]>
Commit | Line | Data |
---|---|---|
1c79356b A |
1 | /* |
2 | * Copyright (c) 2000 Apple Computer, Inc. All rights reserved. | |
3 | * | |
4 | * @APPLE_LICENSE_HEADER_START@ | |
5 | * | |
43866e37 | 6 | * Copyright (c) 1999-2003 Apple Computer, Inc. All Rights Reserved. |
1c79356b | 7 | * |
43866e37 A |
8 | * This file contains Original Code and/or Modifications of Original Code |
9 | * as defined in and that are subject to the Apple Public Source License | |
10 | * Version 2.0 (the 'License'). You may not use this file except in | |
11 | * compliance with the License. Please obtain a copy of the License at | |
12 | * http://www.opensource.apple.com/apsl/ and read it before using this | |
13 | * file. | |
14 | * | |
15 | * The Original Code and all software distributed under the License are | |
16 | * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER | |
1c79356b A |
17 | * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, |
18 | * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY, | |
43866e37 A |
19 | * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. |
20 | * Please see the License for the specific language governing rights and | |
21 | * limitations under the License. | |
1c79356b A |
22 | * |
23 | * @APPLE_LICENSE_HEADER_END@ | |
24 | */ | |
25 | /* | |
26 | * This file is used to maintain the exception save areas | |
27 | * | |
28 | */ | |
29 | ||
30 | #include <cpus.h> | |
31 | #include <debug.h> | |
32 | #include <mach_kgdb.h> | |
33 | #include <mach_vm_debug.h> | |
34 | ||
35 | #include <kern/thread.h> | |
36 | #include <mach/vm_attributes.h> | |
37 | #include <mach/vm_param.h> | |
38 | #include <vm/vm_kern.h> | |
39 | #include <vm/vm_map.h> | |
40 | #include <vm/vm_page.h> | |
41 | #include <mach/ppc/thread_status.h> | |
42 | #include <kern/spl.h> | |
43 | #include <kern/simple_lock.h> | |
44 | ||
45 | #include <kern/misc_protos.h> | |
46 | #include <ppc/misc_protos.h> | |
47 | #include <ppc/proc_reg.h> | |
48 | #include <ppc/mem.h> | |
49 | #include <ppc/pmap.h> | |
de355530 | 50 | #include <ppc/pmap_internals.h> |
1c79356b A |
51 | #include <ppc/Firmware.h> |
52 | #include <ppc/mappings.h> | |
53 | #include <ppc/exception.h> | |
54 | #include <ppc/savearea.h> | |
55 | #include <ddb/db_output.h> | |
56 | ||
57 | ||
58 | extern struct Saveanchor saveanchor; /* Aliged savearea anchor */ | |
9bccf70c | 59 | struct Saveanchor backpocket; /* Emergency saveareas */ |
1c79356b A |
60 | unsigned int debsave0 = 0; /* Debug flag */ |
61 | unsigned int backchain = 0; /* Debug flag */ | |
62 | ||
63 | /* | |
64 | * These routines keep track of exception save areas and keeps the count within specific limits. If there are | |
65 | * too few, more are allocated, too many, and they are released. This savearea is where the PCBs are | |
66 | * stored. They never span a page boundary and are referenced by both virtual and real addresses. | |
67 | * Within the interrupt vectors, the real address is used because at that level, no exceptions | |
68 | * can be tolerated. Save areas can be dynamic or permanent. Permanant saveareas are allocated | |
69 | * at boot time and must be in place before any type of exception occurs. These are never released, | |
70 | * and the number is based upon some arbitrary (yet to be determined) amount times the number of | |
71 | * processors. This represents the minimum number required to process a total system failure without | |
72 | * destroying valuable and ever-so-handy system debugging information. | |
73 | * | |
9bccf70c A |
74 | * We keep two global free lists (the savearea free pool and the savearea free list) and one local |
75 | * list per processor. | |
1c79356b | 76 | * |
9bccf70c A |
77 | * The local lists are small and require no locked access. They are chained using physical addresses |
78 | * and no interruptions are allowed when adding to or removing from the list. Also known as the | |
79 | * qfret list. This list is local to a processor and is intended for use only by very low level | |
80 | * context handling code. | |
81 | * | |
