pmap.c revision 1.94 1 /* $NetBSD: pmap.c,v 1.94 2007/02/21 22:59:54 thorpej 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 1.94 2007/02/21 22:59:54 thorpej 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 /* Safe to use pmap_bootstrap_alloc(). */
273 static bool bootstrap_alloc_enabled = FALSE;
274 /* Temporary virtual pages are in use */
275 int tmp_vpages_inuse;
276
277 /*
278 * XXX: For now, retain the traditional variables that were
279 * used in the old pmap/vm interface (without NONCONTIG).
280 */
281 /* Kernel virtual address space available: */
282 vaddr_t virtual_avail, virtual_end;
283 /* Physical address space available: */
284 paddr_t avail_start, avail_end;
285
286 /* This keep track of the end of the contiguously mapped range. */
287 vaddr_t virtual_contig_end;
288
289 /* Physical address used by pmap_next_page() */
290 paddr_t avail_next;
291
292 /* These are used by pmap_copy_page(), etc. */
293 vaddr_t tmp_vpages[2];
294
295 /* memory pool for pmap structures */
296 struct pool pmap_pmap_pool;
297
298 /*
299 * The 3/80 is the only member of the sun3x family that has non-contiguous
300 * physical memory. Memory is divided into 4 banks which are physically
301 * locatable on the system board. Although the size of these banks varies
302 * with the size of memory they contain, their base addresses are
303 * permenently fixed. The following structure, which describes these
304 * banks, is initialized by pmap_bootstrap() after it reads from a similar
305 * structure provided by the ROM Monitor.
306 *
307 * For the other machines in the sun3x architecture which do have contiguous
308 * RAM, this list will have only one entry, which will describe the entire
309 * range of available memory.
310 */
311 struct pmap_physmem_struct avail_mem[SUN3X_NPHYS_RAM_SEGS];
312 u_int total_phys_mem;
313
314 /*************************************************************************/
315
316 /*
317 * XXX - Should "tune" these based on statistics.
318 *
319 * My first guess about the relative numbers of these needed is
320 * based on the fact that a "typical" process will have several
321 * pages mapped at low virtual addresses (text, data, bss), then
322 * some mapped shared libraries, and then some stack pages mapped
323 * near the high end of the VA space. Each process can use only
324 * one A table, and most will use only two B tables (maybe three)
325 * and probably about four C tables. Therefore, the first guess
326 * at the relative numbers of these needed is 1:2:4 -gwr
327 *
328 * The number of C tables needed is closely related to the amount
329 * of physical memory available plus a certain amount attributable
330 * to the use of double mappings. With a few simulation statistics
331 * we can find a reasonably good estimation of this unknown value.
332 * Armed with that and the above ratios, we have a good idea of what
333 * is needed at each level. -j
334 *
335 * Note: It is not physical memory memory size, but the total mapped
336 * virtual space required by the combined working sets of all the
337 * currently _runnable_ processes. (Sleeping ones don't count.)
338 * The amount of physical memory should be irrelevant. -gwr
339 */
340 #ifdef FIXED_NTABLES
341 #define NUM_A_TABLES 16
342 #define NUM_B_TABLES 32
343 #define NUM_C_TABLES 64
344 #else
345 unsigned int NUM_A_TABLES, NUM_B_TABLES, NUM_C_TABLES;
346 #endif /* FIXED_NTABLES */
347
348 /*
349 * This determines our total virtual mapping capacity.
350 * Yes, it is a FIXED value so we can pre-allocate.
