1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2001-2006 Silicon Graphics, Inc. All rights 4 * reserved. 5 */ 6 7 /* 8 * SN Platform Special Memory (mspec) Support 9 * 10 * This driver exports the SN special memory (mspec) facility to user 11 * processes. 12 * There are three types of memory made available thru this driver: 13 * fetchops, uncached and cached. 14 * 15 * Fetchops are atomic memory operations that are implemented in the 16 * memory controller on SGI SN hardware. 17 * 18 * Uncached are used for memory write combining feature of the ia64 19 * cpu. 20 * 21 * Cached are used for areas of memory that are used as cached addresses 22 * on our partition and used as uncached addresses from other partitions. 23 * Due to a design constraint of the SN2 Shub, you can not have processors 24 * on the same FSB perform both a cached and uncached reference to the 25 * same cache line. These special memory cached regions prevent the 26 * kernel from ever dropping in a TLB entry and therefore prevent the 27 * processor from ever speculating a cache line from this page. 28 */ 29 30 #include <linux/types.h> 31 #include <linux/kernel.h> 32 #include <linux/module.h> 33 #include <linux/init.h> 34 #include <linux/errno.h> 35 #include <linux/miscdevice.h> 36 #include <linux/spinlock.h> 37 #include <linux/mm.h> 38 #include <linux/fs.h> 39 #include <linux/vmalloc.h> 40 #include <linux/string.h> 41 #include <linux/slab.h> 42 #include <linux/numa.h> 43 #include <linux/refcount.h> 44 #include <asm/page.h> 45 #include <asm/pgtable.h> 46 #include <linux/atomic.h> 47 #include <asm/tlbflush.h> 48 #include <asm/uncached.h> 49 #include <asm/sn/addrs.h> 50 #include <asm/sn/arch.h> 51 #include <asm/sn/mspec.h> 52 #include <asm/sn/sn_cpuid.h> 53 #include <asm/sn/io.h> 54 #include <asm/sn/bte.h> 55 #include <asm/sn/shubio.h> 56 57 58 #define FETCHOP_ID "SGI Fetchop," 59 #define CACHED_ID "Cached," 60 #define UNCACHED_ID "Uncached" 61 #define REVISION "4.0" 62 #define MSPEC_BASENAME "mspec" 63 64 /* 65 * Page types allocated by the device. 66 */ 67 enum mspec_page_type { 68 MSPEC_FETCHOP = 1, 69 MSPEC_CACHED, 70 MSPEC_UNCACHED 71 }; 72 73 #ifdef CONFIG_SGI_SN 74 static int is_sn2; 75 #else 76 #define is_sn2 0 77 #endif 78 79 /* 80 * One of these structures is allocated when an mspec region is mmaped. The 81 * structure is pointed to by the vma->vm_private_data field in the vma struct. 82 * This structure is used to record the addresses of the mspec pages. 83 * This structure is shared by all vma's that are split off from the 84 * original vma when split_vma()'s are done. 85 * 86 * The refcnt is incremented atomically because mm->mmap_sem does not 87 * protect in fork case where multiple tasks share the vma_data. 88 */ 89 struct vma_data { 90 refcount_t refcnt; /* Number of vmas sharing the data. */ 91 spinlock_t lock; /* Serialize access to this structure. */ 92 int count; /* Number of pages allocated. */ 93 enum mspec_page_type type; /* Type of pages allocated. */ 94 unsigned long vm_start; /* Original (unsplit) base. */ 95 unsigned long vm_end; /* Original (unsplit) end. */ 96 unsigned long maddr[0]; /* Array of MSPEC addresses. */ 97 }; 98 99 /* used on shub2 to clear FOP cache in the HUB */ 100 static unsigned long scratch_page[MAX_NUMNODES]; 101 #define SH2_AMO_CACHE_ENTRIES 4 102 103 static inline int 104 mspec_zero_block(unsigned long addr, int len) 105 { 106 int status; 107 108 if (is_sn2) { 109 if (is_shub2()) { 110 int nid; 111 void *p; 112 int i; 113 114 nid = nasid_to_cnodeid(get_node_number(__pa(addr))); 115 p = (void *)TO_AMO(scratch_page[nid]); 116 117 for (i=0; i < SH2_AMO_CACHE_ENTRIES; i++) { 118 FETCHOP_LOAD_OP(p, FETCHOP_LOAD); 119 p += FETCHOP_VAR_SIZE; 120 } 121 } 122 123 status = bte_copy(0, addr & ~__IA64_UNCACHED_OFFSET, len, 124 BTE_WACQUIRE | BTE_ZERO_FILL, NULL); 125 } else { 126 memset((char *) addr, 0, len); 127 status = 0; 128 } 129 return status; 130 } 131 132 /* 133 * mspec_open 134 * 135 * Called when a device mapping is created by a means other than mmap 136 * (via fork, munmap, etc.). Increments the reference count on the 137 * underlying mspec data so it is not freed prematurely. 