user uses a non-standard calling convention; see
59 * include/asm-mips/uaccess.h
61 * When an exception happens on a load, the handler must
62 # ensure that all of the destination buffer is overwritten to prevent
63 * leaking information to user mode programs.
71 * The exception handler for loads requires that:
72 * 1- AT contain the address of the byte just past the end of the source
74 * 2- src_entry <= src < AT, and
75 * 3- (dst - src) == (dst_entry - src_entry),
76 * The _entry suffix denotes values when __copy_user was called.
78 * (1) is set up up by uaccess.h and maintained by not writing AT in copy_user
79 * (2) is met by incrementing src by the number of bytes copied
80 * (3) is met by not doing loads between a pair of increments of dst and src
82 * The exception handlers for stores adjust len (if necessary) and return.
83 * These handlers do not need to overwrite any data.
85 * For __rmemcpy and memmove an exception is always a kernel bug, therefore
86 * they're not protected.
89 /* Instruction type */
93 #define SRC_PREFETCH 1
94 #define DST_PREFETCH 2
101 * Wrapper to add an entry in the exception table
102 * in case the insn causes a memory exception.
104 * insn : Load/store instruction
105 * type : Instruction type
108 * handler : Exception handler
111 #define EXC(insn, type, reg, addr, handler) \
112 .if \mode == LEGACY_MODE; \
114 .section __ex_table,"a"; \
117 /* This is assembled in EVA mode */ \
119 /* If loading from user or storing to user */ \
120 .if ((\from == USEROP) && (type == LD_INSN)) || \
121 ((\to == USEROP) && (type == ST_INSN)); \
122 9: __BUILD_EVA_INSN(insn##e, reg, addr); \
123 .section __ex_table,"a"; \
128 * Still in EVA, but no need for \
129 * exception handler or EVA insn \
136 * Only on the 64-bit kernel we can made use of 64-bit registers.
144 #define LOADK ld /* No exception */
145 #define LOAD(reg, addr, handler) EXC(ld, LD_INSN, reg, addr, handler)
146 #define LOADL(reg, addr, handler) EXC(ldl, LD_INSN, reg, addr, handler)
147 #define LOADR(reg, addr, handler) EXC(ldr, LD_INSN, reg, addr, handler)
148 #define STOREL(reg, addr, handler) EXC(sdl, ST_INSN, reg, addr, handler)
149 #define STORER(reg, addr, handler) EXC(sdr, ST_INSN, reg, addr, handler)
150 #define STORE(reg, addr, handler) EXC(sd, ST_INSN, reg, addr, handler)
162 * As we are sharing code base with the mips32 tree (which use the o32 ABI
163 * register definitions). We need to redefine the register definitions from
164 * the n64 ABI register naming to the o32 ABI register naming.
181 #define LOADK lw /* No exception */
182 #define LOAD(reg, addr, handler) EXC(lw, LD_INSN, reg, addr, handler)
183 #define LOADL(reg, addr, handler) EXC(lwl, LD_INSN, reg, addr, handler)
184 #define LOADR(reg, addr, handler) EXC(lwr, LD_INSN, reg, addr, handler)
185 #define STOREL(reg, addr, handler) EXC(swl, ST_INSN, reg, addr, handler)
186 #define STORER(reg, addr, handler) EXC(swr, ST_INSN, reg, addr, handler)
187 #define STORE(reg, addr, handler) EXC(sw, ST_INSN, reg, addr, handler)
198 #endif /* USE_DOUBLE */
200 #define LOADB(reg, addr, handler) EXC(lb, LD_INSN, reg, addr, handler)
201 #define STOREB(reg, addr, handler) EXC(sb, ST_INSN, reg, addr, handler)
203 #define _PREF(hint, addr, type) \
204 .if \mode == LEGACY_MODE; \
207 .if ((\from == USEROP) && (type == SRC_PREFETCH)) || \
208 ((\to == USEROP) && (type == DST_PREFETCH)); \
210 * PREFE has only 9 bits for the offset \
211 * compared to PREF which has 16, so it may \
212 * need to use the $at register but this \
213 * register should remain intact because it's \
214 * used later on. Therefore use $v1. \
224 #define PREFS(hint, addr) _PREF(hint, addr, SRC_PREFETCH)
225 #define PREFD(hint, addr) _PREF(hint, addr, DST_PREFETCH)
227 #ifdef CONFIG_CPU_LITTLE_ENDIAN
228 #define LDFIRST LOADR
230 #define STFIRST STORER
231 #define STREST STOREL
232 #define SHIFT_DISCARD SLLV
234 #define LDFIRST LOADL
236 #define STFIRST STOREL
237 #define STREST STORER
238 #define SHIFT_DISCARD SRLV
241 #define FIRST(unit) ((unit)*NBYTES)
242 #define REST(unit) (FIRST(unit)+NBYTES-1)
243 #define UNIT(unit) FIRST(unit)
245 #define ADDRMASK (NBYTES-1)
249 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
258 * Macro to build the __copy_user common code
260 * mode : LEGACY_MODE or EVA_MODE
261 * from : Source operand. USEROP or KERNELOP
262 * to : Destination operand. USEROP or KERNELOP
264 .macro __BUILD_COPY_USER mode, from, to
266 /* initialize __memcpy if this the first time we execute this macro */
269 .hidden __memcpy /* make sure it does not leak */
273 * Note: dst & src may be unaligned, len may be 0
280 * The "issue break"s below are very approximate.