351 */
352 #define NUM_USER_PTES (NUM_C_TABLES * MMU_C_TBL_SIZE)
353
354 /*
355 * The size of the Kernel Virtual Address Space (KVAS)
356 * for purposes of MMU table allocation is -KERNBASE
357 * (length from KERNBASE to 0xFFFFffff)
358 */
359 #define KVAS_SIZE (-KERNBASE)
360
361 /* Numbers of kernel MMU tables to support KVAS_SIZE. */
362 #define KERN_B_TABLES (KVAS_SIZE >> MMU_TIA_SHIFT)
363 #define KERN_C_TABLES (KVAS_SIZE >> MMU_TIB_SHIFT)
364 #define NUM_KERN_PTES (KVAS_SIZE >> MMU_TIC_SHIFT)
365
366 /*************************** MISCELANEOUS MACROS *************************/
367 #define pmap_lock(pmap) simple_lock(&pmap->pm_lock)
368 #define pmap_unlock(pmap) simple_unlock(&pmap->pm_lock)
369 #define pmap_add_ref(pmap) ++pmap->pm_refcount
370 #define pmap_del_ref(pmap) --pmap->pm_refcount
371 #define pmap_refcount(pmap) pmap->pm_refcount
372
373 void *pmap_bootstrap_alloc(int);
374
375 static INLINE void *mmu_ptov(paddr_t);
376 static INLINE paddr_t mmu_vtop(void *);
377
378 #if 0
379 static INLINE a_tmgr_t *mmuA2tmgr(mmu_long_dte_t *);
380 #endif /* 0 */
381 static INLINE b_tmgr_t *mmuB2tmgr(mmu_short_dte_t *);
382 static INLINE c_tmgr_t *mmuC2tmgr(mmu_short_pte_t *);
383
384 static INLINE pv_t *pa2pv(paddr_t);
385 static INLINE int pteidx(mmu_short_pte_t *);
386 static INLINE pmap_t current_pmap(void);
387
388 /*
389 * We can always convert between virtual and physical addresses
390 * for anything in the range [KERNBASE ... avail_start] because
391 * that range is GUARANTEED to be mapped linearly.
392 * We rely heavily upon this feature!
393 */
394 static INLINE void *
395 mmu_ptov(paddr_t pa)
396 {
397 vaddr_t va;
398
399 va = (pa + KERNBASE);
400 #ifdef PMAP_DEBUG
401 if ((va < KERNBASE) || (va >= virtual_contig_end))
402 panic("mmu_ptov");
403 #endif
404 return (void *)va;
405 }
406
407 static INLINE paddr_t
408 mmu_vtop(void *vva)
409 {
410 vaddr_t va;
411
412 va = (vaddr_t)vva;
413 #ifdef PMAP_DEBUG
414 if ((va < KERNBASE) || (va >= virtual_contig_end))
415 panic("mmu_vtop");
416 #endif
417 return va - KERNBASE;
418 }
419
420 /*
421 * These macros map MMU tables to their corresponding manager structures.
422 * They are needed quite often because many of the pointers in the pmap
423 * system reference MMU tables and not the structures that control them.
424 * There needs to be a way to find one when given the other and these
425 * macros do so by taking advantage of the memory layout described above.
426 * Here's a quick step through the first macro, mmuA2tmgr():
427 *
428 * 1) find the offset of the given MMU A table from the base of its table
429 * pool (table - mmuAbase).
430 * 2) convert this offset into a table index by dividing it by the
431 * size of one MMU 'A' table. (sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE)
432 * 3) use this index to select the corresponding 'A' table manager
433 * structure from the 'A' table manager pool (Atmgrbase[index]).
434 */
435 /* This function is not currently used. */
436 #if 0
437 static INLINE a_tmgr_t *
438 mmuA2tmgr(mmu_long_dte_t *mmuAtbl)
439 {
440 int idx;
441
442 /* Which table is this in? */
443 idx = (mmuAtbl - mmuAbase) / MMU_A_TBL_SIZE;
444 #ifdef PMAP_DEBUG
445 if ((idx < 0) || (idx >= NUM_A_TABLES))
446 panic("mmuA2tmgr");
447 #endif
448 return &Atmgrbase[idx];
449 }
450 #endif /* 0 */
451
452 static INLINE b_tmgr_t *
453 mmuB2tmgr(mmu_short_dte_t *mmuBtbl)
454 {
455 int idx;
456
457 /* Which table is this in? */
458 idx = (mmuBtbl - mmuBbase) / MMU_B_TBL_SIZE;
459 #ifdef PMAP_DEBUG
460 if ((idx < 0) || (idx >= NUM_B_TABLES))
461 panic("mmuB2tmgr");
462 #endif
463 return &Btmgrbase[idx];
464 }
465
466 /* mmuC2tmgr INTERNAL
467 **
468 * Given a pte known to belong to a C table, return the address of
469 * that table's management structure.
470 */
471 static INLINE c_tmgr_t *
472 mmuC2tmgr(mmu_short_pte_t *mmuCtbl)
473 {
474 int idx;
475
476 /* Which table is this in? */
477 idx = (mmuCtbl - mmuCbase) / MMU_C_TBL_SIZE;
478 #ifdef PMAP_DEBUG
479 if ((idx < 0) || (idx >= NUM_C_TABLES))
480 panic("mmuC2tmgr");
481 #endif
482 return &Ctmgrbase[idx];
483 }
484
485 /* This is now a function call below.
