pmap.c revision 1.91 1 /* $NetBSD: pmap.c,v 1.91 2006/09/16 03:35:50 tsutsui 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 process 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/cdefs.h>
115 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.91 2006/09/16 03:35:50 tsutsui Exp $");
116
117 #include "opt_ddb.h"
118 #include "opt_pmap_debug.h"
119
120 #include <sys/param.h>
121 #include <sys/systm.h>
122 #include <sys/proc.h>
123 #include <sys/malloc.h>
124 #include <sys/pool.h>
125 #include <sys/user.h>
126 #include <sys/queue.h>
127 #include <sys/kcore.h>
128
129 #include <uvm/uvm.h>
130
131 #include <machine/cpu.h>
132 #include <machine/kcore.h>
133 #include <machine/mon.h>
134 #include <machine/pmap.h>
135 #include <machine/pte.h>
136 #include <machine/vmparam.h>
137 #include <m68k/cacheops.h>
138
139 #include <sun3/sun3/cache.h>
140 #include <sun3/sun3/machdep.h>
141
142 #include "pmap_pvt.h"
143
144 /* XXX - What headers declare these? */
145 extern struct pcb *curpcb;
146 extern int physmem;
147
148 /* Defined in locore.s */
149 extern char kernel_text[];
150
151 /* Defined by the linker */
152 extern char etext[], edata[], end[];
153 extern char *esym; /* DDB */
154
155 /*************************** DEBUGGING DEFINITIONS ***********************
156 * Macros, preprocessor defines and variables used in debugging can make *
157 * code hard to read. Anything used exclusively for debugging purposes *
158 * is defined here to avoid having such mess scattered around the file. *
159 *************************************************************************/
160 #ifdef PMAP_DEBUG
161 /*
162 * To aid the debugging process, macros should be expanded into smaller steps
163 * that accomplish the same goal, yet provide convenient places for placing
164 * breakpoints. When this code is compiled with PMAP_DEBUG mode defined, the
165 * 'INLINE' keyword is defined to an empty string. This way, any function
166 * defined to be a 'static INLINE' will become 'outlined' and compiled as
167 * a separate function, which is much easier to debug.
168 */
169 #define INLINE /* nothing */
170
171 /*
172 * It is sometimes convenient to watch the activity of a particular table
173 * in the system. The following variables are used for that purpose.
174 */
175 a_tmgr_t *pmap_watch_atbl = 0;
176 b_tmgr_t *pmap_watch_btbl = 0;
177 c_tmgr_t *pmap_watch_ctbl = 0;
178
179 int pmap_debug = 0;
180 #define DPRINT(args) if (pmap_debug) printf args
181
182 #else /********** Stuff below is defined if NOT debugging **************/
183
184 #define INLINE inline
185 #define DPRINT(args) /* nada */
186
187 #endif /* PMAP_DEBUG */
188 /*********************** END OF DEBUGGING DEFINITIONS ********************/
189
190 /*** Management Structure - Memory Layout
191 * For every MMU table in the sun3x pmap system there must be a way to
192 * manage it; we must know which process is using it, what other tables
193 * depend on it, and whether or not it contains any locked pages. This
194 * is solved by the creation of 'table management' or 'tmgr'
195 * structures. One for each MMU table in the system.
196 *
197 * MAP OF MEMORY USED BY THE PMAP SYSTEM
198 *
199 * towards lower memory
200 * kernAbase -> +-------------------------------------------------------+
201 * | Kernel MMU A level table |
202 * kernBbase -> +-------------------------------------------------------+
203 * | Kernel MMU B level tables |
204 * kernCbase -> +-------------------------------------------------------+
205 * | |
206 * | Kernel MMU C level tables |
207 * | |
208 * mmuCbase -> +-------------------------------------------------------+
209 * | User MMU C level tables |
210 * mmuAbase -> +-------------------------------------------------------+
211 * | |
212 * | User MMU A level tables |
213 * | |
214 * mmuBbase -> +-------------------------------------------------------+
215 * | User MMU B level tables |
216 * tmgrAbase -> +-------------------------------------------------------+
217 * | TMGR A level table structures |
218 * tmgrBbase -> +-------------------------------------------------------+
219 * | TMGR B level table structures |
220 * tmgrCbase -> +-------------------------------------------------------+
221 * | TMGR C level table structures |
222 * pvbase -> +-------------------------------------------------------+
223 * | Physical to Virtual mapping table (list heads) |
224 * pvebase -> +-------------------------------------------------------+
225 * | Physical to Virtual mapping table (list elements) |
226 * | |
227 * +-------------------------------------------------------+
228 * towards higher memory
229 *
230 * For every A table in the MMU A area, there will be a corresponding
231 * a_tmgr structure in the TMGR A area. The same will be true for
232 * the B and C tables. This arrangement will make it easy to find the
233 * controling tmgr structure for any table in the system by use of
234 * (relatively) simple macros.
235 */
236
237 /*
238 * Global variables for storing the base addresses for the areas
239 * labeled above.
240 */
241 static vaddr_t kernAphys;
242 static mmu_long_dte_t *kernAbase;
243 static mmu_short_dte_t *kernBbase;
244 static mmu_short_pte_t *kernCbase;
245 static mmu_short_pte_t *mmuCbase;
246 static mmu_short_dte_t *mmuBbase;
247 static mmu_long_dte_t *mmuAbase;
248 static a_tmgr_t *Atmgrbase;
249 static b_tmgr_t *Btmgrbase;
250 static c_tmgr_t *Ctmgrbase;
251 static pv_t *pvbase;
252 static pv_elem_t *pvebase;
253 struct pmap kernel_pmap;
254
255 /*
256 * This holds the CRP currently loaded into the MMU.
257 */
258 struct mmu_rootptr kernel_crp;
259
260 /*
261 * Just all around global variables.
262 */
263 static TAILQ_HEAD(a_pool_head_struct, a_tmgr_struct) a_pool;
264 static TAILQ_HEAD(b_pool_head_struct, b_tmgr_struct) b_pool;
265 static TAILQ_HEAD(c_pool_head_struct, c_tmgr_struct) c_pool;
266
267
268 /*
269 * Flags used to mark the safety/availability of certain operations or
270 * resources.
271 */
272 static boolean_t bootstrap_alloc_enabled = FALSE; /*Safe to use pmap_bootstrap_alloc().*/
273 int tmp_vpages_inuse; /* Temporary virtual pages are in use */
274
275 /*
276 * XXX: For now, retain the traditional variables that were
277 * used in the old pmap/vm interface (without NONCONTIG).
278 */
279 /* Kernel virtual address space available: */
280 vaddr_t virtual_avail, virtual_end;
281 /* Physical address space available: */
282 paddr_t avail_start, avail_end;
283
284 /* This keep track of the end of the contiguously mapped range. */
285 vaddr_t virtual_contig_end;
286
287 /* Physical address used by pmap_next_page() */
288 paddr_t avail_next;
289
290 /* These are used by pmap_copy_page(), etc. */
291 vaddr_t tmp_vpages[2];
292
293 /* memory pool for pmap structures */
294 struct pool pmap_pmap_pool;
295
296 /*
297 * The 3/80 is the only member of the sun3x family that has non-contiguous
298 * physical memory. Memory is divided into 4 banks which are physically
299 * locatable on the system board. Although the size of these banks varies
300 * with the size of memory they contain, their base addresses are
301 * permenently fixed. The following structure, which describes these
302 * banks, is initialized by pmap_bootstrap() after it reads from a similar
303 * structure provided by the ROM Monitor.
304 *
305 * For the other machines in the sun3x architecture which do have contiguous
306 * RAM, this list will have only one entry, which will describe the entire
307 * range of available memory.
308 */
309 struct pmap_physmem_struct avail_mem[SUN3X_NPHYS_RAM_SEGS];
310 u_int total_phys_mem;
311
312 /*************************************************************************/
313
314 /*
315 * XXX - Should "tune" these based on statistics.
316 *
317 * My first guess about the relative numbers of these needed is
318 * based on the fact that a "typical" process will have several
319 * pages mapped at low virtual addresses (text, data, bss), then
320 * some mapped shared libraries, and then some stack pages mapped
321 * near the high end of the VA space. Each process can use only
322 * one A table, and most will use only two B tables (maybe three)
323 * and probably about four C tables. Therefore, the first guess
324 * at the relative numbers of these needed is 1:2:4 -gwr
325 *
326 * The number of C tables needed is closely related to the amount
327 * of physical memory available plus a certain amount attributable
328 * to the use of double mappings. With a few simulation statistics
329 * we can find a reasonably good estimation of this unknown value.
330 * Armed with that and the above ratios, we have a good idea of what
331 * is needed at each level. -j
332 *
333 * Note: It is not physical memory memory size, but the total mapped
334 * virtual space required by the combined working sets of all the
335 * currently _runnable_ processes. (Sleeping ones don't count.)
336 * The amount of physical memory should be irrelevant. -gwr
337 */
338 #ifdef FIXED_NTABLES
339 #define NUM_A_TABLES 16
340 #define NUM_B_TABLES 32
341 #define NUM_C_TABLES 64
342 #else
343 unsigned int NUM_A_TABLES, NUM_B_TABLES, NUM_C_TABLES;
344 #endif /* FIXED_NTABLES */
345
346 /*
347 * This determines our total virtual mapping capacity.
348 * Yes, it is a FIXED value so we can pre-allocate.
349 */
350 #define NUM_USER_PTES (NUM_C_TABLES * MMU_C_TBL_SIZE)
351
352 /*
353 * The size of the Kernel Virtual Address Space (KVAS)
354 * for purposes of MMU table allocation is -KERNBASE
355 * (length from KERNBASE to 0xFFFFffff)
356 */
357 #define KVAS_SIZE (-KERNBASE)
358
359 /* Numbers of kernel MMU tables to support KVAS_SIZE. */
360 #define KERN_B_TABLES (KVAS_SIZE >> MMU_TIA_SHIFT)
361 #define KERN_C_TABLES (KVAS_SIZE >> MMU_TIB_SHIFT)
362 #define NUM_KERN_PTES (KVAS_SIZE >> MMU_TIC_SHIFT)
363
364 /*************************** MISCELANEOUS MACROS *************************/
365 #define pmap_lock(pmap) simple_lock(&pmap->pm_lock)
366 #define pmap_unlock(pmap) simple_unlock(&pmap->pm_lock)
367 #define pmap_add_ref(pmap) ++pmap->pm_refcount
368 #define pmap_del_ref(pmap) --pmap->pm_refcount
369 #define pmap_refcount(pmap) pmap->pm_refcount
370
371 void *pmap_bootstrap_alloc(int);
372
373 static INLINE void *mmu_ptov(paddr_t);
374 static INLINE paddr_t mmu_vtop(void *);
375
376 #if 0
377 static INLINE a_tmgr_t * mmuA2tmgr(mmu_long_dte_t *);
378 #endif /* 0 */
379 static INLINE b_tmgr_t * mmuB2tmgr(mmu_short_dte_t *);
380 static INLINE c_tmgr_t * mmuC2tmgr(mmu_short_pte_t *);
381
382 static INLINE pv_t *pa2pv(paddr_t);
383 static INLINE int pteidx(mmu_short_pte_t *);
384 static INLINE pmap_t current_pmap(void);
385
386 /*
387 * We can always convert between virtual and physical addresses
388 * for anything in the range [KERNBASE ... avail_start] because
389 * that range is GUARANTEED to be mapped linearly.
390 * We rely heavily upon this feature!
391 */
392 static INLINE void *
393 mmu_ptov(paddr_t pa)
394 {
395 vaddr_t va;
396
397 va = (pa + KERNBASE);
398 #ifdef PMAP_DEBUG
399 if ((va < KERNBASE) || (va >= virtual_contig_end))
400 panic("mmu_ptov");
401 #endif
402 return ((void*)va);
403 }
404
405 static INLINE paddr_t
406 mmu_vtop(void *vva)
407 {
408 vaddr_t va;
409
410 va = (vaddr_t)vva;
411 #ifdef PMAP_DEBUG
412 if ((va < KERNBASE) || (va >= virtual_contig_end))
413 panic("mmu_vtop");
414 #endif
415 return (va - KERNBASE);
416 }
417
418 /*
419 * These macros map MMU tables to their corresponding manager structures.
420 * They are needed quite often because many of the pointers in the pmap
421 * system reference MMU tables and not the structures that control them.
422 * There needs to be a way to find one when given the other and these
423 * macros do so by taking advantage of the memory layout described above.
424 * Here's a quick step through the first macro, mmuA2tmgr():
425 *
426 * 1) find the offset of the given MMU A table from the base of its table
427 * pool (table - mmuAbase).
428 * 2) convert this offset into a table index by dividing it by the
429 * size of one MMU 'A' table. (sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE)
430 * 3) use this index to select the corresponding 'A' table manager
431 * structure from the 'A' table manager pool (Atmgrbase[index]).
432 */
433 /* This function is not currently used. */
434 #if 0
435 static INLINE a_tmgr_t *
436 mmuA2tmgr(mmu_long_dte_t *mmuAtbl)
437 {
438 int idx;
439
440 /* Which table is this in? */
441 idx = (mmuAtbl - mmuAbase) / MMU_A_TBL_SIZE;
442 #ifdef PMAP_DEBUG
443 if ((idx < 0) || (idx >= NUM_A_TABLES))
444 panic("mmuA2tmgr");
445 #endif
446 return (&Atmgrbase[idx]);
447 }
448 #endif /* 0 */
449
450 static INLINE b_tmgr_t *
451 mmuB2tmgr(mmu_short_dte_t *mmuBtbl)
452 {
453 int idx;
454
455 /* Which table is this in? */
456 idx = (mmuBtbl - mmuBbase) / MMU_B_TBL_SIZE;
457 #ifdef PMAP_DEBUG
458 if ((idx < 0) || (idx >= NUM_B_TABLES))
459 panic("mmuB2tmgr");
460 #endif
461 return (&Btmgrbase[idx]);
462 }
463
464 /* mmuC2tmgr INTERNAL
465 **
466 * Given a pte known to belong to a C table, return the address of
467 * that table's management structure.
468 */
469 static INLINE c_tmgr_t *
470 mmuC2tmgr(mmu_short_pte_t *mmuCtbl)
471 {
472 int idx;
473
474 /* Which table is this in? */
475 idx = (mmuCtbl - mmuCbase) / MMU_C_TBL_SIZE;
476 #ifdef PMAP_DEBUG
477 if ((idx < 0) || (idx >= NUM_C_TABLES))
478 panic("mmuC2tmgr");
479 #endif
480 return (&Ctmgrbase[idx]);
481 }
482
483 /* This is now a function call below.
484 * #define pa2pv(pa) \
485 * (&pvbase[(unsigned long)\
486 * m68k_btop(pa)\
487 * ])
488 */
489
490 /* pa2pv INTERNAL
491 **
492 * Return the pv_list_head element which manages the given physical
493 * address.
494 */
495 static INLINE pv_t *
496 pa2pv(paddr_t pa)
497 {
498 struct pmap_physmem_struct *bank;
499 int idx;
500
501 bank = &avail_mem[0];
502 while (pa >= bank->pmem_end)
503 bank = bank->pmem_next;
504
505 pa -= bank->pmem_start;
506 idx = bank->pmem_pvbase + m68k_btop(pa);
507 #ifdef PMAP_DEBUG
508 if ((idx < 0) || (idx >= physmem))
509 panic("pa2pv");
510 #endif
511 return &pvbase[idx];
512 }
513
514 /* pteidx INTERNAL
515 **
516 * Return the index of the given PTE within the entire fixed table of
517 * PTEs.
518 */
519 static INLINE int
520 pteidx(mmu_short_pte_t *pte)
521 {
522 return (pte - kernCbase);
523 }
524
525 /*
526 * This just offers a place to put some debugging checks,
527 * and reduces the number of places "curlwp" appears...
528 */
529 static INLINE pmap_t
530 current_pmap(void)
531 {
532 struct vmspace *vm;
533 struct vm_map *map;
534 pmap_t pmap;
535
536 if (curlwp == NULL)
537 pmap = &kernel_pmap;
538 else {
539 vm = curproc->p_vmspace;
540 map = &vm->vm_map;
541 pmap = vm_map_pmap(map);
542 }
543
544 return (pmap);
545 }
546
547
548 /*************************** FUNCTION DEFINITIONS ************************
549 * These appear here merely for the compiler to enforce type checking on *
550 * all function calls. *
551 *************************************************************************/
552
553 /** Internal functions
554 ** Most functions used only within this module are defined in
555 ** pmap_pvt.h (why not here if used only here?)
556 **/
557 static void pmap_page_upload(void);
558
559 /** Interface functions
560 ** - functions required by the Mach VM Pmap interface, with MACHINE_CONTIG
561 ** defined.
