1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright 2006 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 #pragma ident "%Z%%M% %I% %E% SMI" 27 28 #include <sys/types.h> 29 #include <sys/systm.h> 30 #include <sys/archsystm.h> 31 #include <sys/machsystm.h> 32 #include <sys/t_lock.h> 33 #include <sys/vmem.h> 34 #include <sys/mman.h> 35 #include <sys/vm.h> 36 #include <sys/cpu.h> 37 #include <sys/cmn_err.h> 38 #include <sys/cpuvar.h> 39 #include <sys/atomic.h> 40 #include <vm/as.h> 41 #include <vm/hat.h> 42 #include <vm/as.h> 43 #include <vm/page.h> 44 #include <vm/seg.h> 45 #include <vm/seg_kmem.h> 46 #include <vm/hat_sfmmu.h> 47 #include <sys/debug.h> 48 #include <sys/cpu_module.h> 49 50 /* 51 * A quick way to generate a cache consistent address to map in a page. 52 * users: ppcopy, pagezero, /proc, dev/mem 53 * 54 * The ppmapin/ppmapout routines provide a quick way of generating a cache 55 * consistent address by reserving a given amount of kernel address space. 56 * The base is PPMAPBASE and its size is PPMAPSIZE. This memory is divided 57 * into x number of sets, where x is the number of colors for the virtual 58 * cache. The number of colors is how many times a page can be mapped 59 * simulatenously in the cache. For direct map caches this translates to 60 * the number of pages in the cache. 61 * Each set will be assigned a group of virtual pages from the reserved memory 62 * depending on its virtual color. 63 * When trying to assign a virtual address we will find out the color for the 64 * physical page in question (if applicable). Then we will try to find an 65 * available virtual page from the set of the appropiate color. 66 */ 67 68 int pp_slots = 4; /* small default, tuned by cpu module */ 69 70 /* tuned by cpu module, default is "safe" */ 71 int pp_consistent_coloring = PPAGE_STORES_POLLUTE | PPAGE_LOADS_POLLUTE; 72 73 static caddr_t ppmap_vaddrs[PPMAPSIZE / MMU_PAGESIZE]; 74 static int nsets; /* number of sets */ 75 static int ppmap_shift; /* set selector */ 76 77 #ifdef PPDEBUG 78 #define MAXCOLORS 16 /* for debug only */ 79 static int ppalloc_noslot = 0; /* # of allocations from kernelmap */ 80 static int align_hits; 81 static int pp_allocs; /* # of ppmapin requests */ 82 #endif /* PPDEBUG */ 83 84 /* 85 * There are only 64 TLB entries on spitfire, 16 on cheetah 86 * (fully-associative TLB) so we allow the cpu module to tune the 87 * number to use here via pp_slots. 88 */ 89 static struct ppmap_va { 90 caddr_t ppmap_slots[MAXPP_SLOTS]; 91 } ppmap_va[NCPU]; 92 93 /* prevent compilation with VAC defined */ 94 #ifdef VAC 95 #error "sun4v ppmapin and ppmapout do not support VAC" 96 #endif 97 98 void 99 ppmapinit(void) 100 { 101 int nset; 102 caddr_t va; 103 104 ASSERT(pp_slots <= MAXPP_SLOTS); 105 106 va = (caddr_t)PPMAPBASE; 107 108 /* 109 * sun4v does not have a virtual indexed cache and simply 110 * has only one set containing all pages. 