pmap.c revision 1.23 1 /* $NetBSD: pmap.c,v 1.23 1997/05/28 04:28:52 jeremy Exp $ */
2
3 /*-
4 * Copyright (c) 1996, 1997 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Jeremy Cooper.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the NetBSD
21 * Foundation, Inc. and its contributors.
22 * 4. Neither the name of The NetBSD Foundation nor the names of its
23 * contributors may be used to endorse or promote products derived
24 * from this software without specific prior written permission.
25 *
26 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
28 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
29 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
30 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
31 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
32 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
33 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
34 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
35 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
36 * POSSIBILITY OF SUCH DAMAGE.
37 */
38
39 /*
40 * XXX These comments aren't quite accurate. Need to change.
41 * The sun3x uses the MC68851 Memory Management Unit, which is built
42 * into the CPU. The 68851 maps virtual to physical addresses using
43 * a multi-level table lookup, which is stored in the very memory that
44 * it maps. The number of levels of lookup is configurable from one
45 * to four. In this implementation, we use three, named 'A' through 'C'.
46 *
47 * The MMU translates virtual addresses into physical addresses by
48 * traversing these tables in a proccess called a 'table walk'. The most
49 * significant 7 bits of the Virtual Address ('VA') being translated are
50 * used as an index into the level A table, whose base in physical memory
51 * is stored in a special MMU register, the 'CPU Root Pointer' or CRP. The
52 * address found at that index in the A table is used as the base
53 * address for the next table, the B table. The next six bits of the VA are
54 * used as an index into the B table, which in turn gives the base address
55 * of the third and final C table.
56 *
57 * The next six bits of the VA are used as an index into the C table to
58 * locate a Page Table Entry (PTE). The PTE is a physical address in memory
59 * to which the remaining 13 bits of the VA are added, producing the
60 * mapped physical address.
61 *
62 * To map the entire memory space in this manner would require 2114296 bytes
63 * of page tables per process - quite expensive. Instead we will
64 * allocate a fixed but considerably smaller space for the page tables at
65 * the time the VM system is initialized. When the pmap code is asked by
66 * the kernel to map a VA to a PA, it allocates tables as needed from this
67 * pool. When there are no more tables in the pool, tables are stolen
68 * from the oldest mapped entries in the tree. This is only possible
69 * because all memory mappings are stored in the kernel memory map
70 * structures, independent of the pmap structures. A VA which references
71 * one of these invalidated maps will cause a page fault. The kernel
72 * will determine that the page fault was caused by a task using a valid
73 * VA, but for some reason (which does not concern it), that address was
74 * not mapped. It will ask the pmap code to re-map the entry and then
75 * it will resume executing the faulting task.
76 *
77 * In this manner the most efficient use of the page table space is
78 * achieved. Tasks which do not execute often will have their tables
79 * stolen and reused by tasks which execute more frequently. The best
80 * size for the page table pool will probably be determined by
81 * experimentation.
82 *
83 * You read all of the comments so far. Good for you.
84 * Now go play!
85 */
86
87 /*** A Note About the 68851 Address Translation Cache
88 * The MC68851 has a 64 entry cache, called the Address Translation Cache
89 * or 'ATC'. This cache stores the most recently used page descriptors
90 * accessed by the MMU when it does translations. Using a marker called a
91 * 'task alias' the MMU can store the descriptors from 8 different table
92 * spaces concurrently. The task alias is associated with the base
93 * address of the level A table of that address space. When an address
94 * space is currently active (the CRP currently points to its A table)
95 * the only cached descriptors that will be obeyed are ones which have a
96 * matching task alias of the current space associated with them.
97 *
98 * Since the cache is always consulted before any table lookups are done,
99 * it is important that it accurately reflect the state of the MMU tables.
100 * Whenever a change has been made to a table that has been loaded into
101 * the MMU, the code must be sure to flush any cached entries that are
102 * affected by the change. These instances are documented in the code at
103 * various points.
104 */
105 /*** A Note About the Note About the 68851 Address Translation Cache
106 * 4 months into this code I discovered that the sun3x does not have
107 * a MC68851 chip. Instead, it has a version of this MMU that is part of the
108 * the 68030 CPU.
109 * All though it behaves very similarly to the 68851, it only has 1 task
110 * alias and a 22 entry cache. So sadly (or happily), the first paragraph
111 * of the previous note does not apply to the sun3x pmap.
112 */
113
114 #include <sys/param.h>
115 #include <sys/systm.h>
116 #include <sys/proc.h>
117 #include <sys/malloc.h>
118 #include <sys/user.h>
119 #include <sys/queue.h>
120 #include <sys/kcore.h>
121
122 #include <vm/vm.h>
123 #include <vm/vm_kern.h>
124 #include <vm/vm_page.h>
125
126 #include <machine/cpu.h>
127 #include <machine/kcore.h>
128 #include <machine/pmap.h>
129 #include <machine/pte.h>
130 #include <machine/machdep.h>
131 #include <machine/mon.h>
132
133 #include "pmap_pvt.h"
134
135 /* XXX - What headers declare these? */
136 extern struct pcb *curpcb;
137 extern int physmem;
138
139 extern void copypage __P((const void*, void*));
140 extern void zeropage __P((void*));
141
142 /* Defined in locore.s */
143 extern char kernel_text[];
144
145 /* Defined by the linker */
146 extern char etext[], edata[], end[];
147 extern char *esym; /* DDB */
148
149 /*************************** DEBUGGING DEFINITIONS ***********************
150 * Macros, preprocessor defines and variables used in debugging can make *
151 * code hard to read. Anything used exclusively for debugging purposes *
152 * is defined here to avoid having such mess scattered around the file. *
153 *************************************************************************/
154 #ifdef PMAP_DEBUG
155 /*
156 * To aid the debugging process, macros should be expanded into smaller steps
157 * that accomplish the same goal, yet provide convenient places for placing
158 * breakpoints. When this code is compiled with PMAP_DEBUG mode defined, the
159 * 'INLINE' keyword is defined to an empty string. This way, any function
160 * defined to be a 'static INLINE' will become 'outlined' and compiled as
161 * a separate function, which is much easier to debug.
162 */
163 #define INLINE /* nothing */
164
165 /*
166 * It is sometimes convenient to watch the activity of a particular table
167 * in the system. The following variables are used for that purpose.
168 */
169 a_tmgr_t *pmap_watch_atbl = 0;
170 b_tmgr_t *pmap_watch_btbl = 0;
171 c_tmgr_t *pmap_watch_ctbl = 0;
172
173 int pmap_debug = 0;
174 #define DPRINT(args) if (pmap_debug) printf args
175
176 #else /********** Stuff below is defined if NOT debugging **************/
177
178 #define INLINE inline
179 #define DPRINT(args) /* nada */
180
181 #endif /* PMAP_DEBUG */
182 /*********************** END OF DEBUGGING DEFINITIONS ********************/
183
184 /*** Management Structure - Memory Layout
185 * For every MMU table in the sun3x pmap system there must be a way to
186 * manage it; we must know which process is using it, what other tables
187 * depend on it, and whether or not it contains any locked pages. This
188 * is solved by the creation of 'table management' or 'tmgr'
189 * structures. One for each MMU table in the system.
190 *
191 * MAP OF MEMORY USED BY THE PMAP SYSTEM
192 *
193 * towards lower memory
194 * kernAbase -> +-------------------------------------------------------+
195 * | Kernel MMU A level table |
196 * kernBbase -> +-------------------------------------------------------+
197 * | Kernel MMU B level tables |
198 * kernCbase -> +-------------------------------------------------------+
199 * | |
200 * | Kernel MMU C level tables |
201 * | |
202 * mmuCbase -> +-------------------------------------------------------+
203 * | User MMU C level tables |
204 * mmuAbase -> +-------------------------------------------------------+
205 * | |
206 * | User MMU A level tables |
207 * | |
208 * mmuBbase -> +-------------------------------------------------------+
209 * | User MMU B level tables |
210 * tmgrAbase -> +-------------------------------------------------------+
211 * | TMGR A level table structures |
212 * tmgrBbase -> +-------------------------------------------------------+
213 * | TMGR B level table structures |
214 * tmgrCbase -> +-------------------------------------------------------+
215 * | TMGR C level table structures |
216 * pvbase -> +-------------------------------------------------------+
217 * | Physical to Virtual mapping table (list heads) |
218 * pvebase -> +-------------------------------------------------------+
219 * | Physical to Virtual mapping table (list elements) |
220 * | |
221 * +-------------------------------------------------------+
222 * towards higher memory
223 *
224 * For every A table in the MMU A area, there will be a corresponding
225 * a_tmgr structure in the TMGR A area. The same will be true for
226 * the B and C tables. This arrangement will make it easy to find the
227 * controling tmgr structure for any table in the system by use of
228 * (relatively) simple macros.
229 */
230
231 /*
232 * Global variables for storing the base addresses for the areas
233 * labeled above.
234 */
235 static vm_offset_t kernAphys;
236 static mmu_long_dte_t *kernAbase;
237 static mmu_short_dte_t *kernBbase;
238 static mmu_short_pte_t *kernCbase;
239 static mmu_short_pte_t *mmuCbase;
240 static mmu_short_dte_t *mmuBbase;
241 static mmu_long_dte_t *mmuAbase;
242 static a_tmgr_t *Atmgrbase;
243 static b_tmgr_t *Btmgrbase;
244 static c_tmgr_t *Ctmgrbase;
245 static pv_t *pvbase;
246 static pv_elem_t *pvebase;
247 struct pmap kernel_pmap;
248
249 /*
250 * This holds the CRP currently loaded into the MMU.
251 */
252 struct mmu_rootptr kernel_crp;
253
254 /*
255 * Just all around global variables.
256 */
257 static TAILQ_HEAD(a_pool_head_struct, a_tmgr_struct) a_pool;
258 static TAILQ_HEAD(b_pool_head_struct, b_tmgr_struct) b_pool;
259 static TAILQ_HEAD(c_pool_head_struct, c_tmgr_struct) c_pool;
260
261
262 /*
263 * Flags used to mark the safety/availability of certain operations or
264 * resources.
265 */
266 static boolean_t pv_initialized = FALSE, /* PV system has been initialized. */
267 bootstrap_alloc_enabled = FALSE; /*Safe to use pmap_bootstrap_alloc().*/
268 u_char tmp_vpage0_inuse, /* Temporary virtual page 0 is in use */
269 tmp_vpage1_inuse; /* Temporary virtual page 1 is in use */
270
271 /*
272 * XXX: For now, retain the traditional variables that were
273 * used in the old pmap/vm interface (without NONCONTIG).
274 */
275 /* Kernel virtual address space available: */
276 vm_offset_t virtual_avail, virtual_end;
277 /* Physical address space available: */
278 vm_offset_t avail_start, avail_end;
279
280 /* This keep track of the end of the contiguously mapped range. */
281 vm_offset_t virtual_contig_end;
282
283 /* Physical address used by pmap_next_page() */
284 vm_offset_t avail_next;
285
286 /* These are used by pmap_copy_page(), etc. */
287 vm_offset_t tmp_vpages[2]; /* Note: tmp_vpage[0] MUST be mapped R/O */
288 /* tmp_vpage[1] MUST be mapped R/W */
289
290 /*
291 * The 3/80 is the only member of the sun3x family that has non-contiguous
292 * physical memory. Memory is divided into 4 banks which are physically
293 * locatable on the system board. Although the size of these banks varies
294 * with the size of memory they contain, their base addresses are
295 * permenently fixed. The following structure, which describes these
296 * banks, is initialized by pmap_bootstrap() after it reads from a similar
297 * structure provided by the ROM Monitor.
298 *
299 * For the other machines in the sun3x architecture which do have contiguous
300 * RAM, this list will have only one entry, which will describe the entire
301 * range of available memory.
302 */
303 struct pmap_physmem_struct avail_mem[SUN3X_NPHYS_RAM_SEGS];
304 u_int total_phys_mem;
305
306 /*************************************************************************/
307
308 /*
309 * XXX - Should "tune" these based on statistics.
310 *
311 * My first guess about the relative numbers of these needed is
312 * based on the fact that a "typical" process will have several
313 * pages mapped at low virtual addresses (text, data, bss), then
314 * some mapped shared libraries, and then some stack pages mapped
315 * near the high end of the VA space. Each process can use only
316 * one A table, and most will use only two B tables (maybe three)
317 * and probably about four C tables. Therefore, the first guess
318 * at the relative numbers of these needed is 1:2:4 -gwr
319 *
320 * The number of C tables needed is closely related to the amount
321 * of physical memory available plus a certain amount attributable
322 * to the use of double mappings. With a few simulation statistics
323 * we can find a reasonably good estimation of this unknown value.
324 * Armed with that and the above ratios, we have a good idea of what
325 * is needed at each level. -j
326 *
327 * Note: It is not physical memory memory size, but the total mapped
328 * virtual space required by the combined working sets of all the
329 * currently _runnable_ processes. (Sleeping ones don't count.)
330 * The amount of physical memory should be irrelevant. -gwr
331 */
332 #ifdef FIXED_NTABLES
333 #define NUM_A_TABLES 16
334 #define NUM_B_TABLES 32
335 #define NUM_C_TABLES 64
336 #else
337 unsigned int NUM_A_TABLES, NUM_B_TABLES, NUM_C_TABLES;
338 #endif /* FIXED_NTABLES */
339
340 /*
341 * This determines our total virtual mapping capacity.
342 * Yes, it is a FIXED value so we can pre-allocate.
343 */
344 #define NUM_USER_PTES (NUM_C_TABLES * MMU_C_TBL_SIZE)
345
346 /*
347 * The size of the Kernel Virtual Address Space (KVAS)
348 * for purposes of MMU table allocation is -KERNBASE
349 * (length from KERNBASE to 0xFFFFffff)
350 */
351 #define KVAS_SIZE (-KERNBASE)
352
353 /* Numbers of kernel MMU tables to support KVAS_SIZE. */
354 #define KERN_B_TABLES (KVAS_SIZE >> MMU_TIA_SHIFT)
355 #define KERN_C_TABLES (KVAS_SIZE >> MMU_TIB_SHIFT)
356 #define NUM_KERN_PTES (KVAS_SIZE >> MMU_TIC_SHIFT)
357
358 /*************************** MISCELANEOUS MACROS *************************/
359 #define PMAP_LOCK() ; /* Nothing, for now */
360 #define PMAP_UNLOCK() ; /* same. */
361 #define NULL 0
362
363 static INLINE void * mmu_ptov __P((vm_offset_t pa));
364 static INLINE vm_offset_t mmu_vtop __P((void * va));
365
366 #if 0
367 static INLINE a_tmgr_t * mmuA2tmgr __P((mmu_long_dte_t *));
368 #endif
369 static INLINE b_tmgr_t * mmuB2tmgr __P((mmu_short_dte_t *));
370 static INLINE c_tmgr_t * mmuC2tmgr __P((mmu_short_pte_t *));
371
372 static INLINE pv_t *pa2pv __P((vm_offset_t pa));
373 static INLINE int pteidx __P((mmu_short_pte_t *));
374 static INLINE pmap_t current_pmap __P((void));
375
376 /*
377 * We can always convert between virtual and physical addresses
378 * for anything in the range [KERNBASE ... avail_start] because
379 * that range is GUARANTEED to be mapped linearly.
380 * We rely heavily upon this feature!
381 */
382 static INLINE void *
383 mmu_ptov(pa)
384 vm_offset_t pa;
385 {
386 register vm_offset_t va;
387
388 va = (pa + KERNBASE);
389 #ifdef PMAP_DEBUG
390 if ((va < KERNBASE) || (va >= virtual_contig_end))
391 panic("mmu_ptov");
392 #endif
393 return ((void*)va);
394 }
395 static INLINE vm_offset_t
396 mmu_vtop(vva)
397 void *vva;
398 {
399 register vm_offset_t va;
400
401 va = (vm_offset_t)vva;
402 #ifdef PMAP_DEBUG
403 if ((va < KERNBASE) || (va >= virtual_contig_end))
404 panic("mmu_ptov");
405 #endif
406 return (va - KERNBASE);
407 }
408
409 /*
410 * These macros map MMU tables to their corresponding manager structures.
411 * They are needed quite often because many of the pointers in the pmap
412 * system reference MMU tables and not the structures that control them.
413 * There needs to be a way to find one when given the other and these
414 * macros do so by taking advantage of the memory layout described above.
415 * Here's a quick step through the first macro, mmuA2tmgr():
416 *
417 * 1) find the offset of the given MMU A table from the base of its table
418 * pool (table - mmuAbase).
419 * 2) convert this offset into a table index by dividing it by the
420 * size of one MMU 'A' table. (sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE)
421 * 3) use this index to select the corresponding 'A' table manager
422 * structure from the 'A' table manager pool (Atmgrbase[index]).
423 */
424 /* This function is not currently used. */
425 #if 0
426 static INLINE a_tmgr_t *
427 mmuA2tmgr(mmuAtbl)
428 mmu_long_dte_t *mmuAtbl;
429 {
430 register int idx;
431
432 /* Which table is this in? */
433 idx = (mmuAtbl - mmuAbase) / MMU_A_TBL_SIZE;
434 #ifdef PMAP_DEBUG
435 if ((idx < 0) || (idx >= NUM_A_TABLES))
436 panic("mmuA2tmgr");
437 #endif
438 return (&Atmgrbase[idx]);
439 }
440 #endif /* 0 */
441
442 static INLINE b_tmgr_t *
443 mmuB2tmgr(mmuBtbl)
444 mmu_short_dte_t *mmuBtbl;
445 {
446 register int idx;
447
448 /* Which table is this in? */
449 idx = (mmuBtbl - mmuBbase) / MMU_B_TBL_SIZE;
450 #ifdef PMAP_DEBUG
451 if ((idx < 0) || (idx >= NUM_B_TABLES))
452 panic("mmuB2tmgr");
453 #endif
454 return (&Btmgrbase[idx]);
455 }
456
457 /* mmuC2tmgr INTERNAL
458 **
459 * Given a pte known to belong to a C table, return the address of
460 * that table's management structure.