82 | * The savearea free list is a medium size list that is globally accessible. It is updated | |
83 | * while holding a simple lock. The length of time that the lock is held is kept short. The | |
84 | * longest period of time is when the list is trimmed. Like the qfret lists, this is chained physically | |
85 | * and must be accessed with translation and interruptions disabled. This is where the bulk | |
86 | * of the free entries are located. | |
87 | * | |
88 | * The saveareas are allocated from full pages. A pool element is marked | |
89 | * with an allocation map that shows which "slots" are free. These pages are allocated via the | |
90 | * normal kernel memory allocation functions. Queueing is with physical addresses. The enqueue, | |
91 | * dequeue, and search for free blocks is done under free list lock. | |
92 | * only if there are empty slots in it. | |
93 | * | |
94 | * Saveareas that are counted as "in use" once they are removed from the savearea free list. | |
95 | * This means that all areas on the local qfret list are considered in use. | |
96 | * | |
97 | * There are two methods of obtaining a savearea. The save_get function (which is also inlined | |
98 | * in the low-level exception handler) attempts to get an area from the local qfret list. This is | |
99 | * done completely without locks. If qfret is exahusted (or maybe just too low) an area is allocated | |
100 | * from the savearea free list. If the free list is empty, we install the back pocket areas and | |
101 | * panic. | |
102 | * | |
103 | * The save_alloc function is designed to be called by high level routines, e.g., thread creation, | |
104 | * etc. It will allocate from the free list. After allocation, it will compare the free count | |
105 | * to the target value. If outside of the range, it will adjust the size either upwards or | |
106 | * downwards. | |
107 | * | |
108 | * If we need to shrink the list, it will be trimmed to the target size and unlocked. The code | |
109 | * will walk the chain and return each savearea to its pool page. If a pool page becomes | |
110 | * completely empty, it is dequeued from the free pool list and enqueued (atomic queue | |
111 | * function) to be released. | |
112 | * | |
113 | * Once the trim list is finished, the pool release queue is checked to see if there are pages | |
114 | * waiting to be released. If so, they are released one at a time. | |
115 | * | |
116 | * If the free list needed to be grown rather than shrunken, we will first attempt to recover | |
117 | * a page from the pending release queue (built when we trim the free list). If we find one, | |
118 | * it is allocated, otherwise, a page of kernel memory is allocated. This loops until there are | |
119 | * enough free saveareas. | |
120 | * | |
1c79356b A |
121 | */ |
122 | ||
9bccf70c A |
123 | |
124 | ||
1c79356b | 125 | /* |
9bccf70c A |
126 | * Allocate our initial context save areas. As soon as we do this, |
127 | * we can take an interrupt. We do the saveareas here, 'cause they're guaranteed | |
128 | * to be at least page aligned. | |
1c79356b A |
129 | */ |
130 | ||
131 | ||
de355530 | 132 | void savearea_init(vm_offset_t *addrx) { |
1c79356b | 133 | |
de355530 A |
134 | savearea_comm *savec, *savec2, *saveprev; |
135 | vm_offset_t save, save2, addr; | |
9bccf70c | 136 | int i; |
1c79356b | 137 | |
9bccf70c A |
138 | |
139 | saveanchor.savetarget = InitialSaveTarget; /* Initial target value */ | |
140 | saveanchor.saveinuse = 0; /* Number of areas in use */ | |
1c79356b | 141 | |
de355530 | 142 | saveanchor.savefree = 0; /* Remember the start of the free chain */ |
9bccf70c | 143 | saveanchor.savefreecnt = 0; /* Remember the length */ |
de355530 A |
144 | saveanchor.savepoolfwd = (unsigned int *)&saveanchor; /* Remember pool forward */ |
145 | saveanchor.savepoolbwd = (unsigned int *)&saveanchor; /* Remember pool backward */ | |
146 | ||
147 | addr = *addrx; /* Make this easier for ourselves */ | |
1c79356b | 148 | |
9bccf70c | 149 | save = addr; /* Point to the whole block of blocks */ |
1c79356b A |
150 | |
151 | /* | |
