1685784aaSDan Williams /* 2685784aaSDan Williams * xor.c : Multiple Devices driver for Linux 3685784aaSDan Williams * 4685784aaSDan Williams * Copyright (C) 1996, 1997, 1998, 1999, 2000, 5685784aaSDan Williams * Ingo Molnar, Matti Aarnio, Jakub Jelinek, Richard Henderson. 6685784aaSDan Williams * 7685784aaSDan Williams * Dispatch optimized RAID-5 checksumming functions. 8685784aaSDan Williams * 9685784aaSDan Williams * This program is free software; you can redistribute it and/or modify 10685784aaSDan Williams * it under the terms of the GNU General Public License as published by 11685784aaSDan Williams * the Free Software Foundation; either version 2, or (at your option) 12685784aaSDan Williams * any later version. 13685784aaSDan Williams * 14685784aaSDan Williams * You should have received a copy of the GNU General Public License 15685784aaSDan Williams * (for example /usr/src/linux/COPYING); if not, write to the Free 16685784aaSDan Williams * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. 17685784aaSDan Williams */ 18685784aaSDan Williams 19685784aaSDan Williams #define BH_TRACE 0 20685784aaSDan Williams #include <linux/module.h> 215a0e3ad6STejun Heo #include <linux/gfp.h> 22685784aaSDan Williams #include <linux/raid/xor.h> 23bff61975SNeilBrown #include <linux/jiffies.h> 2456a51991SJim Kukunas #include <linux/preempt.h> 25685784aaSDan Williams #include <asm/xor.h> 26685784aaSDan Williams 27*27c4d548SHerbert Xu #ifndef XOR_SELECT_TEMPLATE 28*27c4d548SHerbert Xu #define XOR_SELECT_TEMPLATE(x) (x) 29*27c4d548SHerbert Xu #endif 30*27c4d548SHerbert Xu 31685784aaSDan Williams /* The xor routines to use. */ 32685784aaSDan Williams static struct xor_block_template *active_template; 33685784aaSDan Williams 34685784aaSDan Williams void 359bc89cd8SDan Williams xor_blocks(unsigned int src_count, unsigned int bytes, void *dest, void **srcs) 36685784aaSDan Williams { 379bc89cd8SDan Williams unsigned long *p1, *p2, *p3, *p4; 38685784aaSDan Williams 399bc89cd8SDan Williams p1 = (unsigned long *) srcs[0]; 409bc89cd8SDan Williams if (src_count == 1) { 419bc89cd8SDan Williams active_template->do_2(bytes, dest, p1); 42685784aaSDan Williams return; 43685784aaSDan Williams } 44685784aaSDan Williams 459bc89cd8SDan Williams p2 = (unsigned long *) srcs[1]; 469bc89cd8SDan Williams if (src_count == 2) { 479bc89cd8SDan Williams active_template->do_3(bytes, dest, p1, p2); 48685784aaSDan Williams return; 49685784aaSDan Williams } 50685784aaSDan Williams 519bc89cd8SDan Williams p3 = (unsigned long *) srcs[2]; 529bc89cd8SDan Williams if (src_count == 3) { 539bc89cd8SDan Williams active_template->do_4(bytes, dest, p1, p2, p3); 54685784aaSDan Williams return; 55685784aaSDan Williams } 56685784aaSDan Williams 579bc89cd8SDan Williams p4 = (unsigned long *) srcs[3]; 589bc89cd8SDan Williams active_template->do_5(bytes, dest, p1, p2, p3, p4); 59685784aaSDan Williams } 60685784aaSDan Williams EXPORT_SYMBOL(xor_blocks); 61685784aaSDan Williams 62685784aaSDan Williams /* Set of all registered templates. */ 63af7cf25dSJan Beulich static struct xor_block_template *__initdata template_list; 64685784aaSDan Williams 65685784aaSDan Williams #define BENCH_SIZE (PAGE_SIZE) 66685784aaSDan Williams 67af7cf25dSJan Beulich static void __init 68685784aaSDan Williams do_xor_speed(struct xor_block_template *tmpl, void *b1, void *b2) 69685784aaSDan Williams { 70685784aaSDan Williams int speed; 716a328475SJim Kukunas unsigned long now, j; 72685784aaSDan Williams int i, count, max; 73685784aaSDan Williams 74685784aaSDan Williams tmpl->next = template_list; 75685784aaSDan Williams template_list = tmpl; 76685784aaSDan Williams 7756a51991SJim Kukunas preempt_disable(); 7856a51991SJim Kukunas 79685784aaSDan Williams /* 80685784aaSDan Williams * Count the number of XORs done during a whole jiffy, and use 81685784aaSDan Williams * this to calculate the speed of checksumming. We use a 2-page 82685784aaSDan Williams * allocation to have guaranteed color L1-cache layout. 