osfmk/ppc/savearea.c

   1 /*
   2  * Copyright (c) 2000 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 <cpus.h>
  28 #include <debug.h>
  29 #include <mach_kgdb.h>
  30 #include <mach_vm_debug.h>
  31
  32 #include <kern/thread.h>
  33 #include <mach/vm_attributes.h>
  34 #include <mach/vm_param.h>
  35 #include <vm/vm_kern.h>
  36 #include <vm/vm_map.h>
  37 #include <vm/vm_page.h>
  38 #include <mach/ppc/thread_status.h>
  39 #include <kern/spl.h>
  40 #include <kern/simple_lock.h>
  41
  42 #include <kern/misc_protos.h>
  43 #include <ppc/misc_protos.h>
  44 #include <ppc/proc_reg.h>
  45 #include <ppc/mem.h>
  46 #include <ppc/pmap.h>
  47 #include <ppc/Firmware.h>
  48 #include <ppc/mappings.h>
  49 #include <ppc/exception.h>
  50 #include <ppc/savearea.h>
  51 #include <ddb/db_output.h>
  52
  53
  54 extern struct   Saveanchor saveanchor;                                                  /* Aliged savearea anchor */
  55 struct Saveanchor backpocket;                                                                   /* Emergency saveareas */
  56 unsigned int    debsave0 = 0;                                                                   /* Debug flag */
  57 unsigned int    backchain = 0;                                                                  /* Debug flag */
  58
  59 /*
  60  *              These routines keep track of exception save areas and keeps the count within specific limits.  If there are
  61  *              too few, more are allocated, too many, and they are released. This savearea is where the PCBs are
  62  *              stored.  They never span a page boundary and are referenced by both virtual and real addresses.
  63  *              Within the interrupt vectors, the real address is used because at that level, no exceptions
  64  *              can be tolerated.  Save areas can be dynamic or permanent.  Permanant saveareas are allocated
  65  *              at boot time and must be in place before any type of exception occurs.  These are never released,
  66  *              and the number is based upon some arbitrary (yet to be determined) amount times the number of
  67  *              processors.  This represents the minimum number required to process a total system failure without
  68  *              destroying valuable and ever-so-handy system debugging information.
  69  *
  70  *              We keep two global free lists (the savearea free pool and the savearea free list) and one local
  71  *              list per processor.
  72  *
  73  *              The local lists are small and require no locked access.  They are chained using physical addresses
  74  *              and no interruptions are allowed when adding to or removing from the list. Also known as the
  75  *              qfret list. This list is local to a processor and is intended for use only by very low level
  76  *              context handling code.
  77  *
  78  *              The savearea free list is a medium size list that is globally accessible.  It is updated
  79  *              while holding a simple lock. The length of time that the lock is held is kept short.  The
  80  *              longest period of time is when the list is trimmed. Like the qfret lists, this is chained physically
  81  *              and must be accessed with translation and interruptions disabled. This is where the bulk
  82  *              of the free entries are located.
  83  *
  84  *              The saveareas are allocated from full pages.  A pool element is marked
  85  *              with an allocation map that shows which "slots" are free.  These pages are allocated via the
  86  *              normal kernel memory allocation functions. Queueing is with physical addresses.  The enqueue,
  87  *              dequeue, and search for free blocks is done under free list lock.
  88  *              only if there are empty slots in it.
  89  *
  90  *              Saveareas that are counted as "in use" once they are removed from the savearea free list.
  91  *              This means that all areas on the local qfret list are considered in use.
  92  *
  93  *              There are two methods of obtaining a savearea.  The save_get function (which is also inlined
  94  *              in the low-level exception handler) attempts to get an area from the local qfret list.  This is
  95  *              done completely without locks.  If qfret is exahusted (or maybe just too low) an area is allocated
  96  *              from the savearea free list. If the free list is empty, we install the back pocket areas and
  97  *              panic.
  98  *
  99  *              The save_alloc function is designed to be called by high level routines, e.g., thread creation,
 100  *              etc.  It will allocate from the free list.  After allocation, it will compare the free count
 101  *              to the target value.  If outside of the range, it will adjust the size either upwards or
 102  *              downwards.
 103  *
 104  *              If we need to shrink the list, it will be trimmed to the target size and unlocked.  The code
 105  *              will walk the chain and return each savearea to its pool page.  If a pool page becomes
 106  *              completely empty, it is dequeued from the free pool list and enqueued (atomic queue
 107  *              function) to be released.
 108  *
 109  *              Once the trim list is finished, the pool release queue is checked to see if there are pages
 110  *              waiting to be released. If so, they are released one at a time.
