osfmk/ppc/savearea.c

   1 /*
   2  * Copyright (c) 2000-2004 Apple Computer, Inc. All rights reserved.
   3  *
   4  * @APPLE_LICENSE_HEADER_START@
   5  *
   6  * The contents of this file constitute Original Code as defined in and
   7  * are subject to the Apple Public Source License Version 1.1 (the
   8  * "License").  You may not use this file except in compliance with the
   9  * License.  Please obtain a copy of the License at
  10  * http://www.apple.com/publicsource and read it before using this file.
  11  *
  12  * This Original Code and all software distributed under the License are
  13  * distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  14  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  15  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE OR NON-INFRINGEMENT.  Please see the
  17  * License for the specific language governing rights and limitations
  18  * under the License.
  19  *
  20  * @APPLE_LICENSE_HEADER_END@
  21  */
  22 /*
  23  *      This file is used to maintain the exception save areas
  24  *
  25  */
  26
  27 #include <debug.h>
  28 #include <mach_kgdb.h>
  29 #include <mach_vm_debug.h>
  30
  31 #include <kern/thread.h>
  32 #include <mach/vm_attributes.h>
  33 #include <mach/vm_param.h>
  34 #include <vm/vm_kern.h>
  35 #include <vm/vm_map.h>
  36 #include <vm/vm_page.h>
  37 #include <mach/ppc/thread_status.h>
  38 #include <kern/spl.h>
  39 #include <kern/simple_lock.h>
  40
  41 #include <kern/misc_protos.h>
  42 #include <ppc/misc_protos.h>
  43 #include <ppc/proc_reg.h>
  44 #include <ppc/mem.h>
  45 #include <ppc/pmap.h>
  46 #include <ppc/Firmware.h>
  47 #include <ppc/mappings.h>
  48 #include <ppc/exception.h>
  49 #include <ppc/savearea.h>
  50 #include <ddb/db_output.h>
  51
  52
  53 extern struct   Saveanchor saveanchor;                                                  /* Aliged savearea anchor */
  54 struct Saveanchor backpocket;                                                                   /* Emergency saveareas */
  55 unsigned int    debsave0 = 0;                                                                   /* Debug flag */
  56 unsigned int    backchain = 0;                                                                  /* Debug flag */
  57
  58 /*
  59  *              These routines keep track of exception save areas and keeps the count within specific limits.  If there are
  60  *              too few, more are allocated, too many, and they are released. This savearea is where the PCBs are
  61  *              stored.  They never span a page boundary and are referenced by both virtual and real addresses.
  62  *              Within the interrupt vectors, the real address is used because at that level, no exceptions
  63  *              can be tolerated.  Save areas can be dynamic or permanent.  Permanant saveareas are allocated
  64  *              at boot time and must be in place before any type of exception occurs.  These are never released,
  65  *              and the number is based upon some arbitrary (yet to be determined) amount times the number of
  66  *              processors.  This represents the minimum number required to process a total system failure without
  67  *              destroying valuable and ever-so-handy system debugging information.
  68  *
  69  *              We keep two global free lists (the savearea free pool and the savearea free list) and one local
  70  *              list per processor.
  71  *
  72  *              The local lists are small and require no locked access.  They are chained using physical addresses
  73  *              and no interruptions are allowed when adding to or removing from the list. Also known as the
  74  *              qfret list. This list is local to a processor and is intended for use only by very low level
  75  *              context handling code.
  76  *
  77  *              The savearea free list is a medium size list that is globally accessible.  It is updated
  78  *              while holding a simple lock. The length of time that the lock is held is kept short.  The
  79  *              longest period of time is when the list is trimmed. Like the qfret lists, this is chained physically
  80  *              and must be accessed with translation and interruptions disabled. This is where the bulk
  81  *              of the free entries are located.
  82  *
  83  *              The saveareas are allocated from full pages.  A pool element is marked
  84  *              with an allocation map that shows which "slots" are free.  These pages are allocated via the
  85  *              normal kernel memory allocation functions. Queueing is with physical addresses.  The enqueue,
  86  *              dequeue, and search for free blocks is done under free list lock.
  87  *              only if there are empty slots in it.
  88  *
  89  *              Saveareas that are counted as "in use" once they are removed from the savearea free list.
  90  *              This means that all areas on the local qfret list are considered in use.
  91  *
  92  *              There are two methods of obtaining a savearea.  The save_get function (which is also inlined
  93  *              in the low-level exception handler) attempts to get an area from the local qfret list.  This is
  94  *              done completely without locks.  If qfret is exahusted (or maybe just too low) an area is allocated
  95  *              from the savearea free list. If the free list is empty, we install the back pocket areas and
  96  *              panic.
  97  *
  98  *              The save_alloc function is designed to be called by high level routines, e.g., thread creation,
  99  *              etc.  It will allocate from the free list.  After allocation, it will compare the free count
 100  *              to the target value.  If outside of the range, it will adjust the size either upwards or
 101  *              downwards.
