1/* 2 * Copyright (C) 2001-2006 Silicon Graphics, Inc. All rights 3 * reserved. 4 * 5 * This program is free software; you can redistribute it and/or modify it 6 * under the terms of version 2 of the GNU General Public License 7 * as published by the Free Software Foundation. 8 */ 9 10/* 11 * SN Platform Special Memory (mspec) Support 12 * 13 * This driver exports the SN special memory (mspec) facility to user 14 * processes. 15 * There are three types of memory made available thru this driver: 16 * fetchops, uncached and cached. 17 * 18 * Fetchops are atomic memory operations that are implemented in the 19 * memory controller on SGI SN hardware. 20 * 21 * Uncached are used for memory write combining feature of the ia64 22 * cpu. 23 * 24 * Cached are used for areas of memory that are used as cached addresses 25 * on our partition and used as uncached addresses from other partitions. 26 * Due to a design constraint of the SN2 Shub, you can not have processors 27 * on the same FSB perform both a cached and uncached reference to the 28 * same cache line. These special memory cached regions prevent the 29 * kernel from ever dropping in a TLB entry and therefore prevent the 30 * processor from ever speculating a cache line from this page. 31 */ 32 33#include <linux/types.h> 34#include <linux/kernel.h> 35#include <linux/module.h> 36#include <linux/init.h> 37#include <linux/errno.h> 38#include <linux/miscdevice.h> 39#include <linux/spinlock.h> 40#include <linux/mm.h> 41#include <linux/fs.h> 42#include <linux/vmalloc.h> 43#include <linux/string.h> 44#include <linux/slab.h> 45#include <linux/numa.h> 46#include <asm/page.h> 47#include <asm/system.h> 48#include <asm/pgtable.h> 49#include <asm/atomic.h> 50#include <asm/tlbflush.h> 51#include <asm/uncached.h> 52#include <asm/sn/addrs.h> 53#include <asm/sn/arch.h> 54#include <asm/sn/mspec.h> 55#include <asm/sn/sn_cpuid.h> 56#include <asm/sn/io.h> 57#include <asm/sn/bte.h> 58#include <asm/sn/shubio.h> 59 60 61#define FETCHOP_ID "SGI Fetchop," 62#define CACHED_ID "Cached," 63#define UNCACHED_ID "Uncached" 64#define REVISION "4.0" 65#define MSPEC_BASENAME "mspec" 66 67/* 68 * Page types allocated by the device. 69 */ 70enum mspec_page_type { 71 MSPEC_FETCHOP = 1, 72 MSPEC_CACHED, 73 MSPEC_UNCACHED 74}; 75 76#ifdef CONFIG_SGI_SN 77static int is_sn2; 78#else 79#define is_sn2 0 80#endif 81 82/* 83 * One of these structures is allocated when an mspec region is mmaped. The 84 * structure is pointed to by the vma->vm_private_data field in the vma struct. 85 * This structure is used to record the addresses of the mspec pages. 86 * This structure is shared by all vma's that are split off from the 87 * original vma when split_vma()'s are done. 88 * 89 * The refcnt is incremented atomically because mm->mmap_sem does not 90 * protect in fork case where multiple tasks share the vma_data. 91 */ 92struct vma_data { 93 atomic_t refcnt; /* Number of vmas sharing the data. */ 94 spinlock_t lock; /* Serialize access to this structure. */ 95 int count; /* Number of pages allocated. */ 96 enum mspec_page_type type; /* Type of pages allocated. */ 97 int flags; /* See VMD_xxx below. */ 98 unsigned long vm_start; /* Original (unsplit) base. */ 99 unsigned long vm_end; /* Original (unsplit) end. */ 100 unsigned long maddr[0]; /* Array of MSPEC addresses. */ 101}; 102 103#define VMD_VMALLOCED 0x1 /* vmalloc'd rather than kmalloc'd */ 104 105/* used on shub2 to clear FOP cache in the HUB */ 106static unsigned long scratch_page[MAX_NUMNODES]; 107#define SH2_AMO_CACHE_ENTRIES 4 108 109static inline int 110mspec_zero_block(unsigned long addr, int len) 111{ 112 int status; 113 114 if (is_sn2) { 115 if (is_shub2()) { 116 int nid; 117 void *p; 118 int i; 119 120 nid = nasid_to_cnodeid(get_node_number(__pa(addr))); 121 p = (void *)TO_AMO(scratch_page[nid]); 122 123 for (i=0; i < SH2_AMO_CACHE_ENTRIES; i++) { 124 FETCHOP_LOAD_OP(p, FETCHOP_LOAD); 125 p += FETCHOP_VAR_SIZE; 126 } 127 } 128 129 status = bte_copy(0, addr & ~__IA64_UNCACHED_OFFSET, len, 130 BTE_WACQUIRE | BTE_ZERO_FILL, NULL); 131 } else { 132 memset((char *) addr, 0, len); 133 status = 0; 134 } 135 return status; 136} 137 138/* 139 * mspec_open 140 * 141 * Called when a device mapping is created by a means other than mmap 142 * (via fork, munmap, etc.). Increments the reference count on the 143 * underlying mspec data so it is not freed prematurely. 