461 */
462 static INLINE c_tmgr_t *
463 mmuC2tmgr(mmuCtbl)
464 mmu_short_pte_t *mmuCtbl;
465 {
466 register int idx;
467
468 /* Which table is this in? */
469 idx = (mmuCtbl - mmuCbase) / MMU_C_TBL_SIZE;
470 #ifdef PMAP_DEBUG
471 if ((idx < 0) || (idx >= NUM_C_TABLES))
472 panic("mmuC2tmgr");
473 #endif
474 return (&Ctmgrbase[idx]);
475 }
476
477 /* This is now a function call below.
478 * #define pa2pv(pa) \
479 * (&pvbase[(unsigned long)\
480 * _btop(pa)\
481 * ])
482 */
483
484 /* pa2pv INTERNAL
485 **
486 * Return the pv_list_head element which manages the given physical
487 * address.
488 */
489 static INLINE pv_t *
490 pa2pv(pa)
491 vm_offset_t pa;
492 {
493 register struct pmap_physmem_struct *bank;
494 register int idx;
495
496 bank = &avail_mem[0];
497 while (pa >= bank->pmem_end)
498 bank = bank->pmem_next;
499
500 pa -= bank->pmem_start;
501 idx = bank->pmem_pvbase + _btop(pa);
502 #ifdef PMAP_DEBUG
503 if ((idx < 0) || (idx >= physmem))
504 panic("pa2pv");
505 #endif
506 return &pvbase[idx];
507 }
508
509 /* pteidx INTERNAL
510 **
511 * Return the index of the given PTE within the entire fixed table of
512 * PTEs.
513 */
514 static INLINE int
515 pteidx(pte)
516 mmu_short_pte_t *pte;
517 {
518 return (pte - kernCbase);
519 }
520
521 /*
522 * This just offers a place to put some debugging checks,
523 * and reduces the number of places "curproc" appears...
524 */
525 static INLINE pmap_t
526 current_pmap()
527 {
528 struct proc *p;
529 struct vmspace *vm;
530 vm_map_t map;
531 pmap_t pmap;
532
533 p = curproc; /* XXX */
534 if (p == NULL)
535 pmap = &kernel_pmap;
536 else {
537 vm = p->p_vmspace;
538 map = &vm->vm_map;
539 pmap = vm_map_pmap(map);
540 }
541
542 return (pmap);
543 }
544
545
546 /*************************** FUNCTION DEFINITIONS ************************
547 * These appear here merely for the compiler to enforce type checking on *
548 * all function calls. *
549 *************************************************************************/
550
551 /** External functions
552 ** - functions used within this module but written elsewhere.
553 ** both of these functions are in locore.s
554 ** XXX - These functions were later replaced with their more cryptic
555 ** hp300 counterparts. They may be removed now.
556 **/
557 #if 0 /* deprecated mmu */
558 void mmu_seturp __P((vm_offset_t));
559 void mmu_flush __P((int, vm_offset_t));
560 void mmu_flusha __P((void));
561 #endif /* 0 */
562
563 /** Internal functions
564 ** - all functions used only within this module are defined in
565 ** pmap_pvt.h
566 **/
567
568 /** Interface functions
569 ** - functions required by the Mach VM Pmap interface, with MACHINE_CONTIG
570 ** defined.
571 **/
572 #ifdef INCLUDED_IN_PMAP_H
573 void pmap_bootstrap __P((void));
574 void *pmap_bootstrap_alloc __P((int));
575 void pmap_enter __P((pmap_t, vm_offset_t, vm_offset_t, vm_prot_t, boolean_t));
576 pmap_t pmap_create __P((vm_size_t));
577 void pmap_destroy __P((pmap_t));
578 void pmap_reference __P((pmap_t));
579 boolean_t pmap_is_referenced __P((vm_offset_t));
580 boolean_t pmap_is_modified __P((vm_offset_t));
581 void pmap_clear_modify __P((vm_offset_t));
582 vm_offset_t pmap_extract __P((pmap_t, vm_offset_t));
583 void pmap_activate __P((pmap_t));
584 int pmap_page_index __P((vm_offset_t));
585 u_int pmap_free_pages __P((void));
586 #endif /* INCLUDED_IN_PMAP_H */
587
588 /********************************** CODE ********************************
589 * Functions that are called from other parts of the kernel are labeled *
590 * as 'INTERFACE' functions. Functions that are only called from *
591 * within the pmap module are labeled as 'INTERNAL' functions. *
592 * Functions that are internal, but are not (currently) used at all are *
593 * labeled 'INTERNAL_X'. *
594 ************************************************************************/
595
596 /* pmap_bootstrap INTERNAL
597 **
598 * Initializes the pmap system. Called at boot time from _vm_init()
599 * in _startup.c.
600 *
601 * Reminder: having a pmap_bootstrap_alloc() and also having the VM
602 * system implement pmap_steal_memory() is redundant.
603 * Don't release this code without removing one or the other!
604 */
605 void
606 pmap_bootstrap(nextva)
607 vm_offset_t nextva;
608 {
609 struct physmemory *membank;
610 struct pmap_physmem_struct *pmap_membank;
611 vm_offset_t va, pa, eva;
612 int b, c, i, j; /* running table counts */
613 int size;
614
615 /*
616 * This function is called by __bootstrap after it has
617 * determined the type of machine and made the appropriate
618 * patches to the ROM vectors (XXX- I don't quite know what I meant
619 * by that.) It allocates and sets up enough of the pmap system
620 * to manage the kernel's address space.
621 */
622
623 /*
624 * Determine the range of kernel virtual and physical
625 * space available. Note that we ABSOLUTELY DEPEND on
626 * the fact that the first bank of memory (4MB) is
627 * mapped linearly to KERNBASE (which we guaranteed in
628 * the first instructions of locore.s).
629 * That is plenty for our bootstrap work.
630 */
631 virtual_avail = _round_page(nextva);
632 virtual_contig_end = KERNBASE + 0x400000; /* +4MB */
633 virtual_end = VM_MAX_KERNEL_ADDRESS;
634 /* Don't need avail_start til later. */
635
636 /* We may now call pmap_bootstrap_alloc(). */
637 bootstrap_alloc_enabled = TRUE;
638
639 /*
640 * This is a somewhat unwrapped loop to deal with
641 * copying the PROM's 'phsymem' banks into the pmap's
642 * banks. The following is always assumed:
643 * 1. There is always at least one bank of memory.
644 * 2. There is always a last bank of memory, and its
645 * pmem_next member must be set to NULL.
646 * XXX - Use: do { ... } while (membank->next) instead?
647 * XXX - Why copy this stuff at all? -gwr
648 * - It is needed in pa2pv().
649 */
650 membank = romVectorPtr->v_physmemory;
651 pmap_membank = avail_mem;
652 total_phys_mem = 0;
653
654 while (membank->next) {
655 pmap_membank->pmem_start = membank->address;
656 pmap_membank->pmem_end = membank->address + membank->size;
657 total_phys_mem += membank->size;
658 /* This silly syntax arises because pmap_membank
659 * is really a pre-allocated array, but it is put into
660 * use as a linked list.
661 */
662 pmap_membank->pmem_next = pmap_membank + 1;
663 pmap_membank = pmap_membank->pmem_next;
664 membank = membank->next;
665 }
666
667 /*
668 * XXX The last bank of memory should be reduced to exclude the
669 * physical pages needed by the PROM monitor from being used
670 * in the VM system. XXX - See below - Fix!
671 */
672 pmap_membank->pmem_start = membank->address;
673 pmap_membank->pmem_end = membank->address + membank->size;
674 pmap_membank->pmem_next = NULL;
675
676 #if 0 /* XXX - Need to integrate this! */
677 /*
678 * The last few pages of physical memory are "owned" by
679 * the PROM. The total amount of memory we are allowed
680 * to use is given by the romvec pointer. -gwr
681 *
682 * We should dedicate different variables for 'useable'
683 * and 'physically available'. Most users are used to the
684 * kernel reporting the amount of memory 'physically available'
685 * as opposed to 'useable by the kernel' at boot time. -j
686 */
687 total_phys_mem = *romVectorPtr->memoryAvail;
688 #endif /* XXX */
689
690 total_phys_mem += membank->size; /* XXX see above */
691 physmem = btoc(total_phys_mem);
692
693 /*
694 * Avail_end is set to the first byte of physical memory
695 * after the end of the last bank. We use this only to
696 * determine if a physical address is "managed" memory.
697 *
698 * XXX - The setting of avail_end is a temporary ROM saving hack.
699 */
700 avail_end = pmap_membank->pmem_end -
701 (total_phys_mem - *romVectorPtr->memoryAvail);
702 avail_end = _trunc_page(avail_end);
703
704 /*
705 * First allocate enough kernel MMU tables to map all
706 * of kernel virtual space from KERNBASE to 0xFFFFFFFF.
707 * Note: All must be aligned on 256 byte boundaries.
708 * Start with the level-A table (one of those).
709 */
710 size = sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE;
711 kernAbase = pmap_bootstrap_alloc(size);
712 bzero(kernAbase, size);
713
714 /* Now the level-B kernel tables... */
715 size = sizeof(mmu_short_dte_t) * MMU_B_TBL_SIZE * KERN_B_TABLES;
716 kernBbase = pmap_bootstrap_alloc(size);
717 bzero(kernBbase, size);
718
719 /* Now the level-C kernel tables... */
720 size = sizeof(mmu_short_pte_t) * MMU_C_TBL_SIZE * KERN_C_TABLES;
721 kernCbase = pmap_bootstrap_alloc(size);
722 bzero(kernCbase, size);
723 /*
724 * Note: In order for the PV system to work correctly, the kernel
725 * and user-level C tables must be allocated contiguously.
726 * Nothing should be allocated between here and the allocation of
727 * mmuCbase below. XXX: Should do this as one allocation, and
728 * then compute a pointer for mmuCbase instead of this...
729 *
730 * Allocate user MMU tables.
731 * These must be contiguous with the preceeding.
732 */
733
734 #ifndef FIXED_NTABLES
735 /*
736 * The number of user-level C tables that should be allocated is
737 * related to the size of physical memory. In general, there should
738 * be enough tables to map four times the amount of available RAM.
739 * The extra amount is needed because some table space is wasted by
740 * fragmentation.
741 */
742 NUM_C_TABLES = (total_phys_mem * 4) / (MMU_C_TBL_SIZE * MMU_PAGE_SIZE);
743 NUM_B_TABLES = NUM_C_TABLES / 2;
744 NUM_A_TABLES = NUM_B_TABLES / 2;
745 #endif /* !FIXED_NTABLES */
746
747 size = sizeof(mmu_short_pte_t) * MMU_C_TBL_SIZE * NUM_C_TABLES;
748 mmuCbase = pmap_bootstrap_alloc(size);
749
750 size = sizeof(mmu_short_dte_t) * MMU_B_TBL_SIZE * NUM_B_TABLES;
751 mmuBbase = pmap_bootstrap_alloc(size);
752
753 size = sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE * NUM_A_TABLES;
754 mmuAbase = pmap_bootstrap_alloc(size);
755
756 /*
757 * Fill in the never-changing part of the kernel tables.
758 * For simplicity, the kernel's mappings will be editable as a
759 * flat array of page table entries at kernCbase. The
760 * higher level 'A' and 'B' tables must be initialized to point
761 * to this lower one.
762 */
763 b = c = 0;
764
765 /*
766 * Invalidate all mappings below KERNBASE in the A table.
767 * This area has already been zeroed out, but it is good
768 * practice to explicitly show that we are interpreting
769 * it as a list of A table descriptors.
770 */
771 for (i = 0; i < MMU_TIA(KERNBASE); i++) {
772 kernAbase[i].addr.raw = 0;
773 }
774
775 /*
776 * Set up the kernel A and B tables so that they will reference the
777 * correct spots in the contiguous table of PTEs allocated for the
778 * kernel's virtual memory space.
779 */
780 for (i = MMU_TIA(KERNBASE); i < MMU_A_TBL_SIZE; i++) {
781 kernAbase[i].attr.raw =
782 MMU_LONG_DTE_LU | MMU_LONG_DTE_SUPV | MMU_DT_SHORT;
783 kernAbase[i].addr.raw = mmu_vtop(&kernBbase[b]);
784
785 for (j=0; j < MMU_B_TBL_SIZE; j++) {
786 kernBbase[b + j].attr.raw = mmu_vtop(&kernCbase[c])
787 | MMU_DT_SHORT;
788 c += MMU_C_TBL_SIZE;
789 }
790 b += MMU_B_TBL_SIZE;
791 }
792
793 /* XXX - Doing kernel_pmap a little further down. */
794
795 pmap_alloc_usermmu(); /* Allocate user MMU tables. */
796 pmap_alloc_usertmgr(); /* Allocate user MMU table managers.*/
797 pmap_alloc_pv(); /* Allocate physical->virtual map. */
798
799 /*
800 * We are now done with pmap_bootstrap_alloc(). Round up
801 * `virtual_avail' to the nearest page, and set the flag
802 * to prevent use of pmap_bootstrap_alloc() hereafter.
803 */
804 pmap_bootstrap_aalign(NBPG);
805 bootstrap_alloc_enabled = FALSE;
806
807 /*
808 * Now that we are done with pmap_bootstrap_alloc(), we
809 * must save the virtual and physical addresses of the
810 * end of the linearly mapped range, which are stored in
811 * virtual_contig_end and avail_start, respectively.
812 * These variables will never change after this point.
813 */
814 virtual_contig_end = virtual_avail;
815 avail_start = virtual_avail - KERNBASE;
816
817 /*
818 * `avail_next' is a running pointer used by pmap_next_page() to
819 * keep track of the next available physical page to be handed
820 * to the VM system during its initialization, in which it
821 * asks for physical pages, one at a time.
822 */
823 avail_next = avail_start;
824
825 /*
826 * Now allocate some virtual addresses, but not the physical pages
827 * behind them. Note that virtual_avail is already page-aligned.
828 *
829 * tmp_vpages[] is an array of two virtual pages used for temporary
830 * kernel mappings in the pmap module to facilitate various physical
831 * address-oritented operations.
832 */
833 tmp_vpages[0] = virtual_avail;
834 virtual_avail += NBPG;
835 tmp_vpages[1] = virtual_avail;
836 virtual_avail += NBPG;
837 tmp_vpage0_inuse = tmp_vpage1_inuse = 0;
838
839 /** Initialize the PV system **/
840 pmap_init_pv();
841
842 /*
843 * Fill in the kernel_pmap structure and kernel_crp.
844 */
845 kernAphys = mmu_vtop(kernAbase);
846 kernel_pmap.pm_a_tmgr = NULL;
847 kernel_pmap.pm_a_phys = kernAphys;
848 kernel_pmap.pm_refcount = 1; /* always in use */
849
850 kernel_crp.rp_attr = MMU_LONG_DTE_LU | MMU_DT_LONG;
851 kernel_crp.rp_addr = kernAphys;
852
853 /*
854 * Now pmap_enter_kernel() may be used safely and will be
855 * the main interface used hereafter to modify the kernel's
856 * virtual address space. Note that since we are still running
857 * under the PROM's address table, none of these table modifications
858 * actually take effect until pmap_takeover_mmu() is called.
859 *
860 * Note: Our tables do NOT have the PROM linear mappings!
861 * Only the mappings created here exist in our tables, so
862 * remember to map anything we expect to use.
863 */
864 va = (vm_offset_t) KERNBASE;
865 pa = 0;
866
867 /*
868 * The first page of the kernel virtual address space is the msgbuf
869 * page. The page attributes (data, non-cached) are set here, while
870 * the address is assigned to this global pointer in cpu_startup().
871 * XXX - Make it non-cached?
872 */
873 pmap_enter_kernel(va, pa|PMAP_NC, VM_PROT_ALL);
874 va += NBPG; pa += NBPG;
875
876 /* Next page is used as the temporary stack. */
877 pmap_enter_kernel(va, pa, VM_PROT_ALL);
878 va += NBPG; pa += NBPG;
879
880 /*
881 * Map all of the kernel's text segment as read-only and cacheable.
882 * (Cacheable is implied by default). Unfortunately, the last bytes
883 * of kernel text and the first bytes of kernel data will often be
884 * sharing the same page. Therefore, the last page of kernel text
885 * has to be mapped as read/write, to accomodate the data.
886 */
887 eva = _trunc_page((vm_offset_t)etext);
888 for (; va < eva; va += NBPG, pa += NBPG)
889 pmap_enter_kernel(va, pa, VM_PROT_READ|VM_PROT_EXECUTE);
890
891 /*
892 * Map all of the kernel's data as read/write and cacheable.
893 * This includes: data, BSS, symbols, and everything in the
894 * contiguous memory used by pmap_bootstrap_alloc()
895 */
896 for (; pa < avail_start; va += NBPG, pa += NBPG)
897 pmap_enter_kernel(va, pa, VM_PROT_READ|VM_PROT_WRITE);
898
899 /*
900 * At this point we are almost ready to take over the MMU. But first
901 * we must save the PROM's address space in our map, as we call its
902 * routines and make references to its data later in the kernel.
903 */
904 pmap_bootstrap_copyprom();
905 pmap_takeover_mmu();
906 pmap_bootstrap_setprom();
907
908 /* Notify the VM system of our page size. */
909 PAGE_SIZE = NBPG;
910 vm_set_page_size();
911 }
912
913
914 /* pmap_alloc_usermmu INTERNAL
915 **
916 * Called from pmap_bootstrap() to allocate MMU tables that will
917 * eventually be used for user mappings.