 102  *
 103  *              If we need to shrink the list, it will be trimmed to the target size and unlocked.  The code
 104  *              will walk the chain and return each savearea to its pool page.  If a pool page becomes
 105  *              completely empty, it is dequeued from the free pool list and enqueued (atomic queue
 106  *              function) to be released.
 107  *
 108  *              Once the trim list is finished, the pool release queue is checked to see if there are pages
 109  *              waiting to be released. If so, they are released one at a time.
 110  *
 111  *              If the free list needed to be grown rather than shrunken, we will first attempt to recover
 112  *              a page from the pending release queue (built when we trim the free list).  If we find one,
 113  *              it is allocated, otherwise, a page of kernel memory is allocated.  This loops until there are
 114  *              enough free saveareas.
 115  *
 116  */
 117
 118
 119
 120 /*
 121  *              Allocate our initial context save areas.  As soon as we do this,
 122  *              we can take an interrupt. We do the saveareas here, 'cause they're guaranteed
 123  *              to be at least page aligned.
 124  *
 125  *              Note: these initial saveareas are all to be allocated from V=R, less than 4GB
 126  *              space.
 127  */
 128
 129
 130 void savearea_init(vm_offset_t addr) {
 131
 132         savearea_comm   *savec;
 133         vm_offset_t     save;
 134         int i;
 135
 136
 137         saveanchor.savetarget   = InitialSaveTarget;            /* Initial target value */
 138         saveanchor.saveinuse    = 0;                                            /* Number of areas in use */
 139
 140         saveanchor.savefree    = 0;                                                     /* Remember the start of the free chain */
 141         saveanchor.savefreecnt = 0;                                                     /* Remember the length */
 142         saveanchor.savepoolfwd = (addr64_t)&saveanchor;         /* Remember pool forward */
 143         saveanchor.savepoolbwd = (addr64_t)&saveanchor;         /* Remember pool backward */
 144
 145         save =  addr;                                                                           /* Point to the whole block of blocks */
 146
 147 /*
 148  *      First we allocate the back pocket in case of emergencies
 149  */
 150
 151
 152         for(i=0; i < BackPocketSaveBloks; i++) {                        /* Initialize the back pocket saveareas */
 153
 154                 savec = (savearea_comm *)save;                                  /* Get the control area for this one */
 155
 156                 savec->sac_alloc = 0;                                                   /* Mark it allocated */
 157                 savec->sac_vrswap = 0;                                                  /* V=R, so the translation factor is 0 */
 158                 savec->sac_flags = sac_perm;                                    /* Mark it permanent */
 159                 savec->sac_flags |= 0x0000EE00;                                 /* Debug eyecatcher */
 160                 save_queue((uint32_t)savec >> 12);                              /* Add page to savearea lists */
 161                 save += PAGE_SIZE;                                                              /* Jump up to the next one now */
 162
 163         }
 164
 165         backpocket = saveanchor;                                                        /* Save this for emergencies */
 166
 167
 168 /*
 169  *      We've saved away the back pocket savearea info, so reset it all and
 170  *      now allocate for real
 171  */
 172
 173
 174         saveanchor.savefree = 0;                                                        /* Remember the start of the free chain */
 175         saveanchor.savefreecnt = 0;                                                     /* Remember the length */
 176         saveanchor.saveadjust = 0;                                                      /* Set none needed yet */
 177         saveanchor.savepoolfwd = (addr64_t)&saveanchor;         /* Remember pool forward */
 178         saveanchor.savepoolbwd = (addr64_t)&saveanchor;         /* Remember pool backward */
 179
 180         for(i=0; i < InitialSaveBloks; i++) {                           /* Initialize the saveareas */
 181
 182                 savec = (savearea_comm *)save;                                  /* Get the control area for this one */
 183
 184                 savec->sac_alloc = 0;                                                   /* Mark it allocated */
 185                 savec->sac_vrswap = 0;                                                  /* V=R, so the translation factor is 0 */
 186                 savec->sac_flags = sac_perm;                                    /* Mark it permanent */
 187                 savec->sac_flags |= 0x0000EE00;                                 /* Debug eyecatcher */
 188                 save_queue((uint32_t)savec >> 12);                              /* Add page to savearea lists */
 189                 save += PAGE_SIZE;                                                              /* Jump up to the next one now */
 190
 191         }
 192
 193 /*
 194  *      We now have a free list that has our initial number of entries
 195  *      The local qfret lists is empty.  When we call save_get below it will see that
 196  *      the local list is empty and fill it for us.
 197  *
 198  *      It is ok to call save_get here because all initial saveareas are V=R in less
 199  *  than 4GB space, so 32-bit addressing is ok.
 200  *
 201  */
 202
 203 /*
 204  * This will populate the local list  and get the first one for the system
 205  */
 206         getPerProc()->next_savearea = (vm_offset_t)save_get();
 207
 208 /*
 209  *      The system is now able to take interruptions
 210  */
 211         return;
 212 }
 213
 214
 215
 216
 217 /*
 218  *              Obtains a savearea.  If the free list needs size adjustment it happens here.