144 */ 145static void 146mspec_open(struct vm_area_struct *vma) 147{ 148 struct vma_data *vdata; 149 150 vdata = vma->vm_private_data; 151 atomic_inc(&vdata->refcnt); 152} 153 154/* 155 * mspec_close 156 * 157 * Called when unmapping a device mapping. Frees all mspec pages 158 * belonging to all the vma's sharing this vma_data structure. 159 */ 160static void 161mspec_close(struct vm_area_struct *vma) 162{ 163 struct vma_data *vdata; 164 int index, last_index; 165 unsigned long my_page; 166 167 vdata = vma->vm_private_data; 168 169 if (!atomic_dec_and_test(&vdata->refcnt)) 170 return; 171 172 last_index = (vdata->vm_end - vdata->vm_start) >> PAGE_SHIFT; 173 for (index = 0; index < last_index; index++) { 174 if (vdata->maddr[index] == 0) 175 continue; 176 /* 177 * Clear the page before sticking it back 178 * into the pool. 179 */ 180 my_page = vdata->maddr[index]; 181 vdata->maddr[index] = 0; 182 if (!mspec_zero_block(my_page, PAGE_SIZE)) 183 uncached_free_page(my_page, 1); 184 else 185 printk(KERN_WARNING "mspec_close(): " 186 "failed to zero page %ld\n", my_page); 187 } 188 189 if (vdata->flags & VMD_VMALLOCED) 190 vfree(vdata); 191 else 192 kfree(vdata); 193} 194 195/* 196 * mspec_fault 197 * 198 * Creates a mspec page and maps it to user space. 199 */ 200static int 201mspec_fault(struct vm_area_struct *vma, struct vm_fault *vmf) 202{ 203 unsigned long paddr, maddr; 204 unsigned long pfn; 205 pgoff_t index = vmf->pgoff; 206 struct vma_data *vdata = vma->vm_private_data; 207 208 maddr = (volatile unsigned long) vdata->maddr[index]; 209 if (maddr == 0) { 210 maddr = uncached_alloc_page(numa_node_id(), 1); 211 if (maddr == 0) 212 return VM_FAULT_OOM; 213 214 spin_lock(&vdata->lock); 215 if (vdata->maddr[index] == 0) { 216 vdata->count++; 217 vdata->maddr[index] = maddr; 218 } else { 219 uncached_free_page(maddr, 1); 220 maddr = vdata->maddr[index]; 221 } 222 spin_unlock(&vdata->lock); 223 } 224 225 if (vdata->type == MSPEC_FETCHOP) 226 paddr = TO_AMO(maddr); 227 else 228 paddr = maddr & ~__IA64_UNCACHED_OFFSET; 229 230 pfn = paddr >> PAGE_SHIFT; 231 232 /* 233 * vm_insert_pfn can fail with -EBUSY, but in that case it will 234 * be because another thread has installed the pte first, so it 235 * is no problem. 236 */ 237 vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn); 238 239 return VM_FAULT_NOPAGE; 240} 241 242static const struct vm_operations_struct mspec_vm_ops = { 243 .open = mspec_open, 244 .close = mspec_close, 245 .fault = mspec_fault, 246}; 247 248/* 249 * mspec_mmap 250 * 251 * Called when mmapping the device. Initializes the vma with a fault handler 252 * and private data structure necessary to allocate, track, and free the 253 * underlying pages. 254 */ 255static int 256mspec_mmap(struct file *file, struct vm_area_struct *vma, 257 enum mspec_page_type type) 258{ 259 struct vma_data *vdata; 260 int pages, vdata_size, flags = 0; 261 262 if (vma->vm_pgoff != 0) 263 return -EINVAL; 264 265 if ((vma->vm_flags & VM_SHARED) == 0) 266 return -EINVAL; 267 268 if ((vma->vm_flags & VM_WRITE) == 0) 269 return -EPERM; 270 271 pages = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT; 272 vdata_size = sizeof(struct vma_data) + pages * sizeof(long); 273 if (vdata_size <= PAGE_SIZE) 274 vdata = kmalloc(vdata_size, GFP_KERNEL); 275 else { 276 vdata = vmalloc(vdata_size); 277 flags = VMD_VMALLOCED; 278 } 279 if (!vdata) 280 return -ENOMEM; 281 memset(vdata, 0, vdata_size); 282 283 vdata->vm_start = vma->vm_start; 284 vdata->vm_end = vma->vm_end; 285 vdata->flags = flags; 286 vdata->type = type; 287 