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         SEGKPM_BADOP(int),      /* dup */
 115         SEGKPM_BADOP(int),      /* unmap */
 116         SEGKPM_BADOP(void),     /* free */
 117         segkpm_fault,
 118         SEGKPM_BADOP(int),      /* faulta */
 119         SEGKPM_BADOP(int),      /* setprot */
 120         SEGKPM_BADOP(int),      /* checkprot */
 121         SEGKPM_BADOP(int),      /* kluster */
 122         SEGKPM_BADOP(size_t),   /* swapout */
 123         SEGKPM_BADOP(int),      /* sync */
 124         SEGKPM_BADOP(size_t),   /* incore */
 125         SEGKPM_BADOP(int),      /* lockop */
 126         SEGKPM_BADOP(int),      /* getprot */
 127         SEGKPM_BADOP(u_offset_t), /* getoffset */
 128         SEGKPM_BADOP(int),      /* gettype */
 129         SEGKPM_BADOP(int),      /* getvp */
 130         SEGKPM_BADOP(int),      /* advise */
 131         segkpm_dump,            /* dump */
 132         SEGKPM_NOTSUP,          /* pagelock */
 133         SEGKPM_BADOP(int),      /* setpgsz */
 134         SEGKPM_BADOP(int),      /* getmemid */
 135         SEGKPM_BADOP(lgrp_mem_policy_info_t *), /* getpolicy */
 136         segkpm_capable,         /* capable */
 137         seg_inherit_notsup      /* inherit */
 138 };
 139 
 140 /*
 141  * kpm_pgsz and kpm_pgshft are set by platform layer.
 142  */
 143 size_t          kpm_pgsz;       /* kpm page size */
 144 uint_t          kpm_pgshft;     /* kpm page shift */
 145 u_offset_t      kpm_pgoff;      /* kpm page offset mask */
 146 uint_t          kpmp2pshft;     /* kpm page to page shift */
 147 pgcnt_t         kpmpnpgs;       /* how many pages per kpm page */
 148 
 149 
 150 #ifdef  SEGKPM_SUPPORT
 151 
 152 int
 153 segkpm_create(struct seg *seg, void *argsp)
 154 {
 155         struct segkpm_data *skd;
 156         struct segkpm_crargs *b = (struct segkpm_crargs *)argsp;
 157         ushort_t *p;
 158         int i, j;
 159 
 160         ASSERT(seg->s_as && RW_WRITE_HELD(&seg->s_as->a_lock));
 161         ASSERT(btokpmp(seg->s_size) >= 1 &&
 162             kpmpageoff((uintptr_t)seg->s_base) == 0 &&
 163             kpmpageoff((uintptr_t)seg->s_base + seg->s_size) == 0);
 164 
 165         skd = kmem_zalloc(sizeof (struct segkpm_data), KM_SLEEP);
 166 
 167         seg->s_data = (void *)skd;
 168         seg->s_ops = &segkpm_ops;
 169         skd->skd_prot = b->prot;
 170 
 171         /*
 172          * (1) Segkpm virtual addresses are based on physical adresses.
 173          * From this and in opposite to other segment drivers it is
 174          * often required to allocate a page first to be able to
 175          * calculate the final segkpm virtual address.
 176          * (2) Page  allocation is done by calling page_create_va(),
 177          * one important input argument is a virtual address (also
 178          * expressed by the "va" in the function name). This function
 179          * is highly optimized to select the right page for an optimal
 180          * processor and platform support (e.g. virtual addressed
 181          * caches (VAC), physical addressed caches, NUMA).
 182          *
 183          * Because of (1) the approach is to generate a faked virtual
 184          * address for calling page_create_va(). In order to exploit
 185          * the abilities of (2), especially to utilize the cache
 186          * hierarchy (3) and to avoid VAC alias conflicts (4) the
 187          * selection has to be done carefully. For each virtual color
 188          * a separate counter is provided (4). The count values are
 189          * used for the utilization of all cache lines (3) and are
 190          * corresponding to the cache bins.
 191          */
 192         skd->skd_nvcolors = b->nvcolors;
 193 
 194         p = skd->skd_va_select =
 195             kmem_zalloc(NCPU * b->nvcolors * sizeof (ushort_t), KM_SLEEP);
 196 
 197         for (i = 0; i < NCPU; i++)
 198                 for (j = 0; j < b->nvcolors; j++, p++)
 199                         *p = j;
 200 
 201         return (0);
 202 }
 203 
 204 /*
 205  * This routine is called via a machine specific fault handling
 206  * routine.