486 * #define pa2pv(pa) \
487 * (&pvbase[(unsigned long)\
488 * m68k_btop(pa)\
489 * ])
490 */
491
492 /* pa2pv INTERNAL
493 **
494 * Return the pv_list_head element which manages the given physical
495 * address.
496 */
497 static INLINE pv_t *
498 pa2pv(paddr_t pa)
499 {
500 struct pmap_physmem_struct *bank;
501 int idx;
502
503 bank = &avail_mem[0];
504 while (pa >= bank->pmem_end)
505 bank = bank->pmem_next;
506
507 pa -= bank->pmem_start;
508 idx = bank->pmem_pvbase + m68k_btop(pa);
509 #ifdef PMAP_DEBUG
510 if ((idx < 0) || (idx >= physmem))
511 panic("pa2pv");
512 #endif
513 return &pvbase[idx];
514 }
515
516 /* pteidx INTERNAL
517 **
518 * Return the index of the given PTE within the entire fixed table of
519 * PTEs.
520 */
521 static INLINE int
522 pteidx(mmu_short_pte_t *pte)
523 {
524
525 return pte - kernCbase;
526 }
527
528 /*
529 * This just offers a place to put some debugging checks,
530 * and reduces the number of places "curlwp" appears...
531 */
532 static INLINE pmap_t
533 current_pmap(void)
534 {
535 struct vmspace *vm;
536 struct vm_map *map;
537 pmap_t pmap;
538
539 if (curlwp == NULL)
540 pmap = &kernel_pmap;
541 else {
542 vm = curproc->p_vmspace;
543 map = &vm->vm_map;
544 pmap = vm_map_pmap(map);
545 }
546
547 return pmap;
548 }
549
550
551 /*************************** FUNCTION DEFINITIONS ************************
552 * These appear here merely for the compiler to enforce type checking on *
553 * all function calls. *
554 *************************************************************************/
555
556 /*
557 * Internal functions
558 */
559 a_tmgr_t *get_a_table(void);
560 b_tmgr_t *get_b_table(void);
561 c_tmgr_t *get_c_table(void);
562 int free_a_table(a_tmgr_t *, bool);
563 int free_b_table(b_tmgr_t *, bool);
564 int free_c_table(c_tmgr_t *, bool);
565
566 void pmap_bootstrap_aalign(int);
567 void pmap_alloc_usermmu(void);
568 void pmap_alloc_usertmgr(void);
569 void pmap_alloc_pv(void);
570 void pmap_init_a_tables(void);
571 void pmap_init_b_tables(void);
572 void pmap_init_c_tables(void);
573 void pmap_init_pv(void);
574 void pmap_clear_pv(paddr_t, int);
575 static INLINE bool is_managed(paddr_t);
576
577 bool pmap_remove_a(a_tmgr_t *, vaddr_t, vaddr_t);
578 bool pmap_remove_b(b_tmgr_t *, vaddr_t, vaddr_t);
579 bool pmap_remove_c(c_tmgr_t *, vaddr_t, vaddr_t);
580 void pmap_remove_pte(mmu_short_pte_t *);
581
582 void pmap_enter_kernel(vaddr_t, paddr_t, vm_prot_t);
583 static INLINE void pmap_remove_kernel(vaddr_t, vaddr_t);
584 static INLINE void pmap_protect_kernel(vaddr_t, vaddr_t, vm_prot_t);
585 static INLINE bool pmap_extract_kernel(vaddr_t, paddr_t *);
586 vaddr_t pmap_get_pteinfo(u_int, pmap_t *, c_tmgr_t **);
587 static INLINE int pmap_dereference(pmap_t);
588
589 bool pmap_stroll(pmap_t, vaddr_t, a_tmgr_t **, b_tmgr_t **, c_tmgr_t **,
590 mmu_short_pte_t **, int *, int *, int *);
591 void pmap_bootstrap_copyprom(void);
592 void pmap_takeover_mmu(void);
593 void pmap_bootstrap_setprom(void);
594 static void pmap_page_upload(void);
595
596 #ifdef PMAP_DEBUG
597 /* Debugging function definitions */
598 void pv_list(paddr_t, int);
599 #endif /* PMAP_DEBUG */
600
601 /** Interface functions
602 ** - functions required by the Mach VM Pmap interface, with MACHINE_CONTIG
603 ** defined.