111 */ 112 nsets = mmu_btop(PPMAPSIZE); 113 ppmap_shift = MMU_PAGESHIFT; 114 115 for (nset = 0; nset < nsets; nset++) { 116 ppmap_vaddrs[nset] = 117 (caddr_t)((uintptr_t)va + (nset * MMU_PAGESIZE)); 118 } 119 } 120 121 /* 122 * Allocate a cache consistent virtual address to map a page, pp, 123 * with protection, vprot; and map it in the MMU, using the most 124 * efficient means possible. The argument avoid is a virtual address 125 * hint which when masked yields an offset into a virtual cache 126 * that should be avoided when allocating an address to map in a 127 * page. An avoid arg of -1 means you don't care, for instance pagezero. 128 * 129 * machine dependent, depends on virtual address space layout, 130 * understands that all kernel addresses have bit 31 set. 131 * 132 * NOTE: For sun4 platforms the meaning of the hint argument is opposite from 133 * that found in other architectures. In other architectures the hint 134 * (called avoid) was used to ask ppmapin to NOT use the specified cache color. 135 * This was used to avoid virtual cache trashing in the bcopy. Unfortunately 136 * in the case of a COW, this later on caused a cache aliasing conflict. In 137 * sun4, the bcopy routine uses the block ld/st instructions so we don't have 138 * to worry about virtual cache trashing. Actually, by using the hint to choose 139 * the right color we can almost guarantee a cache conflict will not occur. 140 */ 141 142 /*ARGSUSED2*/ 143 caddr_t 144 ppmapin(page_t *pp, uint_t vprot, caddr_t hint) 145 { 146 int nset; 147 caddr_t va; 148 149 #ifdef PPDEBUG 150 pp_allocs++; 151 #endif /* PPDEBUG */ 152 153 /* 154 * For sun4v caches are physical caches, we can pick any address 155 * we want. 156 */ 157 for (nset = 0; nset < nsets; nset++) { 158 va = ppmap_vaddrs[nset]; 159 if (va != NULL) { 160 #ifdef PPDEBUG 161 align_hits++; 162 #endif /* PPDEBUG */ 163 if (casptr(&ppmap_vaddrs[nset], va, NULL) == va) { 164 hat_memload(kas.a_hat, va, pp, 165 vprot | HAT_NOSYNC, 166 HAT_LOAD_LOCK); 167 return (va); 168 } 169 } 170 } 171 172 #ifdef PPDEBUG 173 ppalloc_noslot++; 174 #endif /* PPDEBUG */ 175 176 /* 177 * No free slots; get a random one from the kernel heap area. 178 */ 179 va = vmem_alloc(heap_arena, PAGESIZE, VM_SLEEP); 180 181 hat_memload(kas.a_hat, va, pp, vprot | HAT_NOSYNC, HAT_LOAD_LOCK); 182 183 return (va); 184 185 } 186 187 void 188 ppmapout(caddr_t va) 189 { 190 int nset; 191 192 if (va >= kernelheap && va < ekernelheap) { 193 /* 194 * Space came from kernelmap, flush the page and 195 * return the space. 196 */ 197 hat_unload(kas.a_hat, va, PAGESIZE, 198 (HAT_UNLOAD_NOSYNC | HAT_UNLOAD_UNLOCK)); 199 vmem_free(heap_arena, va, PAGESIZE); 200 } else { 201 /* 202 * Space came from ppmap_vaddrs[], give it back. 203 */ 204 nset = ((uintptr_t)va >> ppmap_shift) & (nsets - 1); 205 hat_unload(kas.a_hat, va, PAGESIZE, 206 (HAT_UNLOAD_NOSYNC | HAT_UNLOAD_UNLOCK)); 207 208 ASSERT(ppmap_vaddrs[nset] == NULL); 209 ppmap_vaddrs[nset] = va; 210 } 211 } 212 213 #ifdef DEBUG 214 #define PP_STAT_ADD(stat) (stat)++ 215 uint_t pload, ploadfail; 216 uint_t ppzero, ppzero_short; 217 #else 218 #define PP_STAT_ADD(stat) 219 #endif /* DEBUG */ 220 221 static void 222 pp_unload_tlb(caddr_t *pslot, caddr_t va) 223 { 224 ASSERT(*pslot == va); 225 226 vtag_flushpage(va, (uint64_t)ksfmmup); 227 *pslot = NULL; /* release the slot */ 228 } 229 230 /* 231 * Routine to copy kernel pages during relocation. It will copy one 232 * PAGESIZE page to another PAGESIZE page. This function may be called 233 * above LOCK_LEVEL so it should not grab any locks. 