138 */ 139 static void 140 mspec_open(struct vm_area_struct *vma) 141 { 142 struct vma_data *vdata; 143 144 vdata = vma->vm_private_data; 145 refcount_inc(&vdata->refcnt); 146 } 147 148 /* 149 * mspec_close 150 * 151 * Called when unmapping a device mapping. Frees all mspec pages 152 * belonging to all the vma's sharing this vma_data structure. 153 */ 154 static void 155 mspec_close(struct vm_area_struct *vma) 156 { 157 struct vma_data *vdata; 158 int index, last_index; 159 unsigned long my_page; 160 161 vdata = vma->vm_private_data; 162 163 if (!refcount_dec_and_test(&vdata->refcnt)) 164 return; 165 166 last_index = (vdata->vm_end - vdata->vm_start) >> PAGE_SHIFT; 167 for (index = 0; index < last_index; index++) { 168 if (vdata->maddr[index] == 0) 169 continue; 170 /* 171 * Clear the page before sticking it back 172 * into the pool. 173 */ 174 my_page = vdata->maddr[index]; 175 vdata->maddr[index] = 0; 176 if (!mspec_zero_block(my_page, PAGE_SIZE)) 177 uncached_free_page(my_page, 1); 178 else 179 printk(KERN_WARNING "mspec_close(): " 180 "failed to zero page %ld\n", my_page); 181 } 182 183 kvfree(vdata); 184 } 185 186 /* 187 * mspec_fault 188 * 189 * Creates a mspec page and maps it to user space. 190 */ 191 static vm_fault_t 192 mspec_fault(struct vm_fault *vmf) 193 { 194 unsigned long paddr, maddr; 195 unsigned long pfn; 196 pgoff_t index = vmf->pgoff; 197 struct vma_data *vdata = vmf->vma->vm_private_data; 198 199 maddr = (volatile unsigned long) vdata->maddr[index]; 200 if (maddr == 0) { 201 maddr = uncached_alloc_page(numa_node_id(), 1); 202 if (maddr == 0) 203 return VM_FAULT_OOM; 204 205 spin_lock(&vdata->lock); 206 if (vdata->maddr[index] == 0) { 207 vdata->count++; 208 vdata->maddr[index] = maddr; 209 } else { 210 uncached_free_page(maddr, 1); 211 maddr = vdata->maddr[index]; 212 } 213 spin_unlock(&vdata->lock); 214 } 215 216 if (vdata->type == MSPEC_FETCHOP) 217 paddr = TO_AMO(maddr); 218 else 219 paddr = maddr & ~__IA64_UNCACHED_OFFSET; 220 221 pfn = paddr >> PAGE_SHIFT; 222 223 return vmf_insert_pfn(vmf->vma, vmf->address, pfn); 224 } 225 226 static const struct vm_operations_struct mspec_vm_ops = { 227 .open = mspec_open, 228 .close = mspec_close, 229 .fault = mspec_fault, 230 }; 231 232 /* 233 * mspec_mmap 234 * 235 * Called when mmapping the device. Initializes the vma with a fault handler 236 * and private data structure necessary to allocate, track, and free the 237 * underlying pages. 238 */ 239 static int 240 mspec_mmap(struct file *file, struct vm_area_struct *vma, 241 enum mspec_page_type type) 242 { 243 struct vma_data *vdata; 244 int pages, vdata_size; 245 246 if (vma->vm_pgoff != 0) 247 return -EINVAL; 248 249 if ((vma->vm_flags & VM_SHARED) == 0) 250 return -EINVAL; 251 252 if ((vma->vm_flags & VM_WRITE) == 0) 253 return -EPERM; 254 255 pages = vma_pages(vma); 256 vdata_size = sizeof(struct vma_data) + pages * sizeof(long); 257 if (vdata_size <= PAGE_SIZE) 258 vdata = kzalloc(vdata_size, GFP_KERNEL); 259 else 260 vdata = vzalloc(vdata_size); 261 if (!vdata) 262 return -ENOMEM; 263 264 vdata->vm_start = vma->vm_start; 265 vdata->vm_end = vma->vm_end; 266 vdata->type = type; 267 spin_lock_init(&vdata->lock); 268 refcount_set(&vdata->refcnt, 1); 269 vma->vm_private_data = vdata; 270 271 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; 272 if (vdata->type == MSPEC_FETCHOP || vdata->type == MSPEC_UNCACHED) 273 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 