 207  */
 208 /* ARGSUSED */
 209 faultcode_t
 210 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
 211         enum fault_type type, enum seg_rw rw)
 212 {
 213         ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
 214 
 215         switch (type) {
 216         case F_INVAL:
 217                 return (hat_kpm_fault(hat, addr));
 218         case F_SOFTLOCK:
 219         case F_SOFTUNLOCK:
 220                 return (0);
 221         default:
 222                 return (FC_NOSUPPORT);
 223         }
 224         /*NOTREACHED*/
 225 }
 226 
 227 #define addr_to_vcolor(addr, vcolors) \
 228         ((int)(((uintptr_t)(addr) & ((vcolors << PAGESHIFT) - 1)) >> PAGESHIFT))
 229 
 230 /*
 231  * Create a virtual address that can be used for invocations of
 232  * page_create_va. Goal is to utilize the cache hierarchy (round
 233  * robin bins) and to select the right color for virtual indexed
 234  * caches. It isn't exact since we also increment the bin counter
 235  * when the caller uses VOP_GETPAGE and gets a hit in the page
 236  * cache, but we keep the bins turning for cache distribution
 237  * (see also segkpm_create block comment).
 238  */
 239 caddr_t
 240 segkpm_create_va(u_offset_t off)
 241 {
 242         int vcolor;
 243         ushort_t *p;
 244         struct segkpm_data *skd = (struct segkpm_data *)segkpm->s_data;
 245         int nvcolors = skd->skd_nvcolors;
 246         caddr_t va;
 247 
 248         vcolor = (nvcolors > 1) ? addr_to_vcolor(off, nvcolors) : 0;
 249         p = &skd->skd_va_select[(CPU->cpu_id * nvcolors) + vcolor];
 250         va = (caddr_t)ptob(*p);
 251 
 252         atomic_add_16(p, nvcolors);
 253 
 254         return (va);
 255 }
 256 
 257 /*
 258  * Unload mapping if the instance has an active kpm mapping.
 259  */
 260 void
 261 segkpm_mapout_validkpme(struct kpme *kpme)
 262 {
 263         caddr_t vaddr;
 264         page_t *pp;
 265 
 266 retry:
 267         if ((pp = kpme->kpe_page) == NULL) {
 268                 return;
 269         }
 270 
 271         if (page_lock(pp, SE_SHARED, (kmutex_t *)NULL, P_RECLAIM) == 0)
 272                 goto retry;
 273 
 274         /*
 275          * Check if segkpm mapping is not unloaded in the meantime
 276          */
 277         if (kpme->kpe_page == NULL) {
 278                 page_unlock(pp);
 279                 return;
 280         }
 281 
 282         vaddr = hat_kpm_page2va(pp, 1);
 283         hat_kpm_mapout(pp, kpme, vaddr);
 284         page_unlock(pp);
 285 }
 286 
 287 static void
 288 segkpm_badop()
 289 {
 290         panic("segkpm_badop");
 291 }
 292 
 293 #else   /* SEGKPM_SUPPORT */
 294 
 295 /* segkpm stubs */
 296 
 297 /*ARGSUSED*/
 298 int segkpm_create(struct seg *seg, void *argsp) { return (0); }
 299 
 300 /* ARGSUSED */
 301 faultcode_t
 302 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
 303         enum fault_type type, enum seg_rw rw)
 304 {
 305         return ((faultcode_t)0);
 306 }
 307 
 308 /* ARGSUSED */
 309 caddr_t segkpm_create_va(u_offset_t off) { return (NULL); }
 310 
 311 /* ARGSUSED */
 312 void segkpm_mapout_validkpme(struct kpme *kpme) {}
 313 
 314 static void
 315 segkpm_badop() {}
 316 
 317 #endif  /* SEGKPM_SUPPORT */
 318 
 319 static int
 320 segkpm_notsup()
 321 {
 322         return (ENOTSUP);
 323 }
 324 
 325 /*
 326  * segkpm pages are not dumped, so we just return
 327  */
 328 /*ARGSUSED*/
 329 static void
 330 segkpm_dump(struct seg *seg)
 331 {}
 332 
 333 /*
 334  * We claim to have no special capabilities.
 335  */
 336 /*ARGSUSED*/
 337 static int
 338 segkpm_capable(struct seg *seg, segcapability_t capability)
 339 {
 340         return (0);
 341 }