9bccf70c | 152 | * First we allocate the back pocket in case of emergencies |
1c79356b A |
153 | */ |
154 | ||
155 | ||
de355530 | 156 | for(i=0; i < 8; i++) { /* Initialize the back pocket saveareas */ |
1c79356b | 157 | |
9bccf70c | 158 | savec = (savearea_comm *)save; /* Get the control area for this one */ |
1c79356b | 159 | |
9bccf70c A |
160 | savec->sac_alloc = 0; /* Mark it allocated */ |
161 | savec->sac_vrswap = 0; /* V=R, so the translation factor is 0 */ | |
162 | savec->sac_flags = sac_perm; /* Mark it permanent */ | |
163 | savec->sac_flags |= 0x0000EE00; /* Debug eyecatcher */ | |
de355530 | 164 | save_queue((savearea *)savec); /* Add page to savearea lists */ |
9bccf70c | 165 | save += PAGE_SIZE; /* Jump up to the next one now */ |
1c79356b | 166 | |
1c79356b | 167 | } |
1c79356b | 168 | |
9bccf70c | 169 | backpocket = saveanchor; /* Save this for emergencies */ |
1c79356b A |
170 | |
171 | ||
172 | /* | |
9bccf70c A |
173 | * We've saved away the back pocket savearea info, so reset it all and |
174 | * now allocate for real | |
1c79356b A |
175 | */ |
176 | ||
177 | ||
9bccf70c A |
178 | saveanchor.savefree = 0; /* Remember the start of the free chain */ |
179 | saveanchor.savefreecnt = 0; /* Remember the length */ | |
180 | saveanchor.saveadjust = 0; /* Set none needed yet */ | |
de355530 A |
181 | saveanchor.savepoolfwd = (unsigned int *)&saveanchor; /* Remember pool forward */ |
182 | saveanchor.savepoolbwd = (unsigned int *)&saveanchor; /* Remember pool backward */ | |
1c79356b | 183 | |
9bccf70c | 184 | for(i=0; i < InitialSaveBloks; i++) { /* Initialize the saveareas */ |
1c79356b | 185 | |
9bccf70c | 186 | savec = (savearea_comm *)save; /* Get the control area for this one */ |
1c79356b | 187 | |
9bccf70c A |
188 | savec->sac_alloc = 0; /* Mark it allocated */ |
189 | savec->sac_vrswap = 0; /* V=R, so the translation factor is 0 */ | |
190 | savec->sac_flags = sac_perm; /* Mark it permanent */ | |
191 | savec->sac_flags |= 0x0000EE00; /* Debug eyecatcher */ | |
de355530 | 192 | save_queue((savearea *)savec); /* Add page to savearea lists */ |
9bccf70c | 193 | save += PAGE_SIZE; /* Jump up to the next one now */ |
1c79356b A |
194 | |
195 | } | |
9bccf70c | 196 | |
de355530 A |
197 | *addrx = save; /* Move the free storage lowwater mark */ |
198 | ||
9bccf70c A |
199 | /* |
200 | * We now have a free list that has our initial number of entries | |
201 | * The local qfret lists is empty. When we call save_get below it will see that | |
202 | * the local list is empty and fill it for us. | |
203 | * | |
de355530 A |
204 | * It is ok to call save_get_phys here because even though if we are translation on, we are still V=R and |
205 | * running with BAT registers so no interruptions. Regular interruptions will be off. Using save_get | |
206 | * would be wrong if the tracing was enabled--it would cause an exception. | |
9bccf70c A |
207 | */ |
208 | ||
de355530 A |
209 | save2 = (vm_offset_t)save_get_phys(); /* This will populate the local list |
210 | and get the first one for the system */ | |
211 | per_proc_info[0].next_savearea = (unsigned int)save2; /* Tell the exception handler about it */ | |
212 | ||
9bccf70c A |
213 | /* |
214 | * The system is now able to take interruptions | |
215 | */ | |
de355530 | 216 | |
1c79356b | 217 | return; |
de355530 | 218 | |
1c79356b A |
219 | } |
220 | ||
9bccf70c A |
221 | |
222 | ||
223 | ||
1c79356b | 224 | /* |
de355530 | 225 | * Returns a savearea. If the free list needs size adjustment it happens here. |
9bccf70c | 226 | * Don't actually allocate the savearea until after the adjustment is done. |
1c79356b A |
227 | */ |
228 | ||
9bccf70c A |
229 | struct savearea *save_alloc(void) { /* Reserve a save area */ |
230 | ||
231 | ||
232 | if(saveanchor.saveadjust) save_adjust(); /* If size need adjustment, do it now */ | |
233 | ||
234 | return save_get(); /* Pass the baby... */ | |
235 | } | |
236 | ||
237 | ||
238 | /* | |
239 | * This routine releases a save area to the free queue. If after that, we have more than our maximum target, | |