604 ** The new UVM doesn't require them so now INTERNAL.
605 **/
606 static INLINE void pmap_pinit(pmap_t);
607 static INLINE void pmap_release(pmap_t);
608
609 /********************************** CODE ********************************
610 * Functions that are called from other parts of the kernel are labeled *
611 * as 'INTERFACE' functions. Functions that are only called from *
612 * within the pmap module are labeled as 'INTERNAL' functions. *
613 * Functions that are internal, but are not (currently) used at all are *
614 * labeled 'INTERNAL_X'. *
615 ************************************************************************/
616
617 /* pmap_bootstrap INTERNAL
618 **
619 * Initializes the pmap system. Called at boot time from
620 * locore2.c:_vm_init()
621 *
622 * Reminder: having a pmap_bootstrap_alloc() and also having the VM
623 * system implement pmap_steal_memory() is redundant.
624 * Don't release this code without removing one or the other!
625 */
626 void
627 pmap_bootstrap(vaddr_t nextva)
628 {
629 struct physmemory *membank;
630 struct pmap_physmem_struct *pmap_membank;
631 vaddr_t va, eva;
632 paddr_t pa;
633 int b, c, i, j; /* running table counts */
634 int size, resvmem;
635
636 /*
637 * This function is called by __bootstrap after it has
638 * determined the type of machine and made the appropriate
639 * patches to the ROM vectors (XXX- I don't quite know what I meant
640 * by that.) It allocates and sets up enough of the pmap system
641 * to manage the kernel's address space.
642 */
643
644 /*
645 * Determine the range of kernel virtual and physical
646 * space available. Note that we ABSOLUTELY DEPEND on
647 * the fact that the first bank of memory (4MB) is
648 * mapped linearly to KERNBASE (which we guaranteed in
649 * the first instructions of locore.s).
650 * That is plenty for our bootstrap work.
651 */
652 virtual_avail = m68k_round_page(nextva);
653 virtual_contig_end = KERNBASE + 0x400000; /* +4MB */
654 virtual_end = VM_MAX_KERNEL_ADDRESS;
655 /* Don't need avail_start til later. */
656
657 /* We may now call pmap_bootstrap_alloc(). */
658 bootstrap_alloc_enabled = TRUE;
659
660 /*
661 * This is a somewhat unwrapped loop to deal with
662 * copying the PROM's 'phsymem' banks into the pmap's
663 * banks. The following is always assumed:
664 * 1. There is always at least one bank of memory.
665 * 2. There is always a last bank of memory, and its
666 * pmem_next member must be set to NULL.
667 */
668 membank = romVectorPtr->v_physmemory;
669 pmap_membank = avail_mem;
670 total_phys_mem = 0;
671
672 for (;;) { /* break on !membank */
673 pmap_membank->pmem_start = membank->address;
674 pmap_membank->pmem_end = membank->address + membank->size;
675 total_phys_mem += membank->size;
676 membank = membank->next;
677 if (!membank)
678 break;
679 /* This silly syntax arises because pmap_membank
680 * is really a pre-allocated array, but it is put into
681 * use as a linked list.
682 */
683 pmap_membank->pmem_next = pmap_membank + 1;
684 pmap_membank = pmap_membank->pmem_next;
685 }
686 /* This is the last element. */
687 pmap_membank->pmem_next = NULL;
688
689 /*
690 * Note: total_phys_mem, physmem represent
691 * actual physical memory, including that
692 * reserved for the PROM monitor.
693 */
694 physmem = btoc(total_phys_mem);
695
696 /*
697 * Avail_end is set to the first byte of physical memory
698 * after the end of the last bank. We use this only to
699 * determine if a physical address is "managed" memory.
700 * This address range should be reduced to prevent the
701 * physical pages needed by the PROM monitor from being used
702 * in the VM system.
703 */
704 resvmem = total_phys_mem - *(romVectorPtr->memoryAvail);
705 resvmem = m68k_round_page(resvmem);
706 avail_end = pmap_membank->pmem_end - resvmem;
707
708 /*
709 * First allocate enough kernel MMU tables to map all
710 * of kernel virtual space from KERNBASE to 0xFFFFFFFF.
711 * Note: All must be aligned on 256 byte boundaries.
712 * Start with the level-A table (one of those).