562 **/
563 void pmap_pinit(pmap_t);
564 void pmap_release(pmap_t);
565
566 /********************************** CODE ********************************
567 * Functions that are called from other parts of the kernel are labeled *
568 * as 'INTERFACE' functions. Functions that are only called from *
569 * within the pmap module are labeled as 'INTERNAL' functions. *
570 * Functions that are internal, but are not (currently) used at all are *
571 * labeled 'INTERNAL_X'. *
572 ************************************************************************/
573
574 /* pmap_bootstrap INTERNAL
575 **
576 * Initializes the pmap system. Called at boot time from
577 * locore2.c:_vm_init()
578 *
579 * Reminder: having a pmap_bootstrap_alloc() and also having the VM
580 * system implement pmap_steal_memory() is redundant.
581 * Don't release this code without removing one or the other!
582 */
583 void
584 pmap_bootstrap(vaddr_t nextva)
585 {
586 struct physmemory *membank;
587 struct pmap_physmem_struct *pmap_membank;
588 vaddr_t va, eva;
589 paddr_t pa;
590 int b, c, i, j; /* running table counts */
591 int size, resvmem;
592
593 /*
594 * This function is called by __bootstrap after it has
595 * determined the type of machine and made the appropriate
596 * patches to the ROM vectors (XXX- I don't quite know what I meant
597 * by that.) It allocates and sets up enough of the pmap system
598 * to manage the kernel's address space.
599 */
600
601 /*
602 * Determine the range of kernel virtual and physical
603 * space available. Note that we ABSOLUTELY DEPEND on
604 * the fact that the first bank of memory (4MB) is
605 * mapped linearly to KERNBASE (which we guaranteed in
606 * the first instructions of locore.s).
607 * That is plenty for our bootstrap work.
608 */
609 virtual_avail = m68k_round_page(nextva);
610 virtual_contig_end = KERNBASE + 0x400000; /* +4MB */
611 virtual_end = VM_MAX_KERNEL_ADDRESS;
612 /* Don't need avail_start til later. */
613
614 /* We may now call pmap_bootstrap_alloc(). */
615 bootstrap_alloc_enabled = TRUE;
616
617 /*
618 * This is a somewhat unwrapped loop to deal with
619 * copying the PROM's 'phsymem' banks into the pmap's
620 * banks. The following is always assumed:
621 * 1. There is always at least one bank of memory.
622 * 2. There is always a last bank of memory, and its
623 * pmem_next member must be set to NULL.
624 */
625 membank = romVectorPtr->v_physmemory;
626 pmap_membank = avail_mem;
627 total_phys_mem = 0;
628
629 for (;;) { /* break on !membank */
630 pmap_membank->pmem_start = membank->address;
631 pmap_membank->pmem_end = membank->address + membank->size;
632 total_phys_mem += membank->size;
633 membank = membank->next;
634 if (!membank)
635 break;
636 /* This silly syntax arises because pmap_membank
637 * is really a pre-allocated array, but it is put into
638 * use as a linked list.
639 */
640 pmap_membank->pmem_next = pmap_membank + 1;
641 pmap_membank = pmap_membank->pmem_next;
642 }
643 /* This is the last element. */
644 pmap_membank->pmem_next = NULL;
645
646 /*
647 * Note: total_phys_mem, physmem represent
648 * actual physical memory, including that
649 * reserved for the PROM monitor.
650 */
651 physmem = btoc(total_phys_mem);
652
653 /*
654 * Avail_end is set to the first byte of physical memory
655 * after the end of the last bank. We use this only to
656 * determine if a physical address is "managed" memory.
657 * This address range should be reduced to prevent the
658 * physical pages needed by the PROM monitor from being used
659 * in the VM system.
660 */
661 resvmem = total_phys_mem - *(romVectorPtr->memoryAvail);
662 resvmem = m68k_round_page(resvmem);
663 avail_end = pmap_membank->pmem_end - resvmem;
664
665 /*
666 * First allocate enough kernel MMU tables to map all
667 * of kernel virtual space from KERNBASE to 0xFFFFFFFF.
668 * Note: All must be aligned on 256 byte boundaries.
669 * Start with the level-A table (one of those).
670 */
671 size = sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE;
672 kernAbase = pmap_bootstrap_alloc(size);
673 memset(kernAbase, 0, size);
674
675 /* Now the level-B kernel tables... */
676 size = sizeof(mmu_short_dte_t) * MMU_B_TBL_SIZE * KERN_B_TABLES;
677 kernBbase = pmap_bootstrap_alloc(size);
678 memset(kernBbase, 0, size);
679
680 /* Now the level-C kernel tables... */
681 size = sizeof(mmu_short_pte_t) * MMU_C_TBL_SIZE * KERN_C_TABLES;
682 kernCbase = pmap_bootstrap_alloc(size);
683 memset(kernCbase, 0, size);
684 /*
685 * Note: In order for the PV system to work correctly, the kernel
686 * and user-level C tables must be allocated contiguously.
687 * Nothing should be allocated between here and the allocation of
688 * mmuCbase below. XXX: Should do this as one allocation, and
689 * then compute a pointer for mmuCbase instead of this...
690 *
691 * Allocate user MMU tables.
692 * These must be contiguous with the preceding.
693 */
694
695 #ifndef FIXED_NTABLES
696 /*
697 * The number of user-level C tables that should be allocated is
698 * related to the size of physical memory. In general, there should
699 * be enough tables to map four times the amount of available RAM.
700 * The extra amount is needed because some table space is wasted by
701 * fragmentation.
702 */
703 NUM_C_TABLES = (total_phys_mem * 4) / (MMU_C_TBL_SIZE * MMU_PAGE_SIZE);
704 NUM_B_TABLES = NUM_C_TABLES / 2;
705 NUM_A_TABLES = NUM_B_TABLES / 2;
706 #endif /* !FIXED_NTABLES */
707
708 size = sizeof(mmu_short_pte_t) * MMU_C_TBL_SIZE * NUM_C_TABLES;
709 mmuCbase = pmap_bootstrap_alloc(size);
710
711 size = sizeof(mmu_short_dte_t) * MMU_B_TBL_SIZE * NUM_B_TABLES;
712 mmuBbase = pmap_bootstrap_alloc(size);
713
714 size = sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE * NUM_A_TABLES;
715 mmuAbase = pmap_bootstrap_alloc(size);
716
717 /*
718 * Fill in the never-changing part of the kernel tables.
719 * For simplicity, the kernel's mappings will be editable as a
720 * flat array of page table entries at kernCbase. The
721 * higher level 'A' and 'B' tables must be initialized to point
722 * to this lower one.
723 */
724 b = c = 0;
725
726 /*
727 * Invalidate all mappings below KERNBASE in the A table.
728 * This area has already been zeroed out, but it is good
729 * practice to explicitly show that we are interpreting
730 * it as a list of A table descriptors.
731 */
732 for (i = 0; i < MMU_TIA(KERNBASE); i++) {
733 kernAbase[i].addr.raw = 0;
734 }
735
736 /*
737 * Set up the kernel A and B tables so that they will reference the
738 * correct spots in the contiguous table of PTEs allocated for the
739 * kernel's virtual memory space.
740 */
741 for (i = MMU_TIA(KERNBASE); i < MMU_A_TBL_SIZE; i++) {
742 kernAbase[i].attr.raw =
743 MMU_LONG_DTE_LU | MMU_LONG_DTE_SUPV | MMU_DT_SHORT;
744 kernAbase[i].addr.raw = mmu_vtop(&kernBbase[b]);
745
746 for (j=0; j < MMU_B_TBL_SIZE; j++) {
747 kernBbase[b + j].attr.raw = mmu_vtop(&kernCbase[c])
748 | MMU_DT_SHORT;
749 c += MMU_C_TBL_SIZE;
750 }
751 b += MMU_B_TBL_SIZE;
752 }
753
754 pmap_alloc_usermmu(); /* Allocate user MMU tables. */
755 pmap_alloc_usertmgr(); /* Allocate user MMU table managers.*/
756 pmap_alloc_pv(); /* Allocate physical->virtual map. */
757
758 /*
759 * We are now done with pmap_bootstrap_alloc(). Round up
760 * `virtual_avail' to the nearest page, and set the flag
761 * to prevent use of pmap_bootstrap_alloc() hereafter.
762 */
763 pmap_bootstrap_aalign(PAGE_SIZE);
764 bootstrap_alloc_enabled = FALSE;
765
766 /*
767 * Now that we are done with pmap_bootstrap_alloc(), we
768 * must save the virtual and physical addresses of the
769 * end of the linearly mapped range, which are stored in
770 * virtual_contig_end and avail_start, respectively.
771 * These variables will never change after this point.
772 */
773 virtual_contig_end = virtual_avail;
774 avail_start = virtual_avail - KERNBASE;
775
776 /*
777 * `avail_next' is a running pointer used by pmap_next_page() to
778 * keep track of the next available physical page to be handed
779 * to the VM system during its initialization, in which it
780 * asks for physical pages, one at a time.
781 */
782 avail_next = avail_start;
783
784 /*
785 * Now allocate some virtual addresses, but not the physical pages
786 * behind them. Note that virtual_avail is already page-aligned.
787 *
788 * tmp_vpages[] is an array of two virtual pages used for temporary
789 * kernel mappings in the pmap module to facilitate various physical
790 * address-oritented operations.
791 */
792 tmp_vpages[0] = virtual_avail;
793 virtual_avail += PAGE_SIZE;
794 tmp_vpages[1] = virtual_avail;
795 virtual_avail += PAGE_SIZE;
796
797 /** Initialize the PV system **/
798 pmap_init_pv();
799
800 /*
801 * Fill in the kernel_pmap structure and kernel_crp.
802 */
803 kernAphys = mmu_vtop(kernAbase);
804 kernel_pmap.pm_a_tmgr = NULL;
805 kernel_pmap.pm_a_phys = kernAphys;
806 kernel_pmap.pm_refcount = 1; /* always in use */
807 simple_lock_init(&kernel_pmap.pm_lock);
808
809 kernel_crp.rp_attr = MMU_LONG_DTE_LU | MMU_DT_LONG;
810 kernel_crp.rp_addr = kernAphys;
811
812 /*
813 * Now pmap_enter_kernel() may be used safely and will be
814 * the main interface used hereafter to modify the kernel's
815 * virtual address space. Note that since we are still running
816 * under the PROM's address table, none of these table modifications
817 * actually take effect until pmap_takeover_mmu() is called.
818 *
819 * Note: Our tables do NOT have the PROM linear mappings!
820 * Only the mappings created here exist in our tables, so
821 * remember to map anything we expect to use.
822 */
823 va = (vaddr_t)KERNBASE;
824 pa = 0;
825
826 /*
827 * The first page of the kernel virtual address space is the msgbuf
828 * page. The page attributes (data, non-cached) are set here, while
829 * the address is assigned to this global pointer in cpu_startup().
830 * It is non-cached, mostly due to paranoia.
831 */
832 pmap_enter_kernel(va, pa|PMAP_NC, VM_PROT_ALL);
833 va += PAGE_SIZE; pa += PAGE_SIZE;
834
835 /* Next page is used as the temporary stack. */
836 pmap_enter_kernel(va, pa, VM_PROT_ALL);
837 va += PAGE_SIZE; pa += PAGE_SIZE;
838
839 /*
840 * Map all of the kernel's text segment as read-only and cacheable.
841 * (Cacheable is implied by default). Unfortunately, the last bytes
842 * of kernel text and the first bytes of kernel data will often be
843 * sharing the same page. Therefore, the last page of kernel text
844 * has to be mapped as read/write, to accomodate the data.
845 */
846 eva = m68k_trunc_page((vaddr_t)etext);
847 for (; va < eva; va += PAGE_SIZE, pa += PAGE_SIZE)
848 pmap_enter_kernel(va, pa, VM_PROT_READ|VM_PROT_EXECUTE);
849
850 /*
851 * Map all of the kernel's data as read/write and cacheable.
852 * This includes: data, BSS, symbols, and everything in the
853 * contiguous memory used by pmap_bootstrap_alloc()
854 */
855 for (; pa < avail_start; va += PAGE_SIZE, pa += PAGE_SIZE)
856 pmap_enter_kernel(va, pa, VM_PROT_READ|VM_PROT_WRITE);
857
858 /*
859 * At this point we are almost ready to take over the MMU. But first
860 * we must save the PROM's address space in our map, as we call its
861 * routines and make references to its data later in the kernel.
862 */
863 pmap_bootstrap_copyprom();
864 pmap_takeover_mmu();
865 pmap_bootstrap_setprom();
866
867 /* Notify the VM system of our page size. */
868 uvmexp.pagesize = PAGE_SIZE;
869 uvm_setpagesize();
870
871 pmap_page_upload();
872 }
873
874
875 /* pmap_alloc_usermmu INTERNAL
876 **
877 * Called from pmap_bootstrap() to allocate MMU tables that will
878 * eventually be used for user mappings.
879 */
880 void
881 pmap_alloc_usermmu(void)
882 {
883 /* XXX: Moved into caller. */
884 }
885
886 /* pmap_alloc_pv INTERNAL
887 **
888 * Called from pmap_bootstrap() to allocate the physical
889 * to virtual mapping list. Each physical page of memory
890 * in the system has a corresponding element in this list.
891 */
892 void
893 pmap_alloc_pv(void)
894 {
895 int i;
896 unsigned int total_mem;
897
898 /*
899 * Allocate a pv_head structure for every page of physical
900 * memory that will be managed by the system. Since memory on
901 * the 3/80 is non-contiguous, we cannot arrive at a total page
902 * count by subtraction of the lowest available address from the
903 * highest, but rather we have to step through each memory
904 * bank and add the number of pages in each to the total.
905 *
906 * At this time we also initialize the offset of each bank's
907 * starting pv_head within the pv_head list so that the physical
908 * memory state routines (pmap_is_referenced(),
909 * pmap_is_modified(), et al.) can quickly find coresponding
910 * pv_heads in spite of the non-contiguity.
911 */
912 total_mem = 0;
913 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
914 avail_mem[i].pmem_pvbase = m68k_btop(total_mem);
915 total_mem += avail_mem[i].pmem_end -
916 avail_mem[i].pmem_start;
917 if (avail_mem[i].pmem_next == NULL)
918 break;
919 }
920 pvbase = (pv_t *) pmap_bootstrap_alloc(sizeof(pv_t) *
921 m68k_btop(total_phys_mem));
922 }
923
924 /* pmap_alloc_usertmgr INTERNAL
925 **
926 * Called from pmap_bootstrap() to allocate the structures which
927 * facilitate management of user MMU tables. Each user MMU table
928 * in the system has one such structure associated with it.
929 */
930 void
931 pmap_alloc_usertmgr(void)
932 {
933 /* Allocate user MMU table managers */
934 /* It would be a lot simpler to just make these BSS, but */
935 /* we may want to change their size at boot time... -j */
936 Atmgrbase = (a_tmgr_t *) pmap_bootstrap_alloc(sizeof(a_tmgr_t)
937 * NUM_A_TABLES);
938 Btmgrbase = (b_tmgr_t *) pmap_bootstrap_alloc(sizeof(b_tmgr_t)
939 * NUM_B_TABLES);
940 Ctmgrbase = (c_tmgr_t *) pmap_bootstrap_alloc(sizeof(c_tmgr_t)
941 * NUM_C_TABLES);
942
943 /*
944 * Allocate PV list elements for the physical to virtual
945 * mapping system.
946 */
947 pvebase = (pv_elem_t *) pmap_bootstrap_alloc(
948 sizeof(pv_elem_t) * (NUM_USER_PTES + NUM_KERN_PTES));
949 }
950
951 /* pmap_bootstrap_copyprom() INTERNAL
952 **
953 * Copy the PROM mappings into our own tables. Note, we
954 * can use physical addresses until __bootstrap returns.
955 */
956 void
957 pmap_bootstrap_copyprom(void)
958 {
959 struct sunromvec *romp;
960 int *mon_ctbl;
961 mmu_short_pte_t *kpte;
962 int i, len;
963
964 romp = romVectorPtr;
965
966 /*
967 * Copy the mappings in SUN3X_MON_KDB_BASE...SUN3X_MONEND
968 * Note: mon_ctbl[0] maps SUN3X_MON_KDB_BASE
969 */
970 mon_ctbl = *romp->monptaddr;
971 i = m68k_btop(SUN3X_MON_KDB_BASE - KERNBASE);
972 kpte = &kernCbase[i];
973 len = m68k_btop(SUN3X_MONEND - SUN3X_MON_KDB_BASE);
974
975 for (i = 0; i < len; i++) {
976 kpte[i].attr.raw = mon_ctbl[i];
977 }
978
979 /*
980 * Copy the mappings at MON_DVMA_BASE (to the end).
981 * Note, in here, mon_ctbl[0] maps MON_DVMA_BASE.