234 */ 235 void 236 ppcopy_kernel__relocatable(page_t *fm_pp, page_t *to_pp) 237 { 238 uint64_t fm_pa, to_pa; 239 size_t nbytes; 240 241 fm_pa = (uint64_t)(fm_pp->p_pagenum) << MMU_PAGESHIFT; 242 to_pa = (uint64_t)(to_pp->p_pagenum) << MMU_PAGESHIFT; 243 244 nbytes = MMU_PAGESIZE; 245 246 for (; nbytes > 0; fm_pa += 32, to_pa += 32, nbytes -= 32) 247 hw_pa_bcopy32(fm_pa, to_pa); 248 } 249 250 /* 251 * Copy the data from the physical page represented by "frompp" to 252 * that represented by "topp". 253 * 254 * Try to use per cpu mapping first, if that fails then call pp_mapin 255 * to load it. 256 * Returns one on success or zero on some sort of fault while doing the copy. 257 */ 258 int 259 ppcopy(page_t *fm_pp, page_t *to_pp) 260 { 261 caddr_t fm_va; 262 caddr_t to_va; 263 boolean_t fast; 264 label_t ljb; 265 int ret = 1; 266 267 ASSERT(PAGE_LOCKED(fm_pp)); 268 ASSERT(PAGE_LOCKED(to_pp)); 269 270 /* 271 * Try to map using KPM. If it fails, fall back to 272 * ppmapin/ppmapout. 273 */ 274 if ((fm_va = hat_kpm_mapin(fm_pp, NULL)) == NULL || 275 (to_va = hat_kpm_mapin(to_pp, NULL)) == NULL) { 276 if (fm_va != NULL) 277 hat_kpm_mapout(fm_pp, NULL, fm_va); 278 fm_va = ppmapin(fm_pp, PROT_READ, (caddr_t)-1); 279 to_va = ppmapin(to_pp, PROT_READ | PROT_WRITE, fm_va); 280 fast = B_FALSE; 281 } else 282 fast = B_TRUE; 283 284 if (on_fault(&ljb)) { 285 ret = 0; 286 goto faulted; 287 } 288 bcopy(fm_va, to_va, PAGESIZE); 289 no_fault(); 290 faulted: 291 292 /* Unmap */ 293 if (fast) { 294 hat_kpm_mapout(fm_pp, NULL, fm_va); 295 hat_kpm_mapout(to_pp, NULL, to_va); 296 } else { 297 ppmapout(fm_va); 298 ppmapout(to_va); 299 } 300 return (ret); 301 } 302 303 /* 304 * Zero the physical page from off to off + len given by `pp' 305 * without changing the reference and modified bits of page. 306 * 307 * Again, we'll try per cpu mapping first. 308 */ 309 310 void 311 pagezero(page_t *pp, uint_t off, uint_t len) 312 { 313 caddr_t va; 314 extern int hwblkclr(void *, size_t); 315 extern int use_hw_bzero; 316 boolean_t fast; 317 318 ASSERT((int)len > 0 && (int)off >= 0 && off + len <= PAGESIZE); 319 ASSERT(PAGE_LOCKED(pp)); 320 321 PP_STAT_ADD(ppzero); 322 323 if (len != MMU_PAGESIZE || !use_hw_bzero) { 324 PP_STAT_ADD(ppzero_short); 325 } 326 327 kpreempt_disable(); 328 329 /* 330 * Try to use KPM. If that fails, fall back to 331 * ppmapin/ppmapout. 332 */ 333 fast = B_TRUE; 334 va = hat_kpm_mapin(pp, NULL); 335 if (va == NULL) { 336 fast = B_FALSE; 337 va = ppmapin(pp, PROT_READ | PROT_WRITE, (caddr_t)-1); 338 } 339 340 if (!use_hw_bzero) { 341 bzero(va + off, len); 342 sync_icache(va + off, len); 343 } else if (hwblkclr(va + off, len)) { 344 /* 345 * We may not have used block commit asi. 346 * So flush the I-$ manually 347 */ 348 sync_icache(va + off, len); 349 } else { 350 /* 351 * We have used blk commit, and flushed the I-$. 352 * However we still may have an instruction in the 353 * pipeline. Only a flush will invalidate that. 354 */ 355 doflush(va); 356 } 357 358 if (fast) { 359 hat_kpm_mapout(pp, NULL, va); 360 } else { 361 ppmapout(va); 362 } 363 kpreempt_enable(); 364 } 365