diff options
author | Linus Torvalds <torvalds@ppc970.osdl.org> | 2005-04-16 15:20:36 -0700 |
---|---|---|
committer | Linus Torvalds <torvalds@ppc970.osdl.org> | 2005-04-16 15:20:36 -0700 |
commit | 1da177e4c3f41524e886b7f1b8a0c1fc7321cac2 (patch) | |
tree | 0bba044c4ce775e45a88a51686b5d9f90697ea9d /arch/cris/arch-v10/lib | |
download | op-kernel-dev-1da177e4c3f41524e886b7f1b8a0c1fc7321cac2.zip op-kernel-dev-1da177e4c3f41524e886b7f1b8a0c1fc7321cac2.tar.gz |
Linux-2.6.12-rc2v2.6.12-rc2
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!
Diffstat (limited to 'arch/cris/arch-v10/lib')
-rw-r--r-- | arch/cris/arch-v10/lib/Makefile | 9 | ||||
-rw-r--r-- | arch/cris/arch-v10/lib/checksum.S | 124 | ||||
-rw-r--r-- | arch/cris/arch-v10/lib/checksumcopy.S | 132 | ||||
-rw-r--r-- | arch/cris/arch-v10/lib/csumcpfruser.S | 64 | ||||
-rw-r--r-- | arch/cris/arch-v10/lib/dmacopy.c | 43 | ||||
-rw-r--r-- | arch/cris/arch-v10/lib/dram_init.S | 205 | ||||
-rw-r--r-- | arch/cris/arch-v10/lib/hw_settings.S | 62 | ||||
-rw-r--r-- | arch/cris/arch-v10/lib/memset.c | 252 | ||||
-rw-r--r-- | arch/cris/arch-v10/lib/old_checksum.c | 85 | ||||
-rw-r--r-- | arch/cris/arch-v10/lib/string.c | 225 | ||||
-rw-r--r-- | arch/cris/arch-v10/lib/usercopy.c | 523 |
11 files changed, 1724 insertions, 0 deletions
diff --git a/arch/cris/arch-v10/lib/Makefile b/arch/cris/arch-v10/lib/Makefile new file mode 100644 index 0000000..36e9a9c --- /dev/null +++ b/arch/cris/arch-v10/lib/Makefile @@ -0,0 +1,9 @@ +# +# Makefile for Etrax-specific library files.. +# + + +EXTRA_AFLAGS := -traditional + +lib-y = checksum.o checksumcopy.o string.o usercopy.o memset.o csumcpfruser.o + diff --git a/arch/cris/arch-v10/lib/checksum.S b/arch/cris/arch-v10/lib/checksum.S new file mode 100644 index 0000000..85c48f0 --- /dev/null +++ b/arch/cris/arch-v10/lib/checksum.S @@ -0,0 +1,124 @@ +/* $Id: checksum.S,v 1.1 2001/12/17 13:59:27 bjornw Exp $ + * A fast checksum routine using movem + * Copyright (c) 1998-2001 Axis Communications AB + * + * csum_partial(const unsigned char * buff, int len, unsigned int sum) + */ + + .globl csum_partial +csum_partial: + + ;; r10 - src + ;; r11 - length + ;; r12 - checksum + + ;; check for breakeven length between movem and normal word looping versions + ;; we also do _NOT_ want to compute a checksum over more than the + ;; actual length when length < 40 + + cmpu.w 80,$r11 + blo _word_loop + nop + + ;; need to save the registers we use below in the movem loop + ;; this overhead is why we have a check above for breakeven length + ;; only r0 - r8 have to be saved, the other ones are clobber-able + ;; according to the ABI + + subq 9*4,$sp + movem $r8,[$sp] + + ;; do a movem checksum + + subq 10*4,$r11 ; update length for the first loop + +_mloop: movem [$r10+],$r9 ; read 10 longwords + + ;; perform dword checksumming on the 10 longwords + + add.d $r0,$r12 + ax + add.d $r1,$r12 + ax + add.d $r2,$r12 + ax + add.d $r3,$r12 + ax + add.d $r4,$r12 + ax + add.d $r5,$r12 + ax + add.d $r6,$r12 + ax + add.d $r7,$r12 + ax + add.d $r8,$r12 + ax + add.d $r9,$r12 + + ;; fold the carry into the checksum, to avoid having to loop the carry + ;; back into the top + + ax + addq 0,$r12 + ax ; do it again, since we might have generated a carry + addq 0,$r12 + + subq 10*4,$r11 + bge _mloop + nop + + addq 10*4,$r11 ; compensate for last loop underflowing length + + movem [$sp+],$r8 ; restore regs + +_word_loop: + ;; only fold if there is anything to fold. + + cmpq 0,$r12 + beq _no_fold + + ;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below. + ;; r9 and r13 can be used as temporaries. + + moveq -1,$r9 ; put 0xffff in r9, faster than move.d 0xffff,r9 + lsrq 16,$r9 + + move.d $r12,$r13 + lsrq 16,$r13 ; r13 = checksum >> 16 + and.d $r9,$r12 ; checksum = checksum & 0xffff + add.d $r13,$r12 ; checksum += r13 + move.d $r12,$r13 ; do the same again, maybe we got a carry last add + lsrq 16,$r13 + and.d $r9,$r12 + add.d $r13,$r12 + +_no_fold: + cmpq 2,$r11 + blt _no_words + nop + + ;; checksum the rest of the words + + subq 2,$r11 + +_wloop: subq 2,$r11 + bge _wloop + addu.w [$r10+],$r12 + + addq 2,$r11 + +_no_words: + ;; see if we have one odd byte more + cmpq 1,$r11 + beq _do_byte + nop + ret + move.d $r12, $r10 + +_do_byte: + ;; copy and checksum the last byte + addu.b [$r10],$r12 + ret + move.d $r12, $r10 + diff --git a/arch/cris/arch-v10/lib/checksumcopy.S b/arch/cris/arch-v10/lib/checksumcopy.S new file mode 100644 index 0000000..35cbffb --- /dev/null +++ b/arch/cris/arch-v10/lib/checksumcopy.S @@ -0,0 +1,132 @@ +/* $Id: checksumcopy.S,v 1.1 2001/12/17 13:59:27 bjornw Exp $ + * A fast checksum+copy routine using movem + * Copyright (c) 1998, 2001 Axis Communications AB + * + * Authors: Bjorn Wesen + * + * csum_partial_copy_nocheck(const char *src, char *dst, + * int len, unsigned int sum) + */ + + .globl csum_partial_copy_nocheck +csum_partial_copy_nocheck: + + ;; r10 - src + ;; r11 - dst + ;; r12 - length + ;; r13 - checksum + + ;; check for breakeven length between movem and normal word looping versions + ;; we also do _NOT_ want to compute a checksum over more than the + ;; actual length when length < 40 + + cmpu.w 80, $r12 + blo _word_loop + nop + + ;; need to save the registers we use below in the movem loop + ;; this overhead is why we have a check above for breakeven length + ;; only r0 - r8 have to be saved, the other ones are clobber-able + ;; according to the ABI + + subq 9*4, $sp + movem $r8, [$sp] + + ;; do a movem copy and checksum + + subq 10*4, $r12 ; update length for the first loop + +_mloop: movem [$r10+],$r9 ; read 10 longwords +1: ;; A failing userspace access will have this as PC. + movem $r9,[$r11+] ; write 10 longwords + + ;; perform dword checksumming on the 10 longwords + + add.d $r0,$r13 + ax + add.d $r1,$r13 + ax + add.d $r2,$r13 + ax + add.d $r3,$r13 + ax + add.d $r4,$r13 + ax + add.d $r5,$r13 + ax + add.d $r6,$r13 + ax + add.d $r7,$r13 + ax + add.d $r8,$r13 + ax + add.d $r9,$r13 + + ;; fold the carry into the checksum, to avoid having to loop the carry + ;; back into the top + + ax + addq 0,$r13 + ax ; do it again, since we might have generated a carry + addq 0,$r13 + + subq 10*4,$r12 + bge _mloop + nop + + addq 10*4,$r12 ; compensate for last loop underflowing length + + movem [$sp+],$r8 ; restore regs + +_word_loop: + ;; only fold if there is anything to fold. + + cmpq 0,$r13 + beq _no_fold + + ;; fold 32-bit checksum into a 16-bit checksum, to avoid carries below + ;; r9 can be used as temporary. + + move.d $r13,$r9 + lsrq 16,$r9 ; r0 = checksum >> 16 + and.d 0xffff,$r13 ; checksum = checksum & 0xffff + add.d $r9,$r13 ; checksum += r0 + move.d $r13,$r9 ; do the same again, maybe we got a carry last add + lsrq 16,$r9 + and.d 0xffff,$r13 + add.d $r9,$r13 + +_no_fold: + cmpq 2,$r12 + blt _no_words + nop + + ;; copy and checksum the rest of the words + + subq 2,$r12 + +_wloop: move.w [$r10+],$r9 +2: ;; A failing userspace access will have this as PC. + addu.w $r9,$r13 + subq 2,$r12 + bge _wloop + move.w $r9,[$r11+] + + addq 2,$r12 + +_no_words: + ;; see if we have one odd byte more + cmpq 1,$r12 + beq _do_byte + nop + ret + move.d $r13, $r10 + +_do_byte: + ;; copy and checksum the last byte + move.b [$r10],$r9 +3: ;; A failing userspace access will have this as PC. + addu.b $r9,$r13 + move.b $r9,[$r11] + ret + move.d $r13, $r10 diff --git a/arch/cris/arch-v10/lib/csumcpfruser.S b/arch/cris/arch-v10/lib/csumcpfruser.S new file mode 100644 index 0000000..5f41ccd --- /dev/null +++ b/arch/cris/arch-v10/lib/csumcpfruser.S @@ -0,0 +1,64 @@ +/* + * Add-on to transform csum_partial_copy_nocheck in checksumcopy.S into + * csum_partial_copy_from_user by adding exception records. + * + * Copyright (C) 2001 Axis Communications AB. + * + * Author: Hans-Peter Nilsson. + */ + +#include <asm/errno.h> + +/* Same function body, but a different name. If we just added exception + records to _csum_partial_copy_nocheck and made it generic, we wouldn't + know a user fault from a kernel fault and we would have overhead in + each kernel caller for the error-pointer argument. + + unsigned int csum_partial_copy_from_user + (const char *src, char *dst, int len, unsigned int sum, int *errptr); + + Note that the errptr argument is only set if we encounter an error. + It is conveniently located on the stack, so the normal function body + does not have to handle it. */ + +#define csum_partial_copy_nocheck csum_partial_copy_from_user + +/* There are local labels numbered 1, 2 and 3 present to mark the + different from-user accesses. */ +#include "checksumcopy.S" + + .section .fixup,"ax" + +;; Here from the movem loop; restore stack. +4: + movem [$sp+],$r8 +;; r12 is already decremented. Add back chunk_size-2. + addq 40-2,$r12 + +;; Here from the word loop; r12 is off by 2; add it back. +5: + addq 2,$r12 + +;; Here from a failing single byte. +6: + +;; Signal in *errptr that we had a failing access. + moveq -EFAULT,$r9 + move.d $r9,[[$sp]] + +;; Clear the rest of the destination area using memset. Preserve the +;; checksum for the readable bytes. + push $srp + push $r13 + move.d $r11,$r10 + clear.d $r11 + jsr memset + pop $r10 + jump [$sp+] + + .previous + .section __ex_table,"a" + .dword 1b,4b + .dword 2b,5b + .dword 3b,6b + .previous diff --git a/arch/cris/arch-v10/lib/dmacopy.c b/arch/cris/arch-v10/lib/dmacopy.c new file mode 100644 index 0000000..e5fb44f --- /dev/null +++ b/arch/cris/arch-v10/lib/dmacopy.c @@ -0,0 +1,43 @@ +/* $Id: dmacopy.c,v 1.1 2001/12/17 13:59:27 bjornw Exp $ + * + * memcpy for large blocks, using memory-memory DMA channels 6 and 7 in Etrax + */ + +#include <asm/svinto.h> +#include <asm/io.h> + +#define D(x) + +void *dma_memcpy(void *pdst, + const void *psrc, + unsigned int pn) +{ + static etrax_dma_descr indma, outdma; + + D(printk("dma_memcpy %d bytes... ", pn)); + +#if 0 + *R_GEN_CONFIG = genconfig_shadow = + (genconfig_shadow & ~0x3c0000) | + IO_STATE(R_GEN_CONFIG, dma6, intdma7) | + IO_STATE(R_GEN_CONFIG, dma7, intdma6); +#endif + indma.sw_len = outdma.sw_len = pn; + indma.ctrl = d_eol | d_eop; + outdma.ctrl = d_eol; + indma.buf = psrc; + outdma.buf = pdst; + + *R_DMA_CH6_FIRST = &indma; + *R_DMA_CH7_FIRST = &outdma; + *R_DMA_CH6_CMD = IO_STATE(R_DMA_CH6_CMD, cmd, start); + *R_DMA_CH7_CMD = IO_STATE(R_DMA_CH7_CMD, cmd, start); + + while(*R_DMA_CH7_CMD == 1) /* wait for completion */ ; + + D(printk("done\n")); + +} + + + diff --git a/arch/cris/arch-v10/lib/dram_init.S b/arch/cris/arch-v10/lib/dram_init.S new file mode 100644 index 0000000..2ef4ad57 --- /dev/null +++ b/arch/cris/arch-v10/lib/dram_init.S @@ -0,0 +1,205 @@ +/* $Id: dram_init.S,v 1.4 2003/09/22 09:21:59 starvik Exp $ + * + * DRAM/SDRAM initialization - alter with care + * This file is intended to be included from other assembler files + * + * Note: This file may not modify r9 because r9 is used to carry + * information from the decompresser to the kernel + * + * Copyright (C) 2000, 2001 Axis Communications AB + * + * Authors: Mikael Starvik (starvik@axis.com) + * + * $Log: dram_init.S,v $ + * Revision 1.4 2003/09/22 09:21:59 starvik + * Decompresser is linked to 0x407xxxxx and sdram commands are at 0x000xxxxx + * so we need to mask off 12 bits. + * + * Revision 1.3 2003/03/31 09:38:37 starvik + * Corrected calculation of end of sdram init commands + * + * Revision 1.2 2002/11/19 13:33:29 starvik + * Changes from Linux 2.4 + * + * Revision 1.13 2002/10/30 07:42:28 starvik + * Always read SDRAM command sequence from flash + * + * Revision 1.12 2002/08/09 11:37:37 orjanf + * Added double initialization work-around for Samsung SDRAMs. + * + * Revision 1.11 2002/06/04 11:43:21 starvik + * Check if mrs_data is specified in kernelconfig (necessary for MCM) + * + * Revision 1.10 2001/10/04 12:00:21 martinnn + * Added missing underscores. + * + * Revision 1.9 2001/10/01 14:47:35 bjornw + * Added register prefixes and removed underscores + * + * Revision 1.8 2001/05/15 07:12:45 