982 * Actually, we only want the last page, which the
983 * PROM has set up for use by the "ie" driver.
984 * (The i82686 needs its SCP there.)
985 * If we copy all the mappings, pmap_enter_kernel
986 * may complain about finding valid PTEs that are
987 * not recorded in our PV lists...
988 */
989 mon_ctbl = *romp->shadowpteaddr;
990 i = m68k_btop(SUN3X_MON_DVMA_BASE - KERNBASE);
991 kpte = &kernCbase[i];
992 len = m68k_btop(SUN3X_MON_DVMA_SIZE);
993 for (i = (len-1); i < len; i++) {
994 kpte[i].attr.raw = mon_ctbl[i];
995 }
996 }
997
998 /* pmap_takeover_mmu INTERNAL
999 **
1000 * Called from pmap_bootstrap() after it has copied enough of the
1001 * PROM mappings into the kernel map so that we can use our own
1002 * MMU table.
1003 */
1004 void
1005 pmap_takeover_mmu(void)
1006 {
1007
1008 loadcrp(&kernel_crp);
1009 }
1010
1011 /* pmap_bootstrap_setprom() INTERNAL
1012 **
1013 * Set the PROM mappings so it can see kernel space.
1014 * Note that physical addresses are used here, which
1015 * we can get away with because this runs with the
1016 * low 1GB set for transparent translation.
1017 */
1018 void
1019 pmap_bootstrap_setprom(void)
1020 {
1021 mmu_long_dte_t *mon_dte;
1022 extern struct mmu_rootptr mon_crp;
1023 int i;
1024
1025 mon_dte = (mmu_long_dte_t *) mon_crp.rp_addr;
1026 for (i = MMU_TIA(KERNBASE); i < MMU_TIA(KERN_END); i++) {
1027 mon_dte[i].attr.raw = kernAbase[i].attr.raw;
1028 mon_dte[i].addr.raw = kernAbase[i].addr.raw;
1029 }
1030 }
1031
1032
1033 /* pmap_init INTERFACE
1034 **
1035 * Called at the end of vm_init() to set up the pmap system to go
1036 * into full time operation. All initialization of kernel_pmap
1037 * should be already done by now, so this should just do things
1038 * needed for user-level pmaps to work.
1039 */
1040 void
1041 pmap_init(void)
1042 {
1043 /** Initialize the manager pools **/
1044 TAILQ_INIT(&a_pool);
1045 TAILQ_INIT(&b_pool);
1046 TAILQ_INIT(&c_pool);
1047
1048 /**************************************************************
1049 * Initialize all tmgr structures and MMU tables they manage. *
1050 **************************************************************/
1051 /** Initialize A tables **/
1052 pmap_init_a_tables();
1053 /** Initialize B tables **/
1054 pmap_init_b_tables();
1055 /** Initialize C tables **/
1056 pmap_init_c_tables();
1057
1058 /** Initialize the pmap pools **/
1059 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
1060 &pool_allocator_nointr);
1061 }
1062
1063 /* pmap_init_a_tables() INTERNAL
1064 **
1065 * Initializes all A managers, their MMU A tables, and inserts
1066 * them into the A manager pool for use by the system.
1067 */
1068 void
1069 pmap_init_a_tables(void)
1070 {
1071 int i;
1072 a_tmgr_t *a_tbl;
1073
1074 for (i = 0; i < NUM_A_TABLES; i++) {
1075 /* Select the next available A manager from the pool */
1076 a_tbl = &Atmgrbase[i];
1077
1078 /*
1079 * Clear its parent entry. Set its wired and valid
1080 * entry count to zero.
1081 */
1082 a_tbl->at_parent = NULL;
1083 a_tbl->at_wcnt = a_tbl->at_ecnt = 0;
1084
1085 /* Assign it the next available MMU A table from the pool */
1086 a_tbl->at_dtbl = &mmuAbase[i * MMU_A_TBL_SIZE];
1087
1088 /*
1089 * Initialize the MMU A table with the table in the `proc0',
1090 * or kernel, mapping. This ensures that every process has
1091 * the kernel mapped in the top part of its address space.
1092 */
1093 memcpy(a_tbl->at_dtbl, kernAbase, MMU_A_TBL_SIZE *
1094 sizeof(mmu_long_dte_t));
1095
1096 /*
1097 * Finally, insert the manager into the A pool,
1098 * making it ready to be used by the system.
1099 */
1100 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
1101 }
1102 }
1103
1104 /* pmap_init_b_tables() INTERNAL
1105 **
1106 * Initializes all B table managers, their MMU B tables, and
1107 * inserts them into the B manager pool for use by the system.
1108 */
1109 void
1110 pmap_init_b_tables(void)
1111 {
1112 int i, j;
1113 b_tmgr_t *b_tbl;
1114
1115 for (i = 0; i < NUM_B_TABLES; i++) {
1116 /* Select the next available B manager from the pool */
1117 b_tbl = &Btmgrbase[i];
1118
1119 b_tbl->bt_parent = NULL; /* clear its parent, */
1120 b_tbl->bt_pidx = 0; /* parent index, */
1121 b_tbl->bt_wcnt = 0; /* wired entry count, */
1122 b_tbl->bt_ecnt = 0; /* valid entry count. */
1123
1124 /* Assign it the next available MMU B table from the pool */
1125 b_tbl->bt_dtbl = &mmuBbase[i * MMU_B_TBL_SIZE];
1126
1127 /* Invalidate every descriptor in the table */
1128 for (j=0; j < MMU_B_TBL_SIZE; j++)
1129 b_tbl->bt_dtbl[j].attr.raw = MMU_DT_INVALID;
1130
1131 /* Insert the manager into the B pool */
1132 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
1133 }
1134 }
1135
1136 /* pmap_init_c_tables() INTERNAL
1137 **
1138 * Initializes all C table managers, their MMU C tables, and
1139 * inserts them into the C manager pool for use by the system.
1140 */
1141 void
1142 pmap_init_c_tables(void)
1143 {
1144 int i, j;
1145 c_tmgr_t *c_tbl;
1146
1147 for (i = 0; i < NUM_C_TABLES; i++) {
1148 /* Select the next available C manager from the pool */
1149 c_tbl = &Ctmgrbase[i];
1150
1151 c_tbl->ct_parent = NULL; /* clear its parent, */
1152 c_tbl->ct_pidx = 0; /* parent index, */
1153 c_tbl->ct_wcnt = 0; /* wired entry count, */
1154 c_tbl->ct_ecnt = 0; /* valid entry count, */
1155 c_tbl->ct_pmap = NULL; /* parent pmap, */
1156 c_tbl->ct_va = 0; /* base of managed range */
1157
1158 /* Assign it the next available MMU C table from the pool */
1159 c_tbl->ct_dtbl = &mmuCbase[i * MMU_C_TBL_SIZE];
1160
1161 for (j=0; j < MMU_C_TBL_SIZE; j++)
1162 c_tbl->ct_dtbl[j].attr.raw = MMU_DT_INVALID;
1163
1164 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
1165 }
1166 }
1167
1168 /* pmap_init_pv() INTERNAL
1169 **
1170 * Initializes the Physical to Virtual mapping system.
1171 */
1172 void
1173 pmap_init_pv(void)
1174 {
1175 int i;
1176
1177 /* Initialize every PV head. */
1178 for (i = 0; i < m68k_btop(total_phys_mem); i++) {
1179 pvbase[i].pv_idx = PVE_EOL; /* Indicate no mappings */
1180 pvbase[i].pv_flags = 0; /* Zero out page flags */
1181 }
1182 }
1183
1184 /* get_a_table INTERNAL
1185 **
1186 * Retrieve and return a level A table for use in a user map.
1187 */
1188 a_tmgr_t *
1189 get_a_table(void)
1190 {
1191 a_tmgr_t *tbl;
1192 pmap_t pmap;
1193
1194 /* Get the top A table in the pool */
1195 tbl = TAILQ_FIRST(&a_pool);
1196 if (tbl == NULL) {
1197 /*
1198 * XXX - Instead of panicking here and in other get_x_table
1199 * functions, we do have the option of sleeping on the head of
1200 * the table pool. Any function which updates the table pool
1201 * would then issue a wakeup() on the head, thus waking up any
1202 * processes waiting for a table.
1203 *
1204 * Actually, the place to sleep would be when some process
1205 * asks for a "wired" mapping that would run us short of
1206 * mapping resources. This design DEPENDS on always having
1207 * some mapping resources in the pool for stealing, so we
1208 * must make sure we NEVER let the pool become empty. -gwr
1209 */
1210 panic("get_a_table: out of A tables.");
1211 }
1212
1213 TAILQ_REMOVE(&a_pool, tbl, at_link);
1214 /*
1215 * If the table has a non-null parent pointer then it is in use.
1216 * Forcibly abduct it from its parent and clear its entries.
1217 * No re-entrancy worries here. This table would not be in the
1218 * table pool unless it was available for use.
1219 *
1220 * Note that the second argument to free_a_table() is FALSE. This
1221 * indicates that the table should not be relinked into the A table
1222 * pool. That is a job for the function that called us.
1223 */
1224 if (tbl->at_parent) {
1225 KASSERT(tbl->at_wcnt == 0);
1226 pmap = tbl->at_parent;
1227 free_a_table(tbl, FALSE);
1228 pmap->pm_a_tmgr = NULL;
1229 pmap->pm_a_phys = kernAphys;
1230 }
1231 return tbl;
1232 }
1233
1234 /* get_b_table INTERNAL
1235 **
1236 * Return a level B table for use.
1237 */
1238 b_tmgr_t *
1239 get_b_table(void)
1240 {
1241 b_tmgr_t *tbl;
1242
1243 /* See 'get_a_table' for comments. */
1244 tbl = TAILQ_FIRST(&b_pool);
1245 if (tbl == NULL)
1246 panic("get_b_table: out of B tables.");
1247 TAILQ_REMOVE(&b_pool, tbl, bt_link);
1248 if (tbl->bt_parent) {
1249 KASSERT(tbl->bt_wcnt == 0);
1250 tbl->bt_parent->at_dtbl[tbl->bt_pidx].attr.raw = MMU_DT_INVALID;
1251 tbl->bt_parent->at_ecnt--;
1252 free_b_table(tbl, FALSE);
1253 }
1254 return tbl;
1255 }
1256
1257 /* get_c_table INTERNAL
1258 **
1259 * Return a level C table for use.
1260 */
1261 c_tmgr_t *
1262 get_c_table(void)
1263 {
1264 c_tmgr_t *tbl;
1265
1266 /* See 'get_a_table' for comments */
1267 tbl = TAILQ_FIRST(&c_pool);
1268 if (tbl == NULL)
1269 panic("get_c_table: out of C tables.");
1270 TAILQ_REMOVE(&c_pool, tbl, ct_link);
1271 if (tbl->ct_parent) {
1272 KASSERT(tbl->ct_wcnt == 0);
1273 tbl->ct_parent->bt_dtbl[tbl->ct_pidx].attr.raw = MMU_DT_INVALID;
1274 tbl->ct_parent->bt_ecnt--;
1275 free_c_table(tbl, FALSE);
1276 }
1277 return tbl;
1278 }
1279
1280 /*
1281 * The following 'free_table' and 'steal_table' functions are called to
1282 * detach tables from their current obligations (parents and children) and
1283 * prepare them for reuse in another mapping.
1284 *
1285 * Free_table is used when the calling function will handle the fate
1286 * of the parent table, such as returning it to the free pool when it has
1287 * no valid entries. Functions that do not want to handle this should
1288 * call steal_table, in which the parent table's descriptors and entry
1289 * count are automatically modified when this table is removed.
1290 */
1291
1292 /* free_a_table INTERNAL
1293 **
1294 * Unmaps the given A table and all child tables from their current
1295 * mappings. Returns the number of pages that were invalidated.
1296 * If 'relink' is true, the function will return the table to the head
1297 * of the available table pool.
1298 *
1299 * Cache note: The MC68851 will automatically flush all
1300 * descriptors derived from a given A table from its
1301 * Automatic Translation Cache (ATC) if we issue a
1302 * 'PFLUSHR' instruction with the base address of the
1303 * table. This function should do, and does so.
1304 * Note note: We are using an MC68030 - there is no
1305 * PFLUSHR.
1306 */
1307 int
1308 free_a_table(a_tmgr_t *a_tbl, boolean_t relink)
1309 {
1310 int i, removed_cnt;
1311 mmu_long_dte_t *dte;
1312 mmu_short_dte_t *dtbl;
1313 b_tmgr_t *b_tbl;
1314 uint8_t at_wired, bt_wired;
1315
1316 /*
1317 * Flush the ATC cache of all cached descriptors derived
1318 * from this table.
1319 * Sun3x does not use 68851's cached table feature
1320 * flush_atc_crp(mmu_vtop(a_tbl->dte));
1321 */
1322
1323 /*
1324 * Remove any pending cache flushes that were designated
1325 * for the pmap this A table belongs to.
1326 * a_tbl->parent->atc_flushq[0] = 0;
1327 * Not implemented in sun3x.
1328 */
1329
1330 /*
1331 * All A tables in the system should retain a map for the
1332 * kernel. If the table contains any valid descriptors
1333 * (other than those for the kernel area), invalidate them all,
1334 * stopping short of the kernel's entries.
1335 */
1336 removed_cnt = 0;
1337 at_wired = a_tbl->at_wcnt;
1338 if (a_tbl->at_ecnt) {
1339 dte = a_tbl->at_dtbl;
1340 for (i=0; i < MMU_TIA(KERNBASE); i++) {
1341 /*
1342 * If a table entry points to a valid B table, free
1343 * it and its children.
1344 */
1345 if (MMU_VALID_DT(dte[i])) {
1346 /*
1347 * The following block does several things,
1348 * from innermost expression to the
1349 * outermost:
1350 * 1) It extracts the base (cc 1996)
1351 * address of the B table pointed
1352 * to in the A table entry dte[i].
1353 * 2) It converts this base address into
1354 * the virtual address it can be
1355 * accessed with. (all MMU tables point
1356 * to physical addresses.)
1357 * 3) It finds the corresponding manager
1358 * structure which manages this MMU table.
1359 * 4) It frees the manager structure.
1360 * (This frees the MMU table and all
1361 * child tables. See 'free_b_table' for
1362 * details.)
1363 */
1364 dtbl = mmu_ptov(dte[i].addr.raw);
1365 b_tbl = mmuB2tmgr(dtbl);
1366 bt_wired = b_tbl->bt_wcnt;
1367 removed_cnt += free_b_table(b_tbl, TRUE);
1368 if (bt_wired)
1369 a_tbl->at_wcnt--;
1370 dte[i].attr.raw = MMU_DT_INVALID;
1371 }
1372 }
1373 a_tbl->at_ecnt = 0;
1374 }
1375 KASSERT(a_tbl->at_wcnt == 0);
1376
1377 if (relink) {
1378 a_tbl->at_parent = NULL;
1379 if (!at_wired)
1380 TAILQ_REMOVE(&a_pool, a_tbl, at_link);
1381 TAILQ_INSERT_HEAD(&a_pool, a_tbl, at_link);
1382 }
1383 return removed_cnt;
1384 }
1385
1386 /* free_b_table INTERNAL
1387 **
1388 * Unmaps the given B table and all its children from their current
1389 * mappings. Returns the number of pages that were invalidated.
1390 * (For comments, see 'free_a_table()').
1391 */
1392 int
1393 free_b_table(b_tmgr_t *b_tbl, boolean_t relink)
1394 {
1395 int i, removed_cnt;
1396 mmu_short_dte_t *dte;
1397 mmu_short_pte_t *dtbl;
1398 c_tmgr_t *c_tbl;
1399 uint8_t bt_wired, ct_wired;
1400
1401 removed_cnt = 0;
1402 bt_wired = b_tbl->bt_wcnt;
1403 if (b_tbl->bt_ecnt) {
1404 dte = b_tbl->bt_dtbl;
1405 for (i=0; i < MMU_B_TBL_SIZE; i++) {
1406 if (MMU_VALID_DT(dte[i])) {
1407 dtbl = mmu_ptov(MMU_DTE_PA(dte[i]));
1408 c_tbl = mmuC2tmgr(dtbl);
1409 ct_wired = c_tbl->ct_wcnt;
1410 removed_cnt += free_c_table(c_tbl, TRUE);
1411 if (ct_wired)
1412 b_tbl->bt_wcnt--;
1413 dte[i].attr.raw = MMU_DT_INVALID;
1414 }
1415 }
1416 b_tbl->bt_ecnt = 0;
1417 }
1418 KASSERT(b_tbl->bt_wcnt == 0);
1419
1420 if (relink) {
1421 b_tbl->bt_parent = NULL;
1422 if (!bt_wired)
1423 TAILQ_REMOVE(&b_pool, b_tbl, bt_link);
1424 TAILQ_INSERT_HEAD(&b_pool, b_tbl, bt_link);
1425 }
1426 return removed_cnt;
1427 }
1428
1429 /* free_c_table INTERNAL
1430 **
1431 * Unmaps the given C table from use and returns it to the pool for
1432 * re-use. Returns the number of pages that were invalidated.