 111  *
 112  *              If the free list needed to be grown rather than shrunken, we will first attempt to recover
 113  *              a page from the pending release queue (built when we trim the free list).  If we find one,
 114  *              it is allocated, otherwise, a page of kernel memory is allocated.  This loops until there are
 115  *              enough free saveareas.
 116  *
 117  */
 118
 119
 120
 121 /*
 122  *              Allocate our initial context save areas.  As soon as we do this,
 123  *              we can take an interrupt. We do the saveareas here, 'cause they're guaranteed
 124  *              to be at least page aligned.
 125  *
 126  *              Note: these initial saveareas are all to be allocated from V=R, less than 4GB
 127  *              space.
 128  */
 129
 130
 131 void savearea_init(vm_offset_t addr) {
 132
 133         savearea_comm   *savec;
 134         vm_offset_t     save;
 135         int i;
 136
 137
 138         saveanchor.savetarget   = InitialSaveTarget;            /* Initial target value */
 139         saveanchor.saveinuse    = 0;                                            /* Number of areas in use */
 140
 141         saveanchor.savefree    = 0;                                                     /* Remember the start of the free chain */
 142         saveanchor.savefreecnt = 0;                                                     /* Remember the length */
 143         saveanchor.savepoolfwd = (addr64_t)&saveanchor;         /* Remember pool forward */
 144         saveanchor.savepoolbwd = (addr64_t)&saveanchor;         /* Remember pool backward */
 145
 146         save =  addr;                                                                           /* Point to the whole block of blocks */
 147
 148 /*
 149  *      First we allocate the back pocket in case of emergencies
 150  */
 151
 152
 153         for(i=0; i < BackPocketSaveBloks; i++) {                        /* Initialize the back pocket saveareas */
 154
 155                 savec = (savearea_comm *)save;                                  /* Get the control area for this one */
 156
 157                 savec->sac_alloc = 0;                                                   /* Mark it allocated */
 158                 savec->sac_vrswap = 0;                                                  /* V=R, so the translation factor is 0 */
 159                 savec->sac_flags = sac_perm;                                    /* Mark it permanent */
 160                 savec->sac_flags |= 0x0000EE00;                                 /* Debug eyecatcher */
 161                 save_queue((uint32_t)savec >> 12);                              /* Add page to savearea lists */
 162                 save += PAGE_SIZE;                                                              /* Jump up to the next one now */
 163
 164         }
 165
 166         backpocket = saveanchor;                                                        /* Save this for emergencies */
 167
 168
 169 /*
 170  *      We've saved away the back pocket savearea info, so reset it all and
 171  *      now allocate for real
 172  */
 173
 174
 175         saveanchor.savefree = 0;                                                        /* Remember the start of the free chain */
 176         saveanchor.savefreecnt = 0;                                                     /* Remember the length */
 177         saveanchor.saveadjust = 0;                                                      /* Set none needed yet */
 178         saveanchor.savepoolfwd = (addr64_t)&saveanchor;         /* Remember pool forward */
 179         saveanchor.savepoolbwd = (addr64_t)&saveanchor;         /* Remember pool backward */
 180
 181         for(i=0; i < InitialSaveBloks; i++) {                           /* Initialize the saveareas */
 182
 183                 savec = (savearea_comm *)save;                                  /* Get the control area for this one */
 184
 185                 savec->sac_alloc = 0;                                                   /* Mark it allocated */
 186                 savec->sac_vrswap = 0;                                                  /* V=R, so the translation factor is 0 */
 187                 savec->sac_flags = sac_perm;                                    /* Mark it permanent */
 188                 savec->sac_flags |= 0x0000EE00;                                 /* Debug eyecatcher */
 189                 save_queue((uint32_t)savec >> 12);                              /* Add page to savearea lists */
 190                 save += PAGE_SIZE;                                                              /* Jump up to the next one now */
 191
 192         }
 193
 194 /*
 195  *      We now have a free list that has our initial number of entries
 196  *      The local qfret lists is empty.  When we call save_get below it will see that
 197  *      the local list is empty and fill it for us.
 198  *
 199  *      It is ok to call save_get here because all initial saveareas are V=R in less
 200  *  than 4GB space, so 32-bit addressing is ok.
 201  *
 202  */
 203
 204 /*
 205  * This will populate the local list  and get the first one for the system
 206  */
 207         per_proc_info[0].next_savearea = (vm_offset_t)save_get();
 208
 209 /*
 210  *      The system is now able to take interruptions
 211  */
 212         return;
 213 }
 214
 215
 216
 217
 218 /*
 219  *              Obtains a savearea.  If the free list needs size adjustment it happens here.
 220  *              Don't actually allocate the savearea until after the adjustment is done.