hp + * Copy warning from head.S about r8 and r9 + * + * Revision 1.7 2001/04/18 12:05:39 bjornw + * Fixed comments, and explicitely include config.h to be sure its there + * + * Revision 1.6 2001/04/10 06:20:16 starvik + * Delay should be 200us, not 200ns + * + * Revision 1.5 2001/04/09 06:01:13 starvik + * Added support for 100 MHz SDRAMs + * + * Revision 1.4 2001/03/26 14:24:01 bjornw + * Namechange of some config options + * + * Revision 1.3 2001/03/23 08:29:41 starvik + * Corrected calculation of mrs_data + * + * Revision 1.2 2001/02/08 15:20:00 starvik + * Corrected SDRAM initialization + * Should now be included as inline + * + * Revision 1.1 2001/01/29 13:08:02 starvik + * Initial version + * This file should be included from all assembler files that needs to + * initialize DRAM/SDRAM. + * + */ + +/* Just to be certain the config file is included, we include it here + * explicitely instead of depending on it being included in the file that + * uses this code. + */ + +#include <linux/config.h> + + ;; WARNING! The registers r8 and r9 are used as parameters carrying + ;; information from the decompressor (if the kernel was compressed). + ;; They should not be used in the code below. + +#ifndef CONFIG_SVINTO_SIM + move.d CONFIG_ETRAX_DEF_R_WAITSTATES, $r0 + move.d $r0, [R_WAITSTATES] + + move.d CONFIG_ETRAX_DEF_R_BUS_CONFIG, $r0 + move.d $r0, [R_BUS_CONFIG] + +#ifndef CONFIG_ETRAX_SDRAM + move.d CONFIG_ETRAX_DEF_R_DRAM_CONFIG, $r0 + move.d $r0, [R_DRAM_CONFIG] + + move.d CONFIG_ETRAX_DEF_R_DRAM_TIMING, $r0 + move.d $r0, [R_DRAM_TIMING] +#else + ;; Samsung SDRAMs seem to require to be initialized twice to work properly. + moveq 2, $r6 +_sdram_init: + + ; Refer to ETRAX 100LX Designers Reference for a description of SDRAM initialization + + ; Bank configuration + move.d CONFIG_ETRAX_DEF_R_SDRAM_CONFIG, $r0 + move.d $r0, [R_SDRAM_CONFIG] + + ; Calculate value of mrs_data + ; CAS latency = 2 && bus_width = 32 => 0x40 + ; CAS latency = 3 && bus_width = 32 => 0x60 + ; CAS latency = 2 && bus_width = 16 => 0x20 + ; CAS latency = 3 && bus_width = 16 => 0x30 + + ; Check if value is already supplied in kernel config + move.d CONFIG_ETRAX_DEF_R_SDRAM_TIMING, $r2 + and.d 0x00ff0000, $r2 + bne _set_timing + lsrq 16, $r2 + + move.d 0x40, $r2 ; Assume 32 bits and CAS latency = 2 + move.d CONFIG_ETRAX_DEF_R_SDRAM_TIMING, $r1 + move.d $r1, $r3 + and.d 0x03, $r1 ; Get CAS latency + and.d 0x1000, $r3 ; 50 or 100 MHz? + beq _speed_50 + nop +_speed_100: + cmp.d 0x00, $r1 ; CAS latency = 2? + beq _bw_check + nop + or.d 0x20, $r2 ; CAS latency = 3 + ba _bw_check + nop +_speed_50: + cmp.d 0x01, $r1 ; CAS latency = 2? + beq _bw_check + nop + or.d 0x20, $r2 ; CAS latency = 3 +_bw_check: + move.d CONFIG_ETRAX_DEF_R_SDRAM_CONFIG, $r1 + and.d 0x800000, $r1 ; DRAM width is bit 23 + bne _set_timing + nop + lsrq 1, $r2 ; 16 bits. Shift down value. + + ; Set timing parameters. Starts master clock +_set_timing: + move.d CONFIG_ETRAX_DEF_R_SDRAM_TIMING, $r1 + and.d 0x8000f9ff, $r1 ; Make sure mrs data and command is 0 + or.d 0x80000000, $r1 ; Make sure sdram enable bit is set + move.d $r1, $r5 + or.d 0x0000c000, $r1 ; ref = disable + lslq 16, $r2 ; mrs data starts at bit 16 + or.d $r2, $r1 + move.d $r1, [R_SDRAM_TIMING] + + ; Wait 200us + move.d 10000, $r2 +1: bne 1b + subq 1, $r2 + + ; Issue initialization command sequence + move.d _sdram_commands_start, $r2 + and.d 0x000fffff, $r2 ; Make sure commands are read from flash + move.d _sdram_commands_end, $r3 + and.d 0x000fffff, $r3 +1: clear.d $r4 + move.b [$r2+], $r4 + lslq 9, $r4 ; Command starts at bit 9 + or.d $r1, $r4 + move.d $r4, [R_SDRAM_TIMING] + nop ; Wait five nop cycles between each command + nop + nop + nop + nop + cmp.d $r2, $r3 + bne 1b + nop + move.d $r5, [R_SDRAM_TIMING] + subq 1, $r6 + bne _sdram_init + nop + ba _sdram_commands_end + nop + +_sdram_commands_start: + .byte 3 ; Precharge + .byte 0 ; nop + .byte 2 ; refresh + .byte 0 ; nop + .byte 2 ; refresh + .byte 0 ; nop + .byte 2 ; refresh + .byte 0 ; nop + .byte 2 ; refresh + .byte 0 ; nop + .byte 2 ; refresh + .byte 0 ; nop + .byte 2 ; refresh + .byte 0 ; nop + .byte 2 ; refresh + .byte 0 ; nop + .byte 2 ; refresh + .byte 0 ; nop + .byte 1 ; mrs + .byte 0 ; nop +_sdram_commands_end: +#endif +#endif diff --git a/arch/cris/arch-v10/lib/hw_settings.S b/arch/cris/arch-v10/lib/hw_settings.S new file mode 100644 index 0000000..56905aa --- /dev/null +++ b/arch/cris/arch-v10/lib/hw_settings.S @@ -0,0 +1,62 @@ +/* + * $Id: hw_settings.S,v 1.1 2001/12/17 13:59:27 bjornw Exp $ + * + * This table is used by some tools to extract hardware parameters. + * The table should be included in the kernel and the decompressor. + * Don't forget to update the tools if you change this table. + * + * Copyright (C) 2001 Axis Communications AB + * + * Authors: Mikael Starvik (starvik@axis.com) + */ + +#define PA_SET_VALUE ((CONFIG_ETRAX_DEF_R_PORT_PA_DIR << 8) | \ + (CONFIG_ETRAX_DEF_R_PORT_PA_DATA)) +#define PB_SET_VALUE ((CONFIG_ETRAX_DEF_R_PORT_PB_CONFIG << 16) | \ + (CONFIG_ETRAX_DEF_R_PORT_PB_DIR << 8) | \ + (CONFIG_ETRAX_DEF_R_PORT_PB_DATA)) + + .ascii "HW_PARAM_MAGIC" ; Magic number + .dword 0xc0004000 ; Kernel start address + + ; Debug port +#ifdef CONFIG_ETRAX_DEBUG_PORT0 + .dword 0 +#elif defined(CONFIG_ETRAX_DEBUG_PORT1) + .dword 1 +#elif defined(CONFIG_ETRAX_DEBUG_PORT2) + .dword 2 +#elif defined(CONFIG_ETRAX_DEBUG_PORT3) + .dword 3 +#else + .dword 4 ; No debug +#endif + + ; SDRAM or EDO DRAM? +#ifdef CONFIG_ETRAX_SDRAM + .dword 1 +#else + .dword 0 +#endif + + ; Register values + .dword R_WAITSTATES + .dword CONFIG_ETRAX_DEF_R_WAITSTATES + .dword R_BUS_CONFIG + .dword CONFIG_ETRAX_DEF_R_BUS_CONFIG +#ifdef CONFIG_ETRAX_SDRAM + .dword R_SDRAM_CONFIG + .dword CONFIG_ETRAX_DEF_R_SDRAM_CONFIG + .dword R_SDRAM_TIMING + .dword CONFIG_ETRAX_DEF_R_SDRAM_TIMING +#else + .dword R_DRAM_CONFIG + .dword CONFIG_ETRAX_DEF_R_DRAM_CONFIG + .dword R_DRAM_TIMING + .dword CONFIG_ETRAX_DEF_R_DRAM_TIMING +#endif + .dword R_PORT_PA_SET + .dword PA_SET_VALUE + .dword R_PORT_PB_SET + .dword PB_SET_VALUE + .dword 0 ; No more register values diff --git a/arch/cris/arch-v10/lib/memset.c b/arch/cris/arch-v10/lib/memset.c new file mode 100644 index 0000000..82bb668 --- /dev/null +++ b/arch/cris/arch-v10/lib/memset.c @@ -0,0 +1,252 @@ +/*#************************************************************************#*/ +/*#-------------------------------------------------------------------------*/ +/*# */ +/*# FUNCTION NAME: memset() */ +/*# */ +/*# PARAMETERS: void* dst; Destination address. */ +/*# int c; Value of byte to write. */ +/*# int len; Number of bytes to write. */ +/*# */ +/*# RETURNS: dst. */ +/*# */ +/*# DESCRIPTION: Sets the memory dst of length len bytes to c, as standard. */ +/*# Framework taken from memcpy. This routine is */ +/*# very sensitive to compiler changes in register allocation. */ +/*# Should really be rewritten to avoid this problem. */ +/*# */ +/*#-------------------------------------------------------------------------*/ +/*# */ +/*# HISTORY */ +/*# */ +/*# DATE NAME CHANGES */ +/*# ---- ---- ------- */ +/*# 990713 HP Tired of watching this function (or */ +/*# really, the nonoptimized generic */ +/*# implementation) take up 90% of simulator */ +/*# output. Measurements needed. */ +/*# */ +/*#-------------------------------------------------------------------------*/ + +#include <linux/types.h> + +/* No, there's no macro saying 12*4, since it is "hard" to get it into + the asm in a good way. Thus better to expose the problem everywhere. + */ + +/* Assuming 1 cycle per dword written or read (ok, not really true), and + one per instruction, then 43+3*(n/48-1) <= 24+24*(n/48-1) + so n >= 45.7; n >= 0.9; we win on the first full 48-byte block to set. */ + +#define ZERO_BLOCK_SIZE (1*12*4) + +void *memset(void *pdst, + int c, + size_t plen) +{ + /* Ok. Now we want the parameters put in special registers. + Make sure the compiler is able to make something useful of this. */ + + register char *return_dst __asm__ ("r10") = pdst; + register int n __asm__ ("r12") = plen; + register int lc __asm__ ("r11") = c; + + /* Most apps use memset sanely. Only those memsetting about 3..4 + bytes or less get penalized compared to the generic implementation + - and that's not really sane use. */ + + /* Ugh. This is fragile at best. Check with newer GCC releases, if + they compile cascaded "x |= x << 8" sanely! */ + __asm__("movu.b %0,$r13\n\t" + "lslq 8,$r13\n\t" + "move.b %0,$r13\n\t" + "move.d $r13,%0\n\t" + "lslq 16,$r13\n\t" + "or.d $r13,%0" + : "=r" (lc) : "0" (lc) : "r13"); + + { + register char *dst __asm__ ("r13") = pdst; + + /* This is NONPORTABLE, but since this whole routine is */ + /* grossly nonportable that doesn't matter. */ + + if (((unsigned long) pdst & 3) != 0 + /* Oops! n=0 must be a legal call, regardless of alignment. */ + && n >= 3) + { + if ((unsigned long)dst & 1) + { + *dst = (char) lc; + n--; + dst++; + } + + if ((unsigned long)dst & 2) + { + *(short *)dst = lc; + n -= 2; + dst += 2; + } + } + + /* Now the fun part. For the threshold value of this, check the equation + above. */ + /* Decide which copying method to use. */ + if (n >= ZERO_BLOCK_SIZE) + { + /* For large copies we use 'movem' */ + + /* It is not optimal to tell the compiler about clobbering any + registers; that will move the saving/restoring of those registers + to the function prologue/epilogue, and make non-movem sizes + suboptimal. + + This method is not foolproof; it assumes that the "asm reg" + declarations at the beginning of the function really are used + here (beware: they may be moved to temporary registers). + This way, we do not have to save/move the registers around into + temporaries; we can safely use them straight away. + + If you want to check that the allocation was right; then + check the equalities in the first comment. It should say + "r13=r13, r12=r12, r11=r11" */ + __asm__ volatile (" + ;; Check that the following is true (same register names on + ;; both sides of equal sign, as in r8=r8): + ;; %0=r13, %1=r12, %4=r11 + ;; + ;; Save the registers we'll clobber in the movem process + ;; on the stack. Don't mention them to gcc, it will only be + ;; upset. + subq 11*4,$sp + movem $r10,[$sp] + + move.d $r11,$r0 + move.d $r11,$r1 + move.d $r11,$r2 + move.d $r11,$r3 + move.d $r11,$r4 + move.d $r11,$r5 + move.d $r11,$r6 + move.d $r11,$r7 + move.d $r11,$r8 + move.d $r11,$r9 + move.d $r11,$r10 + + ;; Now we've got this: + ;; r13 - dst + ;; r12 - n + + ;; Update n for the first loop + subq 12*4,$r12 +0: + subq 12*4,$r12 + bge 0b + movem $r11,[$r13+] + + addq 12*4,$r12 ;; compensate for last loop underflowing n + + ;; Restore registers from stack + movem [$sp+],$r10" + + /* Outputs */ : "=r" (dst), "=r" (n) + /* Inputs */ : "0" (dst), "1" (n), "r" (lc)); + + } + + /* Either we directly starts copying, using dword copying + in a loop, or we copy as much as possible with 'movem' + and then the last block (<44 bytes) is copied here. + This will work since 'movem' will have updated src,dst,n. */ + + while ( n >= 16 ) + { + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + n -= 16; + } + + /* A switch() is definitely the fastest although it takes a LOT of code. + * Particularly if you inline code this. + */ + switch (n) + { + case 0: + break; + case 1: + *(char*)dst = (char) lc; + break; + case 2: + *(short*)dst = (short) lc; + break; + case 3: + *((short*)dst)++ = (short) lc; + *(char*)dst = (char) lc; + break; + case 4: + *((long*)dst)++ = lc; + break; + case 5: + *((long*)dst)++ = lc; + *(char*)dst = (char) lc; + break; + case 6: + *((long*)dst)++ = lc; + *(short*)dst = (short) lc; + break; + case 7: + *((long*)dst)++ = lc; + *((short*)dst)++ = (short) lc; + *(char*)dst = (char) lc; + break; + case 8: + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + break; + case 9: + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *(char*)dst = (char) lc; + break; + case 10: + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *(short*)dst = (short) lc; + break; + case 11: + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *((short*)dst)++ = (short) lc; + *(char*)dst = (char) lc; + break; + case 12: + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + break; + case 13: + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *(char*)dst = (char) lc; + break; + case 14: + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *(short*)dst = (short) lc; + break; + case 15: + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *((long*)dst)++ = lc; + *((short*)dst)++ = (short) lc; + *(char*)dst = (char) lc; + break; + } + } + + return return_dst; /* destination pointer. */ +} /* memset() */ diff --git a/arch/cris/arch-v10/lib/old_checksum.c b/arch/cris/arch-v10/lib/old_checksum.c new file mode 100644 index 0000000..22a6f0a --- /dev/null +++ b/arch/cris/arch-v10/lib/old_checksum.c @@ -0,0 +1,85 @@ +/* $Id: old_checksum.c,v 1.3 2003/10/27 08:04:32 starvik Exp $ + * + * INET An implementation of the TCP/IP protocol suite for the LINUX + * operating system. INET is implemented using the BSD Socket + * interface as the means of communication with the user level. + * + * IP/TCP/UDP checksumming routines + * + * Authors: Jorge Cwik, <jorge@laser.satlink.net> + * Arnt Gulbrandsen, <agulbra@nvg.unit.no> + * Tom May, <ftom@netcom.com> + * Lots of code moved from tcp.c and ip.c; see those files + * for more names. + * + * This program is free software; you can redistribute it and/or + * modify it under the terms of the GNU General Public License + * as published by the Free Software Foundation; either version + * 2 of the License, or (at your option) any later version. + */ + +#include <net/checksum.h> +#include <net/module.h> + +#undef PROFILE_CHECKSUM + +#ifdef PROFILE_CHECKSUM +/* these are just for profiling the checksum code with an oscillioscope.. uh */ +#if 0 +#define BITOFF *((unsigned char *)0xb0000030) = 0xff +#define BITON *((unsigned char *)0xb0000030) = 0x0 +#endif +#include <asm/io.h> +#define CBITON LED_ACTIVE_SET(1) +#define CBITOFF LED_ACTIVE_SET(0) +#define BITOFF +#define BITON +#else +#define BITOFF +#define BITON +#define CBITOFF +#define CBITON +#endif + +/* + * computes a partial checksum, e.g. for TCP/UDP fragments + */ + +#include <asm/delay.h> + +unsigned int csum_partial(const unsigned char * buff, int len, unsigned int sum) +{ + /* + * Experiments with ethernet and slip connections show that buff + * is aligned on either a 2-byte or 4-byte boundary. + */ + const unsigned char *endMarker = buff + len; + const unsigned char *marker = endMarker - (len % 16); +#if 0 + if((int)buff & 0x3) + printk("unaligned buff %p\n", buff); + __delay(900); /* extra delay of 90 us to test performance hit */ +#endif + BITON; + while (buff < marker) { + sum += *((unsigned short *)buff)++; + sum += *((unsigned short *)buff)++; + sum += *((unsigned short *)buff)++; + sum += *((unsigned short *)buff)++; + sum += *((unsigned short *)buff)++; + sum += *((unsigned short *)buff)++; + sum += *((unsigned short *)buff)++; + sum += *((unsigned short *)buff)++; + } + marker = endMarker - (len % 2); + while(buff < marker) { + sum += *((unsigned short *)buff)++; + } + if(endMarker - buff > 0) { + sum += *buff; /* add extra byte seperately */ + } + BITOFF; + return(sum); +} + +EXPORT_SYMBOL(csum_partial); diff --git a/arch/cris/arch-v10/lib/string.c b/arch/cris/arch-v10/lib/string.c new file mode 100644 index 0000000..8ffde49 --- /dev/null +++ b/arch/cris/arch-v10/lib/string.c @@ -0,0 +1,225 @@ +/*#************************************************************************#*/ +/*#-------------------------------------------------------------------------*/ +/*# */ +/*# FUNCTION NAME: memcpy() */ +/*# */ +/*# PARAMETERS: void* dst; Destination address. */ +/*# void* src; Source address. */ +/*# int len; Number of bytes to copy. */ +/*# */ +/*# RETURNS: dst. */ +/*# */ +/*# DESCRIPTION: Copies len bytes of memory from src to dst. No guarantees */ +/*# about copying of overlapping memory areas. This routine is */ +/*# very sensitive to compiler changes in register allocation. */ +/*# Should really be rewritten to avoid this problem. */ +/*# */ +/*#-------------------------------------------------------------------------*/ +/*# */ +/*# HISTORY */ +/*# */ +/*# DATE NAME CHANGES */ +/*# ---- ---- ------- */ +/*# 941007 Kenny R Creation */ +/*# 941011 Kenny R Lots of optimizations and inlining. */ +/*# 941129 Ulf A Adapted for use in libc. */ +/*# 950216 HP N==0 forgotten if non-aligned src/dst. */ +/*# Added some optimizations. */ +/*# 001025 HP Make src and dst char *. Align dst to */ +/*# dword, not just word-if-both-src-and-dst- */ +/*# are-misaligned. */ +/*# */ +/*#-------------------------------------------------------------------------*/ + +#include <linux/types.h> + +void *memcpy(void *pdst, + const void *psrc, + size_t pn) +{ + /* Ok. Now we want the parameters put in special registers. + Make sure the compiler is able to make something useful of this. + As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop). + + If gcc was allright, it really would need no temporaries, and no + stack space to save stuff on. */ + + register void *return_dst __asm__ ("r10") = pdst; + register char *dst __asm__ ("r13") = pdst; + register const char *src __asm__ ("r11") = psrc; + register int n __asm__ ("r12") = pn; + + + /* When src is aligned but not dst, this makes a few extra needless + cycles. I believe it would take as many to check that the + re-alignment was unnecessary. */ + if (((unsigned long) dst & 3) != 0 + /* Don't align if we wouldn't copy more than a few bytes; so we + don't have to check further for overflows. */ + && n >= 3) + { + if ((unsigned long) dst & 1) + { + n--; + *(char*)dst = *(char*)src; + src++; + dst++; + } + + if ((unsigned long) dst & 2) + { + n -= 2; + *(short*)dst = *(short*)src; + src += 2; + dst += 