281 * Issue delays for dcache fills will perturb the schedule, as will
282 * load queue full replay traps, etc.
284 * If len < NBYTES use byte operations.
289 and t1, dst, ADDRMASK
290 PREFS( 0, 1*32(src) )
291 PREFD( 1, 1*32(dst) )
292 bnez t2, .Lcopy_bytes_checklen\@
293 and t0, src, ADDRMASK
294 PREFS( 0, 2*32(src) )
295 PREFD( 1, 2*32(dst) )
296 #ifndef CONFIG_CPU_MIPSR6
297 bnez t1, .Ldst_unaligned\@
299 bnez t0, .Lsrc_unaligned_dst_aligned\@
302 bnez t0, .Lcopy_unaligned_bytes\@
305 * use delay slot for fall-through
306 * src and dst are aligned; need to compute rem
309 SRL t0, len, LOG_NBYTES+3 # +3 for 8 units/iter
310 beqz t0, .Lcleanup_both_aligned\@ # len < 8*NBYTES
311 and rem, len, (8*NBYTES-1) # rem = len % (8*NBYTES)
312 PREFS( 0, 3*32(src) )
313 PREFD( 1, 3*32(dst) )
317 LOAD(t0, UNIT(0)(src), .Ll_exc\@)
318 LOAD(t1, UNIT(1)(src), .Ll_exc_copy\@)
319 LOAD(t2, UNIT(2)(src), .Ll_exc_copy\@)
320 LOAD(t3, UNIT(3)(src), .Ll_exc_copy\@)
321 SUB len, len, 8*NBYTES
322 LOAD(t4, UNIT(4)(src), .Ll_exc_copy\@)
323 LOAD(t7, UNIT(5)(src), .Ll_exc_copy\@)
324 STORE(t0, UNIT(0)(dst), .Ls_exc_p8u\@)
325 STORE(t1, UNIT(1)(dst), .Ls_exc_p7u\@)
326 LOAD(t0, UNIT(6)(src), .Ll_exc_copy\@)
327 LOAD(t1, UNIT(7)(src), .Ll_exc_copy\@)
328 ADD src, src, 8*NBYTES
329 ADD dst, dst, 8*NBYTES
330 STORE(t2, UNIT(-6)(dst), .Ls_exc_p6u\@)
331 STORE(t3, UNIT(-5)(dst), .Ls_exc_p5u\@)
332 STORE(t4, UNIT(-4)(dst), .Ls_exc_p4u\@)
333 STORE(t7, UNIT(-3)(dst), .Ls_exc_p3u\@)
334 STORE(t0, UNIT(-2)(dst), .Ls_exc_p2u\@)
335 STORE(t1, UNIT(-1)(dst), .Ls_exc_p1u\@)
336 PREFS( 0, 8*32(src) )
337 PREFD( 1, 8*32(dst) )
342 * len == rem == the number of bytes left to copy < 8*NBYTES
344 .Lcleanup_both_aligned\@:
346 sltu t0, len, 4*NBYTES
347 bnez t0, .Lless_than_4units\@
348 and rem, len, (NBYTES-1) # rem = len % NBYTES
352 LOAD( t0, UNIT(0)(src), .Ll_exc\@)
353 LOAD( t1, UNIT(1)(src), .Ll_exc_copy\@)
354 LOAD( t2, UNIT(2)(src), .Ll_exc_copy\@)
355 LOAD( t3, UNIT(3)(src), .Ll_exc_copy\@)
356 SUB len, len, 4*NBYTES
357 ADD src, src, 4*NBYTES
359 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@)
360 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@)
361 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@)
362 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@)
363 .set reorder /* DADDI_WAR */
364 ADD dst, dst, 4*NBYTES
367 .Lless_than_4units\@:
371 beq rem, len, .Lcopy_bytes\@
375 LOAD(t0, 0(src), .Ll_exc\@)
378 STORE(t0, 0(dst), .Ls_exc_p1u\@)
379 .set reorder /* DADDI_WAR */
384 #ifndef CONFIG_CPU_MIPSR6
386 * src and dst are aligned, need to copy rem bytes (rem < NBYTES)
387 * A loop would do only a byte at a time with possible branch
388 * mispredicts. Can't do an explicit LOAD dst,mask,or,STORE
389 * because can't assume read-access to dst. Instead, use
390 * STREST dst, which doesn't require read access to dst.