 219  *              Don't actually allocate the savearea until after the adjustment is done.
 220  */
 221
 222 struct savearea *save_alloc(void) {                                             /* Reserve a save area */
 223
 224
 225         if(saveanchor.saveadjust) save_adjust();                        /* If size need adjustment, do it now */
 226
 227         return save_get();                                                                      /* Pass the baby... */
 228 }
 229
 230
 231 /*
 232  *              This routine releases a save area to the free queue.  If after that, we have more than our maximum target,
 233  *              we start releasing what we can until we hit the normal target.
 234  */
 235
 236
 237
 238 void save_release(struct savearea *save) {                              /* Release a save area */
 239
 240         save_ret(save);                                                                         /* Return a savearea to the free list */
 241
 242         if(saveanchor.saveadjust) save_adjust();                        /* Adjust the savearea free list and pool size if needed */
 243
 244         return;
 245
 246 }
 247
 248
 249 /*
 250  *              Adjusts the size of the free list.  Can either release or allocate full pages
 251  *              of kernel memory.  This can block.
 252  *
 253  *              Note that we will only run one adjustment and the amount needed may change
 254  *              while we are executing.
 255  *
 256  *              Calling this routine is triggered by saveanchor.saveadjust.  This value is always calculated just before
 257  *              we unlock the saveanchor lock (this keeps it pretty accurate).  If the total of savefreecnt and saveinuse
 258  *              is within the hysteresis range, it is set to 0.  If outside, it is set to the number needed to bring
 259  *              the total to the target value.  Note that there is a minimum size to the free list (FreeListMin) and if
 260  *              savefreecnt falls below that, saveadjust is set to the number needed to bring it to that.
 261  */
 262
 263
 264 void save_adjust(void) {
 265
 266         savearea_comm   *sctl, *sctlnext, *freepage;
 267         kern_return_t ret;
 268         uint64_t vtopmask;
 269         ppnum_t physpage;
 270
 271         if(saveanchor.saveadjust < 0)                                   {       /* Do we need to adjust down? */
 272
 273                 sctl = (savearea_comm *)save_trim_free();               /* Trim list to the need count, return start of trim list */
 274
 275                 while(sctl) {                                                                   /* Release the free pages back to the kernel */
 276                         sctlnext = CAST_DOWN(savearea_comm *, sctl->save_prev); /* Get next in list */
 277                         kmem_free(kernel_map, (vm_offset_t) sctl, PAGE_SIZE);   /* Release the page */
 278                         sctl = sctlnext;                                                        /* Chain onwards */
 279                 }
 280         }
 281         else {                                                                                          /* We need more... */
 282
 283                 if(save_recover()) return;                                              /* If we can recover enough from the pool, return */
 284
 285                 while(saveanchor.saveadjust > 0) {                              /* Keep going until we have enough */
 286
 287                         ret = kmem_alloc_wired(kernel_map, (vm_offset_t *)&freepage, PAGE_SIZE);        /* Get a page for free pool */
 288                         if(ret != KERN_SUCCESS) {                                       /* Did we get some memory? */
 289                                 panic("Whoops...  Not a bit of wired memory left for saveareas\n");
 290                         }
 291
 292                         physpage = pmap_find_phys(kernel_pmap, (vm_offset_t)freepage);  /* Find physical page */
 293                         if(!physpage) {                                                         /* See if we actually have this mapped*/
 294                                 panic("save_adjust: wired page not mapped - va = %08X\n", freepage);    /* Die */
 295                         }
 296
 297                         bzero((void *)freepage, PAGE_SIZE);                     /* Clear it all to zeros */
 298                         freepage->sac_alloc = 0;                                        /* Mark all entries taken */
 299                         freepage->sac_vrswap = ((uint64_t)physpage << 12) ^ (uint64_t)((uintptr_t)freepage);    /* XOR to calculate conversion mask */
 300
 301                         freepage->sac_flags |= 0x0000EE00;                      /* Set debug eyecatcher */
 302
 303                         save_queue(physpage);                                           /* Add all saveareas on page to free list */
 304                 }
 305         }
 306 }
 307
 308 /*
 309  *              Fake up information to make the saveareas look like a zone
 310  */
 311 void
 312 save_fake_zone_info(int *count, vm_size_t *cur_size, vm_size_t *max_size, vm_size_t *elem_size,
 313                     vm_size_t *alloc_size, int *collectable, int *exhaustable)
 314 {
 315         *count      = saveanchor.saveinuse;
 316         *cur_size   = (saveanchor.savefreecnt + saveanchor.saveinuse) * (PAGE_SIZE / sac_cnt);
 317         *max_size   = saveanchor.savemaxcount * (PAGE_SIZE / sac_cnt);
 318         *elem_size  = sizeof(savearea);
 319         *alloc_size = PAGE_SIZE;
 320         *collectable = 1;
 321         *exhaustable = 0;
 322 }
 323
 324