713 */
714 size = sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE;
715 kernAbase = pmap_bootstrap_alloc(size);
716 memset(kernAbase, 0, size);
717
718 /* Now the level-B kernel tables... */
719 size = sizeof(mmu_short_dte_t) * MMU_B_TBL_SIZE * KERN_B_TABLES;
720 kernBbase = pmap_bootstrap_alloc(size);
721 memset(kernBbase, 0, size);
722
723 /* Now the level-C kernel tables... */
724 size = sizeof(mmu_short_pte_t) * MMU_C_TBL_SIZE * KERN_C_TABLES;
725 kernCbase = pmap_bootstrap_alloc(size);
726 memset(kernCbase, 0, size);
727 /*
728 * Note: In order for the PV system to work correctly, the kernel
729 * and user-level C tables must be allocated contiguously.
730 * Nothing should be allocated between here and the allocation of
731 * mmuCbase below. XXX: Should do this as one allocation, and
732 * then compute a pointer for mmuCbase instead of this...
733 *
734 * Allocate user MMU tables.
735 * These must be contiguous with the preceding.
736 */
737
738 #ifndef FIXED_NTABLES
739 /*
740 * The number of user-level C tables that should be allocated is
741 * related to the size of physical memory. In general, there should
742 * be enough tables to map four times the amount of available RAM.
743 * The extra amount is needed because some table space is wasted by
744 * fragmentation.
745 */
746 NUM_C_TABLES = (total_phys_mem * 4) / (MMU_C_TBL_SIZE * MMU_PAGE_SIZE);
747 NUM_B_TABLES = NUM_C_TABLES / 2;
748 NUM_A_TABLES = NUM_B_TABLES / 2;
749 #endif /* !FIXED_NTABLES */
750
751 size = sizeof(mmu_short_pte_t) * MMU_C_TBL_SIZE * NUM_C_TABLES;
752 mmuCbase = pmap_bootstrap_alloc(size);
753
754 size = sizeof(mmu_short_dte_t) * MMU_B_TBL_SIZE * NUM_B_TABLES;
755 mmuBbase = pmap_bootstrap_alloc(size);
756
757 size = sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE * NUM_A_TABLES;
758 mmuAbase = pmap_bootstrap_alloc(size);
759
760 /*
761 * Fill in the never-changing part of the kernel tables.
762 * For simplicity, the kernel's mappings will be editable as a
763 * flat array of page table entries at kernCbase. The
764 * higher level 'A' and 'B' tables must be initialized to point
765 * to this lower one.
766 */
767 b = c = 0;
768
769 /*
770 * Invalidate all mappings below KERNBASE in the A table.
771 * This area has already been zeroed out, but it is good
772 * practice to explicitly show that we are interpreting
773 * it as a list of A table descriptors.
774 */
775 for (i = 0; i < MMU_TIA(KERNBASE); i++) {
776 kernAbase[i].addr.raw = 0;
777 }
778
779 /*
780 * Set up the kernel A and B tables so that they will reference the
781 * correct spots in the contiguous table of PTEs allocated for the
782 * kernel's virtual memory space.
783 */
784 for (i = MMU_TIA(KERNBASE); i < MMU_A_TBL_SIZE; i++) {
785 kernAbase[i].attr.raw =
786 MMU_LONG_DTE_LU | MMU_LONG_DTE_SUPV | MMU_DT_SHORT;
787 kernAbase[i].addr.raw = mmu_vtop(&kernBbase[b]);
788
789 for (j = 0; j < MMU_B_TBL_SIZE; j++) {
790 kernBbase[b + j].attr.raw =
791 mmu_vtop(&kernCbase[c]) | MMU_DT_SHORT;
792 c += MMU_C_TBL_SIZE;
793 }
794 b += MMU_B_TBL_SIZE;
795 }
796
797 pmap_alloc_usermmu(); /* Allocate user MMU tables. */
798 pmap_alloc_usertmgr(); /* Allocate user MMU table managers.*/
799 pmap_alloc_pv(); /* Allocate physical->virtual map. */
800
801 /*
802 * We are now done with pmap_bootstrap_alloc(). Round up
803 * `virtual_avail' to the nearest page, and set the flag
804 * to prevent use of pmap_bootstrap_alloc() hereafter.
805 */
806 pmap_bootstrap_aalign(PAGE_SIZE);
807 bootstrap_alloc_enabled = FALSE;
808
809 /*
810 * Now that we are done with pmap_bootstrap_alloc(), we
811 * must save the virtual and physical addresses of the
812 * end of the linearly mapped range, which are stored in
813 * virtual_contig_end and avail_start, respectively.