918 */
919 void
920 pmap_alloc_usermmu()
921 {
922 /* XXX: Moved into caller. */
923 }
924
925 /* pmap_alloc_pv INTERNAL
926 **
927 * Called from pmap_bootstrap() to allocate the physical
928 * to virtual mapping list. Each physical page of memory
929 * in the system has a corresponding element in this list.
930 */
931 void
932 pmap_alloc_pv()
933 {
934 int i;
935 unsigned int total_mem;
936
937 /*
938 * Allocate a pv_head structure for every page of physical
939 * memory that will be managed by the system. Since memory on
940 * the 3/80 is non-contiguous, we cannot arrive at a total page
941 * count by subtraction of the lowest available address from the
942 * highest, but rather we have to step through each memory
943 * bank and add the number of pages in each to the total.
944 *
945 * At this time we also initialize the offset of each bank's
946 * starting pv_head within the pv_head list so that the physical
947 * memory state routines (pmap_is_referenced(),
948 * pmap_is_modified(), et al.) can quickly find coresponding
949 * pv_heads in spite of the non-contiguity.
950 */
951 total_mem = 0;
952 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
953 avail_mem[i].pmem_pvbase = _btop(total_mem);
954 total_mem += avail_mem[i].pmem_end -
955 avail_mem[i].pmem_start;
956 if (avail_mem[i].pmem_next == NULL)
957 break;
958 }
959 #ifdef PMAP_DEBUG
960 if (total_mem != total_phys_mem)
961 panic("pmap_alloc_pv did not arrive at correct page count");
962 #endif
963
964 pvbase = (pv_t *) pmap_bootstrap_alloc(sizeof(pv_t) *
965 _btop(total_phys_mem));
966 }
967
968 /* pmap_alloc_usertmgr INTERNAL
969 **
970 * Called from pmap_bootstrap() to allocate the structures which
971 * facilitate management of user MMU tables. Each user MMU table
972 * in the system has one such structure associated with it.
973 */
974 void
975 pmap_alloc_usertmgr()
976 {
977 /* Allocate user MMU table managers */
978 /* It would be a lot simpler to just make these BSS, but */
979 /* we may want to change their size at boot time... -j */
980 Atmgrbase = (a_tmgr_t *) pmap_bootstrap_alloc(sizeof(a_tmgr_t)
981 * NUM_A_TABLES);
982 Btmgrbase = (b_tmgr_t *) pmap_bootstrap_alloc(sizeof(b_tmgr_t)
983 * NUM_B_TABLES);
984 Ctmgrbase = (c_tmgr_t *) pmap_bootstrap_alloc(sizeof(c_tmgr_t)
985 * NUM_C_TABLES);
986
987 /*
988 * Allocate PV list elements for the physical to virtual
989 * mapping system.
990 */
991 pvebase = (pv_elem_t *) pmap_bootstrap_alloc(
992 sizeof(pv_elem_t) * (NUM_USER_PTES + NUM_KERN_PTES));
993 }
994
995 /* pmap_bootstrap_copyprom() INTERNAL
996 **
997 * Copy the PROM mappings into our own tables. Note, we
998 * can use physical addresses until __bootstrap returns.
999 */
1000 void
1001 pmap_bootstrap_copyprom()
1002 {
1003 MachMonRomVector *romp;
1004 int *mon_ctbl;
1005 mmu_short_pte_t *kpte;
1006 int i, len;
1007
1008 romp = romVectorPtr;
1009
1010 /*
1011 * Copy the mappings in MON_KDB_START...MONEND
1012 * Note: mon_ctbl[0] maps MON_KDB_START
1013 */
1014 mon_ctbl = *romp->monptaddr;
1015 i = _btop(MON_KDB_START - KERNBASE);
1016 kpte = &kernCbase[i];
1017 len = _btop(MONEND - MON_KDB_START);
1018
1019 for (i = 0; i < len; i++) {
1020 kpte[i].attr.raw = mon_ctbl[i];
1021 }
1022
1023 /*
1024 * Copy the mappings at MON_DVMA_BASE (to the end).
1025 * Note, in here, mon_ctbl[0] maps MON_DVMA_BASE.
1026 * XXX - This does not appear to be necessary, but
1027 * I'm not sure yet if it is or not. -gwr
1028 */
1029 mon_ctbl = *romp->shadowpteaddr;
1030 i = _btop(MON_DVMA_BASE - KERNBASE);
1031 kpte = &kernCbase[i];
1032 len = _btop(MON_DVMA_SIZE);
1033
1034 for (i = 0; i < len; i++) {
1035 kpte[i].attr.raw = mon_ctbl[i];
1036 }
1037 }
1038
1039 /* pmap_takeover_mmu INTERNAL
1040 **
1041 * Called from pmap_bootstrap() after it has copied enough of the
1042 * PROM mappings into the kernel map so that we can use our own
1043 * MMU table.
1044 */
1045 void
1046 pmap_takeover_mmu()
1047 {
1048
1049 loadcrp(&kernel_crp);
1050 }
1051
1052 /* pmap_bootstrap_setprom() INTERNAL
1053 **
1054 * Set the PROM mappings so it can see kernel space.
1055 * Note that physical addresses are used here, which
1056 * we can get away with because this runs with the
1057 * low 1GB set for transparent translation.
1058 */
1059 void
1060 pmap_bootstrap_setprom()
1061 {
1062 mmu_long_dte_t *mon_dte;
1063 extern struct mmu_rootptr mon_crp;
1064 int i;
1065
1066 mon_dte = (mmu_long_dte_t *) mon_crp.rp_addr;
1067 for (i = MMU_TIA(KERNBASE); i < MMU_TIA(KERN_END); i++) {
1068 mon_dte[i].attr.raw = kernAbase[i].attr.raw;
1069 mon_dte[i].addr.raw = kernAbase[i].addr.raw;
1070 }
1071 }
1072
1073
1074 /* pmap_init INTERFACE
1075 **
1076 * Called at the end of vm_init() to set up the pmap system to go
1077 * into full time operation. All initialization of kernel_pmap
1078 * should be already done by now, so this should just do things
1079 * needed for user-level pmaps to work.
1080 */
1081 void
1082 pmap_init()
1083 {
1084 /** Initialize the manager pools **/
1085 TAILQ_INIT(&a_pool);
1086 TAILQ_INIT(&b_pool);
1087 TAILQ_INIT(&c_pool);
1088
1089 /**************************************************************
1090 * Initialize all tmgr structures and MMU tables they manage. *
1091 **************************************************************/
1092 /** Initialize A tables **/
1093 pmap_init_a_tables();
1094 /** Initialize B tables **/
1095 pmap_init_b_tables();
1096 /** Initialize C tables **/
1097 pmap_init_c_tables();
1098 }
1099
1100 /* pmap_init_a_tables() INTERNAL
1101 **
1102 * Initializes all A managers, their MMU A tables, and inserts
1103 * them into the A manager pool for use by the system.
1104 */
1105 void
1106 pmap_init_a_tables()
1107 {
1108 int i;
1109 a_tmgr_t *a_tbl;
1110
1111 for (i=0; i < NUM_A_TABLES; i++) {
1112 /* Select the next available A manager from the pool */
1113 a_tbl = &Atmgrbase[i];
1114
1115 /*
1116 * Clear its parent entry. Set its wired and valid
1117 * entry count to zero.
1118 */
1119 a_tbl->at_parent = NULL;
1120 a_tbl->at_wcnt = a_tbl->at_ecnt = 0;
1121
1122 /* Assign it the next available MMU A table from the pool */
1123 a_tbl->at_dtbl = &mmuAbase[i * MMU_A_TBL_SIZE];
1124
1125 /*
1126 * Initialize the MMU A table with the table in the `proc0',
1127 * or kernel, mapping. This ensures that every process has
1128 * the kernel mapped in the top part of its address space.
1129 */
1130 bcopy(kernAbase, a_tbl->at_dtbl, MMU_A_TBL_SIZE *
1131 sizeof(mmu_long_dte_t));
1132
1133 /*
1134 * Finally, insert the manager into the A pool,
1135 * making it ready to be used by the system.
1136 */
1137 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
1138 }
1139 }
1140
1141 /* pmap_init_b_tables() INTERNAL
1142 **
1143 * Initializes all B table managers, their MMU B tables, and
1144 * inserts them into the B manager pool for use by the system.
1145 */
1146 void
1147 pmap_init_b_tables()
1148 {
1149 int i,j;
1150 b_tmgr_t *b_tbl;
1151
1152 for (i=0; i < NUM_B_TABLES; i++) {
1153 /* Select the next available B manager from the pool */
1154 b_tbl = &Btmgrbase[i];
1155
1156 b_tbl->bt_parent = NULL; /* clear its parent, */
1157 b_tbl->bt_pidx = 0; /* parent index, */
1158 b_tbl->bt_wcnt = 0; /* wired entry count, */
1159 b_tbl->bt_ecnt = 0; /* valid entry count. */
1160
1161 /* Assign it the next available MMU B table from the pool */
1162 b_tbl->bt_dtbl = &mmuBbase[i * MMU_B_TBL_SIZE];
1163
1164 /* Invalidate every descriptor in the table */
1165 for (j=0; j < MMU_B_TBL_SIZE; j++)
1166 b_tbl->bt_dtbl[j].attr.raw = MMU_DT_INVALID;
1167
1168 /* Insert the manager into the B pool */
1169 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
1170 }
1171 }
1172
1173 /* pmap_init_c_tables() INTERNAL
1174 **
1175 * Initializes all C table managers, their MMU C tables, and
1176 * inserts them into the C manager pool for use by the system.
1177 */
1178 void
1179 pmap_init_c_tables()
1180 {
1181 int i,j;
1182 c_tmgr_t *c_tbl;
1183
1184 for (i=0; i < NUM_C_TABLES; i++) {
1185 /* Select the next available C manager from the pool */
1186 c_tbl = &Ctmgrbase[i];
1187
1188 c_tbl->ct_parent = NULL; /* clear its parent, */
1189 c_tbl->ct_pidx = 0; /* parent index, */
1190 c_tbl->ct_wcnt = 0; /* wired entry count, */
1191 c_tbl->ct_ecnt = 0; /* valid entry count. */
1192
1193 /* Assign it the next available MMU C table from the pool */
1194 c_tbl->ct_dtbl = &mmuCbase[i * MMU_C_TBL_SIZE];
1195
1196 for (j=0; j < MMU_C_TBL_SIZE; j++)
1197 c_tbl->ct_dtbl[j].attr.raw = MMU_DT_INVALID;
1198
1199 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
1200 }
1201 }
1202
1203 /* pmap_init_pv() INTERNAL
1204 **
1205 * Initializes the Physical to Virtual mapping system.
1206 */
1207 void
1208 pmap_init_pv()
1209 {
1210 int i;
1211
1212 /* Initialize every PV head. */
1213 for (i = 0; i < _btop(total_phys_mem); i++) {
1214 pvbase[i].pv_idx = PVE_EOL; /* Indicate no mappings */
1215 pvbase[i].pv_flags = 0; /* Zero out page flags */
1216 }
1217
1218 pv_initialized = TRUE;
1219 }
1220
1221 /* get_a_table INTERNAL
1222 **
1223 * Retrieve and return a level A table for use in a user map.
1224 */
1225 a_tmgr_t *
1226 get_a_table()
1227 {
1228 a_tmgr_t *tbl;
1229 pmap_t pmap;
1230
1231 /* Get the top A table in the pool */
1232 tbl = a_pool.tqh_first;
1233 if (tbl == NULL) {
1234 /*
1235 * XXX - Instead of panicing here and in other get_x_table
1236 * functions, we do have the option of sleeping on the head of
1237 * the table pool. Any function which updates the table pool
1238 * would then issue a wakeup() on the head, thus waking up any
1239 * processes waiting for a table.
1240 *
1241 * Actually, the place to sleep would be when some process
1242 * asks for a "wired" mapping that would run us short of
1243 * mapping resources. This design DEPENDS on always having
1244 * some mapping resources in the pool for stealing, so we
1245 * must make sure we NEVER let the pool become empty. -gwr
1246 */
1247 panic("get_a_table: out of A tables.");
1248 }
1249
1250 TAILQ_REMOVE(&a_pool, tbl, at_link);
1251 /*
1252 * If the table has a non-null parent pointer then it is in use.
1253 * Forcibly abduct it from its parent and clear its entries.
1254 * No re-entrancy worries here. This table would not be in the
1255 * table pool unless it was available for use.
1256 *
1257 * Note that the second argument to free_a_table() is FALSE. This
1258 * indicates that the table should not be relinked into the A table
1259 * pool. That is a job for the function that called us.
1260 */
1261 if (tbl->at_parent) {
1262 pmap = tbl->at_parent;
1263 free_a_table(tbl, FALSE);
1264 pmap->pm_a_tmgr = NULL;
1265 pmap->pm_a_phys = kernAphys;
1266 }
1267 #ifdef NON_REENTRANT
1268 /*
1269 * If the table isn't to be wired down, re-insert it at the
1270 * end of the pool.
1271 */
1272 if (!wired)
1273 /*
1274 * Quandary - XXX
1275 * Would it be better to let the calling function insert this
1276 * table into the queue? By inserting it here, we are allowing
1277 * it to be stolen immediately. The calling function is
1278 * probably not expecting to use a table that it is not
1279 * assured full control of.
1280 * Answer - In the intrest of re-entrancy, it is best to let
1281 * the calling function determine when a table is available
1282 * for use. Therefore this code block is not used.
1283 */
1284 TAILQ_INSERT_TAIL(&a_pool, tbl, at_link);
1285 #endif /* NON_REENTRANT */
1286 return tbl;
1287 }
1288
1289 /* get_b_table INTERNAL
1290 **
1291 * Return a level B table for use.
1292 */
1293 b_tmgr_t *
1294 get_b_table()
1295 {
1296 b_tmgr_t *tbl;
1297
1298 /* See 'get_a_table' for comments. */
1299 tbl = b_pool.tqh_first;
1300 if (tbl == NULL)
1301 panic("get_b_table: out of B tables.");
1302 TAILQ_REMOVE(&b_pool, tbl, bt_link);
1303 if (tbl->bt_parent) {
1304 tbl->bt_parent->at_dtbl[tbl->bt_pidx].attr.raw = MMU_DT_INVALID;
1305 tbl->bt_parent->at_ecnt--;
1306 free_b_table(tbl, FALSE);
1307 }
1308 #ifdef NON_REENTRANT
1309 if (!wired)
1310 /* XXX see quandary in get_b_table */
1311 /* XXX start lock */
1312 TAILQ_INSERT_TAIL(&b_pool, tbl, bt_link);
1313 /* XXX end lock */
1314 #endif /* NON_REENTRANT */
1315 return tbl;
1316 }
1317
1318 /* get_c_table INTERNAL
1319 **
1320 * Return a level C table for use.
1321 */
1322 c_tmgr_t *
1323 get_c_table()
1324 {
1325 c_tmgr_t *tbl;
1326
1327 /* See 'get_a_table' for comments */
1328 tbl = c_pool.tqh_first;
1329 if (tbl == NULL)
1330 panic("get_c_table: out of C tables.");
1331 TAILQ_REMOVE(&c_pool, tbl, ct_link);
1332 if (tbl->ct_parent) {
1333 tbl->ct_parent->bt_dtbl[tbl->ct_pidx].attr.raw = MMU_DT_INVALID;
1334 tbl->ct_parent->bt_ecnt--;
1335 free_c_table(tbl, FALSE);
1336 }
1337 #ifdef NON_REENTRANT
1338 if (!wired)
1339 /* XXX See quandary in get_a_table */
1340 /* XXX start lock */
1341 TAILQ_INSERT_TAIL(&c_pool, tbl, c_link);
1342 /* XXX end lock */
1343 #endif /* NON_REENTRANT */
1344
1345 return tbl;
1346 }
1347
1348 /*
1349 * The following 'free_table' and 'steal_table' functions are called to
1350 * detach tables from their current obligations (parents and children) and
1351 * prepare them for reuse in another mapping.
1352 *
1353 * Free_table is used when the calling function will handle the fate
1354 * of the parent table, such as returning it to the free pool when it has
1355 * no valid entries. Functions that do not want to handle this should
1356 * call steal_table, in which the parent table's descriptors and entry
1357 * count are automatically modified when this table is removed.
1358 */
1359
1360 /* free_a_table INTERNAL
1361 **
1362 * Unmaps the given A table and all child tables from their current
1363 * mappings. Returns the number of pages that were invalidated.
1364 * If 'relink' is true, the function will return the table to the head
1365 * of the available table pool.
1366 *
1367 * Cache note: The MC68851 will automatically flush all
1368 * descriptors derived from a given A table from its
1369 * Automatic Translation Cache (ATC) if we issue a
1370 * 'PFLUSHR' instruction with the base address of the
1371 * table. This function should do, and does so.
1372 * Note note: We are using an MC68030 - there is no
1373 * PFLUSHR.
1374 */
1375 int
1376 free_a_table(a_tbl, relink)
1377 a_tmgr_t *a_tbl;
1378 boolean_t relink;
1379 {
1380 int i, removed_cnt;
1381 mmu_long_dte_t *dte;
1382 mmu_short_dte_t *dtbl;
1383 b_tmgr_t *tmgr;
1384
1385 /*
1386 * Flush the ATC cache of all cached descriptors derived
1387 * from this table.
1388 * Sun3x does not use 68851's cached table feature
1389 * flush_atc_crp(mmu_vtop(a_tbl->dte));
1390 */
1391
1392 /*
1393 * Remove any pending cache flushes that were designated
1394 * for the pmap this A table belongs to.