2; + } + } + + /* Decide which copying method to use. */ + if (n >= 44*2) /* Break even between movem and + move16 is at 38.7*2, but modulo 44. */ + { + /* For large copies we use 'movem' */ + + /* It is not optimal to tell the compiler about clobbering any + registers; that will move the saving/restoring of those registers + to the function prologue/epilogue, and make non-movem sizes + suboptimal. + + This method is not foolproof; it assumes that the "asm reg" + declarations at the beginning of the function really are used + here (beware: they may be moved to temporary registers). + This way, we do not have to save/move the registers around into + temporaries; we can safely use them straight away. + + If you want to check that the allocation was right; then + check the equalities in the first comment. It should say + "r13=r13, r11=r11, r12=r12" */ + __asm__ volatile (" + ;; Check that the following is true (same register names on + ;; both sides of equal sign, as in r8=r8): + ;; %0=r13, %1=r11, %2=r12 + ;; + ;; Save the registers we'll use in the movem process + ;; on the stack. + subq 11*4,$sp + movem $r10,[$sp] + + ;; Now we've got this: + ;; r11 - src + ;; r13 - dst + ;; r12 - n + + ;; Update n for the first loop + subq 44,$r12 +0: + movem [$r11+],$r10 + subq 44,$r12 + bge 0b + movem $r10,[$r13+] + + addq 44,$r12 ;; compensate for last loop underflowing n + + ;; Restore registers from stack + movem [$sp+],$r10" + + /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n) + /* Inputs */ : "0" (dst), "1" (src), "2" (n)); + + } + + /* Either we directly starts copying, using dword copying + in a loop, or we copy as much as possible with 'movem' + and then the last block (<44 bytes) is copied here. + This will work since 'movem' will have updated src,dst,n. */ + + while ( n >= 16 ) + { + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + n -= 16; + } + + /* A switch() is definitely the fastest although it takes a LOT of code. + * Particularly if you inline code this. + */ + switch (n) + { + case 0: + break; + case 1: + *(char*)dst = *(char*)src; + break; + case 2: + *(short*)dst = *(short*)src; + break; + case 3: + *((short*)dst)++ = *((short*)src)++; + *(char*)dst = *(char*)src; + break; + case 4: + *((long*)dst)++ = *((long*)src)++; + break; + case 5: + *((long*)dst)++ = *((long*)src)++; + *(char*)dst = *(char*)src; + break; + case 6: + *((long*)dst)++ = *((long*)src)++; + *(short*)dst = *(short*)src; + break; + case 7: + *((long*)dst)++ = *((long*)src)++; + *((short*)dst)++ = *((short*)src)++; + *(char*)dst = *(char*)src; + break; + case 8: + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + break; + case 9: + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *(char*)dst = *(char*)src; + break; + case 10: + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *(short*)dst = *(short*)src; + break; + case 11: + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *((short*)dst)++ = *((short*)src)++; + *(char*)dst = *(char*)src; + break; + case 12: + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + break; + case 13: + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *(char*)dst = *(char*)src; + break; + case 14: + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *(short*)dst = *(short*)src; + break; + case 15: + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *((long*)dst)++ = *((long*)src)++; + *((short*)dst)++ = *((short*)src)++; + *(char*)dst = *(char*)src; + break; + } + + return return_dst; /* destination pointer. */ +} /* memcpy() */ diff --git a/arch/cris/arch-v10/lib/usercopy.c b/arch/cris/arch-v10/lib/usercopy.c new file mode 100644 index 0000000..43778d5 --- /dev/null +++ b/arch/cris/arch-v10/lib/usercopy.c @@ -0,0 +1,523 @@ +/* + * User address space access functions. + * The non-inlined parts of asm-cris/uaccess.h are here. + * + * Copyright (C) 2000, Axis Communications AB. + * + * Written by Hans-Peter Nilsson. + * Pieces used from memcpy, originally by Kenny Ranerup long time ago. + */ + +#include <asm/uaccess.h> + +/* Asm:s have been tweaked (within the domain of correctness) to give + satisfactory results for "gcc version 2.96 20000427 (experimental)". + + Check regularly... + + Note that the PC saved at a bus-fault is the address *after* the + faulting instruction, which means the branch-target for instructions in + delay-slots for taken branches. Note also that the postincrement in + the instruction is performed regardless of bus-fault; the register is + seen updated in fault handlers. + + Oh, and on the code formatting issue, to whomever feels like "fixing + it" to Conformity: I'm too "lazy", but why don't you go ahead and "fix" + string.c too. I just don't think too many people will hack this file + for the code format to be an issue. */ + + +/* Copy to userspace. This is based on the memcpy used for + kernel-to-kernel copying; see "string.c". */ + +unsigned long +__copy_user (void __user *pdst, const void *psrc, unsigned long pn) +{ + /* We want the parameters put in special registers. + Make sure the compiler is able to make something useful of this. + As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop). + + FIXME: Comment for old gcc version. Check. + If gcc was allright, it really would need no temporaries, and no + stack space to save stuff on. */ + + register char *dst __asm__ ("r13") = pdst; + register const char *src __asm__ ("r11") = psrc; + register int n __asm__ ("r12") = pn; + register int retn __asm__ ("r10") = 0; + + + /* When src is aligned but not dst, this makes a few extra needless + cycles. I believe it would take as many to check that the + re-alignment was unnecessary. */ + if (((unsigned long) dst & 3) != 0 + /* Don't align if we wouldn't copy more than a few bytes; so we + don't have to check further for overflows. */ + && n >= 3) + { + if ((unsigned long) dst & 1) + { + __asm_copy_to_user_1 (dst, src, retn); + n--; + } + + if ((unsigned long) dst & 2) + { + __asm_copy_to_user_2 (dst, src, retn); + n -= 2; + } + } + + /* Decide which copying method to use. */ + if (n >= 44*2) /* Break even between movem and + move16 is at 38.7*2, but modulo 44. */ + { + /* For large copies we use 'movem'. */ + + /* It is not optimal to tell the compiler about clobbering any + registers; that will move the saving/restoring of those registers + to the function prologue/epilogue, and make non-movem sizes + suboptimal. + + This method is not foolproof; it assumes that the "asm reg" + declarations at the beginning of the function really are used + here (beware: they may be moved to temporary registers). + This way, we do not have to save/move the registers around into + temporaries; we can safely use them straight away. + + If you want to check that the allocation was right; then + check the equalities in the first comment. It should say + "r13=r13, r11=r11, r12=r12". */ + __asm__ volatile ("\ + .ifnc %0%1%2%3,$r13$r11$r12$r10 \n\ + .err \n\ + .endif \n\ + + ;; Save the registers we'll use in the movem process + ;; on the stack. + subq 11*4,$sp + movem $r10,[$sp] + + ;; Now we've got this: + ;; r11 - src + ;; r13 - dst + ;; r12 - n + + ;; Update n for the first loop + subq 44,$r12 + +; Since the noted PC of a faulting instruction in a delay-slot of a taken +; branch, is that of the branch target, we actually point at the from-movem +; for this case. There is no ambiguity here; if there was a fault in that +; instruction (meaning a kernel oops), the faulted PC would be the address +; after *that* movem. + +0: + movem [$r11+],$r10 + subq 44,$r12 + bge 0b + movem $r10,[$r13+] +1: + addq 44,$r12 ;; compensate for last loop underflowing n + + ;; Restore registers from stack + movem [$sp+],$r10 +2: + .section .fixup,\"ax\" + +; To provide a correct count in r10 of bytes that failed to be copied, +; we jump back into the loop if the loop-branch was taken. There is no +; performance penalty for sany use; the program will segfault soon enough. + +3: + move.d [$sp],$r10 + addq 44,$r10 + move.d $r10,[$sp] + jump 0b +4: + movem [$sp+],$r10 + addq 44,$r10 + addq 44,$r12 + jump 2b + + .previous + .section __ex_table,\"a\" + .dword 0b,3b + .dword 1b,4b + .previous" + + /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn) + /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn)); + + } + + /* Either we directly start copying, using dword copying in a loop, or + we copy as much as possible with 'movem' and then the last block (<44 + bytes) is copied here. This will work since 'movem' will have + updated SRC, DST and N. */ + + while (n >= 16) + { + __asm_copy_to_user_16 (dst, src, retn); + n -= 16; + } + + /* Having a separate by-four loops cuts down on cache footprint. + FIXME: Test with and without; increasing switch to be 0..15. */ + while (n >= 4) + { + __asm_copy_to_user_4 (dst, src, retn); + n -= 4; + } + + switch (n) + { + case 0: + break; + case 1: + __asm_copy_to_user_1 (dst, src, retn); + break; + case 2: + __asm_copy_to_user_2 (dst, src, retn); + break; + case 3: + __asm_copy_to_user_3 (dst, src, retn); + break; + } + + return retn; +} + +/* Copy from user to kernel, zeroing the bytes that were inaccessible in + userland. The return-value is the number of bytes that were + inaccessible. */ + +unsigned long +__copy_user_zeroing (void __user *pdst, const void *psrc, unsigned long pn) +{ + /* We want the parameters put in special registers. + Make sure the compiler is able to make something useful of this. + As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop). + + FIXME: Comment for old gcc version. Check. + If gcc was allright, it really would need no temporaries, and no + stack space to save stuff on. */ + + register char *dst __asm__ ("r13") = pdst; + register const char *src __asm__ ("r11") = psrc; + register int n __asm__ ("r12") = pn; + register int retn __asm__ ("r10") = 0; + + /* The best reason to align src is that we then know that a read-fault + was for aligned bytes; there's no 1..3 remaining good bytes to + pickle. */ + if (((unsigned long) src & 3) != 0) + { + if (((unsigned long) src & 1) && n != 0) + { + __asm_copy_from_user_1 (dst, src, retn); + n--; + } + + if (((unsigned long) src & 2) && n >= 2) + { + __asm_copy_from_user_2 (dst, src, retn); + n -= 2; + } + + /* We only need one check after the unalignment-adjustments, because + if both adjustments were done, either both or neither reference + had an exception. */ + if (retn != 0) + goto copy_exception_bytes; + } + + /* Decide which copying method to use. */ + if (n >= 44*2) /* Break even between movem and + move16 is at 38.7*2, but modulo 44. + FIXME: We use move4 now. */ + { + /* For large copies we use 'movem' */ + + /* It is not optimal to tell the compiler about clobbering any + registers; that will move the saving/restoring of those registers + to the function prologue/epilogue, and make non-movem sizes + suboptimal. + + This method is not foolproof; it assumes that the "asm reg" + declarations at the beginning of the function really are used + here (beware: they may be moved to temporary registers). + This way, we do not have to save/move the registers around into + temporaries; we can safely use them straight away. + + If you want to check that the allocation was right; then + check the equalities in the first comment. It should say + "r13=r13, r11=r11, r12=r12" */ + __asm__ volatile (" + .ifnc %0%1%2%3,$r13$r11$r12$r10 \n\ + .err \n\ + .endif \n\ + + ;; Save the registers we'll use in the movem process + ;; on the stack. + subq 11*4,$sp + movem $r10,[$sp] + + ;; Now we've got this: + ;; r11 - src + ;; r13 - dst + ;; r12 - n + + ;; Update n for the first loop + subq 44,$r12 +0: + movem [$r11+],$r10 +1: + subq 44,$r12 + bge 0b + movem $r10,[$r13+] + + addq 44,$r12 ;; compensate for last loop underflowing n + + ;; Restore registers from stack + movem [$sp+],$r10 +4: + .section .fixup,\"ax\" + +;; Do not jump back into the loop if we fail. For some uses, we get a +;; page fault