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_badop(void); 106 static int segkpm_notsup(void); 107 108 #define SEGKPM_BADOP(t) (t(*)())segkpm_badop 109 #define SEGKPM_NOTSUP (int(*)())segkpm_notsup 110 111 static struct seg_ops segkpm_ops = { 112 .dup = SEGKPM_BADOP(int), 113 .unmap = SEGKPM_BADOP(int), 114 .free = SEGKPM_BADOP(void), 115 .fault = segkpm_fault, 116 .faulta = SEGKPM_BADOP(int), 117 .setprot = SEGKPM_BADOP(int), 118 .checkprot = SEGKPM_BADOP(int), 119 .kluster = SEGKPM_BADOP(int), 120 .swapout = SEGKPM_BADOP(size_t), 121 .sync = SEGKPM_BADOP(int), 122 .incore = SEGKPM_BADOP(size_t), 123 .lockop = SEGKPM_BADOP(int), 124 .getprot = SEGKPM_BADOP(int), 125 .getoffset = SEGKPM_BADOP(u_offset_t), 126 .gettype = SEGKPM_BADOP(int), 127 .getvp = SEGKPM_BADOP(int), 128 .advise = SEGKPM_BADOP(int), 129 .pagelock = SEGKPM_NOTSUP, 130 .setpagesize = SEGKPM_BADOP(int), 131 .getmemid = SEGKPM_BADOP(int), 132 .getpolicy = SEGKPM_BADOP(lgrp_mem_policy_info_t *), 133 }; 134 135 /* 136 * kpm_pgsz and kpm_pgshft are set by platform layer. 137 */ 138 size_t kpm_pgsz; /* kpm page size */ 139 uint_t kpm_pgshft; /* kpm page shift */ 140 u_offset_t kpm_pgoff; /* kpm page offset mask */ 141 uint_t kpmp2pshft; /* kpm page to page shift */ 142 pgcnt_t kpmpnpgs; /* how many pages per kpm page */ 143 144 145 #ifdef SEGKPM_SUPPORT 146 147 int 148 segkpm_create(struct seg *seg, void *argsp) 149 { 150 struct segkpm_data *skd; 151 struct segkpm_crargs *b = (struct segkpm_crargs *)argsp; 152 ushort_t *p; 153 int i, j; 154 155 ASSERT(seg->s_as && RW_WRITE_HELD(&seg->s_as->a_lock)); 156 ASSERT(btokpmp(seg->s_size) >= 1 && 157 kpmpageoff((uintptr_t)seg->s_base) == 0 && 158 kpmpageoff((uintptr_t)seg->s_base + seg->s_size) == 0); 159 160 skd = kmem_zalloc(sizeof (struct segkpm_data), KM_SLEEP); 161 162 seg->s_data = (void *)skd; 163 seg->s_ops = &segkpm_ops; 164 skd->skd_prot = b->prot; 165 166 /* 167 * (1) Segkpm virtual addresses are based on physical adresses. 168 * From this and in opposite to other segment drivers it is 169 * often required to allocate a page first to be able to 170 * calculate the final segkpm virtual address. 171 * (2) Page allocation is done by calling page_create_va(), 172 * one important input argument is a virtual address (also 173 * expressed by the "va" in the function name). This function 174 * is highly optimized to select the right page for an optimal 175 * processor and platform support (e.g. virtual addressed 176 * caches (VAC), physical addressed caches, NUMA). 177 * 178 * Because of (1) the approach is to generate a faked virtual 179 * address for calling page_create_va(). In order to exploit 180 * the abilities of (2), especially to utilize the cache 181 * hierarchy (3) and to avoid VAC alias conflicts (4) the 182 * selection has to be done carefully. For each virtual color 183 * a separate counter is provided (4). The count values are 184 * used for the utilization of all cache lines (3) and are 185 * corresponding to the cache bins. 186 */ 187 skd->skd_nvcolors = b->nvcolors; 188 189 p = skd->skd_va_select = 190 kmem_zalloc(NCPU * b->nvcolors * sizeof (ushort_t), KM_SLEEP); 191 192 for (i = 0; i < NCPU; i++) 193 for (j = 0; j < b->nvcolors; j++, p++) 194 *p = j; 195 196 return (0); 197 } 198 199 /* 200 * This routine is called via a machine specific fault handling 201 * routine. 