1395 * a_tbl->parent->atc_flushq[0] = 0;
1396 * Not implemented in sun3x.
1397 */
1398
1399 /*
1400 * All A tables in the system should retain a map for the
1401 * kernel. If the table contains any valid descriptors
1402 * (other than those for the kernel area), invalidate them all,
1403 * stopping short of the kernel's entries.
1404 */
1405 removed_cnt = 0;
1406 if (a_tbl->at_ecnt) {
1407 dte = a_tbl->at_dtbl;
1408 for (i=0; i < MMU_TIA(KERNBASE); i++) {
1409 /*
1410 * If a table entry points to a valid B table, free
1411 * it and its children.
1412 */
1413 if (MMU_VALID_DT(dte[i])) {
1414 /*
1415 * The following block does several things,
1416 * from innermost expression to the
1417 * outermost:
1418 * 1) It extracts the base (cc 1996)
1419 * address of the B table pointed
1420 * to in the A table entry dte[i].
1421 * 2) It converts this base address into
1422 * the virtual address it can be
1423 * accessed with. (all MMU tables point
1424 * to physical addresses.)
1425 * 3) It finds the corresponding manager
1426 * structure which manages this MMU table.
1427 * 4) It frees the manager structure.
1428 * (This frees the MMU table and all
1429 * child tables. See 'free_b_table' for
1430 * details.)
1431 */
1432 dtbl = mmu_ptov(dte[i].addr.raw);
1433 tmgr = mmuB2tmgr(dtbl);
1434 removed_cnt += free_b_table(tmgr, TRUE);
1435 dte[i].attr.raw = MMU_DT_INVALID;
1436 }
1437 }
1438 a_tbl->at_ecnt = 0;
1439 }
1440 if (relink) {
1441 a_tbl->at_parent = NULL;
1442 TAILQ_REMOVE(&a_pool, a_tbl, at_link);
1443 TAILQ_INSERT_HEAD(&a_pool, a_tbl, at_link);
1444 }
1445 return removed_cnt;
1446 }
1447
1448 /* free_b_table INTERNAL
1449 **
1450 * Unmaps the given B table and all its children from their current
1451 * mappings. Returns the number of pages that were invalidated.
1452 * (For comments, see 'free_a_table()').
1453 */
1454 int
1455 free_b_table(b_tbl, relink)
1456 b_tmgr_t *b_tbl;
1457 boolean_t relink;
1458 {
1459 int i, removed_cnt;
1460 mmu_short_dte_t *dte;
1461 mmu_short_pte_t *dtbl;
1462 c_tmgr_t *tmgr;
1463
1464 removed_cnt = 0;
1465 if (b_tbl->bt_ecnt) {
1466 dte = b_tbl->bt_dtbl;
1467 for (i=0; i < MMU_B_TBL_SIZE; i++) {
1468 if (MMU_VALID_DT(dte[i])) {
1469 dtbl = mmu_ptov(MMU_DTE_PA(dte[i]));
1470 tmgr = mmuC2tmgr(dtbl);
1471 removed_cnt += free_c_table(tmgr, TRUE);
1472 dte[i].attr.raw = MMU_DT_INVALID;
1473 }
1474 }
1475 b_tbl->bt_ecnt = 0;
1476 }
1477
1478 if (relink) {
1479 b_tbl->bt_parent = NULL;
1480 TAILQ_REMOVE(&b_pool, b_tbl, bt_link);
1481 TAILQ_INSERT_HEAD(&b_pool, b_tbl, bt_link);
1482 }
1483 return removed_cnt;
1484 }
1485
1486 /* free_c_table INTERNAL
1487 **
1488 * Unmaps the given C table from use and returns it to the pool for
1489 * re-use. Returns the number of pages that were invalidated.
1490 *
1491 * This function preserves any physical page modification information
1492 * contained in the page descriptors within the C table by calling
1493 * 'pmap_remove_pte().'
1494 */
1495 int
1496 free_c_table(c_tbl, relink)
1497 c_tmgr_t *c_tbl;
1498 boolean_t relink;
1499 {
1500 int i, removed_cnt;
1501
1502 removed_cnt = 0;
1503 if (c_tbl->ct_ecnt) {
1504 for (i=0; i < MMU_C_TBL_SIZE; i++) {
1505 if (MMU_VALID_DT(c_tbl->ct_dtbl[i])) {
1506 pmap_remove_pte(&c_tbl->ct_dtbl[i]);
1507 removed_cnt++;
1508 }
1509 }
1510 c_tbl->ct_ecnt = 0;
1511 }
1512
1513 if (relink) {
1514 c_tbl->ct_parent = NULL;
1515 TAILQ_REMOVE(&c_pool, c_tbl, ct_link);
1516 TAILQ_INSERT_HEAD(&c_pool, c_tbl, ct_link);
1517 }
1518 return removed_cnt;
1519 }
1520
1521 #if 0
1522 /* free_c_table_novalid INTERNAL
1523 **
1524 * Frees the given C table manager without checking to see whether
1525 * or not it contains any valid page descriptors as it is assumed
1526 * that it does not.
1527 */
1528 void
1529 free_c_table_novalid(c_tbl)
1530 c_tmgr_t *c_tbl;
1531 {
1532 TAILQ_REMOVE(&c_pool, c_tbl, ct_link);
1533 TAILQ_INSERT_HEAD(&c_pool, c_tbl, ct_link);
1534 c_tbl->ct_parent->bt_dtbl[c_tbl->ct_pidx].attr.raw = MMU_DT_INVALID;
1535 c_tbl->ct_parent->bt_ecnt--;
1536 /*
1537 * XXX - Should call equiv. of 'free_b_table_novalid' here if
1538 * we just removed the last entry of the parent B table.
1539 * But I want to insure that this will not endanger pmap_enter()
1540 * with sudden removal of tables it is working with.
1541 *
1542 * We should probably add another field to each table, indicating
1543 * whether or not it is 'locked', ie. in the process of being
1544 * modified.
1545 */
1546 c_tbl->ct_parent = NULL;
1547 }
1548 #endif
1549
1550 /* pmap_remove_pte INTERNAL
1551 **
1552 * Unmap the given pte and preserve any page modification
1553 * information by transfering it to the pv head of the
1554 * physical page it maps to. This function does not update
1555 * any reference counts because it is assumed that the calling
1556 * function will do so.
1557 */
1558 void
1559 pmap_remove_pte(pte)
1560 mmu_short_pte_t *pte;
1561 {
1562 u_short pv_idx, targ_idx;
1563 int s;
1564 vm_offset_t pa;
1565 pv_t *pv;
1566
1567 pa = MMU_PTE_PA(*pte);
1568 if (is_managed(pa)) {
1569 pv = pa2pv(pa);
1570 targ_idx = pteidx(pte); /* Index of PTE being removed */
1571
1572 /*
1573 * If the PTE being removed is the first (or only) PTE in
1574 * the list of PTEs currently mapped to this page, remove the
1575 * PTE by changing the index found on the PV head. Otherwise
1576 * a linear search through the list will have to be executed
1577 * in order to find the PVE which points to the PTE being
1578 * removed, so that it may be modified to point to its new
1579 * neighbor.
1580 */
1581 s = splimp();
1582 pv_idx = pv->pv_idx; /* Index of first PTE in PV list */
1583 if (pv_idx == targ_idx) {
1584 pv->pv_idx = pvebase[targ_idx].pve_next;
1585 } else {
1586 /*
1587 * Find the PV element which points to the target
1588 * element.
1589 */
1590 while (pvebase[pv_idx].pve_next != targ_idx) {
1591 pv_idx = pvebase[pv_idx].pve_next;
1592 #ifdef DIAGNOSTIC
1593 if (pv_idx == PVE_EOL)
1594 panic("pmap_remove_pte: pv list end!");
1595 #endif
1596 }
1597
1598 /*
1599 * At this point, pv_idx is the index of the PV
1600 * element just before the target element in the list.
1601 * Unlink the target.
1602 */
1603 pvebase[pv_idx].pve_next = pvebase[targ_idx].pve_next;
1604 }
1605 /*
1606 * Save the mod/ref bits of the pte by simply
1607 * ORing the entire pte onto the pv_flags member
1608 * of the pv structure.
1609 * There is no need to use a separate bit pattern
1610 * for usage information on the pv head than that
1611 * which is used on the MMU ptes.
1612 */
1613 pv->pv_flags |= (u_short) pte->attr.raw;
1614 splx(s);
1615 }
1616
1617 pte->attr.raw = MMU_DT_INVALID;
1618 }
1619
1620 /* pmap_stroll INTERNAL
1621 **
1622 * Retrieve the addresses of all table managers involved in the mapping of
1623 * the given virtual address. If the table walk completed sucessfully,
1624 * return TRUE. If it was only partially sucessful, return FALSE.
1625 * The table walk performed by this function is important to many other
1626 * functions in this module.
1627 *
1628 * Note: This function ought to be easier to read.
1629 */
1630 boolean_t
1631 pmap_stroll(pmap, va, a_tbl, b_tbl, c_tbl, pte, a_idx, b_idx, pte_idx)
1632 pmap_t pmap;
1633 vm_offset_t va;
1634 a_tmgr_t **a_tbl;
1635 b_tmgr_t **b_tbl;
1636 c_tmgr_t **c_tbl;
1637 mmu_short_pte_t **pte;
1638 int *a_idx, *b_idx, *pte_idx;
1639 {
1640 mmu_long_dte_t *a_dte; /* A: long descriptor table */
1641 mmu_short_dte_t *b_dte; /* B: short descriptor table */
1642
1643 if (pmap == pmap_kernel())
1644 return FALSE;
1645
1646 /* Does the given pmap have its own A table? */
1647 *a_tbl = pmap->pm_a_tmgr;
1648 if (*a_tbl == NULL)
1649 return FALSE; /* No. Return unknown. */
1650 /* Does the A table have a valid B table
1651 * under the corresponding table entry?
1652 */
1653 *a_idx = MMU_TIA(va);
1654 a_dte = &((*a_tbl)->at_dtbl[*a_idx]);
1655 if (!MMU_VALID_DT(*a_dte))
1656 return FALSE; /* No. Return unknown. */
1657 /* Yes. Extract B table from the A table. */
1658 *b_tbl = mmuB2tmgr(mmu_ptov(a_dte->addr.raw));
1659 /* Does the B table have a valid C table
1660 * under the corresponding table entry?
1661 */
1662 *b_idx = MMU_TIB(va);
1663 b_dte = &((*b_tbl)->bt_dtbl[*b_idx]);
1664 if (!MMU_VALID_DT(*b_dte))
1665 return FALSE; /* No. Return unknown. */
1666 /* Yes. Extract C table from the B table. */
1667 *c_tbl = mmuC2tmgr(mmu_ptov(MMU_DTE_PA(*b_dte)));
1668 *pte_idx = MMU_TIC(va);
1669 *pte = &((*c_tbl)->ct_dtbl[*pte_idx]);
1670
1671 return TRUE;
1672 }
1673
1674 /* pmap_enter INTERFACE
1675 **
1676 * Called by the kernel to map a virtual address
1677 * to a physical address in the given process map.
1678 *
1679 * Note: this function should apply an exclusive lock
1680 * on the pmap system for its duration. (it certainly
1681 * would save my hair!!)
1682 * This function ought to be easier to read.
1683 */
1684 void
1685 pmap_enter(pmap, va, pa, prot, wired)
1686 pmap_t pmap;
1687 vm_offset_t va;
1688 vm_offset_t pa;
1689 vm_prot_t prot;
1690 boolean_t wired;
1691 {
1692 boolean_t insert, managed; /* Marks the need for PV insertion.*/
1693 u_short nidx; /* PV list index */
1694 int s; /* Used for splimp()/splx() */
1695 int flags; /* Mapping flags. eg. Cache inhibit */
1696 u_int a_idx, b_idx, pte_idx; /* table indices */
1697 a_tmgr_t *a_tbl; /* A: long descriptor table manager */
1698 b_tmgr_t *b_tbl; /* B: short descriptor table manager */
1699 c_tmgr_t *c_tbl; /* C: short page table manager */
1700 mmu_long_dte_t *a_dte; /* A: long descriptor table */
1701 mmu_short_dte_t *b_dte; /* B: short descriptor table */
1702 mmu_short_pte_t *c_pte; /* C: short page descriptor table */
1703 pv_t *pv; /* pv list head */
1704 enum {NONE, NEWA, NEWB, NEWC} llevel; /* used at end */
1705
1706 if (pmap == NULL)
1707 return;
1708 if (pmap == pmap_kernel()) {
1709 pmap_enter_kernel(va, pa, prot);
1710 return;
1711 }
1712
1713 flags = (pa & ~MMU_PAGE_MASK);
1714 pa &= MMU_PAGE_MASK;
1715
1716 /*
1717 * Determine if the physical address being mapped is on-board RAM.
1718 * Any other area of the address space is likely to belong to a
1719 * device and hence it would be disasterous to cache its contents.
1720 */
1721 if ((managed = is_managed(pa)) == FALSE)
1722 flags |= PMAP_NC;
1723
1724 /*
1725 * For user mappings we walk along the MMU tables of the given
1726 * pmap, reaching a PTE which describes the virtual page being
1727 * mapped or changed. If any level of the walk ends in an invalid
1728 * entry, a table must be allocated and the entry must be updated
1729 * to point to it.
1730 * There is a bit of confusion as to whether this code must be
1731 * re-entrant. For now we will assume it is. To support
1732 * re-entrancy we must unlink tables from the table pool before
1733 * we assume we may use them. Tables are re-linked into the pool
1734 * when we are finished with them at the end of the function.
1735 * But I don't feel like doing that until we have proof that this
1736 * needs to be re-entrant.
1737 * 'llevel' records which tables need to be relinked.
1738 */
1739 llevel = NONE;
1740
1741 /*
1742 * Step 1 - Retrieve the A table from the pmap. If it has no
1743 * A table, allocate a new one from the available pool.
1744 */
1745
1746 a_tbl = pmap->pm_a_tmgr;
1747 if (a_tbl == NULL) {
1748 /*
1749 * This pmap does not currently have an A table. Allocate
1750 * a new one.
1751 */
1752 a_tbl = get_a_table();
1753 a_tbl->at_parent = pmap;
1754
1755 /*
1756 * Assign this new A table to the pmap, and calculate its
1757 * physical address so that loadcrp() can be used to make
1758 * the table active.
1759 */
1760 pmap->pm_a_tmgr = a_tbl;
1761 pmap->pm_a_phys = mmu_vtop(a_tbl->at_dtbl);
1762
1763 /*
1764 * If the process receiving a new A table is the current
1765 * process, we are responsible for setting the MMU so that
1766 * it becomes the current address space. This only adds
1767 * new mappings, so no need to flush anything.
1768 */
1769 if (pmap == current_pmap()) {
1770 kernel_crp.rp_addr = pmap->pm_a_phys;
1771 loadcrp(&kernel_crp);
1772 }
1773
1774 if (!wired)
1775 llevel = NEWA;
1776 } else {
1777 /*
1778 * Use the A table already allocated for this pmap.
1779 * Unlink it from the A table pool if necessary.
1780 */
1781 if (wired && !a_tbl->at_wcnt)
1782 TAILQ_REMOVE(&a_pool, a_tbl, at_link);
1783 }
1784
1785 /*
1786 * Step 2 - Walk into the B table. If there is no valid B table,
1787 * allocate one.
1788 */
1789
1790 a_idx = MMU_TIA(va); /* Calculate the TIA of the VA. */
1791 a_dte = &a_tbl->at_dtbl[a_idx]; /* Retrieve descriptor from table */
1792 if (MMU_VALID_DT(*a_dte)) { /* Is the descriptor valid? */
1793 /* The descriptor is valid. Use the B table it points to. */
1794 /*************************************
1795 * a_idx *
1796 * v *
1797 * a_tbl -> +-+-+-+-+-+-+-+-+-+-+-+- *
1798 * | | | | | | | | | | | | *
1799 * +-+-+-+-+-+-+-+-+-+-+-+- *
1800 * | *
1801 * \- b_tbl -> +-+- *
1802 * | | *
1803 * +-+- *
1804 *************************************/
1805 b_dte = mmu_ptov(a_dte->addr.raw);
1806 b_tbl = mmuB2tmgr(b_dte);
1807
1808 /*
1809 * If the requested mapping must be wired, but this table
1810 * being used to map it is not, the table must be removed
1811 * from the available pool and its wired entry count
1812 * incremented.
1813 */
1814 if (wired && !b_tbl->bt_wcnt) {
1815 TAILQ_REMOVE(&b_pool, b_tbl, bt_link);
1816 a_tbl->at_wcnt++;
1817 }
1818 } else {
1819 /* The descriptor is invalid. Allocate a new B table. */
1820 b_tbl = get_b_table();
1821
1822 /* Point the parent A table descriptor to this new B table. */
1823 a_dte->addr.raw = mmu_vtop(b_tbl->bt_dtbl);
1824 a_dte->attr.raw = MMU_LONG_DTE_LU | MMU_DT_SHORT;
1825 a_tbl->at_ecnt++; /* Update parent's valid entry count */
1826
1827 /* Create the necessary back references to the parent table */
1828 b_tbl->bt_parent = a_tbl;
1829 b_tbl->bt_pidx = a_idx;
1830
1831 /*
1832 * If this table is to be wired, make sure the parent A table
1833 * wired count is updated to reflect that it has another wired
1834 * entry.
1835 */
1836 if (wired)
1837 a_tbl->at_wcnt++;
1838 else if (llevel == NONE)
1839 llevel = NEWB;
1840 }
1841
1842 /*
1843 * Step 3 - Walk into the C table, if there is no valid C table,
1844 * allocate one.
