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
14 * Mnemonic names for arguments to memcpy/__copy_user
18 * Hack to resolve longstanding prefetch issue
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.
24 #ifdef CONFIG_DMA_NONCOHERENT
25 #undef CONFIG_CPU_HAS_PREFETCH
27 #ifdef CONFIG_MIPS_MALTA
28 #undef CONFIG_CPU_HAS_PREFETCH
32 #include <asm/asm-offsets.h>
33 #include <asm/regdef.h>
42 * memcpy copies len bytes from src to dst and sets v0 to dst.
44 * - src and dst don't overlap
47 * memcpy uses the standard calling convention
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:
54 * - src is readable (no exceptions when reading src)
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
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.
70 * The exception handler for loads requires that:
71 * 1- AT contain the address of the byte just past the end of the source
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.
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
81 * The exception handlers for stores adjust len (if necessary) and return.
82 * These handlers do not need to overwrite any data.
84 * For __rmemcpy and memmove an exception is always a kernel bug, therefore
85 * they're not protected.
88 #define EXC(inst_reg,addr,handler) \
90 .section __ex_table,"a"; \
95 * Only on the 64-bit kernel we can made use of 64-bit registers.
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.
155 #endif /* USE_DOUBLE */
157 #ifdef CONFIG_CPU_LITTLE_ENDIAN
158 #define LDFIRST LOADR
160 #define STFIRST STORER
161 #define STREST STOREL
162 #define SHIFT_DISCARD SLLV
164 #define LDFIRST LOADL
166 #define STFIRST STOREL
167 #define STREST STORER
168 #define SHIFT_DISCARD SRLV
171 #define FIRST(unit) ((unit)*NBYTES)
172 #define REST(unit) (FIRST(unit)+NBYTES-1)
173 #define UNIT(unit) FIRST(unit)
175 #define ADDRMASK (NBYTES-1)
179 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
186 * t6 is used as a flag to note inatomic mode.
188 LEAF(__copy_user_inatomic)
191 END(__copy_user_inatomic)
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.
200 LEAF(memcpy) /* a0=dst a1=src a2=len */
201 move v0, dst /* return value */
204 li t6, 0 /* not inatomic */
207 * Note: dst & src may be unaligned, len may be 0
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.
218 * If len < NBYTES use byte operations.
223 and t1, dst, ADDRMASK
226 bnez t2, .Lcopy_bytes_checklen
227 and t0, src, ADDRMASK
230 bnez t1, .Ldst_unaligned
232 bnez t0, .Lsrc_unaligned_dst_aligned
234 * use delay slot for fall-through
235 * src and dst are aligned; need to compute rem
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)
246 EXC( LOAD t0, UNIT(0)(src), .Ll_exc)
247 EXC( LOAD t1, UNIT(1)(src), .Ll_exc_copy)
248 EXC( LOAD t2, UNIT(2)(src), .Ll_exc_copy)
249 EXC( LOAD t3, UNIT(3)(src), .Ll_exc_copy)
250 SUB len, len, 8*NBYTES
251 EXC( LOAD t4, UNIT(4)(src), .Ll_exc_copy)
252 EXC( LOAD t7, UNIT(5)(src), .Ll_exc_copy)
253 EXC( STORE t0, UNIT(0)(dst), .Ls_exc_p8u)
254 EXC( STORE t1, UNIT(1)(dst), .Ls_exc_p7u)
255 EXC( LOAD t0, UNIT(6)(src), .Ll_exc_copy)
256 EXC( LOAD t1, UNIT(7)(src), .Ll_exc_copy)
257 ADD src, src, 8*NBYTES
258 ADD dst, dst, 8*NBYTES
259 EXC( STORE t2, UNIT(-6)(dst), .Ls_exc_p6u)
260 EXC( STORE t3, UNIT(-5)(dst), .Ls_exc_p5u)
261 EXC( STORE t4, UNIT(-4)(dst), .Ls_exc_p4u)
262 EXC( STORE t7, UNIT(-3)(dst), .Ls_exc_p3u)
263 EXC( STORE t0, UNIT(-2)(dst), .Ls_exc_p2u)
264 EXC( STORE t1, UNIT(-1)(dst), .Ls_exc_p1u)
271 * len == rem == the number of bytes left to copy < 8*NBYTES
273 .Lcleanup_both_aligned:
275 sltu t0, len, 4*NBYTES
276 bnez t0, .Lless_than_4units
277 and rem, len, (NBYTES-1) # rem = len % NBYTES
281 EXC( LOAD t0, UNIT(0)(src), .Ll_exc)
282 EXC( LOAD t1, UNIT(1)(src), .Ll_exc_copy)
283 EXC( LOAD t2, UNIT(2)(src), .Ll_exc_copy)
284 EXC( LOAD t3, UNIT(3)(src), .Ll_exc_copy)
285 SUB len, len, 4*NBYTES
286 ADD src, src, 4*NBYTES
288 EXC( STORE t0, UNIT(0)(dst), .Ls_exc_p4u)
289 EXC( STORE t1, UNIT(1)(dst), .Ls_exc_p3u)
290 EXC( STORE t2, UNIT(2)(dst), .Ls_exc_p2u)
291 EXC( STORE t3, UNIT(3)(dst), .Ls_exc_p1u)
292 .set reorder /* DADDI_WAR */
293 ADD dst, dst, 4*NBYTES
300 beq rem, len, .Lcopy_bytes
304 EXC( LOAD t0, 0(src), .Ll_exc)
307 EXC( STORE t0, 0(dst), .Ls_exc_p1u)
308 .set reorder /* DADDI_WAR */
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.
