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 <linux/refcount.h> 47 #include <asm/page.h> 48 #include <asm/pgtable.h> 49 #include <linux/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 */ 70 enum mspec_page_type { 71 MSPEC_FETCHOP = 1, 72 MSPEC_CACHED, 73 MSPEC_UNCACHED 74 }; 75 76 #ifdef CONFIG_SGI_SN 77 static 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 */ 92 struct vma_data { 93 refcount_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 unsigned long vm_start; /* Original (unsplit) base. */ 98 unsigned long vm_end; /* Original (unsplit) end. */ 99 unsigned long maddr[0]; /* Array of MSPEC addresses. */ 100 }; 101 102 /* used on shub2 to clear FOP cache in the HUB */ 103 static unsigned long scratch_page[MAX_NUMNODES]; 104 #define SH2_AMO_CACHE_ENTRIES 4 105 106 static inline int 107 mspec_zero_block(unsigned long addr, int len) 108 { 109 int status; 110 111 if (is_sn2) { 112 if (is_shub2()) { 113 int nid; 114 void *p; 115 int i; 116 117 nid = nasid_to_cnodeid(get_node_number(__pa(addr))); 118 p = (void *)TO_AMO(scratch_page[nid]); 119 120 for (i=0; i < SH2_AMO_CACHE_ENTRIES; i++) { 121 FETCHOP_LOAD_OP(p, FETCHOP_LOAD); 122 p += FETCHOP_VAR_SIZE; 123 } 124 } 125 126 status = bte_copy(0, addr & ~__IA64_UNCACHED_OFFSET, len, 127 BTE_WACQUIRE | BTE_ZERO_FILL, NULL); 128 } else { 129 memset((char *) addr, 0, len); 130 status = 0; 131 } 132 return status; 133 } 134 135 /* 136 * mspec_open 137 * 138 * Called when a device mapping is created by a means other than mmap 139 * (via fork, munmap, etc.). Increments the reference count on the 140 * underlying mspec data so it is not freed prematurely. 141 */ 142 static void 143 mspec_open(struct vm_area_struct *vma) 144 { 145 struct vma_data *vdata; 146 147 vdata = vma->vm_private_data; 148 refcount_inc(&vdata->refcnt); 149 } 150 151 /* 152 * mspec_close 153 * 154 * Called when unmapping a device mapping. Frees all mspec pages 155 * belonging to all the vma's sharing this vma_data structure. 156 */ 157 static void 158 mspec_close(struct vm_area_struct *vma) 159 { 160 struct vma_data *vdata; 161 int index, last_index; 162 unsigned long my_page; 163 164 vdata = vma->vm_private_data; 165 166 if (!refcount_dec_and_test(&vdata->refcnt)) 167 return; 168 169 last_index = (vdata->vm_end - vdata->vm_start) >> PAGE_SHIFT; 170 for (index = 0; index < last_index; index++) { 171 if (vdata->maddr[index] == 0) 172 continue; 173 /* 174 * Clear the page before sticking it back 175 * into the pool. 176 */ 177 my_page = vdata->maddr[index]; 178 vdata->maddr[index] = 0; 179 if (!mspec_zero_block(my_page, PAGE_SIZE)) 180 uncached_free_page(my_page, 1); 181 else 182 printk(KERN_WARNING "mspec_close(): " 183 "failed to zero page %ld\n", my_page); 184 } 185 186 kvfree(vdata); 187 } 188 189 /* 190 * mspec_fault 191 * 192 * Creates a mspec page and maps it to user space. 