 221  */
 222
 223 struct savearea *save_alloc(void) {                                             /* Reserve a save area */
 224
 225
 226         if(saveanchor.saveadjust) save_adjust();                        /* If size need adjustment, do it now */
 227
 228         return save_get();                                                                      /* Pass the baby... */
 229 }
 230
 231
 232 /*
 233  *              This routine releases a save area to the free queue.  If after that, we have more than our maximum target,
 234  *              we start releasing what we can until we hit the normal target.
 235  */
 236
 237
 238
 239 void save_release(struct savearea *save) {                              /* Release a save area */
 240
 241         save_ret(save);                                                                         /* Return a savearea to the free list */
 242
 243         if(saveanchor.saveadjust) save_adjust();                        /* Adjust the savearea free list and pool size if needed */
 244
 245         return;
 246
 247 }
 248
 249
 250 /*
 251  *              Adjusts the size of the free list.  Can either release or allocate full pages
 252  *              of kernel memory.  This can block.
 253  *
 254  *              Note that we will only run one adjustment and the amount needed may change
 255  *              while we are executing.
 256  *
 257  *              Calling this routine is triggered by saveanchor.saveadjust.  This value is always calculated just before
 258  *              we unlock the saveanchor lock (this keeps it pretty accurate).  If the total of savefreecnt and saveinuse
 259  *              is within the hysteresis range, it is set to 0.  If outside, it is set to the number needed to bring
 260  *              the total to the target value.  Note that there is a minimum size to the free list (FreeListMin) and if
 261  *              savefreecnt falls below that, saveadjust is set to the number needed to bring it to that.
 262  */
 263
 264
 265 void save_adjust(void) {
 266
 267         savearea_comm   *sctl, *sctlnext, *freepage;
 268         kern_return_t ret;
 269         uint64_t vtopmask;
 270         ppnum_t physpage;
 271
 272         if(saveanchor.saveadjust < 0)                                   {       /* Do we need to adjust down? */
 273
 274                 sctl = (savearea_comm *)save_trim_free();               /* Trim list to the need count, return start of trim list */
 275
 276                 while(sctl) {                                                                   /* Release the free pages back to the kernel */
 277                         sctlnext = CAST_DOWN(savearea_comm *, sctl->save_prev); /* Get next in list */
 278                         kmem_free(kernel_map, (vm_offset_t) sctl, PAGE_SIZE);   /* Release the page */
 279                         sctl = sctlnext;                                                        /* Chain onwards */
 280                 }
 281         }
 282         else {                                                                                          /* We need more... */
 283
 284                 if(save_recover()) return;                                              /* If we can recover enough from the pool, return */
 285
 286                 while(saveanchor.saveadjust > 0) {                              /* Keep going until we have enough */
 287
 288                         ret = kmem_alloc_wired(kernel_map, (vm_offset_t *)&freepage, PAGE_SIZE);        /* Get a page for free pool */
 289                         if(ret != KERN_SUCCESS) {                                       /* Did we get some memory? */
 290                                 panic("Whoops...  Not a bit of wired memory left for saveareas\n");
 291                         }
 292
 293                         physpage = pmap_find_phys(kernel_pmap, (vm_offset_t)freepage);  /* Find physical page */
 294                         if(!physpage) {                                                         /* See if we actually have this mapped*/
 295                                 panic("save_adjust: wired page not mapped - va = %08X\n", freepage);    /* Die */
 296                         }
 297
 298                         bzero((void *)freepage, PAGE_SIZE);                     /* Clear it all to zeros */
 299                         freepage->sac_alloc = 0;                                        /* Mark all entries taken */
 300                         freepage->sac_vrswap = ((uint64_t)physpage << 12) ^ (uint64_t)((uintptr_t)freepage);    /* XOR to calculate conversion mask */
 301
 302                         freepage->sac_flags |= 0x0000EE00;                      /* Set debug eyecatcher */
 303
 304                         save_queue(physpage);                                           /* Add all saveareas on page to free list */
 305                 }
 306         }
 307 }
 308
 309 /*
 310  *              Fake up information to make the saveareas look like a zone
 311  */
 312
 313 save_fake_zone_info(int *count, vm_size_t *cur_size, vm_size_t *max_size, vm_size_t *elem_size,
 314                     vm_size_t *alloc_size, int *collectable, int *exhaustable)
 315 {
 316         *count      = saveanchor.saveinuse;
 317         *cur_size   = (saveanchor.savefreecnt + saveanchor.saveinuse) * (PAGE_SIZE / sac_cnt);
 318         *max_size   = saveanchor.savemaxcount * (PAGE_SIZE / sac_cnt);
 319         *elem_size  = sizeof(savearea);
 320         *alloc_size = PAGE_SIZE;
 321         *collectable = 1;
 322         *exhaustable = 0;
 323 }
 324
 325