83685784aaSDan Williams */ 84685784aaSDan Williams max = 0; 85685784aaSDan Williams for (i = 0; i < 5; i++) { 866a328475SJim Kukunas j = jiffies; 87685784aaSDan Williams count = 0; 886a328475SJim Kukunas while ((now = jiffies) == j) 896a328475SJim Kukunas cpu_relax(); 906a328475SJim Kukunas while (time_before(jiffies, now + 1)) { 91685784aaSDan Williams mb(); /* prevent loop optimzation */ 92685784aaSDan Williams tmpl->do_2(BENCH_SIZE, b1, b2); 93685784aaSDan Williams mb(); 94685784aaSDan Williams count++; 95685784aaSDan Williams mb(); 96685784aaSDan Williams } 97685784aaSDan Williams if (count > max) 98685784aaSDan Williams max = count; 99685784aaSDan Williams } 100685784aaSDan Williams 10156a51991SJim Kukunas preempt_enable(); 10256a51991SJim Kukunas 103685784aaSDan Williams speed = max * (HZ * BENCH_SIZE / 1024); 104685784aaSDan Williams tmpl->speed = speed; 105685784aaSDan Williams 106685784aaSDan Williams printk(KERN_INFO " %-10s: %5d.%03d MB/sec\n", tmpl->name, 107685784aaSDan Williams speed / 1000, speed % 1000); 108685784aaSDan Williams } 109685784aaSDan Williams 110685784aaSDan Williams static int __init 111685784aaSDan Williams calibrate_xor_blocks(void) 112685784aaSDan Williams { 113685784aaSDan Williams void *b1, *b2; 114685784aaSDan Williams struct xor_block_template *f, *fastest; 115685784aaSDan Williams 116*27c4d548SHerbert Xu fastest = XOR_SELECT_TEMPLATE(NULL); 11739457acdSMartin Schwidefsky 11839457acdSMartin Schwidefsky if (fastest) { 11939457acdSMartin Schwidefsky printk(KERN_INFO "xor: automatically using best " 12039457acdSMartin Schwidefsky "checksumming function %-10s\n", 12139457acdSMartin Schwidefsky fastest->name); 12239457acdSMartin Schwidefsky goto out; 12339457acdSMartin Schwidefsky } 12439457acdSMartin Schwidefsky 12533f65df7SVegard Nossum /* 12633f65df7SVegard Nossum * Note: Since the memory is not actually used for _anything_ but to 12733f65df7SVegard Nossum * test the XOR speed, we don't really want kmemcheck to warn about 12833f65df7SVegard Nossum * reading uninitialized bytes here. 12933f65df7SVegard Nossum */ 13033f65df7SVegard Nossum b1 = (void *) __get_free_pages(GFP_KERNEL | __GFP_NOTRACK, 2); 131685784aaSDan Williams if (!b1) { 132685784aaSDan Williams printk(KERN_WARNING "xor: Yikes! No memory available.\n"); 133685784aaSDan Williams return -ENOMEM; 134685784aaSDan Williams } 135685784aaSDan Williams b2 = b1 + 2*PAGE_SIZE + BENCH_SIZE; 136685784aaSDan Williams 137685784aaSDan Williams /* 138685784aaSDan Williams * If this arch/cpu has a short-circuited selection, don't loop through 139685784aaSDan Williams * all the possible functions, just test the best one 140685784aaSDan Williams */ 141685784aaSDan Williams 142685784aaSDan Williams #define xor_speed(templ) do_xor_speed((templ), b1, b2) 143685784aaSDan Williams 1449bc89cd8SDan Williams printk(KERN_INFO "xor: measuring software checksum speed\n"); 145685784aaSDan Williams XOR_TRY_TEMPLATES; 146685784aaSDan Williams fastest = template_list; 147685784aaSDan Williams for (f = fastest; f; f = f->next) 148685784aaSDan Williams if (f->speed > fastest->speed) 149685784aaSDan Williams fastest = f; 150685784aaSDan Williams 151685784aaSDan Williams printk(KERN_INFO "xor: using function: %s (%d.%03d MB/sec)\n", 152685784aaSDan Williams fastest->name, fastest->speed / 1000, fastest->speed % 1000); 153685784aaSDan Williams 154685784aaSDan Williams #undef xor_speed 155685784aaSDan Williams 156685784aaSDan Williams free_pages((unsigned long)b1, 2); 15739457acdSMartin Schwidefsky out: 158685784aaSDan Williams active_template = fastest; 159685784aaSDan Williams return 0; 160685784aaSDan Williams } 161685784aaSDan Williams 162685784aaSDan Williams static __exit void xor_exit(void) { } 163685784aaSDan Williams 164685784aaSDan Williams MODULE_LICENSE("GPL"); 165685784aaSDan Williams 166685784aaSDan Williams /* when built-in xor.o must initialize before drivers/md/md.o */ 167685784aaSDan Williams core_initcall(calibrate_xor_blocks); 168685784aaSDan Williams module_exit(xor_exit); 169