1845 */
1846
1847 b_idx = MMU_TIB(va); /* Calculate the TIB of the VA */
1848 b_dte = &b_tbl->bt_dtbl[b_idx]; /* Retrieve descriptor from table */
1849 if (MMU_VALID_DT(*b_dte)) { /* Is the descriptor valid? */
1850 /* The descriptor is valid. Use the C table it points to. */
1851 /**************************************
1852 * c_idx *
1853 * | v *
1854 * \- b_tbl -> +-+-+-+-+-+-+-+-+-+-+- *
1855 * | | | | | | | | | | | *
1856 * +-+-+-+-+-+-+-+-+-+-+- *
1857 * | *
1858 * \- c_tbl -> +-+-- *
1859 * | | | *
1860 * +-+-- *
1861 **************************************/
1862 c_pte = mmu_ptov(MMU_PTE_PA(*b_dte));
1863 c_tbl = mmuC2tmgr(c_pte);
1864
1865 /* If mapping is wired and table is not */
1866 if (wired && !c_tbl->ct_wcnt) {
1867 TAILQ_REMOVE(&c_pool, c_tbl, ct_link);
1868 b_tbl->bt_wcnt++;
1869 }
1870 } else {
1871 /* The descriptor is invalid. Allocate a new C table. */
1872 c_tbl = get_c_table();
1873
1874 /* Point the parent B table descriptor to this new C table. */
1875 b_dte->attr.raw = mmu_vtop(c_tbl->ct_dtbl);
1876 b_dte->attr.raw |= MMU_DT_SHORT;
1877 b_tbl->bt_ecnt++; /* Update parent's valid entry count */
1878
1879 /* Create the necessary back references to the parent table */
1880 c_tbl->ct_parent = b_tbl;
1881 c_tbl->ct_pidx = b_idx;
1882
1883 /*
1884 * If this table is to be wired, make sure the parent B table
1885 * wired count is updated to reflect that it has another wired
1886 * entry.
1887 */
1888 if (wired)
1889 b_tbl->bt_wcnt++;
1890 else if (llevel == NONE)
1891 llevel = NEWC;
1892 }
1893
1894 /*
1895 * Step 4 - Deposit a page descriptor (PTE) into the appropriate
1896 * slot of the C table, describing the PA to which the VA is mapped.
1897 */
1898
1899 pte_idx = MMU_TIC(va);
1900 c_pte = &c_tbl->ct_dtbl[pte_idx];
1901 if (MMU_VALID_DT(*c_pte)) { /* Is the entry currently valid? */
1902 /*
1903 * The PTE is currently valid. This particular call
1904 * is just a synonym for one (or more) of the following
1905 * operations:
1906 * change protection of a page
1907 * change wiring status of a page
1908 * remove the mapping of a page
1909 *
1910 * XXX - Semi critical: This code should unwire the PTE
1911 * and, possibly, associated parent tables if this is a
1912 * change wiring operation. Currently it does not.
1913 *
1914 * This may be ok if pmap_change_wiring() is the only
1915 * interface used to UNWIRE a page.
1916 */
1917
1918 /* First check if this is a wiring operation. */
1919 if (wired && (c_pte->attr.raw & MMU_SHORT_PTE_WIRED)) {
1920 /*
1921 * The PTE is already wired. To prevent it from being
1922 * counted as a new wiring operation, reset the 'wired'
1923 * variable.
1924 */
1925 wired = FALSE;
1926 }
1927
1928 /* Is the new address the same as the old? */
1929 if (MMU_PTE_PA(*c_pte) == pa) {
1930 /*
1931 * Yes, mark that it does not need to be reinserted
1932 * into the PV list.
1933 */
1934 insert = FALSE;
1935
1936 /*
1937 * Clear all but the modified, referenced and wired
1938 * bits on the PTE.
1939 */
1940 c_pte->attr.raw &= (MMU_SHORT_PTE_M
1941 | MMU_SHORT_PTE_USED | MMU_SHORT_PTE_WIRED);
1942 } else {
1943 /* No, remove the old entry */
1944 pmap_remove_pte(c_pte);
1945 insert = TRUE;
1946 }
1947
1948 /*
1949 * TLB flush is only necessary if modifying current map.
1950 * However, in pmap_enter(), the pmap almost always IS
1951 * the current pmap, so don't even bother to check.
1952 */
1953 TBIS(va);
1954 } else {
1955 /*
1956 * The PTE is invalid. Increment the valid entry count in
1957 * the C table manager to reflect the addition of a new entry.
1958 */
1959 c_tbl->ct_ecnt++;
1960
1961 /* XXX - temporarily make sure the PTE is cleared. */
1962 c_pte->attr.raw = 0;
1963
1964 /* It will also need to be inserted into the PV list. */
1965 insert = TRUE;
1966 }
1967
1968 /*
1969 * If page is changing from unwired to wired status, set an unused bit
1970 * within the PTE to indicate that it is wired. Also increment the
1971 * wired entry count in the C table manager.
1972 */
1973 if (wired) {
1974 c_pte->attr.raw |= MMU_SHORT_PTE_WIRED;
1975 c_tbl->ct_wcnt++;
1976 }
1977
1978 /*
1979 * Map the page, being careful to preserve modify/reference/wired
1980 * bits. At this point it is assumed that the PTE either has no bits
1981 * set, or if there are set bits, they are only modified, reference or
1982 * wired bits. If not, the following statement will cause erratic
1983 * behavior.
1984 */
1985 #ifdef PMAP_DEBUG
1986 if (c_pte->attr.raw & ~(MMU_SHORT_PTE_M |
1987 MMU_SHORT_PTE_USED | MMU_SHORT_PTE_WIRED)) {
1988 printf("pmap_enter: junk left in PTE at %p\n", c_pte);
1989 Debugger();
1990 }
1991 #endif
1992 c_pte->attr.raw |= ((u_long) pa | MMU_DT_PAGE);
1993
1994 /*
1995 * If the mapping should be read-only, set the write protect
1996 * bit in the PTE.
1997 */
1998 if (!(prot & VM_PROT_WRITE))
1999 c_pte->attr.raw |= MMU_SHORT_PTE_WP;
2000
2001 /*
2002 * If the mapping should be cache inhibited (indicated by the flag
2003 * bits found on the lower order of the physical address.)
2004 * mark the PTE as a cache inhibited page.
2005 */
2006 if (flags & PMAP_NC)
2007 c_pte->attr.raw |= MMU_SHORT_PTE_CI;
2008
2009 /*
2010 * If the physical address being mapped is managed by the PV
2011 * system then link the pte into the list of pages mapped to that
2012 * address.
2013 */
2014 if (insert && managed) {
2015 pv = pa2pv(pa);
2016 nidx = pteidx(c_pte);
2017
2018 s = splimp();
2019 pvebase[nidx].pve_next = pv->pv_idx;
2020 pv->pv_idx = nidx;
2021 splx(s);
2022 }
2023
2024 /* Move any allocated tables back into the active pool. */
2025
2026 switch (llevel) {
2027 case NEWA:
2028 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
2029 /* FALLTHROUGH */
2030 case NEWB:
2031 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
2032 /* FALLTHROUGH */
2033 case NEWC:
2034 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
2035 /* FALLTHROUGH */
2036 default:
2037 break;
2038 }
2039 }
2040
2041 /* pmap_enter_kernel INTERNAL
2042 **
2043 * Map the given virtual address to the given physical address within the
2044 * kernel address space. This function exists because the kernel map does
2045 * not do dynamic table allocation. It consists of a contiguous array of ptes
2046 * and can be edited directly without the need to walk through any tables.
2047 *
2048 * XXX: "Danger, Will Robinson!"
2049 * Note that the kernel should never take a fault on any page
2050 * between [ KERNBASE .. virtual_avail ] and this is checked in
2051 * trap.c for kernel-mode MMU faults. This means that mappings
2052 * created in that range must be implicily wired. -gwr
2053 */
2054 void
2055 pmap_enter_kernel(va, pa, prot)
2056 vm_offset_t va;
2057 vm_offset_t pa;
2058 vm_prot_t prot;
2059 {
2060 boolean_t was_valid, insert;
2061 u_short pte_idx, pv_idx;
2062 int s, flags;
2063 mmu_short_pte_t *pte;
2064 pv_t *pv;
2065 vm_offset_t old_pa;
2066
2067 flags = (pa & ~MMU_PAGE_MASK);
2068 pa &= MMU_PAGE_MASK;
2069
2070 /*
2071 * Calculate the index of the PTE being modified.
2072 */
2073 pte_idx = (u_long) _btop(va - KERNBASE);
2074
2075 /* This array is traditionally named "Sysmap" */
2076 pte = &kernCbase[pte_idx];
2077
2078 s = splimp();
2079 if (MMU_VALID_DT(*pte)) {
2080 was_valid = TRUE;
2081 /*
2082 * If the PTE is already mapped to an address and it differs
2083 * from the address requested, unlink it from the PV list.
2084 *
2085 * This only applies to mappings within virtual_avail
2086 * and VM_MAX_KERNEL_ADDRESS. All others are not requests
2087 * from the VM system and should not be part of the PV system.
2088 */
2089 if ((va >= virtual_avail) && (va < VM_MAX_KERNEL_ADDRESS)) {
2090 old_pa = MMU_PTE_PA(*pte);
2091 if (pa != old_pa) {
2092 if (is_managed(old_pa)) {
2093 /* XXX - Make this into a function call? */
2094 pv = pa2pv(old_pa);
2095 pv_idx = pv->pv_idx;
2096 if (pv_idx == pte_idx) {
2097 pv->pv_idx = pvebase[pte_idx].pve_next;
2098 } else {
2099 while (pvebase[pv_idx].pve_next != pte_idx)
2100 pv_idx = pvebase[pv_idx].pve_next;
2101 pvebase[pv_idx].pve_next =
2102 pvebase[pte_idx].pve_next;
2103 }
2104 /* Save modified/reference bits */
2105 pv->pv_flags |= (u_short) pte->attr.raw;
2106 }
2107 if (is_managed(pa))
2108 insert = TRUE;
2109 else
2110 insert = FALSE;
2111 /*
2112 * Clear out any old bits in the PTE.
2113 */
2114 pte->attr.raw = MMU_DT_INVALID;
2115 } else {
2116 /*
2117 * Old PA and new PA are the same. No need to relink
2118 * the mapping within the PV list.
2119 */
2120 insert = FALSE;
2121
2122 /*
2123 * Save any mod/ref bits on the PTE.
2124 */
2125 pte->attr.raw &= (MMU_SHORT_PTE_USED|MMU_SHORT_PTE_M);
2126 }
2127 } else {
2128 /*
2129 * If the VA lies below virtual_avail or beyond
2130 * VM_MAX_KERNEL_ADDRESS, it is not a request by the VM
2131 * system and hence does not need to be linked into the PV
2132 * system.
2133 */
2134 insert = FALSE;
2135 pte->attr.raw = MMU_DT_INVALID;
2136 }
2137 } else {
2138 pte->attr.raw = MMU_DT_INVALID;
2139 was_valid = FALSE;
2140 if ((va >= virtual_avail) && (va < VM_MAX_KERNEL_ADDRESS)) {
2141 if (is_managed(pa))
2142 insert = TRUE;
2143 else
2144 insert = FALSE;
2145 } else
2146 insert = FALSE;
2147 }
2148
2149 /*
2150 * Map the page. Being careful to preserve modified/referenced bits
2151 * on the PTE.
2152 */
2153 pte->attr.raw |= (pa | MMU_DT_PAGE);
2154
2155 if (!(prot & VM_PROT_WRITE)) /* If access should be read-only */
2156 pte->attr.raw |= MMU_SHORT_PTE_WP;
2157 if (flags & PMAP_NC)
2158 pte->attr.raw |= MMU_SHORT_PTE_CI;
2159 if (was_valid)
2160 TBIS(va);
2161
2162 /*
2163 * Insert the PTE into the PV system, if need be.
2164 */
2165 if (insert) {
2166 pv = pa2pv(pa);
2167 pvebase[pte_idx].pve_next = pv->pv_idx;
2168 pv->pv_idx = pte_idx;
2169 }
2170 splx(s);
2171
2172 }
2173
2174 /* pmap_protect INTERFACE
2175 **
2176 * Apply the given protection to the given virtual address range within
2177 * the given map.
2178 *
2179 * It is ok for the protection applied to be stronger than what is
2180 * specified. We use this to our advantage when the given map has no
2181 * mapping for the virtual address. By skipping a page when this
2182 * is discovered, we are effectively applying a protection of VM_PROT_NONE,
2183 * and therefore do not need to map the page just to apply a protection
2184 * code. Only pmap_enter() needs to create new mappings if they do not exist.
2185 *
2186 * XXX - This function could be speeded up by using pmap_stroll() for inital
2187 * setup, and then manual scrolling in the for() loop.
2188 */
2189 void
2190 pmap_protect(pmap, startva, endva, prot)
2191 pmap_t pmap;
2192 vm_offset_t startva, endva;
2193 vm_prot_t prot;
2194 {
2195 boolean_t iscurpmap;
2196 int a_idx, b_idx, c_idx;
2197 a_tmgr_t *a_tbl;
2198 b_tmgr_t *b_tbl;
2199 c_tmgr_t *c_tbl;
2200 mmu_short_pte_t *pte;
2201
2202 if (pmap == NULL)
2203 return;
2204 if (pmap == pmap_kernel()) {
2205 pmap_protect_kernel(startva, endva, prot);
2206 return;
2207 }
2208
2209 /*
2210 * In this particular pmap implementation, there are only three
2211 * types of memory protection: 'all' (read/write/execute),
2212 * 'read-only' (read/execute) and 'none' (no mapping.)
2213 * It is not possible for us to treat 'executable' as a separate
2214 * protection type. Therefore, protection requests that seek to
2215 * remove execute permission while retaining read or write, and those
2216 * that make little sense (write-only for example) are ignored.
2217 */
2218 switch (prot) {
2219 case VM_PROT_NONE:
2220 /*
2221 * A request to apply the protection code of
2222 * 'VM_PROT_NONE' is a synonym for pmap_remove().
2223 */
2224 pmap_remove(pmap, startva, endva);
2225 return;
2226 case VM_PROT_EXECUTE:
2227 case VM_PROT_READ:
2228 case VM_PROT_READ|VM_PROT_EXECUTE:
2229 /* continue */
2230 break;
2231 case VM_PROT_WRITE:
2232 case VM_PROT_WRITE|VM_PROT_READ:
2233 case VM_PROT_WRITE|VM_PROT_EXECUTE:
2234 case VM_PROT_ALL:
2235 /* None of these should happen in a sane system. */
2236 return;
2237 }
2238
2239 /*
2240 * If the pmap has no A table, it has no mappings and therefore
2241 * there is nothing to protect.
2242 */
2243 if ((a_tbl = pmap->pm_a_tmgr) == NULL)
2244 return;
2245
2246 a_idx = MMU_TIA(startva);
2247 b_idx = MMU_TIB(startva);
2248 c_idx = MMU_TIC(startva);
2249 b_tbl = (b_tmgr_t *) c_tbl = NULL;
2250
2251 iscurpmap = (pmap == current_pmap());
2252 while (startva < endva) {
2253 if (b_tbl || MMU_VALID_DT(a_tbl->at_dtbl[a_idx])) {
2254 if (b_tbl == NULL) {
2255 b_tbl = (b_tmgr_t *) a_tbl->at_dtbl[a_idx].addr.raw;
2256 b_tbl = mmu_ptov((vm_offset_t) b_tbl);
2257 b_tbl = mmuB2tmgr((mmu_short_dte_t *) b_tbl);
2258 }
2259 if (c_tbl || MMU_VALID_DT(b_tbl->bt_dtbl[b_idx])) {
2260 if (c_tbl == NULL) {
2261 c_tbl = (c_tmgr_t *) MMU_DTE_PA(b_tbl->bt_dtbl[b_idx]);
2262 c_tbl = mmu_ptov((vm_offset_t) c_tbl);
2263 c_tbl = mmuC2tmgr((mmu_short_pte_t *) c_tbl);
2264 }
2265 if (MMU_VALID_DT(c_tbl->ct_dtbl[c_idx])) {
2266 pte = &c_tbl->ct_dtbl[c_idx];
2267 /* make the mapping read-only */
2268 pte->attr.raw |= MMU_SHORT_PTE_WP;
2269 /*
2270 * If we just modified the current address space,
2271 * flush any translations for the modified page from
2272 * the translation cache and any data from it in the
2273 * data cache.
2274 */
2275 if (iscurpmap)
2276 TBIS(startva);
2277 }
2278 startva += NBPG;
2279
2280 if (++c_idx >= MMU_C_TBL_SIZE) { /* exceeded C table? */
2281 c_tbl = NULL;
2282 c_idx = 0;
2283 if (++b_idx >= MMU_B_TBL_SIZE) { /* exceeded B table? */
2284 b_tbl = NULL;
2285 b_idx = 0;
2286 }
2287 }
2288 } else { /* C table wasn't valid */
2289 c_tbl = NULL;
2290 c_idx = 0;
2291 startva += MMU_TIB_RANGE;
2292 if (++b_idx >= MMU_B_TBL_SIZE) { /* exceeded B table? */
2293 b_tbl = NULL;
2294 b_idx = 0;
2295 }
2296 } /* C table */
2297 } else { /* B table wasn't valid */
2298 b_tbl = NULL;
2299 b_idx = 0;
2300 startva += MMU_TIA_RANGE;
2301 a_idx++;
2302 } /* B table */
2303 }
2304 }
2305
2306 /* pmap_protect_kernel INTERNAL
2307 **
2308 * Apply the given protection code to a kernel address range.