1433 *
1434 * This function preserves any physical page modification information
1435 * contained in the page descriptors within the C table by calling
1436 * 'pmap_remove_pte().'
1437 */
1438 int
1439 free_c_table(c_tmgr_t *c_tbl, boolean_t relink)
1440 {
1441 mmu_short_pte_t *c_pte;
1442 int i, removed_cnt;
1443 uint8_t ct_wired;
1444
1445 removed_cnt = 0;
1446 ct_wired = c_tbl->ct_wcnt;
1447 if (c_tbl->ct_ecnt) {
1448 for (i=0; i < MMU_C_TBL_SIZE; i++) {
1449 c_pte = &c_tbl->ct_dtbl[i];
1450 if (MMU_VALID_DT(*c_pte)) {
1451 if (c_pte->attr.raw & MMU_SHORT_PTE_WIRED)
1452 c_tbl->ct_wcnt--;
1453 pmap_remove_pte(c_pte);
1454 removed_cnt++;
1455 }
1456 }
1457 c_tbl->ct_ecnt = 0;
1458 }
1459 KASSERT(c_tbl->ct_wcnt == 0);
1460
1461 if (relink) {
1462 c_tbl->ct_parent = NULL;
1463 if (!ct_wired)
1464 TAILQ_REMOVE(&c_pool, c_tbl, ct_link);
1465 TAILQ_INSERT_HEAD(&c_pool, c_tbl, ct_link);
1466 }
1467 return removed_cnt;
1468 }
1469
1470
1471 /* pmap_remove_pte INTERNAL
1472 **
1473 * Unmap the given pte and preserve any page modification
1474 * information by transfering it to the pv head of the
1475 * physical page it maps to. This function does not update
1476 * any reference counts because it is assumed that the calling
1477 * function will do so.
1478 */
1479 void
1480 pmap_remove_pte(mmu_short_pte_t *pte)
1481 {
1482 u_short pv_idx, targ_idx;
1483 paddr_t pa;
1484 pv_t *pv;
1485
1486 pa = MMU_PTE_PA(*pte);
1487 if (is_managed(pa)) {
1488 pv = pa2pv(pa);
1489 targ_idx = pteidx(pte); /* Index of PTE being removed */
1490
1491 /*
1492 * If the PTE being removed is the first (or only) PTE in
1493 * the list of PTEs currently mapped to this page, remove the
1494 * PTE by changing the index found on the PV head. Otherwise
1495 * a linear search through the list will have to be executed
1496 * in order to find the PVE which points to the PTE being
1497 * removed, so that it may be modified to point to its new
1498 * neighbor.
1499 */
1500
1501 pv_idx = pv->pv_idx; /* Index of first PTE in PV list */
1502 if (pv_idx == targ_idx) {
1503 pv->pv_idx = pvebase[targ_idx].pve_next;
1504 } else {
1505
1506 /*
1507 * Find the PV element pointing to the target
1508 * element. Note: may have pv_idx==PVE_EOL
1509 */
1510
1511 for (;;) {
1512 if (pv_idx == PVE_EOL) {
1513 goto pv_not_found;
1514 }
1515 if (pvebase[pv_idx].pve_next == targ_idx)
1516 break;
1517 pv_idx = pvebase[pv_idx].pve_next;
1518 }
1519
1520 /*
1521 * At this point, pv_idx is the index of the PV
1522 * element just before the target element in the list.
1523 * Unlink the target.
1524 */
1525
1526 pvebase[pv_idx].pve_next = pvebase[targ_idx].pve_next;
1527 }
1528
1529 /*
1530 * Save the mod/ref bits of the pte by simply
1531 * ORing the entire pte onto the pv_flags member
1532 * of the pv structure.
1533 * There is no need to use a separate bit pattern
1534 * for usage information on the pv head than that
1535 * which is used on the MMU ptes.
1536 */
1537
1538 pv_not_found:
1539 pv->pv_flags |= (u_short) pte->attr.raw;
1540 }
1541 pte->attr.raw = MMU_DT_INVALID;
1542 }
1543
1544 /* pmap_stroll INTERNAL
1545 **
1546 * Retrieve the addresses of all table managers involved in the mapping of
1547 * the given virtual address. If the table walk completed successfully,
1548 * return TRUE. If it was only partially successful, return FALSE.
1549 * The table walk performed by this function is important to many other
1550 * functions in this module.
1551 *
1552 * Note: This function ought to be easier to read.
1553 */
1554 boolean_t
1555 pmap_stroll(pmap_t pmap, vaddr_t va, a_tmgr_t **a_tbl, b_tmgr_t **b_tbl,
1556 c_tmgr_t **c_tbl, mmu_short_pte_t **pte, int *a_idx, int *b_idx,
1557 int *pte_idx)
1558 {
1559 mmu_long_dte_t *a_dte; /* A: long descriptor table */
1560 mmu_short_dte_t *b_dte; /* B: short descriptor table */
1561
1562 if (pmap == pmap_kernel())
1563 return FALSE;
1564
1565 /* Does the given pmap have its own A table? */
1566 *a_tbl = pmap->pm_a_tmgr;
1567 if (*a_tbl == NULL)
1568 return FALSE; /* No. Return unknown. */
1569 /* Does the A table have a valid B table
1570 * under the corresponding table entry?
1571 */
1572 *a_idx = MMU_TIA(va);
1573 a_dte = &((*a_tbl)->at_dtbl[*a_idx]);
1574 if (!MMU_VALID_DT(*a_dte))
1575 return FALSE; /* No. Return unknown. */
1576 /* Yes. Extract B table from the A table. */
1577 *b_tbl = mmuB2tmgr(mmu_ptov(a_dte->addr.raw));
1578 /* Does the B table have a valid C table
1579 * under the corresponding table entry?
1580 */
1581 *b_idx = MMU_TIB(va);
1582 b_dte = &((*b_tbl)->bt_dtbl[*b_idx]);
1583 if (!MMU_VALID_DT(*b_dte))
1584 return FALSE; /* No. Return unknown. */
1585 /* Yes. Extract C table from the B table. */
1586 *c_tbl = mmuC2tmgr(mmu_ptov(MMU_DTE_PA(*b_dte)));
1587 *pte_idx = MMU_TIC(va);
1588 *pte = &((*c_tbl)->ct_dtbl[*pte_idx]);
1589
1590 return TRUE;
1591 }
1592
1593 /* pmap_enter INTERFACE
1594 **
1595 * Called by the kernel to map a virtual address
1596 * to a physical address in the given process map.
1597 *
1598 * Note: this function should apply an exclusive lock
1599 * on the pmap system for its duration. (it certainly
1600 * would save my hair!!)
1601 * This function ought to be easier to read.
1602 */
1603 int
1604 pmap_enter(pmap_t pmap, vaddr_t va, paddr_t pa, vm_prot_t prot, int flags)
1605 {
1606 boolean_t insert, managed; /* Marks the need for PV insertion.*/
1607 u_short nidx; /* PV list index */
1608 int mapflags; /* Flags for the mapping (see NOTE1) */
1609 u_int a_idx, b_idx, pte_idx; /* table indices */
1610 a_tmgr_t *a_tbl; /* A: long descriptor table manager */
1611 b_tmgr_t *b_tbl; /* B: short descriptor table manager */
1612 c_tmgr_t *c_tbl; /* C: short page table manager */
1613 mmu_long_dte_t *a_dte; /* A: long descriptor table */
1614 mmu_short_dte_t *b_dte; /* B: short descriptor table */
1615 mmu_short_pte_t *c_pte; /* C: short page descriptor table */
1616 pv_t *pv; /* pv list head */
1617 boolean_t wired; /* is the mapping to be wired? */
1618 enum {NONE, NEWA, NEWB, NEWC} llevel; /* used at end */
1619
1620 if (pmap == pmap_kernel()) {
1621 pmap_enter_kernel(va, pa, prot);
1622 return 0;
1623 }
1624
1625 /*
1626 * Determine if the mapping should be wired.
1627 */
1628 wired = ((flags & PMAP_WIRED) != 0);
1629
1630 /*
1631 * NOTE1:
1632 *
1633 * On November 13, 1999, someone changed the pmap_enter() API such
1634 * that it now accepts a 'flags' argument. This new argument
1635 * contains bit-flags for the architecture-independent (UVM) system to
1636 * use in signalling certain mapping requirements to the architecture-
1637 * dependent (pmap) system. The argument it replaces, 'wired', is now
1638 * one of the flags within it.
1639 *
1640 * In addition to flags signaled by the architecture-independent
1641 * system, parts of the architecture-dependent section of the sun3x
1642 * kernel pass their own flags in the lower, unused bits of the
1643 * physical address supplied to this function. These flags are
1644 * extracted and stored in the temporary variable 'mapflags'.
1645 *
1646 * Extract sun3x specific flags from the physical address.
1647 */
1648 mapflags = (pa & ~MMU_PAGE_MASK);
1649 pa &= MMU_PAGE_MASK;
1650
1651 /*
1652 * Determine if the physical address being mapped is on-board RAM.
1653 * Any other area of the address space is likely to belong to a
1654 * device and hence it would be disasterous to cache its contents.
1655 */
1656 if ((managed = is_managed(pa)) == FALSE)
1657 mapflags |= PMAP_NC;
1658
1659 /*
1660 * For user mappings we walk along the MMU tables of the given
1661 * pmap, reaching a PTE which describes the virtual page being
1662 * mapped or changed. If any level of the walk ends in an invalid
1663 * entry, a table must be allocated and the entry must be updated
1664 * to point to it.
1665 * There is a bit of confusion as to whether this code must be
1666 * re-entrant. For now we will assume it is. To support
1667 * re-entrancy we must unlink tables from the table pool before
1668 * we assume we may use them. Tables are re-linked into the pool
1669 * when we are finished with them at the end of the function.
1670 * But I don't feel like doing that until we have proof that this
1671 * needs to be re-entrant.
1672 * 'llevel' records which tables need to be relinked.
1673 */
1674 llevel = NONE;
1675
1676 /*
1677 * Step 1 - Retrieve the A table from the pmap. If it has no
1678 * A table, allocate a new one from the available pool.
1679 */
1680
1681 a_tbl = pmap->pm_a_tmgr;
1682 if (a_tbl == NULL) {
1683 /*
1684 * This pmap does not currently have an A table. Allocate
1685 * a new one.
1686 */
1687 a_tbl = get_a_table();
1688 a_tbl->at_parent = pmap;
1689
1690 /*
1691 * Assign this new A table to the pmap, and calculate its
1692 * physical address so that loadcrp() can be used to make
1693 * the table active.
1694 */
1695 pmap->pm_a_tmgr = a_tbl;
1696 pmap->pm_a_phys = mmu_vtop(a_tbl->at_dtbl);
1697
1698 /*
1699 * If the process receiving a new A table is the current
1700 * process, we are responsible for setting the MMU so that
1701 * it becomes the current address space. This only adds
1702 * new mappings, so no need to flush anything.
1703 */
1704 if (pmap == current_pmap()) {
1705 kernel_crp.rp_addr = pmap->pm_a_phys;
1706 loadcrp(&kernel_crp);
1707 }
1708
1709 if (!wired)
1710 llevel = NEWA;
1711 } else {
1712 /*
1713 * Use the A table already allocated for this pmap.
1714 * Unlink it from the A table pool if necessary.
1715 */
1716 if (wired && !a_tbl->at_wcnt)
1717 TAILQ_REMOVE(&a_pool, a_tbl, at_link);
1718 }
1719
1720 /*
1721 * Step 2 - Walk into the B table. If there is no valid B table,
1722 * allocate one.
1723 */
1724
1725 a_idx = MMU_TIA(va); /* Calculate the TIA of the VA. */
1726 a_dte = &a_tbl->at_dtbl[a_idx]; /* Retrieve descriptor from table */
1727 if (MMU_VALID_DT(*a_dte)) { /* Is the descriptor valid? */
1728 /* The descriptor is valid. Use the B table it points to. */
1729 /*************************************
1730 * a_idx *
1731 * v *
1732 * a_tbl -> +-+-+-+-+-+-+-+-+-+-+-+- *
1733 * | | | | | | | | | | | | *
1734 * +-+-+-+-+-+-+-+-+-+-+-+- *
1735 * | *
1736 * \- b_tbl -> +-+- *
1737 * | | *
1738 * +-+- *
1739 *************************************/
1740 b_dte = mmu_ptov(a_dte->addr.raw);
1741 b_tbl = mmuB2tmgr(b_dte);
1742
1743 /*
1744 * If the requested mapping must be wired, but this table
1745 * being used to map it is not, the table must be removed
1746 * from the available pool and its wired entry count
1747 * incremented.
1748 */
1749 if (wired && !b_tbl->bt_wcnt) {
1750 TAILQ_REMOVE(&b_pool, b_tbl, bt_link);
1751 a_tbl->at_wcnt++;
1752 }
1753 } else {
1754 /* The descriptor is invalid. Allocate a new B table. */
1755 b_tbl = get_b_table();
1756
1757 /* Point the parent A table descriptor to this new B table. */
1758 a_dte->addr.raw = mmu_vtop(b_tbl->bt_dtbl);
1759 a_dte->attr.raw = MMU_LONG_DTE_LU | MMU_DT_SHORT;
1760 a_tbl->at_ecnt++; /* Update parent's valid entry count */
1761
1762 /* Create the necessary back references to the parent table */
1763 b_tbl->bt_parent = a_tbl;
1764 b_tbl->bt_pidx = a_idx;
1765
1766 /*
1767 * If this table is to be wired, make sure the parent A table
1768 * wired count is updated to reflect that it has another wired
1769 * entry.
1770 */
1771 if (wired)
1772 a_tbl->at_wcnt++;
1773 else if (llevel == NONE)
1774 llevel = NEWB;
1775 }
1776
1777 /*
1778 * Step 3 - Walk into the C table, if there is no valid C table,
1779 * allocate one.
1780 */
1781
1782 b_idx = MMU_TIB(va); /* Calculate the TIB of the VA */
1783 b_dte = &b_tbl->bt_dtbl[b_idx]; /* Retrieve descriptor from table */
1784 if (MMU_VALID_DT(*b_dte)) { /* Is the descriptor valid? */
1785 /* The descriptor is valid. Use the C table it points to. */
1786 /**************************************
1787 * c_idx *
1788 * | v *
1789 * \- b_tbl -> +-+-+-+-+-+-+-+-+-+-+- *
1790 * | | | | | | | | | | | *
1791 * +-+-+-+-+-+-+-+-+-+-+- *
1792 * | *
1793 * \- c_tbl -> +-+-- *
1794 * | | | *
1795 * +-+-- *
1796 **************************************/
1797 c_pte = mmu_ptov(MMU_PTE_PA(*b_dte));
1798 c_tbl = mmuC2tmgr(c_pte);
1799
1800 /* If mapping is wired and table is not */
1801 if (wired && !c_tbl->ct_wcnt) {
1802 TAILQ_REMOVE(&c_pool, c_tbl, ct_link);
1803 b_tbl->bt_wcnt++;
1804 }
1805 } else {
1806 /* The descriptor is invalid. Allocate a new C table. */
1807 c_tbl = get_c_table();
1808
1809 /* Point the parent B table descriptor to this new C table. */
1810 b_dte->attr.raw = mmu_vtop(c_tbl->ct_dtbl);
1811 b_dte->attr.raw |= MMU_DT_SHORT;
1812 b_tbl->bt_ecnt++; /* Update parent's valid entry count */
1813
1814 /* Create the necessary back references to the parent table */
1815 c_tbl->ct_parent = b_tbl;
1816 c_tbl->ct_pidx = b_idx;
1817 /*
1818 * Store the pmap and base virtual managed address for faster
1819 * retrieval in the PV functions.
1820 */
1821 c_tbl->ct_pmap = pmap;
1822 c_tbl->ct_va = (va & (MMU_TIA_MASK|MMU_TIB_MASK));
1823
1824 /*
1825 * If this table is to be wired, make sure the parent B table
1826 * wired count is updated to reflect that it has another wired
1827 * entry.
1828 */
1829 if (wired)
1830 b_tbl->bt_wcnt++;
1831 else if (llevel == NONE)
1832 llevel = NEWC;
1833 }
1834
1835 /*
1836 * Step 4 - Deposit a page descriptor (PTE) into the appropriate
1837 * slot of the C table, describing the PA to which the VA is mapped.
