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