2309 */
2310 void
2311 pmap_protect_kernel(startva, endva, prot)
2312 vm_offset_t startva, endva;
2313 vm_prot_t prot;
2314 {
2315 vm_offset_t va;
2316 mmu_short_pte_t *pte;
2317
2318 pte = &kernCbase[(unsigned long) _btop(startva - KERNBASE)];
2319 for (va = startva; va < endva; va += NBPG, pte++) {
2320 if (MMU_VALID_DT(*pte)) {
2321 switch (prot) {
2322 case VM_PROT_ALL:
2323 break;
2324 case VM_PROT_EXECUTE:
2325 case VM_PROT_READ:
2326 case VM_PROT_READ|VM_PROT_EXECUTE:
2327 pte->attr.raw |= MMU_SHORT_PTE_WP;
2328 break;
2329 case VM_PROT_NONE:
2330 /* this is an alias for 'pmap_remove_kernel' */
2331 pmap_remove_pte(pte);
2332 break;
2333 default:
2334 break;
2335 }
2336 /*
2337 * since this is the kernel, immediately flush any cached
2338 * descriptors for this address.
2339 */
2340 TBIS(va);
2341 }
2342 }
2343 }
2344
2345 /* pmap_change_wiring INTERFACE
2346 **
2347 * Changes the wiring of the specified page.
2348 *
2349 * This function is called from vm_fault.c to unwire
2350 * a mapping. It really should be called 'pmap_unwire'
2351 * because it is never asked to do anything but remove
2352 * wirings.
2353 */
2354 void
2355 pmap_change_wiring(pmap, va, wire)
2356 pmap_t pmap;
2357 vm_offset_t va;
2358 boolean_t wire;
2359 {
2360 int a_idx, b_idx, c_idx;
2361 a_tmgr_t *a_tbl;
2362 b_tmgr_t *b_tbl;
2363 c_tmgr_t *c_tbl;
2364 mmu_short_pte_t *pte;
2365
2366 /* Kernel mappings always remain wired. */
2367 if (pmap == pmap_kernel())
2368 return;
2369
2370 #ifdef PMAP_DEBUG
2371 if (wire == TRUE)
2372 panic("pmap_change_wiring: wire requested.");
2373 #endif
2374
2375 /*
2376 * Walk through the tables. If the walk terminates without
2377 * a valid PTE then the address wasn't wired in the first place.
2378 * Return immediately.
2379 */
2380 if (pmap_stroll(pmap, va, &a_tbl, &b_tbl, &c_tbl, &pte, &a_idx,
2381 &b_idx, &c_idx) == FALSE)
2382 return;
2383
2384
2385 /* Is the PTE wired? If not, return. */
2386 if (!(pte->attr.raw & MMU_SHORT_PTE_WIRED))
2387 return;
2388
2389 /* Remove the wiring bit. */
2390 pte->attr.raw &= ~(MMU_SHORT_PTE_WIRED);
2391
2392 /*
2393 * Decrement the wired entry count in the C table.
2394 * If it reaches zero the following things happen:
2395 * 1. The table no longer has any wired entries and is considered
2396 * unwired.
2397 * 2. It is placed on the available queue.
2398 * 3. The parent table's wired entry count is decremented.
2399 * 4. If it reaches zero, this process repeats at step 1 and
2400 * stops at after reaching the A table.
2401 */
2402 if (--c_tbl->ct_wcnt == 0) {
2403 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
2404 if (--b_tbl->bt_wcnt == 0) {
2405 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
2406 if (--a_tbl->at_wcnt == 0) {
2407 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
2408 }
2409 }
2410 }
2411 }
2412
2413 /* pmap_pageable INTERFACE
2414 **
2415 * Make the specified range of addresses within the given pmap,
2416 * 'pageable' or 'not-pageable'. A pageable page must not cause
2417 * any faults when referenced. A non-pageable page may.
2418 *
2419 * This routine is only advisory. The VM system will call pmap_enter()
2420 * to wire or unwire pages that are going to be made pageable before calling
2421 * this function. By the time this routine is called, everything that needs
2422 * to be done has already been done.
2423 */
2424 void
2425 pmap_pageable(pmap, start, end, pageable)
2426 pmap_t pmap;
2427 vm_offset_t start, end;
2428 boolean_t pageable;
2429 {
2430 /* not implemented. */
2431 }
2432
2433 /* pmap_copy INTERFACE
2434 **
2435 * Copy the mappings of a range of addresses in one pmap, into
2436 * the destination address of another.
2437 *
2438 * This routine is advisory. Should we one day decide that MMU tables
2439 * may be shared by more than one pmap, this function should be used to
2440 * link them together. Until that day however, we do nothing.
2441 */
2442 void
2443 pmap_copy(pmap_a, pmap_b, dst, len, src)
2444 pmap_t pmap_a, pmap_b;
2445 vm_offset_t dst;
2446 vm_size_t len;
2447 vm_offset_t src;
2448 {
2449 /* not implemented. */
2450 }
2451
2452 /* pmap_copy_page INTERFACE
2453 **
2454 * Copy the contents of one physical page into another.
2455 *
2456 * This function makes use of two virtual pages allocated in pmap_bootstrap()
2457 * to map the two specified physical pages into the kernel address space. It
2458 * then uses bcopy() to copy one into the other.
2459 *
2460 * Note: We could use the transparent translation registers to make the
2461 * mappings. If we do so, be sure to disable interrupts before using them.
2462 */
2463 void
2464 pmap_copy_page(src, dst)
2465 vm_offset_t src, dst;
2466 {
2467 int s;
2468 vm_offset_t oldsrc, olddst;
2469
2470 s = splimp();
2471 oldsrc = olddst = 0; /*XXXgcc*/
2472 if (tmp_vpage0_inuse++) {
2473 oldsrc = pmap_extract_kernel(tmp_vpages[0]);
2474 }
2475 if (tmp_vpage1_inuse++) {
2476 olddst = pmap_extract_kernel(tmp_vpages[1]);
2477 }
2478 splx(s);
2479
2480 /* Map pages as non-cacheable to avoid cache polution? */
2481 pmap_enter_kernel(tmp_vpages[0], src, VM_PROT_READ);
2482 pmap_enter_kernel(tmp_vpages[1], dst, VM_PROT_READ|VM_PROT_WRITE);
2483 copypage((char *) tmp_vpages[0], (char *) tmp_vpages[1]);
2484
2485 s = splimp();
2486 if (--tmp_vpage0_inuse) {
2487 pmap_enter_kernel(tmp_vpages[0], oldsrc, VM_PROT_READ);
2488 }
2489 if (--tmp_vpage1_inuse) {
2490 pmap_enter_kernel(tmp_vpages[1], olddst,
2491 VM_PROT_READ|VM_PROT_WRITE);
2492 }
2493 splx(s);
2494 }
2495
2496 /* pmap_zero_page INTERFACE
2497 **
2498 * Zero the contents of the specified physical page.
2499 *
2500 * Uses one of the virtual pages allocated in pmap_boostrap()
2501 * to map the specified page into the kernel address space. Then uses
2502 * bzero() to zero out the page.
2503 */
2504 void
2505 pmap_zero_page(pa)
2506 vm_offset_t pa;
2507 {
2508 int s;
2509 vm_offset_t oldpa;
2510
2511 s = splimp();
2512 oldpa = 0; /*XXXgcc*/
2513 if (tmp_vpage1_inuse++) {
2514 oldpa = pmap_extract_kernel(tmp_vpages[1]);
2515 }
2516 splx(s);
2517
2518 pmap_enter_kernel(tmp_vpages[1], pa, VM_PROT_READ|VM_PROT_WRITE);
2519 zeropage((char *) tmp_vpages[1]);
2520
2521 s = splimp();
2522 if (--tmp_vpage1_inuse) {
2523 pmap_enter_kernel(tmp_vpages[1], oldpa,
2524 VM_PROT_READ|VM_PROT_WRITE);
2525 }
2526 splx(s);
2527 }
2528
2529 /* pmap_collect INTERFACE
2530 **
2531 * Called from the VM system when we are about to swap out
2532 * the process using this pmap. This should give up any
2533 * resources held here, including all its MMU tables.
2534 */
2535 void
2536 pmap_collect(pmap)
2537 pmap_t pmap;
2538 {
2539 /* XXX - todo... */
2540 }
2541
2542 /* pmap_create INTERFACE
2543 **
2544 * Create and return a pmap structure.
2545 */
2546 pmap_t
2547 pmap_create(size)
2548 vm_size_t size;
2549 {
2550 pmap_t pmap;
2551
2552 if (size)
2553 return NULL;
2554
2555 pmap = (pmap_t) malloc(sizeof(struct pmap), M_VMPMAP, M_WAITOK);
2556 pmap_pinit(pmap);
2557
2558 return pmap;
2559 }
2560
2561 /* pmap_pinit INTERNAL
2562 **
2563 * Initialize a pmap structure.
2564 */
2565 void
2566 pmap_pinit(pmap)
2567 pmap_t pmap;
2568 {
2569 bzero(pmap, sizeof(struct pmap));
2570 pmap->pm_a_tmgr = NULL;
2571 pmap->pm_a_phys = kernAphys;
2572 }
2573
2574 /* pmap_release INTERFACE
2575 **
2576 * Release any resources held by the given pmap.
2577 *
2578 * This is the reverse analog to pmap_pinit. It does not
2579 * necessarily mean for the pmap structure to be deallocated,
2580 * as in pmap_destroy.
2581 */
2582 void
2583 pmap_release(pmap)
2584 pmap_t pmap;
2585 {
2586 /*
2587 * As long as the pmap contains no mappings,
2588 * which always should be the case whenever
2589 * this function is called, there really should
2590 * be nothing to do.
2591 */
2592 #ifdef PMAP_DEBUG
2593 if (pmap == NULL)
2594 return;
2595 if (pmap == pmap_kernel())
2596 panic("pmap_release: kernel pmap");
2597 #endif
2598 /*
2599 * XXX - If this pmap has an A table, give it back.
2600 * The pmap SHOULD be empty by now, and pmap_remove
2601 * should have already given back the A table...
2602 * However, I see: pmap->pm_a_tmgr->at_ecnt == 1
2603 * at this point, which means some mapping was not
2604 * removed when it should have been. -gwr
2605 */
2606 if (pmap->pm_a_tmgr != NULL) {
2607 /* First make sure we are not using it! */
2608 if (kernel_crp.rp_addr == pmap->pm_a_phys) {
2609 kernel_crp.rp_addr = kernAphys;
2610 loadcrp(&kernel_crp);
2611 }
2612 #ifdef PMAP_DEBUG /* XXX - todo! */
2613 /* XXX - Now complain... */
2614 printf("pmap_release: still have table\n");
2615 Debugger();
2616 #endif
2617 free_a_table(pmap->pm_a_tmgr, TRUE);
2618 pmap->pm_a_tmgr = NULL;
2619 pmap->pm_a_phys = kernAphys;
2620 }
2621 }
2622
2623 /* pmap_reference INTERFACE
2624 **
2625 * Increment the reference count of a pmap.
2626 */
2627 void
2628 pmap_reference(pmap)
2629 pmap_t pmap;
2630 {
2631 if (pmap == NULL)
2632 return;
2633
2634 /* pmap_lock(pmap); */
2635 pmap->pm_refcount++;
2636 /* pmap_unlock(pmap); */
2637 }
2638
2639 /* pmap_dereference INTERNAL
2640 **
2641 * Decrease the reference count on the given pmap
2642 * by one and return the current count.
2643 */
2644 int
2645 pmap_dereference(pmap)
2646 pmap_t pmap;
2647 {
2648 int rtn;
2649
2650 if (pmap == NULL)
2651 return 0;
2652
2653 /* pmap_lock(pmap); */
2654 rtn = --pmap->pm_refcount;
2655 /* pmap_unlock(pmap); */
2656
2657 return rtn;
2658 }
2659
2660 /* pmap_destroy INTERFACE
2661 **
2662 * Decrement a pmap's reference count and delete
2663 * the pmap if it becomes zero. Will be called
2664 * only after all mappings have been removed.
2665 */
2666 void
2667 pmap_destroy(pmap)
2668 pmap_t pmap;
2669 {
2670 if (pmap == NULL)
2671 return;
2672 if (pmap == &kernel_pmap)
2673 panic("pmap_destroy: kernel_pmap!");
2674 if (pmap_dereference(pmap) == 0) {
2675 pmap_release(pmap);
2676 free(pmap, M_VMPMAP);
2677 }
2678 }
2679
2680 /* pmap_is_referenced INTERFACE
2681 **
2682 * Determine if the given physical page has been
2683 * referenced (read from [or written to.])
2684 */
2685 boolean_t
2686 pmap_is_referenced(pa)
2687 vm_offset_t pa;
2688 {
2689 pv_t *pv;
2690 int idx, s;
2691
2692 if (!pv_initialized)
2693 return FALSE;
2694 /* XXX - this may be unecessary. */
2695 if (!is_managed(pa))
2696 return FALSE;
2697
2698 pv = pa2pv(pa);
2699 /*
2700 * Check the flags on the pv head. If they are set,
2701 * return immediately. Otherwise a search must be done.
2702 */
2703 if (pv->pv_flags & PV_FLAGS_USED)
2704 return TRUE;
2705 else {
2706 s = splimp();
2707 /*
2708 * Search through all pv elements pointing
2709 * to this page and query their reference bits
2710 */
2711 for (idx = pv->pv_idx; idx != PVE_EOL; idx =
2712 pvebase[idx].pve_next)
2713 if (MMU_PTE_USED(kernCbase[idx])) {
2714 splx(s);
2715 return TRUE;
2716 }
2717 splx(s);
2718 }
2719
2720 return FALSE;
2721 }
2722
2723 /* pmap_is_modified INTERFACE
2724 **
2725 * Determine if the given physical page has been
2726 * modified (written to.)
2727 */
2728 boolean_t
2729 pmap_is_modified(pa)
2730 vm_offset_t pa;
2731 {
2732 pv_t *pv;
2733 int idx, s;
2734
2735 if (!pv_initialized)
2736 return FALSE;
2737 /* XXX - this may be unecessary. */
2738 if (!is_managed(pa))
2739 return FALSE;
2740
2741 /* see comments in pmap_is_referenced() */
2742 pv = pa2pv(pa);
2743 if (pv->pv_flags & PV_FLAGS_MDFY) {
2744 return TRUE;
2745 } else {
2746 s = splimp();
2747 for (idx = pv->pv_idx; idx != PVE_EOL; idx =
2748 pvebase[idx].pve_next)
2749 if (MMU_PTE_MODIFIED(kernCbase[idx])) {
2750 splx(s);
2751 return TRUE;
2752 }
2753 splx(s);
2754 }
2755
2756 return FALSE;
2757 }
2758
2759 /* pmap_page_protect INTERFACE
2760 **
2761 * Applies the given protection to all mappings to the given
2762 * physical page.
2763 */
2764 void
2765 pmap_page_protect(pa, prot)
2766 vm_offset_t pa;
2767 vm_prot_t prot;
2768 {
2769 pv_t *pv;
2770 int idx, s;
2771 vm_offset_t va;
2772 struct mmu_short_pte_struct *pte;
2773 c_tmgr_t *c_tbl;
2774 pmap_t pmap, curpmap;
2775
2776 if (!is_managed(pa))
2777 return;
2778
2779 curpmap = current_pmap();
2780 pv = pa2pv(pa);
2781 s = splimp();
2782 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) {
2783 pte = &kernCbase[idx];
2784 switch (prot) {
2785 case VM_PROT_ALL:
2786 /* do nothing */
2787 break;
2788 case VM_PROT_EXECUTE:
2789 case VM_PROT_READ:
2790 case VM_PROT_READ|VM_PROT_EXECUTE:
2791 pte->attr.raw |= MMU_SHORT_PTE_WP;
2792
2793 /*
2794 * Determine the virtual address mapped by
2795 * the PTE and flush ATC entries if necessary.
2796 */
2797 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
2798 if (pmap == curpmap || pmap == pmap_kernel())
2799 TBIS(va);
2800 break;
2801 case VM_PROT_NONE:
2802 /* Save the mod/ref bits. */
2803 pv->pv_flags |= pte->attr.raw;
2804 /* Invalidate the PTE. */
2805 pte->attr.raw = MMU_DT_INVALID;
2806
2807 /*
2808 * Update table counts. And flush ATC entries
2809 * if necessary.
2810 */
2811 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
2812
2813 /*
2814 * If the PTE belongs to the kernel map,
2815 * be sure to flush the page it maps.
2816 */
2817 if (pmap == pmap_kernel()) {
2818 TBIS(va);
2819 } else {
2820 /*
2821 * The PTE belongs to a user map.
2822 * update the entry count in the C
2823 * table to which it belongs and flush
2824 * the ATC if the mapping belongs to
2825 * the current pmap.
2826 */
2827 c_tbl->ct_ecnt--;
2828 if (pmap == curpmap)
2829 TBIS(va);
2830 }
2831 break;
2832 default:
2833 break;
2834 }
2835 }
2836
2837 /*
2838 * If the protection code indicates that all mappings to the page
2839 * be removed, truncate the PV list to zero entries.
2840 */
2841 if (prot == VM_PROT_NONE)
2842 pv->pv_idx = PVE_EOL;
2843 splx(s);
2844 }
2845
2846 /* pmap_get_pteinfo INTERNAL
2847 **
2848 * Called internally to find the pmap and virtual address within that
2849 * map to which the pte at the given index maps. Also includes the PTE's C
2850 * table manager.
2851 *
2852 * Returns the pmap in the argument provided, and the virtual address
2853 * by return value.