1838 */
1839
1840 pte_idx = MMU_TIC(va);
1841 c_pte = &c_tbl->ct_dtbl[pte_idx];
1842 if (MMU_VALID_DT(*c_pte)) { /* Is the entry currently valid? */
1843 /*
1844 * The PTE is currently valid. This particular call
1845 * is just a synonym for one (or more) of the following
1846 * operations:
1847 * change protection of a page
1848 * change wiring status of a page
1849 * remove the mapping of a page
1850 */
1851
1852 /* First check if this is a wiring operation. */
1853 if (c_pte->attr.raw & MMU_SHORT_PTE_WIRED) {
1854 /*
1855 * The existing mapping is wired, so adjust wired
1856 * entry count here. If new mapping is still wired,
1857 * wired entry count will be incremented again later.
1858 */
1859 c_tbl->ct_wcnt--;
1860 if (!wired) {
1861 /*
1862 * The mapping of this PTE is being changed
1863 * from wired to unwired.
1864 * Adjust wired entry counts in each table and
1865 * set llevel flag to put unwired tables back
1866 * into the active pool.
1867 */
1868 if (c_tbl->ct_wcnt == 0) {
1869 llevel = NEWC;
1870 if (--b_tbl->bt_wcnt == 0) {
1871 llevel = NEWB;
1872 if (--a_tbl->at_wcnt == 0) {
1873 llevel = NEWA;
1874 }
1875 }
1876 }
1877 }
1878 }
1879
1880 /* Is the new address the same as the old? */
1881 if (MMU_PTE_PA(*c_pte) == pa) {
1882 /*
1883 * Yes, mark that it does not need to be reinserted
1884 * into the PV list.
1885 */
1886 insert = FALSE;
1887
1888 /*
1889 * Clear all but the modified, referenced and wired
1890 * bits on the PTE.
1891 */
1892 c_pte->attr.raw &= (MMU_SHORT_PTE_M
1893 | MMU_SHORT_PTE_USED | MMU_SHORT_PTE_WIRED);
1894 } else {
1895 /* No, remove the old entry */
1896 pmap_remove_pte(c_pte);
1897 insert = TRUE;
1898 }
1899
1900 /*
1901 * TLB flush is only necessary if modifying current map.
1902 * However, in pmap_enter(), the pmap almost always IS
1903 * the current pmap, so don't even bother to check.
1904 */
1905 TBIS(va);
1906 } else {
1907 /*
1908 * The PTE is invalid. Increment the valid entry count in
1909 * the C table manager to reflect the addition of a new entry.
1910 */
1911 c_tbl->ct_ecnt++;
1912
1913 /* XXX - temporarily make sure the PTE is cleared. */
1914 c_pte->attr.raw = 0;
1915
1916 /* It will also need to be inserted into the PV list. */
1917 insert = TRUE;
1918 }
1919
1920 /*
1921 * If page is changing from unwired to wired status, set an unused bit
1922 * within the PTE to indicate that it is wired. Also increment the
1923 * wired entry count in the C table manager.
1924 */
1925 if (wired) {
1926 c_pte->attr.raw |= MMU_SHORT_PTE_WIRED;
1927 c_tbl->ct_wcnt++;
1928 }
1929
1930 /*
1931 * Map the page, being careful to preserve modify/reference/wired
1932 * bits. At this point it is assumed that the PTE either has no bits
1933 * set, or if there are set bits, they are only modified, reference or
1934 * wired bits. If not, the following statement will cause erratic
1935 * behavior.
1936 */
1937 #ifdef PMAP_DEBUG
1938 if (c_pte->attr.raw & ~(MMU_SHORT_PTE_M |
1939 MMU_SHORT_PTE_USED | MMU_SHORT_PTE_WIRED)) {
1940 printf("pmap_enter: junk left in PTE at %p\n", c_pte);
1941 Debugger();
1942 }
1943 #endif
1944 c_pte->attr.raw |= ((u_long) pa | MMU_DT_PAGE);
1945
1946 /*
1947 * If the mapping should be read-only, set the write protect
1948 * bit in the PTE.
1949 */
1950 if (!(prot & VM_PROT_WRITE))
1951 c_pte->attr.raw |= MMU_SHORT_PTE_WP;
1952
1953 /*
1954 * Mark the PTE as used and/or modified as specified by the flags arg.
1955 */
1956 if (flags & VM_PROT_ALL) {
1957 c_pte->attr.raw |= MMU_SHORT_PTE_USED;
1958 if (flags & VM_PROT_WRITE) {
1959 c_pte->attr.raw |= MMU_SHORT_PTE_M;
1960 }
1961 }
1962
1963 /*
1964 * If the mapping should be cache inhibited (indicated by the flag
1965 * bits found on the lower order of the physical address.)
1966 * mark the PTE as a cache inhibited page.
1967 */
1968 if (mapflags & PMAP_NC)
1969 c_pte->attr.raw |= MMU_SHORT_PTE_CI;
1970
1971 /*
1972 * If the physical address being mapped is managed by the PV
1973 * system then link the pte into the list of pages mapped to that
1974 * address.
1975 */
1976 if (insert && managed) {
1977 pv = pa2pv(pa);
1978 nidx = pteidx(c_pte);
1979
1980 pvebase[nidx].pve_next = pv->pv_idx;
1981 pv->pv_idx = nidx;
1982 }
1983
1984 /* Move any allocated or unwired tables back into the active pool. */
1985
1986 switch (llevel) {
1987 case NEWA:
1988 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
1989 /* FALLTHROUGH */
1990 case NEWB:
1991 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
1992 /* FALLTHROUGH */
1993 case NEWC:
1994 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
1995 /* FALLTHROUGH */
1996 default:
1997 break;
1998 }
1999
2000 return 0;
2001 }
2002
2003 /* pmap_enter_kernel INTERNAL
2004 **
2005 * Map the given virtual address to the given physical address within the
2006 * kernel address space. This function exists because the kernel map does
2007 * not do dynamic table allocation. It consists of a contiguous array of ptes
2008 * and can be edited directly without the need to walk through any tables.
2009 *
2010 * XXX: "Danger, Will Robinson!"
2011 * Note that the kernel should never take a fault on any page
2012 * between [ KERNBASE .. virtual_avail ] and this is checked in
2013 * trap.c for kernel-mode MMU faults. This means that mappings
2014 * created in that range must be implicily wired. -gwr
2015 */
2016 void
2017 pmap_enter_kernel(vaddr_t va, paddr_t pa, vm_prot_t prot)
2018 {
2019 boolean_t was_valid, insert;
2020 u_short pte_idx;
2021 int flags;
2022 mmu_short_pte_t *pte;
2023 pv_t *pv;
2024 paddr_t old_pa;
2025
2026 flags = (pa & ~MMU_PAGE_MASK);
2027 pa &= MMU_PAGE_MASK;
2028
2029 if (is_managed(pa))
2030 insert = TRUE;
2031 else
2032 insert = FALSE;
2033
2034 /*
2035 * Calculate the index of the PTE being modified.
2036 */
2037 pte_idx = (u_long) m68k_btop(va - KERNBASE);
2038
2039 /* This array is traditionally named "Sysmap" */
2040 pte = &kernCbase[pte_idx];
2041
2042 if (MMU_VALID_DT(*pte)) {
2043 was_valid = TRUE;
2044 /*
2045 * If the PTE already maps a different
2046 * physical address, umap and pv_unlink.
2047 */
2048 old_pa = MMU_PTE_PA(*pte);
2049 if (pa != old_pa)
2050 pmap_remove_pte(pte);
2051 else {
2052 /*
2053 * Old PA and new PA are the same. No need to
2054 * relink the mapping within the PV list.
2055 */
2056 insert = FALSE;
2057
2058 /*
2059 * Save any mod/ref bits on the PTE.
2060 */
2061 pte->attr.raw &= (MMU_SHORT_PTE_USED|MMU_SHORT_PTE_M);
2062 }
2063 } else {
2064 pte->attr.raw = MMU_DT_INVALID;
2065 was_valid = FALSE;
2066 }
2067
2068 /*
2069 * Map the page. Being careful to preserve modified/referenced bits
2070 * on the PTE.
2071 */
2072 pte->attr.raw |= (pa | MMU_DT_PAGE);
2073
2074 if (!(prot & VM_PROT_WRITE)) /* If access should be read-only */
2075 pte->attr.raw |= MMU_SHORT_PTE_WP;
2076 if (flags & PMAP_NC)
2077 pte->attr.raw |= MMU_SHORT_PTE_CI;
2078 if (was_valid)
2079 TBIS(va);
2080
2081 /*
2082 * Insert the PTE into the PV system, if need be.
2083 */
2084 if (insert) {
2085 pv = pa2pv(pa);
2086 pvebase[pte_idx].pve_next = pv->pv_idx;
2087 pv->pv_idx = pte_idx;
2088 }
2089 }
2090
2091 void
2092 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot)
2093 {
2094 mmu_short_pte_t *pte;
2095
2096 /* This array is traditionally named "Sysmap" */
2097 pte = &kernCbase[(u_long)m68k_btop(va - KERNBASE)];
2098
2099 KASSERT(!MMU_VALID_DT(*pte));
2100 pte->attr.raw = MMU_DT_INVALID | MMU_DT_PAGE | (pa & MMU_PAGE_MASK);
2101 if (!(prot & VM_PROT_WRITE))
2102 pte->attr.raw |= MMU_SHORT_PTE_WP;
2103 }
2104
2105 void
2106 pmap_kremove(vaddr_t va, vsize_t len)
2107 {
2108 int idx, eidx;
2109
2110 #ifdef PMAP_DEBUG
2111 if ((va & PGOFSET) || (len & PGOFSET))
2112 panic("pmap_kremove: alignment");
2113 #endif
2114
2115 idx = m68k_btop(va - KERNBASE);
2116 eidx = m68k_btop(va + len - KERNBASE);
2117
2118 while (idx < eidx) {
2119 kernCbase[idx++].attr.raw = MMU_DT_INVALID;
2120 TBIS(va);
2121 va += PAGE_SIZE;
2122 }
2123 }
2124
2125 /* pmap_map INTERNAL
2126 **
2127 * Map a contiguous range of physical memory into a contiguous range of
2128 * the kernel virtual address space.
2129 *
2130 * Used for device mappings and early mapping of the kernel text/data/bss.
2131 * Returns the first virtual address beyond the end of the range.
2132 */
2133 vaddr_t
2134 pmap_map(vaddr_t va, paddr_t pa, paddr_t endpa, int prot)
2135 {
2136 int sz;
2137
2138 sz = endpa - pa;
2139 do {
2140 pmap_enter_kernel(va, pa, prot);
2141 va += PAGE_SIZE;
2142 pa += PAGE_SIZE;
2143 sz -= PAGE_SIZE;
2144 } while (sz > 0);
2145 pmap_update(pmap_kernel());
2146 return(va);
2147 }
2148
2149 /* pmap_protect INTERFACE
2150 **
2151 * Apply the given protection to the given virtual address range within
2152 * the given map.
2153 *
2154 * It is ok for the protection applied to be stronger than what is
2155 * specified. We use this to our advantage when the given map has no
2156 * mapping for the virtual address. By skipping a page when this
2157 * is discovered, we are effectively applying a protection of VM_PROT_NONE,
2158 * and therefore do not need to map the page just to apply a protection
2159 * code. Only pmap_enter() needs to create new mappings if they do not exist.
2160 *
2161 * XXX - This function could be speeded up by using pmap_stroll() for inital
2162 * setup, and then manual scrolling in the for() loop.
2163 */
2164 void
2165 pmap_protect(pmap_t pmap, vaddr_t startva, vaddr_t endva, vm_prot_t prot)
2166 {
2167 boolean_t iscurpmap;
2168 int a_idx, b_idx, c_idx;
2169 a_tmgr_t *a_tbl;
2170 b_tmgr_t *b_tbl;
2171 c_tmgr_t *c_tbl;
2172 mmu_short_pte_t *pte;
2173
2174 if (pmap == pmap_kernel()) {
2175 pmap_protect_kernel(startva, endva, prot);
2176 return;
2177 }
2178
2179 /*
2180 * In this particular pmap implementation, there are only three
2181 * types of memory protection: 'all' (read/write/execute),
2182 * 'read-only' (read/execute) and 'none' (no mapping.)
2183 * It is not possible for us to treat 'executable' as a separate
2184 * protection type. Therefore, protection requests that seek to
2185 * remove execute permission while retaining read or write, and those
2186 * that make little sense (write-only for example) are ignored.
2187 */
2188 switch (prot) {
2189 case VM_PROT_NONE:
2190 /*
2191 * A request to apply the protection code of
2192 * 'VM_PROT_NONE' is a synonym for pmap_remove().
2193 */
2194 pmap_remove(pmap, startva, endva);
2195 return;
2196 case VM_PROT_EXECUTE:
2197 case VM_PROT_READ:
2198 case VM_PROT_READ|VM_PROT_EXECUTE:
2199 /* continue */
2200 break;
2201 case VM_PROT_WRITE:
2202 case VM_PROT_WRITE|VM_PROT_READ:
2203 case VM_PROT_WRITE|VM_PROT_EXECUTE:
2204 case VM_PROT_ALL:
2205 /* None of these should happen in a sane system. */
2206 return;
2207 }
2208
2209 /*
2210 * If the pmap has no A table, it has no mappings and therefore
2211 * there is nothing to protect.
2212 */
2213 if ((a_tbl = pmap->pm_a_tmgr) == NULL)
2214 return;
2215
2216 a_idx = MMU_TIA(startva);
2217 b_idx = MMU_TIB(startva);
2218 c_idx = MMU_TIC(startva);
2219 b_tbl = NULL;
2220 c_tbl = NULL;
2221
2222 iscurpmap = (pmap == current_pmap());
2223 while (startva < endva) {
2224 if (b_tbl || MMU_VALID_DT(a_tbl->at_dtbl[a_idx])) {
2225 if (b_tbl == NULL) {
2226 b_tbl = (b_tmgr_t *) a_tbl->at_dtbl[a_idx].addr.raw;
2227 b_tbl = mmu_ptov((vaddr_t)b_tbl);
2228 b_tbl = mmuB2tmgr((mmu_short_dte_t *)b_tbl);
2229 }
2230 if (c_tbl || MMU_VALID_DT(b_tbl->bt_dtbl[b_idx])) {
2231 if (c_tbl == NULL) {
2232 c_tbl = (c_tmgr_t *) MMU_DTE_PA(b_tbl->bt_dtbl[b_idx]);
2233 c_tbl = mmu_ptov((vaddr_t)c_tbl);
2234 c_tbl = mmuC2tmgr((mmu_short_pte_t *)c_tbl);
2235 }
2236 if (MMU_VALID_DT(c_tbl->ct_dtbl[c_idx])) {
2237 pte = &c_tbl->ct_dtbl[c_idx];
2238 /* make the mapping read-only */
2239 pte->attr.raw |= MMU_SHORT_PTE_WP;
2240 /*
2241 * If we just modified the current address space,
2242 * flush any translations for the modified page from
2243 * the translation cache and any data from it in the
2244 * data cache.
2245 */
2246 if (iscurpmap)
2247 TBIS(startva);
2248 }
2249 startva += PAGE_SIZE;
2250
2251 if (++c_idx >= MMU_C_TBL_SIZE) { /* exceeded C table? */
2252 c_tbl = NULL;
2253 c_idx = 0;
2254 if (++b_idx >= MMU_B_TBL_SIZE) { /* exceeded B table? */
2255 b_tbl = NULL;
2256 b_idx = 0;
2257 }
2258 }
2259 } else { /* C table wasn't valid */
2260 c_tbl = NULL;
2261 c_idx = 0;
2262 startva += MMU_TIB_RANGE;
2263 if (++b_idx >= MMU_B_TBL_SIZE) { /* exceeded B table? */
2264 b_tbl = NULL;
2265 b_idx = 0;
2266 }
2267 } /* C table */
2268 } else { /* B table wasn't valid */
2269 b_tbl = NULL;
2270 b_idx = 0;
2271 startva += MMU_TIA_RANGE;
2272 a_idx++;
2273 } /* B table */
2274 }
2275 }
2276
2277 /* pmap_protect_kernel INTERNAL
2278 **
2279 * Apply the given protection code to a kernel address range.
2280 */
2281 void
2282 pmap_protect_kernel(vaddr_t startva, vaddr_t endva, vm_prot_t prot)
2283 {
2284 vaddr_t va;
2285 mmu_short_pte_t *pte;
2286
2287 pte = &kernCbase[(unsigned long) m68k_btop(startva - KERNBASE)];
2288 for (va = startva; va < endva; va += PAGE_SIZE, pte++) {
2289 if (MMU_VALID_DT(*pte)) {
2290 switch (prot) {
2291 case VM_PROT_ALL:
2292 break;
2293 case VM_PROT_EXECUTE:
2294 case VM_PROT_READ:
2295 case VM_PROT_READ|VM_PROT_EXECUTE:
2296 pte->attr.raw |= MMU_SHORT_PTE_WP;
2297 break;
2298 case VM_PROT_NONE:
2299 /* this is an alias for 'pmap_remove_kernel' */
2300 pmap_remove_pte(pte);
2301 break;
2302 default:
2303 break;
2304 }
2305 /*
2306 * since this is the kernel, immediately flush any cached
2307 * descriptors for this address.
