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