193 */ 194 static int 195 mspec_fault(struct vm_fault *vmf) 196 { 197 unsigned long paddr, maddr; 198 unsigned long pfn; 199 pgoff_t index = vmf->pgoff; 200 struct vma_data *vdata = vmf->vma->vm_private_data; 201 202 maddr = (volatile unsigned long) vdata->maddr[index]; 203 if (maddr == 0) { 204 maddr = uncached_alloc_page(numa_node_id(), 1); 205 if (maddr == 0) 206 return VM_FAULT_OOM; 207 208 spin_lock(&vdata->lock); 209 if (vdata->maddr[index] == 0) { 210 vdata->count++; 211 vdata->maddr[index] = maddr; 212 } else { 213 uncached_free_page(maddr, 1); 214 maddr = vdata->maddr[index]; 215 } 216 spin_unlock(&vdata->lock); 217 } 218 219 if (vdata->type == MSPEC_FETCHOP) 220 paddr = TO_AMO(maddr); 221 else 222 paddr = maddr & ~__IA64_UNCACHED_OFFSET; 223 224 pfn = paddr >> PAGE_SHIFT; 225 226 /* 227 * vm_insert_pfn can fail with -EBUSY, but in that case it will 228 * be because another thread has installed the pte first, so it 229 * is no problem. 230 */ 231 vm_insert_pfn(vmf->vma, vmf->address, pfn); 232 233 return VM_FAULT_NOPAGE; 234 } 235 236 static const struct vm_operations_struct mspec_vm_ops = { 237 .open = mspec_open, 238 .close = mspec_close, 239 .fault = mspec_fault, 240 }; 241 242 /* 243 * mspec_mmap 244 * 245 * Called when mmapping the device. Initializes the vma with a fault handler 246 * and private data structure necessary to allocate, track, and free the 247 * underlying pages. 248 */ 249 static int 250 mspec_mmap(struct file *file, struct vm_area_struct *vma, 251 enum mspec_page_type type) 252 { 253 struct vma_data *vdata; 254 int pages, vdata_size; 255 256 if (vma->vm_pgoff != 0) 257 return -EINVAL; 258 259 if ((vma->vm_flags & VM_SHARED) == 0) 260 return -EINVAL; 261 262 if ((vma->vm_flags & VM_WRITE) == 0) 263 return -EPERM; 264 265 pages = vma_pages(vma); 266 vdata_size = sizeof(struct vma_data) + pages * sizeof(long); 267 if (vdata_size <= PAGE_SIZE) 268 vdata = kzalloc(vdata_size, GFP_KERNEL); 269 else 270 vdata = vzalloc(vdata_size); 271 if (!vdata) 272 return -ENOMEM; 273 274 vdata->vm_start = vma->vm_start; 275 vdata->vm_end = vma->vm_end; 276 vdata->type = type; 277 spin_lock_init(&vdata->lock); 278 refcount_set(&vdata->refcnt, 1); 279 vma->vm_private_data = vdata; 280 281 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; 282 if (vdata->type == MSPEC_FETCHOP || vdata->type == MSPEC_UNCACHED) 283 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 284 vma->vm_ops = &mspec_vm_ops; 285 286 return 0; 287 } 288 289 static int 290 fetchop_mmap(struct file *file, struct vm_area_struct *vma) 291 { 292 return mspec_mmap(file, vma, MSPEC_FETCHOP); 293 } 294 295 static int 296 cached_mmap(struct file *file, struct vm_area_struct *vma) 297 { 298 return mspec_mmap(file, vma, MSPEC_CACHED); 299 } 300 301 static int 302 uncached_mmap(struct file *file, struct vm_area_struct *vma) 303 { 304 return mspec_mmap(file, vma, MSPEC_UNCACHED); 305 } 306 307 static const struct file_operations fetchop_fops = { 308 .owner = THIS_MODULE, 309 .mmap = fetchop_mmap, 310 .llseek = noop_llseek, 311 }; 312 313 static struct miscdevice fetchop_miscdev = { 314 .minor = MISC_DYNAMIC_MINOR, 315 .name = "sgi_fetchop", 316 .fops = &fetchop_fops 