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 #if !defined(CONFIG_DMA_COHERENT) || !defined(CONFIG_DMA_IP27)
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 * 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.
192 LEAF(memcpy) /* a0=dst a1=src a2=len */
193 move v0, dst /* return value */
197 * Note: dst & src may be unaligned, len may be 0
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.
208 * If len < NBYTES use byte operations.
213 and t1, dst, ADDRMASK
216 bnez t2, .Lcopy_bytes_checklen
217 and t0, src, ADDRMASK
220 bnez t1, .Ldst_unaligned
222 bnez t0, .Lsrc_unaligned_dst_aligned
224 * use delay slot for fall-through
225 * src and dst are aligned; need to compute rem
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)
236 EXC( LOAD t0, UNIT(0)(src), .Ll_exc)
237 EXC( LOAD t1, UNIT(1)(src), .Ll_exc_copy)
238 EXC( LOAD t2, UNIT(2)(src), .Ll_exc_copy)
239 EXC( LOAD t3, UNIT(3)(src), .Ll_exc_copy)
240 SUB len, len, 8*NBYTES
241 EXC( LOAD t4, UNIT(4)(src), .Ll_exc_copy)
242 EXC( LOAD t7, UNIT(5)(src), .Ll_exc_copy)
243 EXC( STORE t0, UNIT(0)(dst), .Ls_exc_p8u)
244 EXC( STORE t1, UNIT(1)(dst), .Ls_exc_p7u)
245 EXC( LOAD t0, UNIT(6)(src), .Ll_exc_copy)
246 EXC( LOAD t1, UNIT(7)(src), .Ll_exc_copy)
247 ADD src, src, 8*NBYTES
248 ADD dst, dst, 8*NBYTES
249 EXC( STORE t2, UNIT(-6)(dst), .Ls_exc_p6u)
250 EXC( STORE t3, UNIT(-5)(dst), .Ls_exc_p5u)
251 EXC( STORE t4, UNIT(-4)(dst), .Ls_exc_p4u)
252 EXC( STORE t7, UNIT(-3)(dst), .Ls_exc_p3u)
253 EXC( STORE t0, UNIT(-2)(dst), .Ls_exc_p2u)
254 EXC( STORE t1, UNIT(-1)(dst), .Ls_exc_p1u)
261 * len == rem == the number of bytes left to copy < 8*NBYTES
263 .Lcleanup_both_aligned:
265 sltu t0, len, 4*NBYTES
266 bnez t0, .Lless_than_4units
267 and rem, len, (NBYTES-1) # rem = len % NBYTES
271 EXC( LOAD t0, UNIT(0)(src), .Ll_exc)
272 EXC( LOAD t1, UNIT(1)(src), .Ll_exc_copy)
273 EXC( LOAD t2, UNIT(2)(src), .Ll_exc_copy)
274 EXC( LOAD t3, UNIT(3)(src), .Ll_exc_copy)
275 SUB len, len, 4*NBYTES
276 ADD src, src, 4*NBYTES
278 EXC( STORE t0, UNIT(0)(dst), .Ls_exc_p4u)
279 EXC( STORE t1, UNIT(1)(dst), .Ls_exc_p3u)
280 EXC( STORE t2, UNIT(2)(dst), .Ls_exc_p2u)
281 EXC( STORE t3, UNIT(3)(dst), .Ls_exc_p1u)
282 .set reorder /* DADDI_WAR */
283 ADD dst, dst, 4*NBYTES
290 beq rem, len, .Lcopy_bytes
294 EXC( LOAD t0, 0(src), .Ll_exc)
297 EXC( STORE t0, 0(dst), .Ls_exc_p1u)
298 .set reorder /* DADDI_WAR */
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.
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.
316 ADD t1, dst, len # t1 is just past last byte of dst
318 SLL rem, len, 3 # rem = number of bits to keep
319 EXC( LOAD t0, 0(src), .Ll_exc)
320 SUB bits, bits, rem # bits = number of bits to discard
321 SHIFT_DISCARD t0, t0, bits
322 EXC( STREST t0, -1(t1), .Ls_exc)
328 * t0 = src & ADDRMASK
329 * t1 = dst & ADDRMASK; T1 > 0
332 * Copy enough bytes to align dst
333 * Set match = (src and dst have same alignment)
336 EXC( LDFIRST t3, FIRST(0)(src), .Ll_exc)
338 EXC( LDREST t3, REST(0)(src), .Ll_exc_copy)
339 SUB t2, t2, t1 # t2 = number of bytes copied
342 EXC( STFIRST t3, FIRST(0)(dst), .Ls_exc)
346 beqz match, .Lboth_aligned
349 .Lsrc_unaligned_dst_aligned:
350 SRL t0, len, LOG_NBYTES+2 # +2 for 4 units/iter
352 beqz t0, .Lcleanup_src_unaligned
353 and rem, len, (4*NBYTES-1) # rem = len % 4*NBYTES
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).