814 * These variables will never change after this point.
815 */
816 virtual_contig_end = virtual_avail;
817 avail_start = virtual_avail - KERNBASE;
818
819 /*
820 * `avail_next' is a running pointer used by pmap_next_page() to
821 * keep track of the next available physical page to be handed
822 * to the VM system during its initialization, in which it
823 * asks for physical pages, one at a time.
824 */
825 avail_next = avail_start;
826
827 /*
828 * Now allocate some virtual addresses, but not the physical pages
829 * behind them. Note that virtual_avail is already page-aligned.
830 *
831 * tmp_vpages[] is an array of two virtual pages used for temporary
832 * kernel mappings in the pmap module to facilitate various physical
833 * address-oritented operations.
834 */
835 tmp_vpages[0] = virtual_avail;
836 virtual_avail += PAGE_SIZE;
837 tmp_vpages[1] = virtual_avail;
838 virtual_avail += PAGE_SIZE;
839
840 /** Initialize the PV system **/
841 pmap_init_pv();
842
843 /*
844 * Fill in the kernel_pmap structure and kernel_crp.
845 */
846 kernAphys = mmu_vtop(kernAbase);
847 kernel_pmap.pm_a_tmgr = NULL;
848 kernel_pmap.pm_a_phys = kernAphys;
849 kernel_pmap.pm_refcount = 1; /* always in use */
850 simple_lock_init(&kernel_pmap.pm_lock);
851
852 kernel_crp.rp_attr = MMU_LONG_DTE_LU | MMU_DT_LONG;
853 kernel_crp.rp_addr = kernAphys;
854
855 /*
856 * Now pmap_enter_kernel() may be used safely and will be
857 * the main interface used hereafter to modify the kernel's
858 * virtual address space. Note that since we are still running
859 * under the PROM's address table, none of these table modifications
860 * actually take effect until pmap_takeover_mmu() is called.
861 *
862 * Note: Our tables do NOT have the PROM linear mappings!
863 * Only the mappings created here exist in our tables, so
864 * remember to map anything we expect to use.
865 */
866 va = (vaddr_t)KERNBASE;
867 pa = 0;
868
869 /*
870 * The first page of the kernel virtual address space is the msgbuf
871 * page. The page attributes (data, non-cached) are set here, while
872 * the address is assigned to this global pointer in cpu_startup().
873 * It is non-cached, mostly due to paranoia.
874 */
875 pmap_enter_kernel(va, pa|PMAP_NC, VM_PROT_ALL);
876 va += PAGE_SIZE;
877 pa += PAGE_SIZE;
878
879 /* Next page is used as the temporary stack. */
880 pmap_enter_kernel(va, pa, VM_PROT_ALL);
881 va += PAGE_SIZE;
882 pa += PAGE_SIZE;
883
884 /*
885 * Map all of the kernel's text segment as read-only and cacheable.
886 * (Cacheable is implied by default). Unfortunately, the last bytes
887 * of kernel text and the first bytes of kernel data will often be
888 * sharing the same page. Therefore, the last page of kernel text
889 * has to be mapped as read/write, to accommodate the data.
890 */
891 eva = m68k_trunc_page((vaddr_t)etext);
892 for (; va < eva; va += PAGE_SIZE, pa += PAGE_SIZE)
893 pmap_enter_kernel(va, pa, VM_PROT_READ|VM_PROT_EXECUTE);
894
895 /*
896 * Map all of the kernel's data as read/write and cacheable.
897 * This includes: data, BSS, symbols, and everything in the
898 * contiguous memory used by pmap_bootstrap_alloc()
899 */
900 for (; pa < avail_start; va += PAGE_SIZE, pa += PAGE_SIZE)
901 pmap_enter_kernel(va, pa, VM_PROT_READ|VM_PROT_WRITE);
902
903 /*
904 * At this point we are almost ready to take over the MMU. But first
905 * we must save the PROM's address space in our map, as we call its
906 * routines and make references to its data later in the kernel.
907 */
908 pmap_bootstrap_copyprom();
909 pmap_takeover_mmu();
910 pmap_bootstrap_setprom();
911
912 /* Notify the VM system of our page size. */
913 uvmexp.pagesize = PAGE_SIZE;
914 uvm_setpagesize();
915
916 pmap_page_upload();
917 }
918
919
920 /* pmap_alloc_usermmu INTERNAL
921 **
922 * Called from pmap_bootstrap() to allocate MMU tables that will
923 * eventually be used for user mappings.