somewhere on the line. Without checking for page limits, +;; we don't know where, but we need to copy accurately and keep an +;; accurate count; not just clear the whole line. To do that, we fall +;; down in the code below, proceeding with smaller amounts. It should +;; be kept in mind that we have to cater to code like what at one time +;; was in fs/super.c: +;; i = size - copy_from_user((void *)page, data, size); +;; which would cause repeated faults while clearing the remainder of +;; the SIZE bytes at PAGE after the first fault. +;; A caveat here is that we must not fall through from a failing page +;; to a valid page. + +3: + movem [$sp+],$r10 + addq 44,$r12 ;; Get back count before faulting point. + subq 44,$r11 ;; Get back pointer to faulting movem-line. + jump 4b ;; Fall through, pretending the fault didn't happen. + + .previous + .section __ex_table,\"a\" + .dword 1b,3b + .previous" + + /* Outputs */ : "=r" (dst), "=r" (src), "=r" (n), "=r" (retn) + /* Inputs */ : "0" (dst), "1" (src), "2" (n), "3" (retn)); + + } + + /* Either we directly start copying here, using dword copying in a loop, + or we copy as much as possible with 'movem' and then the last block + (<44 bytes) is copied here. This will work since 'movem' will have + updated src, dst and n. (Except with failing src.) + + Since we want to keep src accurate, we can't use + __asm_copy_from_user_N with N != (1, 2, 4); it updates dst and + retn, but not src (by design; it's value is ignored elsewhere). */ + + while (n >= 4) + { + __asm_copy_from_user_4 (dst, src, retn); + n -= 4; + + if (retn) + goto copy_exception_bytes; + } + + /* If we get here, there were no memory read faults. */ + switch (n) + { + /* These copies are at least "naturally aligned" (so we don't have + to check each byte), due to the src alignment code before the + movem loop. The *_3 case *will* get the correct count for retn. */ + case 0: + /* This case deliberately left in (if you have doubts check the + generated assembly code). */ + break; + case 1: + __asm_copy_from_user_1 (dst, src, retn); + break; + case 2: + __asm_copy_from_user_2 (dst, src, retn); + break; + case 3: + __asm_copy_from_user_3 (dst, src, retn); + break; + } + + /* If we get here, retn correctly reflects the number of failing + bytes. */ + return retn; + +copy_exception_bytes: + /* We already have "retn" bytes cleared, and need to clear the + remaining "n" bytes. A non-optimized simple byte-for-byte in-line + memset is preferred here, since this isn't speed-critical code and + we'd rather have this a leaf-function than calling memset. */ + { + char *endp; + for (endp = dst + n; dst < endp; dst++) + *dst = 0; + } + + return retn + n; +} + +/* Zero userspace. */ + +unsigned long +__do_clear_user (void __user *pto, unsigned long pn) +{ + /* We want the parameters put in special registers. + Make sure the compiler is able to make something useful of this. + As it is now: r10 -> r13; r11 -> r11 (nop); r12 -> r12 (nop). + + FIXME: Comment for old gcc version. Check. + If gcc was allright, it really would need no temporaries, and no + stack space to save stuff on. */ + + register char *dst __asm__ ("r13") = pto; + register int n __asm__ ("r12") = pn; + register int retn __asm__ ("r10") = 0; + + + if (((unsigned long) dst & 3) != 0 + /* Don't align if we wouldn't copy more than a few bytes. */ + && n >= 3) + { + if ((unsigned long) dst & 1) + { + __asm_clear_1 (dst, retn); + n--; + } + + if ((unsigned long) dst & 2) + { + __asm_clear_2 (dst, retn); + n -= 2; + } + } + + /* Decide which copying method to use. + FIXME: This number is from the "ordinary" kernel memset. */ + if (n >= (1*48)) + { + /* For large clears we use 'movem' */ + + /* It is not optimal to tell the compiler about clobbering any + call-saved registers; that will move the saving/restoring of + those registers to the function prologue/epilogue, and make + non-movem sizes suboptimal. + + This method is not foolproof; it assumes that the "asm reg" + declarations at the beginning of the function really are used + here (beware: they may be moved to temporary registers). + This way, we do not have to save/move the registers around into + temporaries; we can safely use them straight away. + + If you want to check that the allocation was right; then + check the equalities in the first comment. It should say + something like "r13=r13, r11=r11, r12=r12". */ + __asm__ volatile (" + .ifnc %0%1%2,$r13$r12$r10 \n\ + .err \n\ + .endif \n\ + + ;; Save the registers we'll clobber in the movem process + ;; on the stack. Don't mention them to gcc, it will only be + ;; upset. + subq 11*4,$sp + movem $r10,[$sp] + + clear.d $r0 + clear.d $r1 + clear.d $r2 + clear.d $r3 + clear.d $r4 + clear.d $r5 + clear.d $r6 + clear.d $r7 + clear.d $r8 + clear.d $r9 + clear.d $r10 + clear.d $r11 + + ;; Now we've got this: + ;; r13 - dst + ;; r12 - n + + ;; Update n for the first loop + subq 12*4,$r12 +0: + subq 12*4,$r12 + bge 0b + movem $r11,[$r13+] +1: + addq 12*4,$r12 ;; compensate for last loop underflowing n + + ;; Restore registers from stack + movem [$sp+],$r10 +2: + .section .fixup,\"ax\" +3: + move.d [$sp],$r10 + addq 12*4,$r10 + move.d $r10,[$sp] + clear.d $r10 + jump 0b + +4: + movem [$sp+],$r10 + addq 12*4,$r10 + addq 12*4,$r12 + jump 2b + + .previous + .section __ex_table,\"a\" + .dword 0b,3b + .dword 1b,4b + .previous" + + /* Outputs */ : "=r" (dst), "=r" (n), "=r" (retn) + /* Inputs */ : "0" (dst), "1" (n), "2" (retn) + /* Clobber */ : "r11"); + } + + while (n >= 16) + { + __asm_clear_16 (dst, retn); + n -= 16; + } + + /* Having a separate by-four loops cuts down on cache footprint. + FIXME: Test with and without; increasing switch to be 0..15. */ + while (n >= 4) + { + __asm_clear_4 (dst, retn); + n -= 4; + } + + switch (n) + { + case 0: + break; + case 1: + __asm_clear_1 (dst, retn); + break; + case 2: + __asm_clear_2 (dst, retn); + break; + case 3: + __asm_clear_3 (dst, retn); + break; + } + + return retn; +} |