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 void     segkpm_badop(void);
 107 static int      segkpm_notsup(void);
 108 static int      segkpm_capable(struct seg *, segcapability_t);
 109 
 110 #define SEGKPM_BADOP(t) (t(*)())segkpm_badop
 111 #define SEGKPM_NOTSUP   (int(*)())segkpm_notsup
 112 
 113 static struct seg_ops segkpm_ops = {
 114         .dup            = SEGKPM_BADOP(int),
 115         .unmap          = SEGKPM_BADOP(int),
 116         .free           = SEGKPM_BADOP(void),
 117         .fault          = segkpm_fault,
 118         .faulta         = SEGKPM_BADOP(int),
 119         .setprot        = SEGKPM_BADOP(int),
 120         .checkprot      = SEGKPM_BADOP(int),
 121         .kluster        = SEGKPM_BADOP(int),
 122         .swapout        = SEGKPM_BADOP(size_t),
 123         .sync           = SEGKPM_BADOP(int),
 124         .incore         = SEGKPM_BADOP(size_t),
 125         .lockop         = SEGKPM_BADOP(int),
 126         .getprot        = SEGKPM_BADOP(int),
 127         .getoffset      = SEGKPM_BADOP(u_offset_t),
 128         .gettype        = SEGKPM_BADOP(int),
 129         .getvp          = SEGKPM_BADOP(int),
 130         .advise         = SEGKPM_BADOP(int),
 131         .dump           = segkpm_dump,
 132         .pagelock       = SEGKPM_NOTSUP,
 133         .setpagesize    = SEGKPM_BADOP(int),
 134         .getmemid       = SEGKPM_BADOP(int),
 135         .getpolicy      = SEGKPM_BADOP(lgrp_mem_policy_info_t *),
 136         .capable        = segkpm_capable,
 137 };
 138 
 139 /*
 140  * kpm_pgsz and kpm_pgshft are set by platform layer.
 141  */
 142 size_t          kpm_pgsz;       /* kpm page size */
 143 uint_t          kpm_pgshft;     /* kpm page shift */
 144 u_offset_t      kpm_pgoff;      /* kpm page offset mask */
 145 uint_t          kpmp2pshft;     /* kpm page to page shift */
 146 pgcnt_t         kpmpnpgs;       /* how many pages per kpm page */
 147 
 148 
 149 #ifdef  SEGKPM_SUPPORT
 150 
 151 int
 152 segkpm_create(struct seg *seg, void *argsp)
 153 {
 154         struct segkpm_data *skd;
 155         struct segkpm_crargs *b = (struct segkpm_crargs *)argsp;
 156         ushort_t *p;
 157         int i, j;
 158 
 159         ASSERT(seg->s_as && RW_WRITE_HELD(&seg->s_as->a_lock));
 160         ASSERT(btokpmp(seg->s_size) >= 1 &&
 161             kpmpageoff((uintptr_t)seg->s_base) == 0 &&
 162             kpmpageoff((uintptr_t)seg->s_base + seg->s_size) == 0);
 163 
 164         skd = kmem_zalloc(sizeof (struct segkpm_data), KM_SLEEP);
 165 
 166         seg->s_data = (void *)skd;
 167         seg->s_ops = &segkpm_ops;
 168         skd->skd_prot = b->prot;
 169 
 170         /*
 171          * (1) Segkpm virtual addresses are based on physical adresses.
 172          * From this and in opposite to other segment drivers it is
 173          * often required to allocate a page first to be able to
 174          * calculate the final segkpm virtual address.
 175          * (2) Page  allocation is done by calling page_create_va(),
 176          * one important input argument is a virtual address (also
 177          * expressed by the "va" in the function name). This function
 178          * is highly optimized to select the right page for an optimal
 179          * processor and platform support (e.g. virtual addressed
 180          * caches (VAC), physical addressed caches, NUMA).
 181          *
 182          * Because of (1) the approach is to generate a faked virtual
 183          * address for calling page_create_va(). In order to exploit
 184          * the abilities of (2), especially to utilize the cache
 185          * hierarchy (3) and to avoid VAC alias conflicts (4) the
 186          * selection has to be done carefully. For each virtual color
 187          * a separate counter is provided (4). The count values are
 188          * used for the utilization of all cache lines (3) and are
 189          * corresponding to the cache bins.
 190          */
 191         skd->skd_nvcolors = b->nvcolors;
 192 
 193         p = skd->skd_va_select =
 194             kmem_zalloc(NCPU * b->nvcolors * sizeof (ushort_t), KM_SLEEP);
 195 
 196         for (i = 0; i < NCPU; i++)
 197                 for (j = 0; j < b->nvcolors; j++, p++)
 198                         *p = j;
 199 
 200         return (0);
 201 }
 202 
 203 /*
 204  * This routine is called via a machine specific fault handling
 205  * routine.
