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, Version 1.0 only 6 * (the "License"). You may not use this file except in compliance 7 * with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or http://www.opensolaris.org/os/licensing. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright 2006 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 /* 28 * Kernel Physical Mapping (kpm) segment driver (segkpm). 29 * 30 * This driver delivers along with the hat_kpm* interfaces an alternative 31 * mechanism for kernel mappings within the 64-bit Solaris operating system, 32 * which allows the mapping of all physical memory into the kernel address 33 * space at once. This is feasible in 64 bit kernels, e.g. for Ultrasparc II 34 * and beyond processors, since the available VA range is much larger than 35 * possible physical memory. Momentarily all physical memory is supported, 36 * that is represented by the list of memory segments (memsegs). 37 * 38 * Segkpm mappings have also very low overhead and large pages are used 39 * (when possible) to minimize the TLB and TSB footprint. It is also 40 * extentable for other than Sparc architectures (e.g. AMD64). Main 41 * advantage is the avoidance of the TLB-shootdown X-calls, which are 42 * normally needed when a kernel (global) mapping has to be removed. 43 * 44 * First example of a kernel facility that uses the segkpm mapping scheme 45 * is seg_map, where it is used as an alternative to hat_memload(). 46 * See also hat layer for more information about the hat_kpm* routines. 47 * The kpm facilty can be turned off at boot time (e.g. /etc/system). 48 */ 49 50 #include <sys/types.h> 51 #include <sys/param.h> 52 #include <sys/sysmacros.h> 53 #include <sys/systm.h> 54 #include <sys/vnode.h> 55 #include <sys/cmn_err.h> 56 #include <sys/debug.h> 57 #include <sys/thread.h> 58 #include <sys/cpuvar.h> 59 #include <sys/bitmap.h> 60 #include <sys/atomic.h> 61 #include <sys/lgrp.h> 62 63 #include <vm/seg_kmem.h> 64 #include <vm/seg_kpm.h> 65 #include <vm/hat.h> 66 #include <vm/as.h> 67 #include <vm/seg.h> 68 #include <vm/page.h> 69 70 /* 71 * Global kpm controls. 72 * See also platform and mmu specific controls. 73 * 74 * kpm_enable -- global on/off switch for segkpm. 75 * . Set by default on 64bit platforms that have kpm support. 76 * . Will be disabled from platform layer if not supported. 77 * . Can be disabled via /etc/system. 78 * 79 * kpm_smallpages -- use only regular/system pagesize for kpm mappings. 80 * . Can be useful for critical debugging of kpm clients. 81 * . Set to zero by default for platforms that support kpm large pages. 82 * The use of kpm large pages reduces the footprint of kpm meta data 83 * and has all the other advantages of using large pages (e.g TLB 84 * miss reduction). 85 * . Set by default for platforms that don't support kpm large pages or 86 * where large pages cannot be used for other reasons (e.g. there are 87 * only few full associative TLB entries available for large pages). 88 * 89 * segmap_kpm -- separate on/off switch for segmap using segkpm: 90 * . Set by default. 91 * . Will be disabled when kpm_enable is zero. 92 * . Will be disabled when MAXBSIZE != PAGESIZE. 93 * . Can be disabled via /etc/system. 94 * 95 */ 96 int kpm_enable = 1; 97 int kpm_smallpages = 0; 98 int segmap_kpm = 1; 99 100 /* 101 * Private seg op routines. 102 */ 103 faultcode_t segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, 104 size_t len, enum fault_type type, enum seg_rw rw); 105 static void segkpm_dump(struct seg *); 106 static int segkpm_pagelock(struct seg *seg, caddr_t addr, size_t len, 107 struct page ***page, enum lock_type type, 108 enum seg_rw rw); 109 110 static const struct seg_ops segkpm_ops = { 111 .fault = segkpm_fault, 112 .dump = segkpm_dump, 113 .pagelock = segkpm_pagelock, 114 //#ifndef SEGKPM_SUPPORT 115 #if 0 116 #error FIXME: define nop 117 .dup = nop, 118 .unmap = nop, 119 .free = nop, 120 .faulta = nop, 121 .setprot = nop, 122 .checkprot = nop, 123 .kluster = nop, 124 .sync = nop, 125 .incore = nop, 126 .lockop = nop, 127 .getprot = nop, 128 .getoffset = nop, 129 .gettype = nop, 130 .getvp = nop, 131 .advise = nop, 132 .getpolicy = nop, 133 #endif 134 }; 135 136 /* 137 * kpm_pgsz and kpm_pgshft are set by platform layer. 138 */ 139 size_t kpm_pgsz; /* kpm page size */ 140 uint_t kpm_pgshft; /* kpm page shift */ 141 u_offset_t kpm_pgoff; /* kpm page offset mask */ 142 uint_t kpmp2pshft; /* kpm page to page shift */ 143 pgcnt_t kpmpnpgs; /* how many pages per kpm page */ 144 145 146 #ifdef SEGKPM_SUPPORT 147 148 int 149 segkpm_create(struct seg *seg, void *argsp) 150 { 151 struct segkpm_data *skd; 152 struct segkpm_crargs *b = (struct segkpm_crargs *)argsp; 153 ushort_t *p; 154 int i, j; 155 156 ASSERT(seg->s_as && RW_WRITE_HELD(&seg->s_as->a_lock)); 157 ASSERT(btokpmp(seg->s_size) >= 1 && 158 kpmpageoff((uintptr_t)seg->s_base) == 0 && 159 kpmpageoff((uintptr_t)seg->s_base + seg->s_size) == 0); 160 161 skd = kmem_zalloc(sizeof (struct segkpm_data), KM_SLEEP); 162 163 seg->s_data = (void *)skd; 164 seg->s_ops = &segkpm_ops; 165 skd->skd_prot = b->prot; 166 167 /* 168 * (1) Segkpm virtual addresses are based on physical adresses. 