924 */
925 void
926 pmap_alloc_usermmu(void)
927 {
928
929 /* XXX: Moved into caller. */
930 }
931
932 /* pmap_alloc_pv INTERNAL
933 **
934 * Called from pmap_bootstrap() to allocate the physical
935 * to virtual mapping list. Each physical page of memory
936 * in the system has a corresponding element in this list.
937 */
938 void
939 pmap_alloc_pv(void)
940 {
941 int i;
942 unsigned int total_mem;
943
944 /*
945 * Allocate a pv_head structure for every page of physical
946 * memory that will be managed by the system. Since memory on
947 * the 3/80 is non-contiguous, we cannot arrive at a total page
948 * count by subtraction of the lowest available address from the
949 * highest, but rather we have to step through each memory
950 * bank and add the number of pages in each to the total.
951 *
952 * At this time we also initialize the offset of each bank's
953 * starting pv_head within the pv_head list so that the physical
954 * memory state routines (pmap_is_referenced(),
955 * pmap_is_modified(), et al.) can quickly find coresponding
956 * pv_heads in spite of the non-contiguity.
957 */
958 total_mem = 0;
959 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
960 avail_mem[i].pmem_pvbase = m68k_btop(total_mem);
961 total_mem += avail_mem[i].pmem_end - avail_mem[i].pmem_start;
962 if (avail_mem[i].pmem_next == NULL)
963 break;
964 }
965 pvbase = (pv_t *)pmap_bootstrap_alloc(sizeof(pv_t) *
966 m68k_btop(total_phys_mem));
967 }
968
969 /* pmap_alloc_usertmgr INTERNAL
970 **
971 * Called from pmap_bootstrap() to allocate the structures which
972 * facilitate management of user MMU tables. Each user MMU table
973 * in the system has one such structure associated with it.
974 */
975 void
976 pmap_alloc_usertmgr(void)
977 {
978 /* Allocate user MMU table managers */
979 /* It would be a lot simpler to just make these BSS, but */
980 /* we may want to change their size at boot time... -j */
981 Atmgrbase =
982 (a_tmgr_t *)pmap_bootstrap_alloc(sizeof(a_tmgr_t) * NUM_A_TABLES);
983 Btmgrbase =
984 (b_tmgr_t *)pmap_bootstrap_alloc(sizeof(b_tmgr_t) * NUM_B_TABLES);
985 Ctmgrbase =
986 (c_tmgr_t *)pmap_bootstrap_alloc(sizeof(c_tmgr_t) * NUM_C_TABLES);
987
988 /*
989 * Allocate PV list elements for the physical to virtual
990 * mapping system.
991 */
992 pvebase = (pv_elem_t *)pmap_bootstrap_alloc(sizeof(pv_elem_t) *
993 (NUM_USER_PTES + NUM_KERN_PTES));
994 }
995
996 /* pmap_bootstrap_copyprom() INTERNAL
997 **
998 * Copy the PROM mappings into our own tables. Note, we
999 * can use physical addresses until __bootstrap returns.
1000 */
1001 void
1002 pmap_bootstrap_copyprom(void)
1003 {
1004 struct sunromvec *romp;
1005 int *mon_ctbl;
1006 mmu_short_pte_t *kpte;
1007 int i, len;
1008
1009 romp = romVectorPtr;
1010
1011 /*
1012 * Copy the mappings in SUN3X_MON_KDB_BASE...SUN3X_MONEND
1013 * Note: mon_ctbl[0] maps SUN3X_MON_KDB_BASE
1014 */
1015 mon_ctbl = *romp->monptaddr;
1016 i = m68k_btop(SUN3X_MON_KDB_BASE - KERNBASE);
1017 kpte = &kernCbase[i];
1018 len = m68k_btop(SUN3X_MONEND - SUN3X_MON_KDB_BASE);
1019
1020 for (i = 0; i < len; i++) {
1021 kpte[i].attr.raw = mon_ctbl[i];
1022 }
1023
1024 /*
1025 * Copy the mappings at MON_DVMA_BASE (to the end).
1026 * Note, in here, mon_ctbl[0] maps MON_DVMA_BASE.