2308 */
2309 TBIS(va);
2310 }
2311 }
2312 }
2313
2314 /* pmap_unwire INTERFACE
2315 **
2316 * Clear the wired attribute of the specified page.
2317 *
2318 * This function is called from vm_fault.c to unwire
2319 * a mapping.
2320 */
2321 void
2322 pmap_unwire(pmap_t pmap, vaddr_t va)
2323 {
2324 int a_idx, b_idx, c_idx;
2325 a_tmgr_t *a_tbl;
2326 b_tmgr_t *b_tbl;
2327 c_tmgr_t *c_tbl;
2328 mmu_short_pte_t *pte;
2329
2330 /* Kernel mappings always remain wired. */
2331 if (pmap == pmap_kernel())
2332 return;
2333
2334 /*
2335 * Walk through the tables. If the walk terminates without
2336 * a valid PTE then the address wasn't wired in the first place.
2337 * Return immediately.
2338 */
2339 if (pmap_stroll(pmap, va, &a_tbl, &b_tbl, &c_tbl, &pte, &a_idx,
2340 &b_idx, &c_idx) == FALSE)
2341 return;
2342
2343
2344 /* Is the PTE wired? If not, return. */
2345 if (!(pte->attr.raw & MMU_SHORT_PTE_WIRED))
2346 return;
2347
2348 /* Remove the wiring bit. */
2349 pte->attr.raw &= ~(MMU_SHORT_PTE_WIRED);
2350
2351 /*
2352 * Decrement the wired entry count in the C table.
2353 * If it reaches zero the following things happen:
2354 * 1. The table no longer has any wired entries and is considered
2355 * unwired.
2356 * 2. It is placed on the available queue.
2357 * 3. The parent table's wired entry count is decremented.
2358 * 4. If it reaches zero, this process repeats at step 1 and
2359 * stops at after reaching the A table.
2360 */
2361 if (--c_tbl->ct_wcnt == 0) {
2362 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
2363 if (--b_tbl->bt_wcnt == 0) {
2364 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
2365 if (--a_tbl->at_wcnt == 0) {
2366 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
2367 }
2368 }
2369 }
2370 }
2371
2372 /* pmap_copy INTERFACE
2373 **
2374 * Copy the mappings of a range of addresses in one pmap, into
2375 * the destination address of another.
2376 *
2377 * This routine is advisory. Should we one day decide that MMU tables
2378 * may be shared by more than one pmap, this function should be used to
2379 * link them together. Until that day however, we do nothing.
2380 */
2381 void
2382 pmap_copy(pmap_t pmap_a, pmap_t pmap_b, vaddr_t dst, vsize_t len, vaddr_t src)
2383 {
2384 /* not implemented. */
2385 }
2386
2387 /* pmap_copy_page INTERFACE
2388 **
2389 * Copy the contents of one physical page into another.
2390 *
2391 * This function makes use of two virtual pages allocated in pmap_bootstrap()
2392 * to map the two specified physical pages into the kernel address space.
2393 *
2394 * Note: We could use the transparent translation registers to make the
2395 * mappings. If we do so, be sure to disable interrupts before using them.
2396 */
2397 void
2398 pmap_copy_page(paddr_t srcpa, paddr_t dstpa)
2399 {
2400 vaddr_t srcva, dstva;
2401 int s;
2402
2403 srcva = tmp_vpages[0];
2404 dstva = tmp_vpages[1];
2405
2406 s = splvm();
2407 #ifdef DIAGNOSTIC
2408 if (tmp_vpages_inuse++)
2409 panic("pmap_copy_page: temporary vpages are in use.");
2410 #endif
2411
2412 /* Map pages as non-cacheable to avoid cache polution? */
2413 pmap_kenter_pa(srcva, srcpa, VM_PROT_READ);
2414 pmap_kenter_pa(dstva, dstpa, VM_PROT_READ|VM_PROT_WRITE);
2415
2416 /* Hand-optimized version of bcopy(src, dst, PAGE_SIZE) */
2417 copypage((char *) srcva, (char *) dstva);
2418
2419 pmap_kremove(srcva, PAGE_SIZE);
2420 pmap_kremove(dstva, PAGE_SIZE);
2421
2422 #ifdef DIAGNOSTIC
2423 --tmp_vpages_inuse;
2424 #endif
2425 splx(s);
2426 }
2427
2428 /* pmap_zero_page INTERFACE
2429 **
2430 * Zero the contents of the specified physical page.
2431 *
2432 * Uses one of the virtual pages allocated in pmap_boostrap()
2433 * to map the specified page into the kernel address space.
2434 */
2435 void
2436 pmap_zero_page(paddr_t dstpa)
2437 {
2438 vaddr_t dstva;
2439 int s;
2440
2441 dstva = tmp_vpages[1];
2442 s = splvm();
2443 #ifdef DIAGNOSTIC
2444 if (tmp_vpages_inuse++)
2445 panic("pmap_zero_page: temporary vpages are in use.");
2446 #endif
2447
2448 /* The comments in pmap_copy_page() above apply here also. */
2449 pmap_kenter_pa(dstva, dstpa, VM_PROT_READ|VM_PROT_WRITE);
2450
2451 /* Hand-optimized version of bzero(ptr, PAGE_SIZE) */
2452 zeropage((char *) dstva);
2453
2454 pmap_kremove(dstva, PAGE_SIZE);
2455 #ifdef DIAGNOSTIC
2456 --tmp_vpages_inuse;
2457 #endif
2458 splx(s);
2459 }
2460
2461 /* pmap_collect INTERFACE
2462 **
2463 * Called from the VM system when we are about to swap out
2464 * the process using this pmap. This should give up any
2465 * resources held here, including all its MMU tables.
2466 */
2467 void
2468 pmap_collect(pmap_t pmap)
2469 {
2470 /* XXX - todo... */
2471 }
2472
2473 /* pmap_create INTERFACE
2474 **
2475 * Create and return a pmap structure.
2476 */
2477 pmap_t
2478 pmap_create(void)
2479 {
2480 pmap_t pmap;
2481
2482 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK);
2483 pmap_pinit(pmap);
2484 return pmap;
2485 }
2486
2487 /* pmap_pinit INTERNAL
2488 **
2489 * Initialize a pmap structure.
2490 */
2491 void
2492 pmap_pinit(pmap_t pmap)
2493 {
2494 memset(pmap, 0, sizeof(struct pmap));
2495 pmap->pm_a_tmgr = NULL;
2496 pmap->pm_a_phys = kernAphys;
2497 pmap->pm_refcount = 1;
2498 simple_lock_init(&pmap->pm_lock);
2499 }
2500
2501 /* pmap_release INTERFACE
2502 **
2503 * Release any resources held by the given pmap.
2504 *
2505 * This is the reverse analog to pmap_pinit. It does not
2506 * necessarily mean for the pmap structure to be deallocated,
2507 * as in pmap_destroy.
2508 */
2509 void
2510 pmap_release(pmap_t pmap)
2511 {
2512 /*
2513 * As long as the pmap contains no mappings,
2514 * which always should be the case whenever
2515 * this function is called, there really should
2516 * be nothing to do.
2517 */
2518 #ifdef PMAP_DEBUG
2519 if (pmap == pmap_kernel())
2520 panic("pmap_release: kernel pmap");
2521 #endif
2522 /*
2523 * XXX - If this pmap has an A table, give it back.
2524 * The pmap SHOULD be empty by now, and pmap_remove
2525 * should have already given back the A table...
2526 * However, I see: pmap->pm_a_tmgr->at_ecnt == 1
2527 * at this point, which means some mapping was not
2528 * removed when it should have been. -gwr
2529 */
2530 if (pmap->pm_a_tmgr != NULL) {
2531 /* First make sure we are not using it! */
2532 if (kernel_crp.rp_addr == pmap->pm_a_phys) {
2533 kernel_crp.rp_addr = kernAphys;
2534 loadcrp(&kernel_crp);
2535 }
2536 #ifdef PMAP_DEBUG /* XXX - todo! */
2537 /* XXX - Now complain... */
2538 printf("pmap_release: still have table\n");
2539 Debugger();
2540 #endif
2541 free_a_table(pmap->pm_a_tmgr, TRUE);
2542 pmap->pm_a_tmgr = NULL;
2543 pmap->pm_a_phys = kernAphys;
2544 }
2545 }
2546
2547 /* pmap_reference INTERFACE
2548 **
2549 * Increment the reference count of a pmap.
2550 */
2551 void
2552 pmap_reference(pmap_t pmap)
2553 {
2554 pmap_lock(pmap);
2555 pmap_add_ref(pmap);
2556 pmap_unlock(pmap);
2557 }
2558
2559 /* pmap_dereference INTERNAL
2560 **
2561 * Decrease the reference count on the given pmap
2562 * by one and return the current count.
2563 */
2564 int
2565 pmap_dereference(pmap_t pmap)
2566 {
2567 int rtn;
2568
2569 pmap_lock(pmap);
2570 rtn = pmap_del_ref(pmap);
2571 pmap_unlock(pmap);
2572
2573 return rtn;
2574 }
2575
2576 /* pmap_destroy INTERFACE
2577 **
2578 * Decrement a pmap's reference count and delete
2579 * the pmap if it becomes zero. Will be called
2580 * only after all mappings have been removed.
2581 */
2582 void
2583 pmap_destroy(pmap_t pmap)
2584 {
2585 if (pmap_dereference(pmap) == 0) {
2586 pmap_release(pmap);
2587 pool_put(&pmap_pmap_pool, pmap);
2588 }
2589 }
2590
2591 /* pmap_is_referenced INTERFACE
2592 **
2593 * Determine if the given physical page has been
2594 * referenced (read from [or written to.])
2595 */
2596 boolean_t
2597 pmap_is_referenced(struct vm_page *pg)
2598 {
2599 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2600 pv_t *pv;
2601 int idx;
2602
2603 /*
2604 * Check the flags on the pv head. If they are set,
2605 * return immediately. Otherwise a search must be done.
2606 */
2607
2608 pv = pa2pv(pa);
2609 if (pv->pv_flags & PV_FLAGS_USED)
2610 return TRUE;
2611
2612 /*
2613 * Search through all pv elements pointing
2614 * to this page and query their reference bits
2615 */
2616
2617 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) {
2618 if (MMU_PTE_USED(kernCbase[idx])) {
2619 return TRUE;
2620 }
2621 }
2622 return FALSE;
2623 }
2624
2625 /* pmap_is_modified INTERFACE
2626 **
2627 * Determine if the given physical page has been
2628 * modified (written to.)
2629 */
2630 boolean_t
2631 pmap_is_modified(struct vm_page *pg)
2632 {
2633 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2634 pv_t *pv;
2635 int idx;
2636
2637 /* see comments in pmap_is_referenced() */
2638 pv = pa2pv(pa);
2639 if (pv->pv_flags & PV_FLAGS_MDFY)
2640 return TRUE;
2641
2642 for (idx = pv->pv_idx;
2643 idx != PVE_EOL;
2644 idx = pvebase[idx].pve_next) {
2645
2646 if (MMU_PTE_MODIFIED(kernCbase[idx])) {
2647 return TRUE;
2648 }
2649 }
2650
2651 return FALSE;
2652 }
2653
2654 /* pmap_page_protect INTERFACE
2655 **
2656 * Applies the given protection to all mappings to the given
2657 * physical page.
2658 */
2659 void
2660 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
2661 {
2662 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2663 pv_t *pv;
2664 int idx;
2665 vaddr_t va;
2666 struct mmu_short_pte_struct *pte;
2667 c_tmgr_t *c_tbl;
2668 pmap_t pmap, curpmap;
2669
2670 curpmap = current_pmap();
2671 pv = pa2pv(pa);
2672
2673 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) {
2674 pte = &kernCbase[idx];
2675 switch (prot) {
2676 case VM_PROT_ALL:
2677 /* do nothing */
2678 break;
2679 case VM_PROT_EXECUTE:
2680 case VM_PROT_READ:
2681 case VM_PROT_READ|VM_PROT_EXECUTE:
2682 /*
2683 * Determine the virtual address mapped by
2684 * the PTE and flush ATC entries if necessary.
2685 */
2686 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
2687 pte->attr.raw |= MMU_SHORT_PTE_WP;
2688 if (pmap == curpmap || pmap == pmap_kernel())
2689 TBIS(va);
2690 break;
2691 case VM_PROT_NONE:
2692 /* Save the mod/ref bits. */
2693 pv->pv_flags |= pte->attr.raw;
2694 /* Invalidate the PTE. */
2695 pte->attr.raw = MMU_DT_INVALID;
2696
2697 /*
2698 * Update table counts. And flush ATC entries
2699 * if necessary.
2700 */
2701 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
2702
2703 /*
2704 * If the PTE belongs to the kernel map,
2705 * be sure to flush the page it maps.
2706 */
2707 if (pmap == pmap_kernel()) {
2708 TBIS(va);
2709 } else {
2710 /*
2711 * The PTE belongs to a user map.
2712 * update the entry count in the C
2713 * table to which it belongs and flush
2714 * the ATC if the mapping belongs to
2715 * the current pmap.
2716 */
2717 c_tbl->ct_ecnt--;
2718 if (pmap == curpmap)
2719 TBIS(va);
2720 }
2721 break;
2722 default:
2723 break;
2724 }
2725 }
2726
2727 /*
2728 * If the protection code indicates that all mappings to the page
2729 * be removed, truncate the PV list to zero entries.
2730 */
2731 if (prot == VM_PROT_NONE)
2732 pv->pv_idx = PVE_EOL;
2733 }
2734
2735 /* pmap_get_pteinfo INTERNAL
2736 **
2737 * Called internally to find the pmap and virtual address within that
2738 * map to which the pte at the given index maps. Also includes the PTE's C
2739 * table manager.
2740 *
2741 * Returns the pmap in the argument provided, and the virtual address
2742 * by return value.
2743 */
2744 vaddr_t
2745 pmap_get_pteinfo(u_int idx, pmap_t *pmap, c_tmgr_t **tbl)
2746 {
2747 vaddr_t va = 0;
2748
2749 /*
2750 * Determine if the PTE is a kernel PTE or a user PTE.
2751 */
2752 if (idx >= NUM_KERN_PTES) {
2753 /*
2754 * The PTE belongs to a user mapping.
2755 */
2756 /* XXX: Would like an inline for this to validate idx... */
2757 *tbl = &Ctmgrbase[(idx - NUM_KERN_PTES) / MMU_C_TBL_SIZE];
2758
2759 *pmap = (*tbl)->ct_pmap;
2760 /*
2761 * To find the va to which the PTE maps, we first take
2762 * the table's base virtual address mapping which is stored
2763 * in ct_va. We then increment this address by a page for
2764 * every slot skipped until we reach the PTE.
2765 */
2766 va = (*tbl)->ct_va;
2767 va += m68k_ptob(idx % MMU_C_TBL_SIZE);
2768 } else {
2769 /*
2770 * The PTE belongs to the kernel map.
2771 */
2772 *pmap = pmap_kernel();
2773
2774 va = m68k_ptob(idx);
2775 va += KERNBASE;
2776 }
2777
2778 return va;
2779 }
2780
2781 /* pmap_clear_modify INTERFACE
2782 **
2783 * Clear the modification bit on the page at the specified
2784 * physical address.
2785 *
2786 */
2787 boolean_t
2788 pmap_clear_modify(struct vm_page *pg)
2789 {
2790 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2791 boolean_t rv;
2792
2793 rv = pmap_is_modified(pg);
2794 pmap_clear_pv(pa, PV_FLAGS_MDFY);
2795 return rv;
2796 }
2797
2798 /* pmap_clear_reference INTERFACE
2799 **
2800 * Clear the referenced bit on the page at the specified
2801 * physical address.
2802 */
2803 boolean_t
2804 pmap_clear_reference(struct vm_page *pg)
2805 {
2806 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2807 boolean_t rv;
2808
2809 rv = pmap_is_referenced(pg);
2810 pmap_clear_pv(pa, PV_FLAGS_USED);
2811 return rv;
2812 }
2813
2814 /* pmap_clear_pv INTERNAL
2815 **
2816 * Clears the specified flag from the specified physical address.
2817 * (Used by pmap_clear_modify() and pmap_clear_reference().)
2818 *
2819 * Flag is one of:
2820 * PV_FLAGS_MDFY - Page modified bit.