392 * This code should perform better than a simple loop on modern,
393 * wide-issue mips processors because the code has fewer branches and
394 * more instruction-level parallelism.
398 ADD t1, dst, len # t1 is just past last byte of dst
400 SLL rem, len, 3 # rem = number of bits to keep
401 LOAD(t0, 0(src), .Ll_exc\@)
402 SUB bits, bits, rem # bits = number of bits to discard
403 SHIFT_DISCARD t0, t0, bits
404 STREST(t0, -1(t1), .Ls_exc\@)
410 * t0 = src & ADDRMASK
411 * t1 = dst & ADDRMASK; T1 > 0
414 * Copy enough bytes to align dst
415 * Set match = (src and dst have same alignment)
418 LDFIRST(t3, FIRST(0)(src), .Ll_exc\@)
420 LDREST(t3, REST(0)(src), .Ll_exc_copy\@)
421 SUB t2, t2, t1 # t2 = number of bytes copied
424 STFIRST(t3, FIRST(0)(dst), .Ls_exc\@)
425 beq len, t2, .Ldone\@
428 beqz match, .Lboth_aligned\@
431 .Lsrc_unaligned_dst_aligned\@:
432 SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter
433 PREFS( 0, 3*32(src) )
434 beqz t0, .Lcleanup_src_unaligned\@
435 and rem, len, (4*NBYTES-1) # rem = len % 4*NBYTES
436 PREFD( 1, 3*32(dst) )
439 * Avoid consecutive LD*'s to the same register since some mips
440 * implementations can't issue them in the same cycle.
441 * It's OK to load FIRST(N+1) before REST(N) because the two addresses
442 * are to the same unit (unless src is aligned, but it's not).
445 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
446 LDFIRST(t1, FIRST(1)(src), .Ll_exc_copy\@)
447 SUB len, len, 4*NBYTES
448 LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
449 LDREST(t1, REST(1)(src), .Ll_exc_copy\@)
450 LDFIRST(t2, FIRST(2)(src), .Ll_exc_copy\@)
451 LDFIRST(t3, FIRST(3)(src), .Ll_exc_copy\@)
452 LDREST(t2, REST(2)(src), .Ll_exc_copy\@)
453 LDREST(t3, REST(3)(src), .Ll_exc_copy\@)
454 PREFS( 0, 9*32(src) ) # 0 is PREF_LOAD (not streamed)
455 ADD src, src, 4*NBYTES
456 #ifdef CONFIG_CPU_SB1
457 nop # improves slotting
459 STORE(t0, UNIT(0)(dst), .Ls_exc_p4u\@)
460 STORE(t1, UNIT(1)(dst), .Ls_exc_p3u\@)
461 STORE(t2, UNIT(2)(dst), .Ls_exc_p2u\@)
462 STORE(t3, UNIT(3)(dst), .Ls_exc_p1u\@)
463 PREFD( 1, 9*32(dst) ) # 1 is PREF_STORE (not streamed)
464 .set reorder /* DADDI_WAR */
465 ADD dst, dst, 4*NBYTES
469 .Lcleanup_src_unaligned\@:
471 and rem, len, NBYTES-1 # rem = len % NBYTES
472 beq rem, len, .Lcopy_bytes\@
476 LDFIRST(t0, FIRST(0)(src), .Ll_exc\@)
477 LDREST(t0, REST(0)(src), .Ll_exc_copy\@)
480 STORE(t0, 0(dst), .Ls_exc_p1u\@)
481 .set reorder /* DADDI_WAR */
486 #endif /* !CONFIG_CPU_MIPSR6 */
487 .Lcopy_bytes_checklen\@:
491 /* 0 < len < NBYTES */
493 #define COPY_BYTE(N) \
494 LOADB(t0, N(src), .Ll_exc\@); \
496 beqz len, .Ldone\@; \
497 STOREB(t0, N(dst), .Ls_exc_p1\@)
507 LOADB(t0, NBYTES-2(src), .Ll_exc\@)
510 STOREB(t0, NBYTES-2(dst), .Ls_exc_p1\@)
515 #ifdef CONFIG_CPU_MIPSR6
516 .Lcopy_unaligned_bytes\@:
529 #endif /* CONFIG_CPU_MIPSR6 */
538 * Copy bytes from src until faulting load address (or until a
541 * When reached by a faulting LDFIRST/LDREST, THREAD_BUADDR($28)
542 * may be more than a byte beyond the last address.