2854 */
2855 vm_offset_t
2856 pmap_get_pteinfo(idx, pmap, tbl)
2857 u_int idx;
2858 pmap_t *pmap;
2859 c_tmgr_t **tbl;
2860 {
2861 a_tmgr_t *a_tbl;
2862 b_tmgr_t *b_tbl;
2863 c_tmgr_t *c_tbl;
2864 vm_offset_t va = 0;
2865
2866 /*
2867 * Determine if the PTE is a kernel PTE or a user PTE.
2868 */
2869 if (idx >= NUM_KERN_PTES) {
2870 /*
2871 * The PTE belongs to a user mapping.
2872 * Find the virtual address by decoding table indices.
2873 * Each successive decode will reveal the address from
2874 * least to most significant bit fashion.
2875 *
2876 * 31 0
2877 * +-------------------------------+
2878 * |AAAAAAABBBBBBCCCCCC............|
2879 * +-------------------------------+
2880 */
2881 /* XXX: c_tbl = mmuC2tmgr(pte); */
2882 /* XXX: Would like an inline for this to validate idx... */
2883 c_tbl = &Ctmgrbase[(idx - NUM_KERN_PTES) / MMU_C_TBL_SIZE];
2884 b_tbl = c_tbl->ct_parent;
2885 a_tbl = b_tbl->bt_parent;
2886 *pmap = a_tbl->at_parent;
2887 *tbl = c_tbl;
2888
2889 /* Start with the 'C' bits, then add B and A... */
2890 va |= ((idx % MMU_C_TBL_SIZE) << MMU_TIC_SHIFT);
2891 va |= (c_tbl->ct_pidx << MMU_TIB_SHIFT);
2892 va |= (b_tbl->bt_pidx << MMU_TIA_SHIFT);
2893 } else {
2894 /*
2895 * The PTE belongs to the kernel map.
2896 */
2897 *pmap = pmap_kernel();
2898
2899 va = _ptob(idx);
2900 va += KERNBASE;
2901 }
2902
2903 return va;
2904 }
2905
2906 /* pmap_clear_modify INTERFACE
2907 **
2908 * Clear the modification bit on the page at the specified
2909 * physical address.
2910 *
2911 */
2912 void
2913 pmap_clear_modify(pa)
2914 vm_offset_t pa;
2915 {
2916 if (!is_managed(pa))
2917 return;
2918 pmap_clear_pv(pa, PV_FLAGS_MDFY);
2919 }
2920
2921 /* pmap_clear_reference INTERFACE
2922 **
2923 * Clear the referenced bit on the page at the specified
2924 * physical address.
2925 */
2926 void
2927 pmap_clear_reference(pa)
2928 vm_offset_t pa;
2929 {
2930 if (!is_managed(pa))
2931 return;
2932 pmap_clear_pv(pa, PV_FLAGS_USED);
2933 }
2934
2935 /* pmap_clear_pv INTERNAL
2936 **
2937 * Clears the specified flag from the specified physical address.
2938 * (Used by pmap_clear_modify() and pmap_clear_reference().)
2939 *
2940 * Flag is one of:
2941 * PV_FLAGS_MDFY - Page modified bit.
2942 * PV_FLAGS_USED - Page used (referenced) bit.
2943 *
2944 * This routine must not only clear the flag on the pv list
2945 * head. It must also clear the bit on every pte in the pv
2946 * list associated with the address.
2947 */
2948 void
2949 pmap_clear_pv(pa, flag)
2950 vm_offset_t pa;
2951 int flag;
2952 {
2953 pv_t *pv;
2954 int idx, s;
2955 vm_offset_t va;
2956 pmap_t pmap;
2957 mmu_short_pte_t *pte;
2958 c_tmgr_t *c_tbl;
2959
2960 pv = pa2pv(pa);
2961
2962 s = splimp();
2963 pv->pv_flags &= ~(flag);
2964 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) {
2965 pte = &kernCbase[idx];
2966 pte->attr.raw &= ~(flag);
2967 /*
2968 * The MC68030 MMU will not set the modified or
2969 * referenced bits on any MMU tables for which it has
2970 * a cached descriptor with its modify bit set. To insure
2971 * that it will modify these bits on the PTE during the next
2972 * time it is written to or read from, we must flush it from
2973 * the ATC.
2974 *
2975 * Ordinarily it is only necessary to flush the descriptor
2976 * if it is used in the current address space. But since I
2977 * am not sure that there will always be a notion of
2978 * 'the current address space' when this function is called,
2979 * I will skip the test and always flush the address. It
2980 * does no harm.
2981 */
2982 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
2983 TBIS(va);
2984 }
2985 splx(s);
2986 }
2987
2988 /* pmap_extract INTERFACE
2989 **
2990 * Return the physical address mapped by the virtual address
2991 * in the specified pmap or 0 if it is not known.
2992 *
2993 * Note: this function should also apply an exclusive lock
2994 * on the pmap system during its duration.
2995 */
2996 vm_offset_t
2997 pmap_extract(pmap, va)
2998 pmap_t pmap;
2999 vm_offset_t va;
3000 {
3001 int a_idx, b_idx, pte_idx;
3002 a_tmgr_t *a_tbl;
3003 b_tmgr_t *b_tbl;
3004 c_tmgr_t *c_tbl;
3005 mmu_short_pte_t *c_pte;
3006
3007 if (pmap == pmap_kernel())
3008 return pmap_extract_kernel(va);
3009 if (pmap == NULL)
3010 return 0;
3011
3012 if (pmap_stroll(pmap, va, &a_tbl, &b_tbl, &c_tbl,
3013 &c_pte, &a_idx, &b_idx, &pte_idx) == FALSE)
3014 return 0;
3015
3016 if (!MMU_VALID_DT(*c_pte))
3017 return 0;
3018
3019 return (MMU_PTE_PA(*c_pte));
3020 }
3021
3022 /* pmap_extract_kernel INTERNAL
3023 **
3024 * Extract a translation from the kernel address space.
3025 */
3026 vm_offset_t
3027 pmap_extract_kernel(va)
3028 vm_offset_t va;
3029 {
3030 mmu_short_pte_t *pte;
3031
3032 pte = &kernCbase[(u_int) _btop(va - KERNBASE)];
3033 return MMU_PTE_PA(*pte);
3034 }
3035
3036 /* pmap_remove_kernel INTERNAL
3037 **
3038 * Remove the mapping of a range of virtual addresses from the kernel map.
3039 * The arguments are already page-aligned.
3040 */
3041 void
3042 pmap_remove_kernel(sva, eva)
3043 vm_offset_t sva;
3044 vm_offset_t eva;
3045 {
3046 int idx, eidx;
3047
3048 #ifdef PMAP_DEBUG
3049 if ((sva & PGOFSET) || (eva & PGOFSET))
3050 panic("pmap_remove_kernel: alignment");
3051 #endif
3052
3053 idx = _btop(sva - KERNBASE);
3054 eidx = _btop(eva - KERNBASE);
3055
3056 while (idx < eidx)
3057 pmap_remove_pte(&kernCbase[idx++]);
3058 /* Always flush the ATC when maniplating the kernel address space. */
3059 TBIAS();
3060 }
3061
3062 /* pmap_remove INTERFACE
3063 **
3064 * Remove the mapping of a range of virtual addresses from the given pmap.
3065 *
3066 * If the range contains any wired entries, this function will probably create
3067 * disaster.
3068 */
3069 void
3070 pmap_remove(pmap, start, end)
3071 pmap_t pmap;
3072 vm_offset_t start;
3073 vm_offset_t end;
3074 {
3075
3076 if (pmap == pmap_kernel()) {
3077 pmap_remove_kernel(start, end);
3078 return;
3079 }
3080
3081 /*
3082 * XXX - Temporary(?) statement to prevent panic caused
3083 * by vm_alloc_with_pager() handing us a software map (ie NULL)
3084 * to remove because it couldn't get backing store.
3085 * (I guess.)
3086 */
3087 if (pmap == NULL)
3088 return;
3089
3090 /*
3091 * If the pmap doesn't have an A table of its own, it has no mappings
3092 * that can be removed.
3093 */
3094 if (pmap->pm_a_tmgr == NULL)
3095 return;
3096
3097 /*
3098 * Remove the specified range from the pmap. If the function
3099 * returns true, the operation removed all the valid mappings
3100 * in the pmap and freed its A table. If this happened to the
3101 * currently loaded pmap, the MMU root pointer must be reloaded
3102 * with the default 'kernel' map.
3103 */
3104 if (pmap_remove_a(pmap->pm_a_tmgr, start, end)) {
3105 if (kernel_crp.rp_addr == pmap->pm_a_phys) {
3106 kernel_crp.rp_addr = kernAphys;
3107 loadcrp(&kernel_crp);
3108 /* will do TLB flush below */
3109 }
3110 pmap->pm_a_tmgr = NULL;
3111 pmap->pm_a_phys = kernAphys;
3112 }
3113
3114 /*
3115 * If we just modified the current address space,
3116 * make sure to flush the MMU cache.
3117 *
3118 * XXX - this could be an unecessarily large flush.
3119 * XXX - Could decide, based on the size of the VA range
3120 * to be removed, whether to flush "by pages" or "all".
3121 */
3122 if (pmap == current_pmap())
3123 TBIAU();
3124 }
3125
3126 /* pmap_remove_a INTERNAL
3127 **
3128 * This is function number one in a set of three that removes a range
3129 * of memory in the most efficient manner by removing the highest possible
3130 * tables from the memory space. This particular function attempts to remove
3131 * as many B tables as it can, delegating the remaining fragmented ranges to
3132 * pmap_remove_b().
3133 *
3134 * If the removal operation results in an empty A table, the function returns
3135 * TRUE.
3136 *
3137 * It's ugly but will do for now.
3138 */
3139 boolean_t
3140 pmap_remove_a(a_tbl, start, end)
3141 a_tmgr_t *a_tbl;
3142 vm_offset_t start;
3143 vm_offset_t end;
3144 {
3145 boolean_t empty;
3146 int idx;
3147 vm_offset_t nstart, nend;
3148 b_tmgr_t *b_tbl;
3149 mmu_long_dte_t *a_dte;
3150 mmu_short_dte_t *b_dte;
3151
3152 /*
3153 * The following code works with what I call a 'granularity
3154 * reduction algorithim'. A range of addresses will always have
3155 * the following properties, which are classified according to
3156 * how the range relates to the size of the current granularity
3157 * - an A table entry:
3158 *
3159 * 1 2 3 4
3160 * -+---+---+---+---+---+---+---+-
3161 * -+---+---+---+---+---+---+---+-
3162 *
3163 * A range will always start on a granularity boundary, illustrated
3164 * by '+' signs in the table above, or it will start at some point
3165 * inbetween a granularity boundary, as illustrated by point 1.
3166 * The first step in removing a range of addresses is to remove the
3167 * range between 1 and 2, the nearest granularity boundary. This
3168 * job is handled by the section of code governed by the
3169 * 'if (start < nstart)' statement.
3170 *
3171 * A range will always encompass zero or more intergral granules,
3172 * illustrated by points 2 and 3. Integral granules are easy to
3173 * remove. The removal of these granules is the second step, and
3174 * is handled by the code block 'if (nstart < nend)'.
3175 *
3176 * Lastly, a range will always end on a granularity boundary,
3177 * ill. by point 3, or it will fall just beyond one, ill. by point
3178 * 4. The last step involves removing this range and is handled by
3179 * the code block 'if (nend < end)'.
3180 */
3181 nstart = MMU_ROUND_UP_A(start);
3182 nend = MMU_ROUND_A(end);
3183
3184 if (start < nstart) {
3185 /*
3186 * This block is executed if the range starts between
3187 * a granularity boundary.
3188 *
3189 * First find the DTE which is responsible for mapping
3190 * the start of the range.
3191 */
3192 idx = MMU_TIA(start);
3193 a_dte = &a_tbl->at_dtbl[idx];
3194
3195 /*
3196 * If the DTE is valid then delegate the removal of the sub
3197 * range to pmap_remove_b(), which can remove addresses at
3198 * a finer granularity.
3199 */
3200 if (MMU_VALID_DT(*a_dte)) {
3201 b_dte = mmu_ptov(a_dte->addr.raw);
3202 b_tbl = mmuB2tmgr(b_dte);
3203
3204 /*
3205 * The sub range to be removed starts at the start
3206 * of the full range we were asked to remove, and ends
3207 * at the greater of:
3208 * 1. The end of the full range, -or-
3209 * 2. The end of the full range, rounded down to the
3210 * nearest granularity boundary.
3211 */
3212 if (end < nstart)
3213 empty = pmap_remove_b(b_tbl, start, end);
3214 else
3215 empty = pmap_remove_b(b_tbl, start, nstart);
3216
3217 /*
3218 * If the removal resulted in an empty B table,
3219 * invalidate the DTE that points to it and decrement
3220 * the valid entry count of the A table.
3221 */
3222 if (empty) {
3223 a_dte->attr.raw = MMU_DT_INVALID;
3224 a_tbl->at_ecnt--;
3225 }
3226 }
3227 /*
3228 * If the DTE is invalid, the address range is already non-
3229 * existant and can simply be skipped.
3230 */
3231 }
3232 if (nstart < nend) {
3233 /*
3234 * This block is executed if the range spans a whole number
3235 * multiple of granules (A table entries.)
3236 *
3237 * First find the DTE which is responsible for mapping
3238 * the start of the first granule involved.
3239 */
3240 idx = MMU_TIA(nstart);
3241 a_dte = &a_tbl->at_dtbl[idx];
3242
3243 /*
3244 * Remove entire sub-granules (B tables) one at a time,
3245 * until reaching the end of the range.
3246 */
3247 for (; nstart < nend; a_dte++, nstart += MMU_TIA_RANGE)
3248 if (MMU_VALID_DT(*a_dte)) {
3249 /*
3250 * Find the B table manager for the
3251 * entry and free it.
3252 */
3253 b_dte = mmu_ptov(a_dte->addr.raw);
3254 b_tbl = mmuB2tmgr(b_dte);
3255 free_b_table(b_tbl, TRUE);
3256
3257 /*
3258 * Invalidate the DTE that points to the
3259 * B table and decrement the valid entry
3260 * count of the A table.
3261 */
3262 a_dte->attr.raw = MMU_DT_INVALID;
3263 a_tbl->at_ecnt--;
3264 }
3265 }
3266 if (nend < end) {
3267 /*
3268 * This block is executed if the range ends beyond a
3269 * granularity boundary.
3270 *
3271 * First find the DTE which is responsible for mapping
3272 * the start of the nearest (rounded down) granularity
3273 * boundary.
3274 */
3275 idx = MMU_TIA(nend);
3276 a_dte = &a_tbl->at_dtbl[idx];
3277
3278 /*
3279 * If the DTE is valid then delegate the removal of the sub
3280 * range to pmap_remove_b(), which can remove addresses at
3281 * a finer granularity.
3282 */
3283 if (MMU_VALID_DT(*a_dte)) {
3284 /*
3285 * Find the B table manager for the entry
3286 * and hand it to pmap_remove_b() along with
3287 * the sub range.
3288 */
3289 b_dte = mmu_ptov(a_dte->addr.raw);
3290 b_tbl = mmuB2tmgr(b_dte);
3291
3292 empty = pmap_remove_b(b_tbl, nend, end);
3293
3294 /*
3295 * If the removal resulted in an empty B table,
3296 * invalidate the DTE that points to it and decrement
3297 * the valid entry count of the A table.
3298 */
3299 if (empty) {
3300 a_dte->attr.raw = MMU_DT_INVALID;
3301 a_tbl->at_ecnt--;
3302 }
3303 }
3304 }
3305
3306 /*
3307 * If there are no more entries in the A table, release it
3308 * back to the available pool and return TRUE.
3309 */
3310 if (a_tbl->at_ecnt == 0) {
3311 a_tbl->at_parent = NULL;
3312 TAILQ_REMOVE(&a_pool, a_tbl, at_link);
3313 TAILQ_INSERT_HEAD(&a_pool, a_tbl, at_link);
3314 empty = TRUE;
3315 } else {
3316 empty = FALSE;
3317 }
3318
3319 return empty;
3320 }
3321
3322 /* pmap_remove_b INTERNAL
3323 **
3324 * Remove a range of addresses from an address space, trying to remove entire
3325 * C tables if possible.
3326 *
3327 * If the operation results in an empty B table, the function returns TRUE.