240 | * we start releasing what we can until we hit the normal target. | |
241 | */ | |
1c79356b A |
242 | |
243 | ||
1c79356b | 244 | |
9bccf70c A |
245 | void save_release(struct savearea *save) { /* Release a save area */ |
246 | ||
247 | save_ret(save); /* Return a savearea to the free list */ | |
248 | ||
249 | if(saveanchor.saveadjust) save_adjust(); /* Adjust the savearea free list and pool size if needed */ | |
1c79356b | 250 | |
1c79356b | 251 | return; |
9bccf70c | 252 | |
1c79356b A |
253 | } |
254 | ||
255 | ||
1c79356b | 256 | /* |
9bccf70c A |
257 | * Adjusts the size of the free list. Can either release or allocate full pages |
258 | * of kernel memory. This can block. | |
259 | * | |
260 | * Note that we will only run one adjustment and the amount needed may change | |
261 | * while we are executing. | |
262 | * | |
263 | * Calling this routine is triggered by saveanchor.saveadjust. This value is always calculated just before | |
264 | * we unlock the saveanchor lock (this keeps it pretty accurate). If the total of savefreecnt and saveinuse | |
265 | * is within the hysteresis range, it is set to 0. If outside, it is set to the number needed to bring | |
266 | * the total to the target value. Note that there is a minimum size to the free list (FreeListMin) and if | |
267 | * savefreecnt falls below that, saveadjust is set to the number needed to bring it to that. | |
1c79356b | 268 | */ |
1c79356b | 269 | |
9bccf70c A |
270 | |
271 | void save_adjust(void) { | |
1c79356b | 272 | |
de355530 | 273 | savearea_comm *sctl, *sctlnext, *freepool, *freepage, *realpage; |
9bccf70c A |
274 | kern_return_t ret; |
275 | ||
276 | if(saveanchor.saveadjust < 0) { /* Do we need to adjust down? */ | |
277 | ||
278 | sctl = (savearea_comm *)save_trim_free(); /* Trim list to the need count, return start of trim list */ | |
279 | ||
280 | while(sctl) { /* Release the free pages back to the kernel */ | |
281 | sctlnext = (savearea_comm *)sctl->save_prev; /* Get next in list */ | |
282 | kmem_free(kernel_map, (vm_offset_t) sctl, PAGE_SIZE); /* Release the page */ | |
283 | sctl = sctlnext; /* Chain onwards */ | |
284 | } | |
285 | } | |
286 | else { /* We need more... */ | |
287 | ||
288 | if(save_recover()) return; /* If we can recover enough from the pool, return */ | |
289 | ||
290 | while(saveanchor.saveadjust > 0) { /* Keep going until we have enough */ | |
291 | ||
292 | ret = kmem_alloc_wired(kernel_map, (vm_offset_t *)&freepage, PAGE_SIZE); /* Get a page for free pool */ | |
293 | if(ret != KERN_SUCCESS) { /* Did we get some memory? */ | |
294 | panic("Whoops... Not a bit of wired memory left for saveareas\n"); | |
295 | } | |
296 | ||
de355530 | 297 | realpage = (savearea_comm *)pmap_extract(kernel_pmap, (vm_offset_t)freepage); /* Get the physical */ |
9bccf70c A |
298 | |
299 | bzero((void *)freepage, PAGE_SIZE); /* Clear it all to zeros */ | |
300 | freepage->sac_alloc = 0; /* Mark all entries taken */ | |
de355530 | 301 | freepage->sac_vrswap = (unsigned int)freepage ^ (unsigned int)realpage; /* Form mask to convert V to R and vice versa */ |
9bccf70c A |
302 | |
303 | freepage->sac_flags |= 0x0000EE00; /* Set debug eyecatcher */ | |
304 | ||
de355530 | 305 | save_queue((savearea *)realpage); /* Add all saveareas on page to free list */ |
9bccf70c | 306 | } |
1c79356b | 307 | } |
1c79356b A |
308 | } |
309 | ||
9bccf70c A |
310 | /* |
311 | * Fake up information to make the saveareas look like a zone | |
312 | */ | |
1c79356b | 313 | |
9bccf70c A |
314 | save_fake_zone_info(int *count, vm_size_t *cur_size, vm_size_t *max_size, vm_size_t *elem_size, |
315 | vm_size_t *alloc_size, int *collectable, int *exhaustable) | |
316 | { | |
317 | *count = saveanchor.saveinuse; | |
318 | *cur_size = (saveanchor.savefreecnt + saveanchor.saveinuse) * (PAGE_SIZE / sac_cnt); | |
319 | *max_size = saveanchor.savemaxcount * (PAGE_SIZE / sac_cnt); | |
320 | *elem_size = sizeof(savearea); | |
321 | *alloc_size = PAGE_SIZE; | |
322 | *collectable = 1; | |
323 | *exhaustable = 0; | |
324 | } | |
1c79356b A |
325 | |
326 |