 206  */
 207 /* ARGSUSED */
 208 faultcode_t
 209 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
 210         enum fault_type type, enum seg_rw rw)
 211 {
 212         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 213 
 214         switch (type) {
 215         case F_INVAL:
 216                 return (hat_kpm_fault(hat, addr));
 217         case F_SOFTLOCK:
 218         case F_SOFTUNLOCK:
 219                 return (0);
 220         default:
 221                 return (FC_NOSUPPORT);
 222         }
 223         /*NOTREACHED*/
 224 }
 225 
 226 #define addr_to_vcolor(addr, vcolors) \
 227         ((int)(((uintptr_t)(addr) & ((vcolors << PAGESHIFT) - 1)) >> PAGESHIFT))
 228 
 229 /*
 230  * Create a virtual address that can be used for invocations of
 231  * page_create_va. Goal is to utilize the cache hierarchy (round
 232  * robin bins) and to select the right color for virtual indexed
 233  * caches. It isn't exact since we also increment the bin counter
 234  * when the caller uses VOP_GETPAGE and gets a hit in the page
 235  * cache, but we keep the bins turning for cache distribution
 236  * (see also segkpm_create block comment).
 237  */
 238 caddr_t
 239 segkpm_create_va(u_offset_t off)
 240 {
 241         int vcolor;
 242         ushort_t *p;
 243         struct segkpm_data *skd = (struct segkpm_data *)segkpm->s_data;
 244         int nvcolors = skd->skd_nvcolors;
 245         caddr_t va;
 246 
 247         vcolor = (nvcolors > 1) ? addr_to_vcolor(off, nvcolors) : 0;
 248         p = &skd->skd_va_select[(CPU->cpu_id * nvcolors) + vcolor];
 249         va = (caddr_t)ptob(*p);
 250 
 251         atomic_add_16(p, nvcolors);
 252 
 253         return (va);
 254 }
 255 
 256 /*
 257  * Unload mapping if the instance has an active kpm mapping.
 258  */
 259 void
 260 segkpm_mapout_validkpme(struct kpme *kpme)
 261 {
 262         caddr_t vaddr;
 263         page_t *pp;
 264 
 265 retry:
 266         if ((pp = kpme->kpe_page) == NULL) {
 267                 return;
 268         }
 269 
 270         if (page_lock(pp, SE_SHARED, (kmutex_t *)NULL, P_RECLAIM) == 0)
 271                 goto retry;
 272 
 273         /*
 274          * Check if segkpm mapping is not unloaded in the meantime
 275          */
 276         if (kpme->kpe_page == NULL) {
 277                 page_unlock(pp);
 278                 return;
 279         }
 280 
 281         vaddr = hat_kpm_page2va(pp, 1);
 282         hat_kpm_mapout(pp, kpme, vaddr);
 283         page_unlock(pp);
 284 }
 285 
 286 static void
 287 segkpm_badop()
 288 {
 289         panic("segkpm_badop");
 290 }
 291 
 292 #else   /* SEGKPM_SUPPORT */
 293 
 294 /* segkpm stubs */
 295 
 296 /*ARGSUSED*/
 297 int segkpm_create(struct seg *seg, void *argsp) { return (0); }
 298 
 299 /* ARGSUSED */
 300 faultcode_t
 301 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
 302         enum fault_type type, enum seg_rw rw)
 303 {
 304         return ((faultcode_t)0);
 305 }
 306 
 307 /* ARGSUSED */
 308 caddr_t segkpm_create_va(u_offset_t off) { return (NULL); }
 309 
 310 /* ARGSUSED */
 311 void segkpm_mapout_validkpme(struct kpme *kpme) {}
 312 
 313 static void
 314 segkpm_badop() {}
 315 
 316 #endif  /* SEGKPM_SUPPORT */
 317 
 318 static int
 319 segkpm_notsup()
 320 {
 321         return (ENOTSUP);
 322 }
 323 
 324 /*
 325  * segkpm pages are not dumped, so we just return
 326  */
 327 /*ARGSUSED*/
 328 static void
 329 segkpm_dump(struct seg *seg)
 330 {}
 331 
 332 /*
 333  * We claim to have no special capabilities.
 334  */
 335 /*ARGSUSED*/
 336 static int
 337 segkpm_capable(struct seg *seg, segcapability_t capability)
 338 {
 339         return (0);
 340 }