2821 * PV_FLAGS_USED - Page used (referenced) bit.
2822 *
2823 * This routine must not only clear the flag on the pv list
2824 * head. It must also clear the bit on every pte in the pv
2825 * list associated with the address.
2826 */
2827 void
2828 pmap_clear_pv(paddr_t pa, int flag)
2829 {
2830 pv_t *pv;
2831 int idx;
2832 vaddr_t va;
2833 pmap_t pmap;
2834 mmu_short_pte_t *pte;
2835 c_tmgr_t *c_tbl;
2836
2837 pv = pa2pv(pa);
2838 pv->pv_flags &= ~(flag);
2839 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) {
2840 pte = &kernCbase[idx];
2841 pte->attr.raw &= ~(flag);
2842
2843 /*
2844 * The MC68030 MMU will not set the modified or
2845 * referenced bits on any MMU tables for which it has
2846 * a cached descriptor with its modify bit set. To insure
2847 * that it will modify these bits on the PTE during the next
2848 * time it is written to or read from, we must flush it from
2849 * the ATC.
2850 *
2851 * Ordinarily it is only necessary to flush the descriptor
2852 * if it is used in the current address space. But since I
2853 * am not sure that there will always be a notion of
2854 * 'the current address space' when this function is called,
2855 * I will skip the test and always flush the address. It
2856 * does no harm.
2857 */
2858
2859 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
2860 TBIS(va);
2861 }
2862 }
2863
2864 /* pmap_extract INTERFACE
2865 **
2866 * Return the physical address mapped by the virtual address
2867 * in the specified pmap.
2868 *
2869 * Note: this function should also apply an exclusive lock
2870 * on the pmap system during its duration.
2871 */
2872 boolean_t
2873 pmap_extract(pmap_t pmap, vaddr_t va, paddr_t *pap)
2874 {
2875 int a_idx, b_idx, pte_idx;
2876 a_tmgr_t *a_tbl;
2877 b_tmgr_t *b_tbl;
2878 c_tmgr_t *c_tbl;
2879 mmu_short_pte_t *c_pte;
2880
2881 if (pmap == pmap_kernel())
2882 return pmap_extract_kernel(va, pap);
2883
2884 if (pmap_stroll(pmap, va, &a_tbl, &b_tbl, &c_tbl,
2885 &c_pte, &a_idx, &b_idx, &pte_idx) == FALSE)
2886 return FALSE;
2887
2888 if (!MMU_VALID_DT(*c_pte))
2889 return FALSE;
2890
2891 if (pap != NULL)
2892 *pap = MMU_PTE_PA(*c_pte);
2893 return (TRUE);
2894 }
2895
2896 /* pmap_extract_kernel INTERNAL
2897 **
2898 * Extract a translation from the kernel address space.
2899 */
2900 boolean_t
2901 pmap_extract_kernel(vaddr_t va, paddr_t *pap)
2902 {
2903 mmu_short_pte_t *pte;
2904
2905 pte = &kernCbase[(u_int) m68k_btop(va - KERNBASE)];
2906 if (!MMU_VALID_DT(*pte))
2907 return (FALSE);
2908 if (pap != NULL)
2909 *pap = MMU_PTE_PA(*pte);
2910 return (TRUE);
2911 }
2912
2913 /* pmap_remove_kernel INTERNAL
2914 **
2915 * Remove the mapping of a range of virtual addresses from the kernel map.
2916 * The arguments are already page-aligned.
2917 */
2918 void
2919 pmap_remove_kernel(vaddr_t sva, vaddr_t eva)
2920 {
2921 int idx, eidx;
2922
2923 #ifdef PMAP_DEBUG
2924 if ((sva & PGOFSET) || (eva & PGOFSET))
2925 panic("pmap_remove_kernel: alignment");
2926 #endif
2927
2928 idx = m68k_btop(sva - KERNBASE);
2929 eidx = m68k_btop(eva - KERNBASE);
2930
2931 while (idx < eidx) {
2932 pmap_remove_pte(&kernCbase[idx++]);
2933 TBIS(sva);
2934 sva += PAGE_SIZE;
2935 }
2936 }
2937
2938 /* pmap_remove INTERFACE
2939 **
2940 * Remove the mapping of a range of virtual addresses from the given pmap.
2941 *
2942 */
2943 void
2944 pmap_remove(pmap_t pmap, vaddr_t sva, vaddr_t eva)
2945 {
2946
2947 if (pmap == pmap_kernel()) {
2948 pmap_remove_kernel(sva, eva);
2949 return;
2950 }
2951
2952 /*
2953 * If the pmap doesn't have an A table of its own, it has no mappings
2954 * that can be removed.
2955 */
2956 if (pmap->pm_a_tmgr == NULL)
2957 return;
2958
2959 /*
2960 * Remove the specified range from the pmap. If the function
2961 * returns true, the operation removed all the valid mappings
2962 * in the pmap and freed its A table. If this happened to the
2963 * currently loaded pmap, the MMU root pointer must be reloaded
2964 * with the default 'kernel' map.
2965 */
2966 if (pmap_remove_a(pmap->pm_a_tmgr, sva, eva)) {
2967 if (kernel_crp.rp_addr == pmap->pm_a_phys) {
2968 kernel_crp.rp_addr = kernAphys;
2969 loadcrp(&kernel_crp);
2970 /* will do TLB flush below */
2971 }
2972 pmap->pm_a_tmgr = NULL;
2973 pmap->pm_a_phys = kernAphys;
2974 }
2975
2976 /*
2977 * If we just modified the current address space,
2978 * make sure to flush the MMU cache.
2979 *
2980 * XXX - this could be an unecessarily large flush.
2981 * XXX - Could decide, based on the size of the VA range
2982 * to be removed, whether to flush "by pages" or "all".
2983 */
2984 if (pmap == current_pmap())
2985 TBIAU();
2986 }
2987
2988 /* pmap_remove_a INTERNAL
2989 **
2990 * This is function number one in a set of three that removes a range
2991 * of memory in the most efficient manner by removing the highest possible
2992 * tables from the memory space. This particular function attempts to remove
2993 * as many B tables as it can, delegating the remaining fragmented ranges to
2994 * pmap_remove_b().
2995 *
2996 * If the removal operation results in an empty A table, the function returns
2997 * TRUE.
2998 *
2999 * It's ugly but will do for now.
3000 */
3001 boolean_t
3002 pmap_remove_a(a_tmgr_t *a_tbl, vaddr_t sva, vaddr_t eva)
3003 {
3004 boolean_t empty;
3005 int idx;
3006 vaddr_t nstart, nend;
3007 b_tmgr_t *b_tbl;
3008 mmu_long_dte_t *a_dte;
3009 mmu_short_dte_t *b_dte;
3010 uint8_t at_wired, bt_wired;
3011
3012 /*
3013 * The following code works with what I call a 'granularity
3014 * reduction algorithim'. A range of addresses will always have
3015 * the following properties, which are classified according to
3016 * how the range relates to the size of the current granularity
3017 * - an A table entry:
3018 *
3019 * 1 2 3 4
3020 * -+---+---+---+---+---+---+---+-
3021 * -+---+---+---+---+---+---+---+-
3022 *
3023 * A range will always start on a granularity boundary, illustrated
3024 * by '+' signs in the table above, or it will start at some point
3025 * inbetween a granularity boundary, as illustrated by point 1.
3026 * The first step in removing a range of addresses is to remove the
3027 * range between 1 and 2, the nearest granularity boundary. This
3028 * job is handled by the section of code governed by the
3029 * 'if (start < nstart)' statement.
3030 *
3031 * A range will always encompass zero or more intergral granules,
3032 * illustrated by points 2 and 3. Integral granules are easy to
3033 * remove. The removal of these granules is the second step, and
3034 * is handled by the code block 'if (nstart < nend)'.
3035 *
3036 * Lastly, a range will always end on a granularity boundary,
3037 * ill. by point 3, or it will fall just beyond one, ill. by point
3038 * 4. The last step involves removing this range and is handled by
3039 * the code block 'if (nend < end)'.
3040 */
3041 nstart = MMU_ROUND_UP_A(sva);
3042 nend = MMU_ROUND_A(eva);
3043
3044 at_wired = a_tbl->at_wcnt;
3045
3046 if (sva < nstart) {
3047 /*
3048 * This block is executed if the range starts between
3049 * a granularity boundary.
3050 *
3051 * First find the DTE which is responsible for mapping
3052 * the start of the range.
3053 */
3054 idx = MMU_TIA(sva);
3055 a_dte = &a_tbl->at_dtbl[idx];
3056
3057 /*
3058 * If the DTE is valid then delegate the removal of the sub
3059 * range to pmap_remove_b(), which can remove addresses at
3060 * a finer granularity.
3061 */
3062 if (MMU_VALID_DT(*a_dte)) {
3063 b_dte = mmu_ptov(a_dte->addr.raw);
3064 b_tbl = mmuB2tmgr(b_dte);
3065 bt_wired = b_tbl->bt_wcnt;
3066
3067 /*
3068 * The sub range to be removed starts at the start
3069 * of the full range we were asked to remove, and ends
3070 * at the greater of:
3071 * 1. The end of the full range, -or-
3072 * 2. The end of the full range, rounded down to the
3073 * nearest granularity boundary.
3074 */
3075 if (eva < nstart)
3076 empty = pmap_remove_b(b_tbl, sva, eva);
3077 else
3078 empty = pmap_remove_b(b_tbl, sva, nstart);
3079
3080 /*
3081 * If the child table no longer has wired entries,
3082 * decrement wired entry count.
3083 */
3084 if (bt_wired && b_tbl->bt_wcnt == 0)
3085 a_tbl->at_wcnt--;
3086
3087 /*
3088 * If the removal resulted in an empty B table,
3089 * invalidate the DTE that points to it and decrement
3090 * the valid entry count of the A table.
3091 */
3092 if (empty) {
3093 a_dte->attr.raw = MMU_DT_INVALID;
3094 a_tbl->at_ecnt--;
3095 }
3096 }
3097 /*
3098 * If the DTE is invalid, the address range is already non-
3099 * existent and can simply be skipped.
3100 */
3101 }
3102 if (nstart < nend) {
3103 /*
3104 * This block is executed if the range spans a whole number
3105 * multiple of granules (A table entries.)
3106 *
3107 * First find the DTE which is responsible for mapping
3108 * the start of the first granule involved.
3109 */
3110 idx = MMU_TIA(nstart);
3111 a_dte = &a_tbl->at_dtbl[idx];
3112
3113 /*
3114 * Remove entire sub-granules (B tables) one at a time,
3115 * until reaching the end of the range.
3116 */
3117 for (; nstart < nend; a_dte++, nstart += MMU_TIA_RANGE)
3118 if (MMU_VALID_DT(*a_dte)) {
3119 /*
3120 * Find the B table manager for the
3121 * entry and free it.
3122 */
3123 b_dte = mmu_ptov(a_dte->addr.raw);
3124 b_tbl = mmuB2tmgr(b_dte);
3125 bt_wired = b_tbl->bt_wcnt;
3126
3127 free_b_table(b_tbl, TRUE);
3128
3129 /*
3130 * All child entries has been removed.
3131 * If there were any wired entries in it,
3132 * decrement wired entry count.
3133 */
3134 if (bt_wired)
3135 a_tbl->at_wcnt--;
3136
3137 /*
3138 * Invalidate the DTE that points to the
3139 * B table and decrement the valid entry
3140 * count of the A table.
3141 */
3142 a_dte->attr.raw = MMU_DT_INVALID;
3143 a_tbl->at_ecnt--;
3144 }
3145 }
3146 if (nend < eva) {
3147 /*
3148 * This block is executed if the range ends beyond a
3149 * granularity boundary.
3150 *
3151 * First find the DTE which is responsible for mapping
3152 * the start of the nearest (rounded down) granularity
3153 * boundary.
3154 */
3155 idx = MMU_TIA(nend);
3156 a_dte = &a_tbl->at_dtbl[idx];
3157
3158 /*
3159 * If the DTE is valid then delegate the removal of the sub
3160 * range to pmap_remove_b(), which can remove addresses at
3161 * a finer granularity.
3162 */
3163 if (MMU_VALID_DT(*a_dte)) {
3164 /*
3165 * Find the B table manager for the entry
3166 * and hand it to pmap_remove_b() along with
3167 * the sub range.
3168 */
3169 b_dte = mmu_ptov(a_dte->addr.raw);
3170 b_tbl = mmuB2tmgr(b_dte);
3171 bt_wired = b_tbl->bt_wcnt;
3172
3173 empty = pmap_remove_b(b_tbl, nend, eva);
3174
3175 /*
3176 * If the child table no longer has wired entries,
3177 * decrement wired entry count.
3178 */
3179 if (bt_wired && b_tbl->bt_wcnt == 0)
3180 a_tbl->at_wcnt--;
3181 /*
3182 * If the removal resulted in an empty B table,
3183 * invalidate the DTE that points to it and decrement
3184 * the valid entry count of the A table.
3185 */
3186 if (empty) {
3187 a_dte->attr.raw = MMU_DT_INVALID;
3188 a_tbl->at_ecnt--;
3189 }
3190 }
3191 }
3192
3193 /*
3194 * If there are no more entries in the A table, release it
3195 * back to the available pool and return TRUE.
3196 */
3197 if (a_tbl->at_ecnt == 0) {
3198 KASSERT(a_tbl->at_wcnt == 0);
3199 a_tbl->at_parent = NULL;
3200 if (!at_wired)
3201 TAILQ_REMOVE(&a_pool, a_tbl, at_link);
3202 TAILQ_INSERT_HEAD(&a_pool, a_tbl, at_link);
3203 empty = TRUE;
3204 } else {
3205 /*
3206 * If the table doesn't have wired entries any longer
3207 * but still has unwired entries, put it back into
3208 * the available queue.
3209 */
3210 if (at_wired && a_tbl->at_wcnt == 0)
3211 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
3212 empty = FALSE;
3213 }
3214
3215 return empty;
3216 }
3217
3218 /* pmap_remove_b INTERNAL
3219 **
3220 * Remove a range of addresses from an address space, trying to remove entire
3221 * C tables if possible.
3222 *
3223 * If the operation results in an empty B table, the function returns TRUE.
3224 */
3225 boolean_t
3226 pmap_remove_b(b_tmgr_t *b_tbl, vaddr_t sva, vaddr_t eva)
3227 {
3228 boolean_t empty;
3229 int idx;
3230 vaddr_t nstart, nend, rstart;
3231 c_tmgr_t *c_tbl;
3232 mmu_short_dte_t *b_dte;
3233 mmu_short_pte_t *c_dte;
3234 uint8_t bt_wired, ct_wired;
3235
3236 nstart = MMU_ROUND_UP_B(sva);
3237 nend = MMU_ROUND_B(eva);
3238
3239 bt_wired = b_tbl->bt_wcnt;
3240
3241 if (sva < nstart) {
3242 idx = MMU_TIB(sva);
3243 b_dte = &b_tbl->bt_dtbl[idx];
3244 if (MMU_VALID_DT(*b_dte)) {
3245 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte));
3246 c_tbl = mmuC2tmgr(c_dte);
3247 ct_wired = c_tbl->ct_wcnt;
3248
3249 if (eva < nstart)
3250 empty = pmap_remove_c(c_tbl, sva, eva);
3251 else
3252 empty = pmap_remove_c(c_tbl, sva, nstart);
3253
3254 /*
3255 * If the child table no longer has wired entries,
3256 * decrement wired entry count.
3257 */
3258 if (ct_wired && c_tbl->ct_wcnt == 0)
3259 b_tbl->bt_wcnt--;
3260
3261 if (empty) {
3262 b_dte->attr.raw = MMU_DT_INVALID;
3263 b_tbl->bt_ecnt--;
3264 }
3265 }
3266 }
3267 if (nstart < nend) {
3268 idx = MMU_TIB(nstart);
3269 b_dte = &b_tbl->bt_dtbl[idx];
3270 rstart = nstart;
3271 while (rstart < nend) {
3272 if (MMU_VALID_DT(*b_dte)) {
3273 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte));
3274 c_tbl = mmuC2tmgr(c_dte);
3275 ct_wired = c_tbl->ct_wcnt;
3276
3277 free_c_table(c_tbl, TRUE);
3278
3279 /*
3280 * All child entries has been removed.
3281 * If there were any wired entries in it,
3282 * decrement wired entry count.