317 }; 318 319 static const struct file_operations cached_fops = { 320 .owner = THIS_MODULE, 321 .mmap = cached_mmap, 322 .llseek = noop_llseek, 323 }; 324 325 static struct miscdevice cached_miscdev = { 326 .minor = MISC_DYNAMIC_MINOR, 327 .name = "mspec_cached", 328 .fops = &cached_fops 329 }; 330 331 static const struct file_operations uncached_fops = { 332 .owner = THIS_MODULE, 333 .mmap = uncached_mmap, 334 .llseek = noop_llseek, 335 }; 336 337 static struct miscdevice uncached_miscdev = { 338 .minor = MISC_DYNAMIC_MINOR, 339 .name = "mspec_uncached", 340 .fops = &uncached_fops 341 }; 342 343 /* 344 * mspec_init 345 * 346 * Called at boot time to initialize the mspec facility. 347 */ 348 static int __init 349 mspec_init(void) 350 { 351 int ret; 352 int nid; 353 354 /* 355 * The fetchop device only works on SN2 hardware, uncached and cached 356 * memory drivers should both be valid on all ia64 hardware 357 */ 358 #ifdef CONFIG_SGI_SN 359 if (ia64_platform_is("sn2")) { 360 is_sn2 = 1; 361 if (is_shub2()) { 362 ret = -ENOMEM; 363 for_each_node_state(nid, N_ONLINE) { 364 int actual_nid; 365 int nasid; 366 unsigned long phys; 367 368 scratch_page[nid] = uncached_alloc_page(nid, 1); 369 if (scratch_page[nid] == 0) 370 goto free_scratch_pages; 371 phys = __pa(scratch_page[nid]); 372 nasid = get_node_number(phys); 373 actual_nid = nasid_to_cnodeid(nasid); 374 if (actual_nid != nid) 375 goto free_scratch_pages; 376 } 377 } 378 379 ret = misc_register(&fetchop_miscdev); 380 if (ret) { 381 printk(KERN_ERR 382 "%s: failed to register device %i\n", 383 FETCHOP_ID, ret); 384 goto free_scratch_pages; 385 } 386 } 387 #endif 388 ret = misc_register(&cached_miscdev); 389 if (ret) { 390 printk(KERN_ERR "%s: failed to register device %i\n", 391 CACHED_ID, ret); 392 if (is_sn2) 393 misc_deregister(&fetchop_miscdev); 394 goto free_scratch_pages; 395 } 396 ret = misc_register(&uncached_miscdev); 397 if (ret) { 398 printk(KERN_ERR "%s: failed to register device %i\n", 399 UNCACHED_ID, ret); 400 misc_deregister(&cached_miscdev); 401 if (is_sn2) 402 misc_deregister(&fetchop_miscdev); 403 goto free_scratch_pages; 404 } 405 406 printk(KERN_INFO "%s %s initialized devices: %s %s %s\n", 407 MSPEC_BASENAME, REVISION, is_sn2 ? FETCHOP_ID : "", 408 CACHED_ID, UNCACHED_ID); 409 410 return 0; 411 412 free_scratch_pages: 413 for_each_node(nid) { 414 if (scratch_page[nid] != 0) 415 uncached_free_page(scratch_page[nid], 1); 416 } 417 return ret; 418 } 419 420 static void __exit 421 mspec_exit(void) 422 { 423 int nid; 424 425 misc_deregister(&uncached_miscdev); 426 misc_deregister(&cached_miscdev); 427 if (is_sn2) { 428 misc_deregister(&fetchop_miscdev); 429 430 for_each_node(nid) { 431 if (scratch_page[nid] != 0) 432 uncached_free_page(scratch_page[nid], 1); 433 } 434 } 435 } 436 437 module_init(mspec_init); 438 module_exit(mspec_exit); 439 440 MODULE_AUTHOR("Silicon Graphics, Inc. <linux-altix@sgi.com>"); 441 MODULE_DESCRIPTION("Driver for SGI SN special memory operations"); 442 MODULE_LICENSE("GPL"); 443