spin_lock_init(&vdata->lock); 288 vdata->refcnt = ATOMIC_INIT(1); 289 vma->vm_private_data = vdata; 290 291 vma->vm_flags |= (VM_IO | VM_RESERVED | VM_PFNMAP | VM_DONTEXPAND); 292 if (vdata->type == MSPEC_FETCHOP || vdata->type == MSPEC_UNCACHED) 293 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 294 vma->vm_ops = &mspec_vm_ops; 295 296 return 0; 297} 298 299static int 300fetchop_mmap(struct file *file, struct vm_area_struct *vma) 301{ 302 return mspec_mmap(file, vma, MSPEC_FETCHOP); 303} 304 305static int 306cached_mmap(struct file *file, struct vm_area_struct *vma) 307{ 308 return mspec_mmap(file, vma, MSPEC_CACHED); 309} 310 311static int 312uncached_mmap(struct file *file, struct vm_area_struct *vma) 313{ 314 return mspec_mmap(file, vma, MSPEC_UNCACHED); 315} 316 317static const struct file_operations fetchop_fops = { 318 .owner = THIS_MODULE, 319 .mmap = fetchop_mmap 320}; 321 322static struct miscdevice fetchop_miscdev = { 323 .minor = MISC_DYNAMIC_MINOR, 324 .name = "sgi_fetchop", 325 .fops = &fetchop_fops 326}; 327 328static const struct file_operations cached_fops = { 329 .owner = THIS_MODULE, 330 .mmap = cached_mmap 331}; 332 333static struct miscdevice cached_miscdev = { 334 .minor = MISC_DYNAMIC_MINOR, 335 .name = "mspec_cached", 336 .fops = &cached_fops 337}; 338 339static const struct file_operations uncached_fops = { 340 .owner = THIS_MODULE, 341 .mmap = uncached_mmap 342}; 343 344static struct miscdevice uncached_miscdev = { 345 .minor = MISC_DYNAMIC_MINOR, 346 .name = "mspec_uncached", 347 .fops = &uncached_fops 348}; 349 350/* 351 * mspec_init 352 * 353 * Called at boot time to initialize the mspec facility. 354 */ 355static int __init 356mspec_init(void) 357{ 358 int ret; 359 int nid; 360 361 /* 362 * The fetchop device only works on SN2 hardware, uncached and cached 363 * memory drivers should both be valid on all ia64 hardware 364 */ 365#ifdef CONFIG_SGI_SN 366 if (ia64_platform_is("sn2")) { 367 is_sn2 = 1; 368 if (is_shub2()) { 369 ret = -ENOMEM; 370 for_each_node_state(nid, N_ONLINE) { 371 int actual_nid; 372 int nasid; 373 unsigned long phys; 374 375 scratch_page[nid] = uncached_alloc_page(nid, 1); 376 if (scratch_page[nid] == 0) 377 goto free_scratch_pages; 378 phys = __pa(scratch_page[nid]); 379 nasid = get_node_number(phys); 380 actual_nid = nasid_to_cnodeid(nasid); 381 if (actual_nid != nid) 382 goto free_scratch_pages; 383 } 384 } 385 386 ret = misc_register(&fetchop_miscdev); 387 if (ret) { 388 printk(KERN_ERR 389 "%s: failed to register device %i\n", 390 FETCHOP_ID, ret); 391 goto free_scratch_pages; 392 } 393 } 394#endif 395 ret = misc_register(&cached_miscdev); 396 if (ret) { 397 printk(KERN_ERR "%s: failed to register device %i\n", 398 CACHED_ID, ret); 399 if (is_sn2) 400 misc_deregister(&fetchop_miscdev); 401 goto free_scratch_pages; 402 } 403 ret = misc_register(&uncached_miscdev); 404 if (ret) { 405 printk(KERN_ERR "%s: failed to register device %i\n", 406 UNCACHED_ID, ret); 407 misc_deregister(&cached_miscdev); 408 if (is_sn2) 409 misc_deregister(&fetchop_miscdev); 410 goto free_scratch_pages; 411 } 412 413 printk(KERN_INFO "%s %s initialized devices: %s %s %s\n", 414 MSPEC_BASENAME, REVISION, is_sn2 ? FETCHOP_ID : "", 415 CACHED_ID, UNCACHED_ID); 416 417 return 0; 418 419 free_scratch_pages: 420 for_each_node(nid) { 421 if (scratch_page[nid] != 0) 422 uncached_free_page(scratch_page[nid], 1); 423 } 424 return ret; 425} 426 427static void __exit 428mspec_exit(void) 429{ 430 int nid; 431 432 misc_deregister(&uncached_miscdev); 433 misc_deregister(&cached_miscdev); 434 if (is_sn2) { 435 misc_deregister(&fetchop_miscdev); 436 437 for_each_node(nid) { 438 if (scratch_page[nid] != 0) 439 uncached_free_page(scratch_page[nid], 1); 440 } 441 } 442} 443 444module_init(mspec_init); 445module_exit(mspec_exit); 446 447MODULE_AUTHOR("Silicon Graphics, Inc. <linux-altix@sgi.com>"); 448MODULE_DESCRIPTION("Driver for SGI SN special memory operations"); 449MODULE_LICENSE("GPL"); 450