363 EXC( LDFIRST t0, FIRST(0)(src), .Ll_exc)
364 EXC( LDFIRST t1, FIRST(1)(src), .Ll_exc_copy)
365 SUB len, len, 4*NBYTES
366 EXC( LDREST t0, REST(0)(src), .Ll_exc_copy)
367 EXC( LDREST t1, REST(1)(src), .Ll_exc_copy)
368 EXC( LDFIRST t2, FIRST(2)(src), .Ll_exc_copy)
369 EXC( LDFIRST t3, FIRST(3)(src), .Ll_exc_copy)
370 EXC( LDREST t2, REST(2)(src), .Ll_exc_copy)
371 EXC( 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
377 EXC( STORE t0, UNIT(0)(dst), .Ls_exc_p4u)
378 EXC( STORE t1, UNIT(1)(dst), .Ls_exc_p3u)
379 EXC( STORE t2, UNIT(2)(dst), .Ls_exc_p2u)
380 EXC( 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
387 .Lcleanup_src_unaligned:
389 and rem, len, NBYTES-1 # rem = len % NBYTES
390 beq rem, len, .Lcopy_bytes
394 EXC( LDFIRST t0, FIRST(0)(src), .Ll_exc)
395 EXC( LDREST t0, REST(0)(src), .Ll_exc_copy)
398 EXC( STORE t0, 0(dst), .Ls_exc_p1u)
399 .set reorder /* DADDI_WAR */
404 .Lcopy_bytes_checklen:
408 /* 0 < len < NBYTES */
410 #define COPY_BYTE(N) \
411 EXC( lb t0, N(src), .Ll_exc); \
414 EXC( sb t0, N(dst), .Ls_exc_p1)
424 EXC( lb t0, NBYTES-2(src), .Ll_exc)
427 EXC( sb t0, NBYTES-2(dst), .Ls_exc_p1)
435 * Copy bytes from src until faulting load address (or until a
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.
442 * Assumes src < THREAD_BUADDR($28)
444 LOAD t0, TI_TASK($28)
446 LOAD t0, THREAD_BUADDR(t0)
448 EXC( lb t1, 0(src), .Ll_exc)
450 sb t1, 0(dst) # can't fault -- we're copy_from_user
451 .set reorder /* DADDI_WAR */
456 LOAD t0, TI_TASK($28)
458 LOAD t0, THREAD_BUADDR(t0) # t0 is just past last good address
460 SUB len, AT, t0 # len number of uncopied bytes
462 * Here's where we rely on src and dst being incremented in tandem,
464 * dst += (fault addr - src) to put dst at first byte to clear
466 ADD dst, t0 # compute start address in a1
469 * Clear len bytes starting at dst. Can't call __bzero because it
470 * might modify len. An inefficient loop for these rare times...
472 .set reorder /* DADDI_WAR */
478 #ifndef CONFIG_CPU_DADDI_WORKAROUNDS
494 .set reorder; /* DADDI_WAR */ \
495 .Ls_exc_p ## n ## u: \
496 ADD len, len, n*NBYTES; \
510 .set reorder /* DADDI_WAR */
522 sltu t0, a1, t0 # dst + len <= src -> memcpy
523 sltu t1, a0, t1 # dst >= src + len -> memcpy
526 move v0, a0 /* return value */
530 /* fall through to __rmemcpy */
531 LEAF(__rmemcpy) /* a0=dst a1=src a2=len */
533 beqz t0, .Lr_end_bytes_up # src >= dst
535 ADD a0, a2 # dst = dst + len
536 ADD a1, a2 # src = src + len
544 .set reorder /* DADDI_WAR */
546 bnez a2, .Lr_end_bytes
559 .set reorder /* DADDI_WAR */
561 bnez a2, .Lr_end_bytes_up