3283 */
3284 if (ct_wired)
3285 b_tbl->bt_wcnt--;
3286
3287 b_dte->attr.raw = MMU_DT_INVALID;
3288 b_tbl->bt_ecnt--;
3289 }
3290 b_dte++;
3291 rstart += MMU_TIB_RANGE;
3292 }
3293 }
3294 if (nend < eva) {
3295 idx = MMU_TIB(nend);
3296 b_dte = &b_tbl->bt_dtbl[idx];
3297 if (MMU_VALID_DT(*b_dte)) {
3298 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte));
3299 c_tbl = mmuC2tmgr(c_dte);
3300 ct_wired = c_tbl->ct_wcnt;
3301 empty = pmap_remove_c(c_tbl, nend, eva);
3302
3303 /*
3304 * If the child table no longer has wired entries,
3305 * decrement wired entry count.
3306 */
3307 if (ct_wired && c_tbl->ct_wcnt == 0)
3308 b_tbl->bt_wcnt--;
3309
3310 if (empty) {
3311 b_dte->attr.raw = MMU_DT_INVALID;
3312 b_tbl->bt_ecnt--;
3313 }
3314 }
3315 }
3316
3317 if (b_tbl->bt_ecnt == 0) {
3318 KASSERT(b_tbl->bt_wcnt == 0);
3319 b_tbl->bt_parent = NULL;
3320 if (!bt_wired)
3321 TAILQ_REMOVE(&b_pool, b_tbl, bt_link);
3322 TAILQ_INSERT_HEAD(&b_pool, b_tbl, bt_link);
3323 empty = TRUE;
3324 } else {
3325 /*
3326 * If the table doesn't have wired entries any longer
3327 * but still has unwired entries, put it back into
3328 * the available queue.
3329 */
3330 if (bt_wired && b_tbl->bt_wcnt == 0)
3331 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
3332
3333 empty = FALSE;
3334 }
3335
3336 return empty;
3337 }
3338
3339 /* pmap_remove_c INTERNAL
3340 **
3341 * Remove a range of addresses from the given C table.
3342 */
3343 boolean_t
3344 pmap_remove_c(c_tmgr_t *c_tbl, vaddr_t sva, vaddr_t eva)
3345 {
3346 boolean_t empty;
3347 int idx;
3348 mmu_short_pte_t *c_pte;
3349 uint8_t ct_wired;
3350
3351 ct_wired = c_tbl->ct_wcnt;
3352
3353 idx = MMU_TIC(sva);
3354 c_pte = &c_tbl->ct_dtbl[idx];
3355 for (;sva < eva; sva += MMU_PAGE_SIZE, c_pte++) {
3356 if (MMU_VALID_DT(*c_pte)) {
3357 if (c_pte->attr.raw & MMU_SHORT_PTE_WIRED)
3358 c_tbl->ct_wcnt--;
3359 pmap_remove_pte(c_pte);
3360 c_tbl->ct_ecnt--;
3361 }
3362 }
3363
3364 if (c_tbl->ct_ecnt == 0) {
3365 KASSERT(c_tbl->ct_wcnt == 0);
3366 c_tbl->ct_parent = NULL;
3367 if (!ct_wired)
3368 TAILQ_REMOVE(&c_pool, c_tbl, ct_link);
3369 TAILQ_INSERT_HEAD(&c_pool, c_tbl, ct_link);
3370 empty = TRUE;
3371 } else {
3372 /*
3373 * If the table doesn't have wired entries any longer
3374 * but still has unwired entries, put it back into
3375 * the available queue.
3376 */
3377 if (ct_wired && c_tbl->ct_wcnt == 0)
3378 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
3379 empty = FALSE;
3380 }
3381
3382 return empty;
3383 }
3384
3385 /* is_managed INTERNAL
3386 **
3387 * Determine if the given physical address is managed by the PV system.
3388 * Note that this logic assumes that no one will ask for the status of
3389 * addresses which lie in-between the memory banks on the 3/80. If they
3390 * do so, it will falsely report that it is managed.
3391 *
3392 * Note: A "managed" address is one that was reported to the VM system as
3393 * a "usable page" during system startup. As such, the VM system expects the
3394 * pmap module to keep an accurate track of the useage of those pages.
3395 * Any page not given to the VM system at startup does not exist (as far as
3396 * the VM system is concerned) and is therefore "unmanaged." Examples are
3397 * those pages which belong to the ROM monitor and the memory allocated before
3398 * the VM system was started.
3399 */
3400 boolean_t
3401 is_managed(paddr_t pa)
3402 {
3403 if (pa >= avail_start && pa < avail_end)
3404 return TRUE;
3405 else
3406 return FALSE;
3407 }
3408
3409 /* pmap_bootstrap_alloc INTERNAL
3410 **
3411 * Used internally for memory allocation at startup when malloc is not
3412 * available. This code will fail once it crosses the first memory
3413 * bank boundary on the 3/80. Hopefully by then however, the VM system
3414 * will be in charge of allocation.
3415 */
3416 void *
3417 pmap_bootstrap_alloc(int size)
3418 {
3419 void *rtn;
3420
3421 #ifdef PMAP_DEBUG
3422 if (bootstrap_alloc_enabled == FALSE) {
3423 mon_printf("pmap_bootstrap_alloc: disabled\n");
3424 sunmon_abort();
3425 }
3426 #endif
3427
3428 rtn = (void *) virtual_avail;
3429 virtual_avail += size;
3430
3431 #ifdef PMAP_DEBUG
3432 if (virtual_avail > virtual_contig_end) {
3433 mon_printf("pmap_bootstrap_alloc: out of mem\n");
3434 sunmon_abort();
3435 }
3436 #endif
3437
3438 return rtn;
3439 }
3440
3441 /* pmap_bootstap_aalign INTERNAL
3442 **
3443 * Used to insure that the next call to pmap_bootstrap_alloc() will
3444 * return a chunk of memory aligned to the specified size.
3445 *
3446 * Note: This function will only support alignment sizes that are powers
3447 * of two.
3448 */
3449 void
3450 pmap_bootstrap_aalign(int size)
3451 {
3452 int off;
3453
3454 off = virtual_avail & (size - 1);
3455 if (off) {
3456 (void) pmap_bootstrap_alloc(size - off);
3457 }
3458 }
3459
3460 /* pmap_pa_exists
3461 **
3462 * Used by the /dev/mem driver to see if a given PA is memory
3463 * that can be mapped. (The PA is not in a hole.)
3464 */
3465 int
3466 pmap_pa_exists(paddr_t pa)
3467 {
3468 int i;
3469
3470 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
3471 if ((pa >= avail_mem[i].pmem_start) &&
3472 (pa < avail_mem[i].pmem_end))
3473 return (1);
3474 if (avail_mem[i].pmem_next == NULL)
3475 break;
3476 }
3477 return (0);
3478 }
3479
3480 /* Called only from locore.s and pmap.c */
3481 void _pmap_switch(pmap_t pmap);
3482
3483 /*
3484 * _pmap_switch INTERNAL
3485 *
3486 * This is called by locore.s:cpu_switch() when it is
3487 * switching to a new process. Load new translations.
3488 * Note: done in-line by locore.s unless PMAP_DEBUG
3489 *
3490 * Note that we do NOT allocate a context here, but
3491 * share the "kernel only" context until we really
3492 * need our own context for user-space mappings in
3493 * pmap_enter_user(). [ s/context/mmu A table/ ]
3494 */
3495 void
3496 _pmap_switch(pmap_t pmap)
3497 {
3498 u_long rootpa;
3499
3500 /*
3501 * Only do reload/flush if we have to.
3502 * Note that if the old and new process
3503 * were BOTH using the "null" context,
3504 * then this will NOT flush the TLB.
3505 */
3506 rootpa = pmap->pm_a_phys;
3507 if (kernel_crp.rp_addr != rootpa) {
3508 DPRINT(("pmap_activate(%p)\n", pmap));
3509 kernel_crp.rp_addr = rootpa;
3510 loadcrp(&kernel_crp);
3511 TBIAU();
3512 }
3513 }
3514
3515 /*
3516 * Exported version of pmap_activate(). This is called from the
3517 * machine-independent VM code when a process is given a new pmap.
3518 * If (p == curlwp) do like cpu_switch would do; otherwise just
3519 * take this as notification that the process has a new pmap.
3520 */
3521 void
3522 pmap_activate(struct lwp *l)
3523 {
3524 if (l->l_proc == curproc) {
3525 _pmap_switch(l->l_proc->p_vmspace->vm_map.pmap);
3526 }
3527 }
3528
3529 /*
3530 * pmap_deactivate INTERFACE
3531 **
3532 * This is called to deactivate the specified process's address space.
3533 */
3534 void
3535 pmap_deactivate(struct lwp *l)
3536 {
3537 /* Nothing to do. */
3538 }
3539
3540 /*
3541 * Fill in the sun3x-specific part of the kernel core header
3542 * for dumpsys(). (See machdep.c for the rest.)
3543 */
3544 void
3545 pmap_kcore_hdr(struct sun3x_kcore_hdr *sh)
3546 {
3547 u_long spa, len;
3548 int i;
3549
3550 sh->pg_frame = MMU_SHORT_PTE_BASEADDR;
3551 sh->pg_valid = MMU_DT_PAGE;
3552 sh->contig_end = virtual_contig_end;
3553 sh->kernCbase = (u_long)kernCbase;
3554 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
3555 spa = avail_mem[i].pmem_start;
3556 spa = m68k_trunc_page(spa);
3557 len = avail_mem[i].pmem_end - spa;
3558 len = m68k_round_page(len);
3559 sh->ram_segs[i].start = spa;
3560 sh->ram_segs[i].size = len;
3561 }
3562 }
3563
3564
3565 /* pmap_virtual_space INTERFACE
3566 **
3567 * Return the current available range of virtual addresses in the
3568 * arguuments provided. Only really called once.
3569 */
3570 void
3571 pmap_virtual_space(vaddr_t *vstart, vaddr_t *vend)
3572 {
3573 *vstart = virtual_avail;
3574 *vend = virtual_end;
3575 }
3576
3577 /*
3578 * Provide memory to the VM system.
3579 *
3580 * Assume avail_start is always in the
3581 * first segment as pmap_bootstrap does.
3582 */
3583 static void
3584 pmap_page_upload(void)
3585 {
3586 paddr_t a, b; /* memory range */
3587 int i;
3588
3589 /* Supply the memory in segments. */
3590 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
3591 a = atop(avail_mem[i].pmem_start);
3592 b = atop(avail_mem[i].pmem_end);
3593 if (i == 0)
3594 a = atop(avail_start);
3595 if (avail_mem[i].pmem_end > avail_end)
3596 b = atop(avail_end);
3597
3598 uvm_page_physload(a, b, a, b, VM_FREELIST_DEFAULT);
3599
3600 if (avail_mem[i].pmem_next == NULL)
3601 break;
3602 }
3603 }
3604
3605 /* pmap_count INTERFACE
3606 **
3607 * Return the number of resident (valid) pages in the given pmap.
3608 *
3609 * Note: If this function is handed the kernel map, it will report
3610 * that it has no mappings. Hopefully the VM system won't ask for kernel
3611 * map statistics.
3612 */
3613 segsz_t
3614 pmap_count(pmap_t pmap, int type)
3615 {
3616 u_int count;
3617 int a_idx, b_idx;
3618 a_tmgr_t *a_tbl;
3619 b_tmgr_t *b_tbl;
3620 c_tmgr_t *c_tbl;
3621
3622 /*
3623 * If the pmap does not have its own A table manager, it has no
3624 * valid entires.
3625 */
3626 if (pmap->pm_a_tmgr == NULL)
3627 return 0;
3628
3629 a_tbl = pmap->pm_a_tmgr;
3630
3631 count = 0;
3632 for (a_idx = 0; a_idx < MMU_TIA(KERNBASE); a_idx++) {
3633 if (MMU_VALID_DT(a_tbl->at_dtbl[a_idx])) {
3634 b_tbl = mmuB2tmgr(mmu_ptov(a_tbl->at_dtbl[a_idx].addr.raw));
3635 for (b_idx = 0; b_idx < MMU_B_TBL_SIZE; b_idx++) {
3636 if (MMU_VALID_DT(b_tbl->bt_dtbl[b_idx])) {
3637 c_tbl = mmuC2tmgr(
3638 mmu_ptov(MMU_DTE_PA(b_tbl->bt_dtbl[b_idx])));
3639 if (type == 0)
3640 /*
3641 * A resident entry count has been requested.
3642 */
3643 count += c_tbl->ct_ecnt;
3644 else
3645 /*
3646 * A wired entry count has been requested.
3647 */
3648 count += c_tbl->ct_wcnt;
3649 }
3650 }
3651 }
3652 }
3653
3654 return count;
3655 }
3656
3657 /************************ SUN3 COMPATIBILITY ROUTINES ********************
3658 * The following routines are only used by DDB for tricky kernel text *
3659 * text operations in db_memrw.c. They are provided for sun3 *
3660 * compatibility. *
3661 *************************************************************************/
3662 /* get_pte INTERNAL
3663 **
3664 * Return the page descriptor the describes the kernel mapping
3665 * of the given virtual address.
3666 */
3667 extern u_long ptest_addr(u_long); /* XXX: locore.s */
3668 u_int
3669 get_pte(vaddr_t va)
3670 {
3671 u_long pte_pa;
3672 mmu_short_pte_t *pte;
3673
3674 /* Get the physical address of the PTE */
3675 pte_pa = ptest_addr(va & ~PGOFSET);
3676
3677 /* Convert to a virtual address... */
3678 pte = (mmu_short_pte_t *) (KERNBASE + pte_pa);
3679
3680 /* Make sure it is in our level-C tables... */
3681 if ((pte < kernCbase) ||
3682 (pte >= &mmuCbase[NUM_USER_PTES]))
3683 return 0;
3684
3685 /* ... and just return its contents. */
3686 return (pte->attr.raw);
3687 }
3688
3689
3690 /* set_pte INTERNAL
3691 **
3692 * Set the page descriptor that describes the kernel mapping
3693 * of the given virtual address.
3694 */
3695 void
3696 set_pte(vaddr_t va, u_int pte)
3697 {
3698 u_long idx;
3699
3700 if (va < KERNBASE)
3701 return;
3702
3703 idx = (unsigned long) m68k_btop(va - KERNBASE);
3704 kernCbase[idx].attr.raw = pte;
3705 TBIS(va);
3706 }
3707
3708 /*
3709 * Routine: pmap_procwr
3710 *
3711 * Function:
3712 * Synchronize caches corresponding to [addr, addr+len) in p.
3713 */
3714 void
3715 pmap_procwr(struct proc *p, vaddr_t va, size_t len)
3716 {
3717 (void)cachectl1(0x80000004, va, len, p);
3718 }
3719
3720
3721 #ifdef PMAP_DEBUG
3722 /************************** DEBUGGING ROUTINES **************************
3723 * The following routines are meant to be an aid to debugging the pmap *
3724 * system. They are callable from the DDB command line and should be *
3725 * prepared to be handed unstable or incomplete states of the system. *
3726 ************************************************************************/
3727
3728 /* pv_list
3729 **
3730 * List all pages found on the pv list for the given physical page.
3731 * To avoid endless loops, the listing will stop at the end of the list
3732 * or after 'n' entries - whichever comes first.
3733 */
3734 void
3735 pv_list(paddr_t pa, int n)
3736 {
3737 int idx;
3738 vaddr_t va;
3739 pv_t *pv;
3740 c_tmgr_t *c_tbl;
3741 pmap_t pmap;
3742
3743 pv = pa2pv(pa);
3744 idx = pv->pv_idx;
3745 for (; idx != PVE_EOL && n > 0; idx = pvebase[idx].pve_next, n--) {
3746 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
3747 printf("idx %d, pmap 0x%x, va 0x%x, c_tbl %x\n",
3748 idx, (u_int) pmap, (u_int) va, (u_int) c_tbl);
3749 }
3750 }
3751 #endif /* PMAP_DEBUG */
3752
3753 #ifdef NOT_YET
3754 /* and maybe not ever */
3755 /************************** LOW-LEVEL ROUTINES **************************
3756 * These routines will eventually be re-written into assembly and placed*
3757 * in locore.s. They are here now as stubs so that the pmap module can *
3758 * be linked as a standalone user program for testing. *
3759 ************************************************************************/
3760 /* flush_atc_crp INTERNAL
3761 **
3762 * Flush all page descriptors derived from the given CPU Root Pointer
3763 * (CRP), or 'A' table as it is known here, from the 68851's automatic
3764 * cache.
3765 */
3766 void
3767 flush_atc_crp(int a_tbl)
3768 {
3769 mmu_long_rp_t rp;
3770
3771 /* Create a temporary root table pointer that points to the
3772 * given A table.
3773 */
3774 rp.attr.raw = ~MMU_LONG_RP_LU;
3775 rp.addr.raw = (unsigned int) a_tbl;
3776
3777 mmu_pflushr(&rp);
3778 /* mmu_pflushr:
3779 * movel sp(4)@,a0
3780 * pflushr a0@
3781 * rts
3782 */
3783 }
3784 #endif /* NOT_YET */
3785