169 * From this and in opposite to other segment drivers it is 170 * often required to allocate a page first to be able to 171 * calculate the final segkpm virtual address. 172 * (2) Page allocation is done by calling page_create_va(), 173 * one important input argument is a virtual address (also 174 * expressed by the "va" in the function name). This function 175 * is highly optimized to select the right page for an optimal 176 * processor and platform support (e.g. virtual addressed 177 * caches (VAC), physical addressed caches, NUMA). 178 * 179 * Because of (1) the approach is to generate a faked virtual 180 * address for calling page_create_va(). In order to exploit 181 * the abilities of (2), especially to utilize the cache 182 * hierarchy (3) and to avoid VAC alias conflicts (4) the 183 * selection has to be done carefully. For each virtual color 184 * a separate counter is provided (4). The count values are 185 * used for the utilization of all cache lines (3) and are 186 * corresponding to the cache bins. 187 */ 188 skd->skd_nvcolors = b->nvcolors; 189 190 p = skd->skd_va_select = 191 kmem_zalloc(NCPU * b->nvcolors * sizeof (ushort_t), KM_SLEEP); 192 193 for (i = 0; i < NCPU; i++) 194 for (j = 0; j < b->nvcolors; j++, p++) 195 *p = j; 196 197 return (0); 198 } 199 200 /* 201 * This routine is called via a machine specific fault handling 202 * routine. 203 */ 204 /* ARGSUSED */ 205 faultcode_t 206 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len, 207 enum fault_type type, enum seg_rw rw) 208 { 209 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock)); 210 211 switch (type) { 212 case F_INVAL: 213 return (hat_kpm_fault(hat, addr)); 214 case F_SOFTLOCK: 215 case F_SOFTUNLOCK: 216 return (0); 217 default: 218 return (FC_NOSUPPORT); 219 } 220 /*NOTREACHED*/ 221 } 222 223 #define addr_to_vcolor(addr, vcolors) \ 224 ((int)(((uintptr_t)(addr) & ((vcolors << PAGESHIFT) - 1)) >> PAGESHIFT)) 225 226 /* 227 * Create a virtual address that can be used for invocations of 228 * page_create_va. Goal is to utilize the cache hierarchy (round 229 * robin bins) and to select the right color for virtual indexed 230 * caches. It isn't exact since we also increment the bin counter 231 * when the caller uses VOP_GETPAGE and gets a hit in the page 232 * cache, but we keep the bins turning for cache distribution 233 * (see also segkpm_create block comment). 234 */ 235 caddr_t 236 segkpm_create_va(u_offset_t off) 237 { 238 int vcolor; 239 ushort_t *p; 240 struct segkpm_data *skd = (struct segkpm_data *)segkpm->s_data; 241 int nvcolors = skd->skd_nvcolors; 242 caddr_t va; 243 244 vcolor = (nvcolors > 1) ? addr_to_vcolor(off, nvcolors) : 0; 245 p = &skd->skd_va_select[(CPU->cpu_id * nvcolors) + vcolor]; 246 va = (caddr_t)ptob(*p); 247 248 atomic_add_16(p, nvcolors); 249 250 return (va); 251 } 252 253 /* 254 * Unload mapping if the instance has an active kpm mapping. 255 */ 256 void 257 segkpm_mapout_validkpme(struct kpme *kpme) 258 { 259 caddr_t vaddr; 260 page_t *pp; 261 262 retry: 263 if ((pp = kpme->kpe_page) == NULL) { 264 return; 265 } 266 267 if (page_lock(pp, SE_SHARED, (kmutex_t *)NULL, P_RECLAIM) == 0) 268 goto retry; 269 270 /* 271 * Check if segkpm mapping is not unloaded in the meantime 272 */ 273 if (kpme->kpe_page == NULL) { 274 page_unlock(pp); 275 return; 276 } 277 278 vaddr = hat_kpm_page2va(pp, 1); 279 hat_kpm_mapout(pp, kpme, vaddr); 280 page_unlock(pp); 281 } 282 283 #else /* SEGKPM_SUPPORT */ 284 285 /* segkpm stubs */ 286 287 /*ARGSUSED*/ 288 int segkpm_create(struct seg *seg, void *argsp) 289 { 290 return (0); 291 } 292 293 /* ARGSUSED */ 294 faultcode_t 295 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len, 296 enum fault_type type, enum seg_rw rw) 297 { 298 return (0); 299 } 300 301 /* ARGSUSED */ 302 caddr_t segkpm_create_va(u_offset_t off) 303 { 304 return (NULL); 305 } 306 307 /* ARGSUSED */ 308 void segkpm_mapout_validkpme(struct kpme *kpme) 309 { 310 } 311 312 #endif /* SEGKPM_SUPPORT */ 313 314 /* ARGSUSED */ 315 static int 316 segkpm_pagelock(struct seg *seg, caddr_t addr, size_t len, 317 struct page ***page, enum lock_type type, enum seg_rw rw) 318 { 319 return (ENOTSUP); 320 } 321 322 /* 323 * segkpm pages are not dumped, so we just return 324 */ 325 /*ARGSUSED*/ 326 static void 327 segkpm_dump(struct seg *seg) 328 { 329 }