3328 */
3329 boolean_t
3330 pmap_remove_b(b_tbl, start, end)
3331 b_tmgr_t *b_tbl;
3332 vm_offset_t start;
3333 vm_offset_t end;
3334 {
3335 boolean_t empty;
3336 int idx;
3337 vm_offset_t nstart, nend, rstart;
3338 c_tmgr_t *c_tbl;
3339 mmu_short_dte_t *b_dte;
3340 mmu_short_pte_t *c_dte;
3341
3342
3343 nstart = MMU_ROUND_UP_B(start);
3344 nend = MMU_ROUND_B(end);
3345
3346 if (start < nstart) {
3347 idx = MMU_TIB(start);
3348 b_dte = &b_tbl->bt_dtbl[idx];
3349 if (MMU_VALID_DT(*b_dte)) {
3350 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte));
3351 c_tbl = mmuC2tmgr(c_dte);
3352 if (end < nstart)
3353 empty = pmap_remove_c(c_tbl, start, end);
3354 else
3355 empty = pmap_remove_c(c_tbl, start, nstart);
3356 if (empty) {
3357 b_dte->attr.raw = MMU_DT_INVALID;
3358 b_tbl->bt_ecnt--;
3359 }
3360 }
3361 }
3362 if (nstart < nend) {
3363 idx = MMU_TIB(nstart);
3364 b_dte = &b_tbl->bt_dtbl[idx];
3365 rstart = nstart;
3366 while (rstart < nend) {
3367 if (MMU_VALID_DT(*b_dte)) {
3368 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte));
3369 c_tbl = mmuC2tmgr(c_dte);
3370 free_c_table(c_tbl, TRUE);
3371 b_dte->attr.raw = MMU_DT_INVALID;
3372 b_tbl->bt_ecnt--;
3373 }
3374 b_dte++;
3375 rstart += MMU_TIB_RANGE;
3376 }
3377 }
3378 if (nend < end) {
3379 idx = MMU_TIB(nend);
3380 b_dte = &b_tbl->bt_dtbl[idx];
3381 if (MMU_VALID_DT(*b_dte)) {
3382 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte));
3383 c_tbl = mmuC2tmgr(c_dte);
3384 empty = pmap_remove_c(c_tbl, nend, end);
3385 if (empty) {
3386 b_dte->attr.raw = MMU_DT_INVALID;
3387 b_tbl->bt_ecnt--;
3388 }
3389 }
3390 }
3391
3392 if (b_tbl->bt_ecnt == 0) {
3393 b_tbl->bt_parent = NULL;
3394 TAILQ_REMOVE(&b_pool, b_tbl, bt_link);
3395 TAILQ_INSERT_HEAD(&b_pool, b_tbl, bt_link);
3396 empty = TRUE;
3397 } else {
3398 empty = FALSE;
3399 }
3400
3401 return empty;
3402 }
3403
3404 /* pmap_remove_c INTERNAL
3405 **
3406 * Remove a range of addresses from the given C table.
3407 */
3408 boolean_t
3409 pmap_remove_c(c_tbl, start, end)
3410 c_tmgr_t *c_tbl;
3411 vm_offset_t start;
3412 vm_offset_t end;
3413 {
3414 boolean_t empty;
3415 int idx;
3416 mmu_short_pte_t *c_pte;
3417
3418 idx = MMU_TIC(start);
3419 c_pte = &c_tbl->ct_dtbl[idx];
3420 for (;start < end; start += MMU_PAGE_SIZE, c_pte++) {
3421 if (MMU_VALID_DT(*c_pte)) {
3422 pmap_remove_pte(c_pte);
3423 c_tbl->ct_ecnt--;
3424 }
3425 }
3426
3427 if (c_tbl->ct_ecnt == 0) {
3428 c_tbl->ct_parent = NULL;
3429 TAILQ_REMOVE(&c_pool, c_tbl, ct_link);
3430 TAILQ_INSERT_HEAD(&c_pool, c_tbl, ct_link);
3431 empty = TRUE;
3432 } else {
3433 empty = FALSE;
3434 }
3435
3436 return empty;
3437 }
3438
3439 /* is_managed INTERNAL
3440 **
3441 * Determine if the given physical address is managed by the PV system.
3442 * Note that this logic assumes that no one will ask for the status of
3443 * addresses which lie in-between the memory banks on the 3/80. If they
3444 * do so, it will falsely report that it is managed.
3445 *
3446 * Note: A "managed" address is one that was reported to the VM system as
3447 * a "usable page" during system startup. As such, the VM system expects the
3448 * pmap module to keep an accurate track of the useage of those pages.
3449 * Any page not given to the VM system at startup does not exist (as far as
3450 * the VM system is concerned) and is therefore "unmanaged." Examples are
3451 * those pages which belong to the ROM monitor and the memory allocated before
3452 * the VM system was started.
3453 */
3454 boolean_t
3455 is_managed(pa)
3456 vm_offset_t pa;
3457 {
3458 if (pa >= avail_start && pa < avail_end)
3459 return TRUE;
3460 else
3461 return FALSE;
3462 }
3463
3464 /* pmap_bootstrap_alloc INTERNAL
3465 **
3466 * Used internally for memory allocation at startup when malloc is not
3467 * available. This code will fail once it crosses the first memory
3468 * bank boundary on the 3/80. Hopefully by then however, the VM system
3469 * will be in charge of allocation.
3470 */
3471 void *
3472 pmap_bootstrap_alloc(size)
3473 int size;
3474 {
3475 void *rtn;
3476
3477 #ifdef PMAP_DEBUG
3478 if (bootstrap_alloc_enabled == FALSE) {
3479 mon_printf("pmap_bootstrap_alloc: disabled\n");
3480 sunmon_abort();
3481 }
3482 #endif
3483
3484 rtn = (void *) virtual_avail;
3485 virtual_avail += size;
3486
3487 #ifdef PMAP_DEBUG
3488 if (virtual_avail > virtual_contig_end) {
3489 mon_printf("pmap_bootstrap_alloc: out of mem\n");
3490 sunmon_abort();
3491 }
3492 #endif
3493
3494 return rtn;
3495 }
3496
3497 /* pmap_bootstap_aalign INTERNAL
3498 **
3499 * Used to insure that the next call to pmap_bootstrap_alloc() will
3500 * return a chunk of memory aligned to the specified size.
3501 *
3502 * Note: This function will only support alignment sizes that are powers
3503 * of two.
3504 */
3505 void
3506 pmap_bootstrap_aalign(size)
3507 int size;
3508 {
3509 int off;
3510
3511 off = virtual_avail & (size - 1);
3512 if (off) {
3513 (void) pmap_bootstrap_alloc(size - off);
3514 }
3515 }
3516
3517 /* pmap_pa_exists
3518 **
3519 * Used by the /dev/mem driver to see if a given PA is memory
3520 * that can be mapped. (The PA is not in a hole.)
3521 */
3522 int
3523 pmap_pa_exists(pa)
3524 vm_offset_t pa;
3525 {
3526 register int i;
3527
3528 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
3529 if ((pa >= avail_mem[i].pmem_start) &&
3530 (pa < avail_mem[i].pmem_end))
3531 return (1);
3532 if (avail_mem[i].pmem_next == NULL)
3533 break;
3534 }
3535 return (0);
3536 }
3537
3538 /* pmap_activate INTERFACE
3539 **
3540 * This is called by locore.s:cpu_switch when we are switching to a
3541 * new process. This should load the MMU context for the new proc.
3542 * XXX - Later, this should be done directly in locore.s
3543 */
3544 void
3545 pmap_activate(pmap)
3546 pmap_t pmap;
3547 {
3548 u_long rootpa;
3549
3550 /* Only do reload/flush if we have to. */
3551 rootpa = pmap->pm_a_phys;
3552 if (kernel_crp.rp_addr != rootpa) {
3553 DPRINT(("pmap_activate(%p)\n", pmap));
3554 kernel_crp.rp_addr = rootpa;
3555 loadcrp(&kernel_crp);
3556 TBIAU();
3557 }
3558 }
3559
3560
3561 /* pmap_update
3562 **
3563 * Apply any delayed changes scheduled for all pmaps immediately.
3564 *
3565 * No delayed operations are currently done in this pmap.
3566 */
3567 void
3568 pmap_update()
3569 {
3570 /* not implemented. */
3571 }
3572
3573 /*
3574 * Fill in the cpu_kcore header for dumpsys()
3575 * (See machdep.c)
3576 */
3577 void
3578 pmap_set_kcore_hdr(chdr_p)
3579 cpu_kcore_hdr_t *chdr_p;
3580 {
3581 struct sun3x_kcore_hdr *sh = &chdr_p->un._sun3x;
3582 u_long spa, len;
3583 int i;
3584 extern char machine[];
3585
3586 /*
3587 * Fill in dispatch information.
3588 */
3589 strcpy(chdr_p->name, machine);
3590 chdr_p->page_size = NBPG;
3591 chdr_p->kernbase = KERNBASE;
3592
3593 sh->contig_end = virtual_contig_end;
3594 sh->kernCbase = (u_long) kernCbase;
3595 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
3596 spa = avail_mem[i].pmem_start;
3597 spa = _trunc_page(spa);
3598 len = avail_mem[i].pmem_end - spa;
3599 len = _round_page(len);
3600 sh->ram_segs[i].start = spa;
3601 sh->ram_segs[i].size = len;
3602 }
3603 }
3604
3605
3606 /* pmap_virtual_space INTERFACE
3607 **
3608 * Return the current available range of virtual addresses in the
3609 * arguuments provided. Only really called once.
3610 */
3611 void
3612 pmap_virtual_space(vstart, vend)
3613 vm_offset_t *vstart, *vend;
3614 {
3615 *vstart = virtual_avail;
3616 *vend = virtual_end;
3617 }
3618
3619 /* pmap_free_pages INTERFACE
3620 **
3621 * Return the number of physical pages still available.
3622 *
3623 * This is probably going to be a mess, but it's only called
3624 * once and it's the only function left that I have to implement!
3625 */
3626 u_int
3627 pmap_free_pages()
3628 {
3629 int i;
3630 u_int left;
3631 vm_offset_t avail;
3632
3633 avail = avail_next;
3634 left = 0;
3635 i = 0;
3636 while (avail >= avail_mem[i].pmem_end) {
3637 if (avail_mem[i].pmem_next == NULL)
3638 return 0;
3639 i++;
3640 }
3641 while (i < SUN3X_NPHYS_RAM_SEGS) {
3642 if (avail < avail_mem[i].pmem_start) {
3643 /* Avail is inside a hole, march it
3644 * up to the next bank.
3645 */
3646 avail = avail_mem[i].pmem_start;
3647 }
3648 left += _btop(avail_mem[i].pmem_end - avail);
3649 if (avail_mem[i].pmem_next == NULL)
3650 break;
3651 i++;
3652 }
3653
3654 return left;
3655 }
3656
3657 /* pmap_page_index INTERFACE
3658 **
3659 * Return the index of the given physical page in a list of useable
3660 * physical pages in the system. Holes in physical memory may be counted
3661 * if so desired. As long as pmap_free_pages() and pmap_page_index()
3662 * agree as to whether holes in memory do or do not count as valid pages,
3663 * it really doesn't matter. However, if you like to save a little
3664 * memory, don't count holes as valid pages. This is even more true when
3665 * the holes are large.
3666 *
3667 * We will not count holes as valid pages. We can generate page indices
3668 * that conform to this by using the memory bank structures initialized
3669 * in pmap_alloc_pv().
3670 */
3671 int
3672 pmap_page_index(pa)
3673 vm_offset_t pa;
3674 {
3675 struct pmap_physmem_struct *bank = avail_mem;
3676
3677 /* Search for the memory bank with this page. */
3678 /* XXX - What if it is not physical memory? */
3679 while (pa > bank->pmem_end)
3680 bank = bank->pmem_next;
3681 pa -= bank->pmem_start;
3682
3683 return (bank->pmem_pvbase + _btop(pa));
3684 }
3685
3686 /* pmap_next_page INTERFACE
3687 **
3688 * Place the physical address of the next available page in the
3689 * argument given. Returns FALSE if there are no more pages left.
3690 *
3691 * This function must jump over any holes in physical memory.
3692 * Once this function is used, any use of pmap_bootstrap_alloc()
3693 * is a sin. Sinners will be punished with erratic behavior.
3694 */
3695 boolean_t
3696 pmap_next_page(pa)
3697 vm_offset_t *pa;
3698 {
3699 static struct pmap_physmem_struct *curbank = avail_mem;
3700
3701 /* XXX - temporary ROM saving hack. */
3702 if (avail_next >= avail_end)
3703 return FALSE;
3704
3705 if (avail_next >= curbank->pmem_end)
3706 if (curbank->pmem_next == NULL)
3707 return FALSE;
3708 else {
3709 curbank = curbank->pmem_next;
3710 avail_next = curbank->pmem_start;
3711 }
3712
3713 *pa = avail_next;
3714 avail_next += NBPG;
3715 return TRUE;
3716 }
3717
3718 /* pmap_count INTERFACE
3719 **
3720 * Return the number of resident (valid) pages in the given pmap.
3721 *
3722 * Note: If this function is handed the kernel map, it will report
3723 * that it has no mappings. Hopefully the VM system won't ask for kernel
3724 * map statistics.
3725 */
3726 segsz_t
3727 pmap_count(pmap, type)
3728 pmap_t pmap;
3729 int type;
3730 {
3731 u_int count;
3732 int a_idx, b_idx;
3733 a_tmgr_t *a_tbl;
3734 b_tmgr_t *b_tbl;
3735 c_tmgr_t *c_tbl;
3736
3737 /*
3738 * If the pmap does not have its own A table manager, it has no
3739 * valid entires.
3740 */
3741 if (pmap->pm_a_tmgr == NULL)
3742 return 0;
3743
3744 a_tbl = pmap->pm_a_tmgr;
3745
3746 count = 0;
3747 for (a_idx = 0; a_idx < MMU_TIA(KERNBASE); a_idx++) {
3748 if (MMU_VALID_DT(a_tbl->at_dtbl[a_idx])) {
3749 b_tbl = mmuB2tmgr(mmu_ptov(a_tbl->at_dtbl[a_idx].addr.raw));
3750 for (b_idx = 0; b_idx < MMU_B_TBL_SIZE; b_idx++) {
3751 if (MMU_VALID_DT(b_tbl->bt_dtbl[b_idx])) {
3752 c_tbl = mmuC2tmgr(
3753 mmu_ptov(MMU_DTE_PA(b_tbl->bt_dtbl[b_idx])));
3754 if (type == 0)
3755 /*
3756 * A resident entry count has been requested.
3757 */
3758 count += c_tbl->ct_ecnt;
3759 else
3760 /*
3761 * A wired entry count has been requested.
3762 */
3763 count += c_tbl->ct_wcnt;
3764 }
3765 }
3766 }
3767 }
3768
3769 return count;
3770 }
3771
3772 /************************ SUN3 COMPATIBILITY ROUTINES ********************
3773 * The following routines are only used by DDB for tricky kernel text *
3774 * text operations in db_memrw.c. They are provided for sun3 *
3775 * compatibility. *
3776 *************************************************************************/
3777 /* get_pte INTERNAL
3778 **
3779 * Return the page descriptor the describes the kernel mapping
3780 * of the given virtual address.
3781 */
3782 extern u_long ptest_addr __P((u_long)); /* XXX: locore.s */
3783 u_long
3784 get_pte(va)
3785 vm_offset_t va;
3786 {
3787 u_long pte_pa;
3788 mmu_short_pte_t *pte;
3789
3790 /* Get the physical address of the PTE */
3791 pte_pa = ptest_addr(va & ~PGOFSET);
3792
3793 /* Convert to a virtual address... */
3794 pte = (mmu_short_pte_t *) (KERNBASE + pte_pa);
3795
3796 /* Make sure it is in our level-C tables... */
3797 if ((pte < kernCbase) ||
3798 (pte >= &mmuCbase[NUM_USER_PTES]))
3799 return 0;
3800
3801 /* ... and just return its contents. */
3802 return (pte->attr.raw);
3803 }
3804
3805
3806 /* set_pte INTERNAL
3807 **
3808 * Set the page descriptor that describes the kernel mapping
3809 * of the given virtual address.
3810 */
3811 void
3812 set_pte(va, pte)
3813 vm_offset_t va;
3814 vm_offset_t pte;
3815 {
3816 u_long idx;
3817
3818 if (va < KERNBASE)
3819 return;
3820
3821 idx = (unsigned long) _btop(va - KERNBASE);
3822 kernCbase[idx].attr.raw = pte;
3823 }
3824
3825 #ifdef PMAP_DEBUG
3826 /************************** DEBUGGING ROUTINES **************************
3827 * The following routines are meant to be an aid to debugging the pmap *
3828 * system. They are callable from the DDB command line and should be *
3829 * prepared to be handed unstable or incomplete states of the system. *
3830 ************************************************************************/
3831
3832 /* pv_list
3833 **
3834 * List all pages found on the pv list for the given physical page.
3835 * To avoid endless loops, the listing will stop at the end of the list
3836 * or after 'n' entries - whichever comes first.
3837 */
3838 void
3839 pv_list(pa, n)
3840 vm_offset_t pa;
3841 int n;
3842 {
3843 int idx;
3844 vm_offset_t va;
3845 pv_t *pv;
3846 c_tmgr_t *c_tbl;
3847 pmap_t pmap;
3848
3849 pv = pa2pv(pa);
3850 idx = pv->pv_idx;
3851
3852 for (;idx != PVE_EOL && n > 0; idx=pvebase[idx].pve_next, n--) {
3853 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
3854 printf("idx %d, pmap 0x%x, va 0x%x, c_tbl %x\n",
3855 idx, (u_int) pmap, (u_int) va, (u_int) c_tbl);
3856 }
3857 }
3858 #endif /* PMAP_DEBUG */
3859
3860 #ifdef NOT_YET
3861 /* and maybe not ever */
3862 /************************** LOW-LEVEL ROUTINES **************************
3863 * These routines will eventualy be re-written into assembly and placed *
3864 * in locore.s. They are here now as stubs so that the pmap module can *
3865 * be linked as a standalone user program for testing. *
3866 ************************************************************************/
3867 /* flush_atc_crp INTERNAL
3868 **
3869 * Flush all page descriptors derived from the given CPU Root Pointer
3870 * (CRP), or 'A' table as it is known here, from the 68851's automatic
3871 * cache.
3872 */
3873 void
3874 flush_atc_crp(a_tbl)
3875 {
3876 mmu_long_rp_t rp;
3877
3878 /* Create a temporary root table pointer that points to the
3879 * given A table.
3880 */
3881 rp.attr.raw = ~MMU_LONG_RP_LU;
3882 rp.addr.raw = (unsigned int) a_tbl;
3883
3884 mmu_pflushr(&rp);
3885 /* mmu_pflushr:
3886 * movel sp(4)@,a0
3887 * pflushr a0@
3888 * rts
3889 */
3890 }
3891 #endif /* NOT_YET */
3892