202 */ 203 /* ARGSUSED */ 204 faultcode_t 205 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len, 206 enum fault_type type, enum seg_rw rw) 207 { 208 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock)); 209 210 switch (type) { 211 case F_INVAL: 212 return (hat_kpm_fault(hat, addr)); 213 case F_SOFTLOCK: 214 case F_SOFTUNLOCK: 215 return (0); 216 default: 217 return (FC_NOSUPPORT); 218 } 219 /*NOTREACHED*/ 220 } 221 222 #define addr_to_vcolor(addr, vcolors) \ 223 ((int)(((uintptr_t)(addr) & ((vcolors << PAGESHIFT) - 1)) >> PAGESHIFT)) 224 225 /* 226 * Create a virtual address that can be used for invocations of 227 * page_create_va. Goal is to utilize the cache hierarchy (round 228 * robin bins) and to select the right color for virtual indexed 229 * caches. It isn't exact since we also increment the bin counter 230 * when the caller uses VOP_GETPAGE and gets a hit in the page 231 * cache, but we keep the bins turning for cache distribution 232 * (see also segkpm_create block comment). 233 */ 234 caddr_t 235 segkpm_create_va(u_offset_t off) 236 { 237 int vcolor; 238 ushort_t *p; 239 struct segkpm_data *skd = (struct segkpm_data *)segkpm->s_data; 240 int nvcolors = skd->skd_nvcolors; 241 caddr_t va; 242 243 vcolor = (nvcolors > 1) ? addr_to_vcolor(off, nvcolors) : 0; 244 p = &skd->skd_va_select[(CPU->cpu_id * nvcolors) + vcolor]; 245 va = (caddr_t)ptob(*p); 246 247 atomic_add_16(p, nvcolors); 248 249 return (va); 250 } 251 252 /* 253 * Unload mapping if the instance has an active kpm mapping. 254 */ 255 void 256 segkpm_mapout_validkpme(struct kpme *kpme) 257 { 258 caddr_t vaddr; 259 page_t *pp; 260 261 retry: 262 if ((pp = kpme->kpe_page) == NULL) { 263 return; 264 } 265 266 if (page_lock(pp, SE_SHARED, (kmutex_t *)NULL, P_RECLAIM) == 0) 267 goto retry; 268 269 /* 270 * Check if segkpm mapping is not unloaded in the meantime 271 */ 272 if (kpme->kpe_page == NULL) { 273 page_unlock(pp); 274 return; 275 } 276 277 vaddr = hat_kpm_page2va(pp, 1); 278 hat_kpm_mapout(pp, kpme, vaddr); 279 page_unlock(pp); 280 } 281 282 static void 283 segkpm_badop() 284 { 285 panic("segkpm_badop"); 286 } 287 288 #else /* SEGKPM_SUPPORT */ 289 290 /* segkpm stubs */ 291 292 /*ARGSUSED*/ 293 int segkpm_create(struct seg *seg, void *argsp) { return (0); } 294 295 /* ARGSUSED */ 296 faultcode_t 297 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len, 298 enum fault_type type, enum seg_rw rw) 299 { 300 return ((faultcode_t)0); 301 } 302 303 /* ARGSUSED */ 304 caddr_t segkpm_create_va(u_offset_t off) { return (NULL); } 305 306 /* ARGSUSED */ 307 void segkpm_mapout_validkpme(struct kpme *kpme) {} 308 309 static void 310 segkpm_badop() {} 311 312 #endif /* SEGKPM_SUPPORT */ 313 314 static int 315 segkpm_notsup() 316 { 317 return (ENOTSUP); 318 }