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.
326 ADD t1, dst, len # t1 is just past last byte of dst
328 SLL rem, len, 3 # rem = number of bits to keep
329 EXC( LOAD t0, 0(src), .Ll_exc)
330 SUB bits, bits, rem # bits = number of bits to discard
331 SHIFT_DISCARD t0, t0, bits
332 EXC( STREST t0, -1(t1), .Ls_exc)
338 * t0 = src & ADDRMASK
339 * t1 = dst & ADDRMASK; T1 > 0
342 * Copy enough bytes to align dst
343 * Set match = (src and dst have same alignment)
346 EXC( LDFIRST t3, FIRST(0)(src), .Ll_exc)
348 EXC( LDREST t3, REST(0)(src), .Ll_exc_copy)
349 SUB t2, t2, t1 # t2 = number of bytes copied
352 EXC( STFIRST t3, FIRST(0)(dst), .Ls_exc)
356 beqz match, .Lboth_aligned
359 .Lsrc_unaligned_dst_aligned:
360 SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter
362 beqz t0, .Lcleanup_src_unaligned
363 and rem, len, (4*NBYTES-1) # rem = len % 4*NBYTES
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).
373 EXC( LDFIRST t0, FIRST(0)(src), .Ll_exc)
374 EXC( LDFIRST t1, FIRST(1)(src), .Ll_exc_copy)
375 SUB len, len, 4*NBYTES
376 EXC( LDREST t0, REST(0)(src), .Ll_exc_copy)
377 EXC( LDREST t1, REST(1)(src), .Ll_exc_copy)
378 EXC( LDFIRST t2, FIRST(2)(src), .Ll_exc_copy)
379 EXC( LDFIRST t3, FIRST(3)(src), .Ll_exc_copy)
380 EXC( LDREST t2, REST(2)(src), .Ll_exc_copy)
381 EXC( 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
387 EXC( STORE t0, UNIT(0)(dst), .Ls_exc_p4u)
388 EXC( STORE t1, UNIT(1)(dst), .Ls_exc_p3u)
389 EXC( STORE t2, UNIT(2)(dst), .Ls_exc_p2u)
390 EXC( 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
397 .Lcleanup_src_unaligned:
399 and rem, len, NBYTES-1 # rem = len % NBYTES
400 beq rem, len, .Lcopy_bytes
404 EXC( LDFIRST t0, FIRST(0)(src), .Ll_exc)
405 EXC( LDREST t0, REST(0)(src), .Ll_exc_copy)
408 EXC( STORE t0, 0(dst), .Ls_exc_p1u)
409 .set reorder /* DADDI_WAR */
414 .Lcopy_bytes_checklen:
418 /* 0 < len < NBYTES */
420 #define COPY_BYTE(N) \
421 EXC( lb t0, N(src), .Ll_exc); \
424 EXC( sb t0, N(dst), .Ls_exc_p1)
434 EXC( lb t0, NBYTES-2(src), .Ll_exc)
437 EXC( sb t0, NBYTES-2(dst), .Ls_exc_p1)
445 * Copy bytes from src until faulting load address (or until a
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.
452 * Assumes src < THREAD_BUADDR($28)
454 LOAD t0, TI_TASK($28)
456 LOAD t0, THREAD_BUADDR(t0)
458 EXC( lb t1, 0(src), .Ll_exc)
460 sb t1, 0(dst) # can't fault -- we're copy_from_user
461 .set reorder /* DADDI_WAR */
466 LOAD t0, TI_TASK($28)
468 LOAD t0, THREAD_BUADDR(t0) # t0 is just past last good address
470 SUB len, AT, t0 # len number of uncopied bytes
471 bnez t6, .Ldone /* Skip the zeroing part if inatomic */
473 * Here's where we rely on src and dst being incremented in tandem,
475 * dst += (fault addr - src) to put dst at first byte to clear
477 ADD dst, t0 # compute start address in a1
480 * Clear len bytes starting at dst. Can't call __bzero because it
481 * might modify len. An inefficient loop for these rare times...
483 .set reorder /* DADDI_WAR */
489 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
505 .set reorder; /* DADDI_WAR */ \
506 .Ls_exc_p ## n ## u: \
507 ADD len, len, n*NBYTES; \
521 .set reorder /* DADDI_WAR */
533 sltu t0, a1, t0 # dst + len <= src -> memcpy
534 sltu t1, a0, t1 # dst >= src + len -> memcpy
537 move v0, a0 /* return value */
541 /* fall through to __rmemcpy */
542 LEAF(__rmemcpy) /* a0=dst a1=src a2=len */
544 beqz t0, .Lr_end_bytes_up # src >= dst
546 ADD a0, a2 # dst = dst + len
547 ADD a1, a2 # src = src + len
555 .set reorder /* DADDI_WAR */
557 bnez a2, .Lr_end_bytes
570 .set reorder /* DADDI_WAR */
572 bnez a2, .Lr_end_bytes_up