274 vma->vm_ops = &mspec_vm_ops; 275 276 return 0; 277 } 278 279 static int 280 fetchop_mmap(struct file *file, struct vm_area_struct *vma) 281 { 282 return mspec_mmap(file, vma, MSPEC_FETCHOP); 283 } 284 285 static int 286 cached_mmap(struct file *file, struct vm_area_struct *vma) 287 { 288 return mspec_mmap(file, vma, MSPEC_CACHED); 289 } 290 291 static int 292 uncached_mmap(struct file *file, struct vm_area_struct *vma) 293 { 294 return mspec_mmap(file, vma, MSPEC_UNCACHED); 295 } 296 297 static const struct file_operations fetchop_fops = { 298 .owner = THIS_MODULE, 299 .mmap = fetchop_mmap, 300 .llseek = noop_llseek, 301 }; 302 303 static struct miscdevice fetchop_miscdev = { 304 .minor = MISC_DYNAMIC_MINOR, 305 .name = "sgi_fetchop", 306 .fops = &fetchop_fops 307 }; 308 309 static const struct file_operations cached_fops = { 310 .owner = THIS_MODULE, 311 .mmap = cached_mmap, 312 .llseek = noop_llseek, 313 }; 314 315 static struct miscdevice cached_miscdev = { 316 .minor = MISC_DYNAMIC_MINOR, 317 .name = "mspec_cached", 318 .fops = &cached_fops 319 }; 320 321 static const struct file_operations uncached_fops = { 322 .owner = THIS_MODULE, 323 .mmap = uncached_mmap, 324 .llseek = noop_llseek, 325 }; 326 327 static struct miscdevice uncached_miscdev = { 328 .minor = MISC_DYNAMIC_MINOR, 329 .name = "mspec_uncached", 330 .fops = &uncached_fops 331 }; 332 333 /* 334 * mspec_init 335 * 336 * Called at boot time to initialize the mspec facility. 337 */ 338 static int __init 339 mspec_init(void) 340 { 341 int ret; 342 int nid; 343 344 /* 345 * The fetchop device only works on SN2 hardware, uncached and cached 346 * memory drivers should both be valid on all ia64 hardware 347 */ 348 #ifdef CONFIG_SGI_SN 349 if (ia64_platform_is("sn2")) { 350 is_sn2 = 1; 351 if (is_shub2()) { 352 ret = -ENOMEM; 353 for_each_node_state(nid, N_ONLINE) { 354 int actual_nid; 355 int nasid; 356 unsigned long phys; 357 358 scratch_page[nid] = uncached_alloc_page(nid, 1); 359 if (scratch_page[nid] == 0) 360 goto free_scratch_pages; 361 phys = __pa(scratch_page[nid]); 362 nasid = get_node_number(phys); 363 actual_nid = nasid_to_cnodeid(nasid); 364 if (actual_nid != nid) 365 goto free_scratch_pages; 366 } 367 } 368 369 ret = misc_register(&fetchop_miscdev); 370 if (ret) { 371 printk(KERN_ERR 372 "%s: failed to register device %i\n", 373 FETCHOP_ID, ret); 374 goto free_scratch_pages; 375 } 376 } 377 #endif 378 ret = misc_register(&cached_miscdev); 379 if (ret) { 380 printk(KERN_ERR "%s: failed to register device %i\n", 381 CACHED_ID, ret); 382 if (is_sn2) 383 misc_deregister(&fetchop_miscdev); 384 goto free_scratch_pages; 385 } 386 ret = misc_register(&uncached_miscdev); 387 if (ret) { 388 printk(KERN_ERR "%s: failed to register device %i\n", 389 UNCACHED_ID, ret); 390 misc_deregister(&cached_miscdev); 391 if (is_sn2) 392 misc_deregister(&fetchop_miscdev); 393 goto free_scratch_pages; 394 } 395 396 printk(KERN_INFO "%s %s initialized devices: %s %s %s\n", 397 MSPEC_BASENAME, REVISION, is_sn2 ? FETCHOP_ID : "", 398 CACHED_ID, UNCACHED_ID); 399 400 return 0; 401 402 free_scratch_pages: 403 for_each_node(nid) { 404 if (scratch_page[nid] != 0) 405 uncached_free_page(scratch_page[nid], 1); 406 } 407 return ret; 408 } 409 410 static void __exit 411 mspec_exit(void) 412 { 413 int nid; 414 415 misc_deregister(&uncached_miscdev); 416 misc_deregister(&cached_miscdev); 417 if (is_sn2) { 418 misc_deregister(&fetchop_miscdev); 419 420 for_each_node(nid) { 421 if (scratch_page[nid] != 0) 422 uncached_free_page(scratch_page[nid], 1); 423 } 424 } 425 } 426 427 module_init(mspec_init); 428 module_exit(mspec_exit); 429 430 MODULE_AUTHOR("Silicon Graphics, Inc. <linux-altix@sgi.com>"); 431 MODULE_DESCRIPTION("Driver for SGI SN special memory operations"); 432 MODULE_LICENSE("GPL"); 433