543 * Hence, the lb below may get an exception.
545 * Assumes src < THREAD_BUADDR($28)
547 LOADK t0, TI_TASK($28)
549 LOADK t0, THREAD_BUADDR(t0)
551 LOADB(t1, 0(src), .Ll_exc\@)
553 sb t1, 0(dst) # can't fault -- we're copy_from_user
554 .set reorder /* DADDI_WAR */
559 LOADK t0, TI_TASK($28)
561 LOADK t0, THREAD_BUADDR(t0) # t0 is just past last good address
563 SUB len, AT, t0 # len number of uncopied bytes
564 bnez t6, .Ldone\@ /* Skip the zeroing part if inatomic */
566 * Here's where we rely on src and dst being incremented in tandem,
568 * dst += (fault addr - src) to put dst at first byte to clear
570 ADD dst, t0 # compute start address in a1
573 * Clear len bytes starting at dst. Can't call __bzero because it
574 * might modify len. An inefficient loop for these rare times...
576 .set reorder /* DADDI_WAR */
582 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
598 .set reorder; /* DADDI_WAR */ \
599 .Ls_exc_p ## n ## u\@: \
600 ADD len, len, n*NBYTES; \
614 .set reorder /* DADDI_WAR */
627 sltu t0, a1, t0 # dst + len <= src -> memcpy
628 sltu t1, a0, t1 # dst >= src + len -> memcpy
631 move v0, a0 /* return value */
635 /* fall through to __rmemcpy */
636 LEAF(__rmemcpy) /* a0=dst a1=src a2=len */
638 beqz t0, .Lr_end_bytes_up # src >= dst
640 ADD a0, a2 # dst = dst + len
641 ADD a1, a2 # src = src + len
649 .set reorder /* DADDI_WAR */
651 bnez a2, .Lr_end_bytes
664 .set reorder /* DADDI_WAR */
666 bnez a2, .Lr_end_bytes_up
674 * t6 is used as a flag to note inatomic mode.
676 LEAF(__copy_user_inatomic)
679 END(__copy_user_inatomic)
682 * A combined memcpy/__copy_user
683 * __copy_user sets len to 0 for success; else to an upper bound of
684 * the number of uncopied bytes.
685 * memcpy sets v0 to dst.
688 LEAF(memcpy) /* a0=dst a1=src a2=len */
689 move v0, dst /* return value */
692 li t6, 0 /* not inatomic */
694 /* Legacy Mode, user <-> user */
695 __BUILD_COPY_USER LEGACY_MODE USEROP USEROP
700 * For EVA we need distinct symbols for reading and writing to user space.
701 * This is because we need to use specific EVA instructions to perform the
702 * virtual <-> physical translation when a virtual address is actually in user
706 LEAF(__copy_user_inatomic_eva)
707 b __copy_from_user_common
709 END(__copy_user_inatomic_eva)
712 * __copy_from_user (EVA)
715 LEAF(__copy_from_user_eva)
716 li t6, 0 /* not inatomic */
717 __copy_from_user_common:
718 __BUILD_COPY_USER EVA_MODE USEROP KERNELOP
719 END(__copy_from_user_eva)
724 * __copy_to_user (EVA)
727 LEAF(__copy_to_user_eva)
728 __BUILD_COPY_USER EVA_MODE KERNELOP USEROP
729 END(__copy_to_user_eva)
732 * __copy_in_user (EVA)
735 LEAF(__copy_in_user_eva)
736 __BUILD_COPY_USER EVA_MODE USEROP USEROP
737 END(__copy_in_user_eva)