1027 * Actually, we only want the last page, which the
1028 * PROM has set up for use by the "ie" driver.
1029 * (The i82686 needs its SCP there.)
1030 * If we copy all the mappings, pmap_enter_kernel
1031 * may complain about finding valid PTEs that are
1032 * not recorded in our PV lists...
1033 */
1034 mon_ctbl = *romp->shadowpteaddr;
1035 i = m68k_btop(SUN3X_MON_DVMA_BASE - KERNBASE);
1036 kpte = &kernCbase[i];
1037 len = m68k_btop(SUN3X_MON_DVMA_SIZE);
1038 for (i = (len - 1); i < len; i++) {
1039 kpte[i].attr.raw = mon_ctbl[i];
1040 }
1041 }
1042
1043 /* pmap_takeover_mmu INTERNAL
1044 **
1045 * Called from pmap_bootstrap() after it has copied enough of the
1046 * PROM mappings into the kernel map so that we can use our own
1047 * MMU table.
1048 */
1049 void
1050 pmap_takeover_mmu(void)
1051 {
1052
1053 loadcrp(&kernel_crp);
1054 }
1055
1056 /* pmap_bootstrap_setprom() INTERNAL
1057 **
1058 * Set the PROM mappings so it can see kernel space.
1059 * Note that physical addresses are used here, which
1060 * we can get away with because this runs with the
1061 * low 1GB set for transparent translation.
1062 */
1063 void
1064 pmap_bootstrap_setprom(void)
1065 {
1066 mmu_long_dte_t *mon_dte;
1067 extern struct mmu_rootptr mon_crp;
1068 int i;
1069
1070 mon_dte = (mmu_long_dte_t *)mon_crp.rp_addr;
1071 for (i = MMU_TIA(KERNBASE); i < MMU_TIA(KERN_END); i++) {
1072 mon_dte[i].attr.raw = kernAbase[i].attr.raw;
1073 mon_dte[i].addr.raw = kernAbase[i].addr.raw;
1074 }
1075 }
1076
1077
1078 /* pmap_init INTERFACE
1079 **
1080 * Called at the end of vm_init() to set up the pmap system to go
1081 * into full time operation. All initialization of kernel_pmap
1082 * should be already done by now, so this should just do things
1083 * needed for user-level pmaps to work.
1084 */
1085 void
1086 pmap_init(void)
1087 {
1088
1089 /** Initialize the manager pools **/
1090 TAILQ_INIT(&a_pool);
1091 TAILQ_INIT(&b_pool);
1092 TAILQ_INIT(&c_pool);
1093
1094 /**************************************************************
1095 * Initialize all tmgr structures and MMU tables they manage. *
1096 **************************************************************/
1097 /** Initialize A tables **/
1098 pmap_init_a_tables();
1099 /** Initialize B tables **/
1100 pmap_init_b_tables();
1101 /** Initialize C tables **/
1102 pmap_init_c_tables();
1103
1104 /** Initialize the pmap pools **/
1105 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
1106 &pool_allocator_nointr);
1107 }
1108
1109 /* pmap_init_a_tables() INTERNAL
1110 **
1111 * Initializes all A managers, their MMU A tables, and inserts
1112 * them into the A manager pool for use by the system.
1113 */
1114 void
1115 pmap_init_a_tables(void)
1116 {
1117 int i;
1118 a_tmgr_t *a_tbl;
1119
1120 for (i = 0; i < NUM_A_TABLES; i++) {
1121 /* Select the next available A manager from the pool */
1122 a_tbl = &Atmgrbase[i];
1123
1124 /*
1125 * Clear its parent entry. Set its wired and valid
1126 * entry count to zero.
1127 */
1128 a_tbl->at_parent = NULL;
1129 a_tbl->at_wcnt = a_tbl->at_ecnt = 0;
1130
1131 /* Assign it the next available MMU A table from the pool */
1132 a_tbl->at_dtbl = &mmuAbase[i * MMU_A_TBL_SIZE];
1133
1134 /*
1135 * Initialize the MMU A table with the table in the `proc0',
1136 * or kernel, mapping. This ensures that every process has
1137 * the kernel mapped in the top part of its address space.
1138 */
1139 memcpy(a_tbl->at_dtbl, kernAbase,
1140 MMU_A_TBL_SIZE * sizeof(mmu_long_dte_t));
1141
1142 /*
1143 * Finally, insert the manager into the A pool,
1144 * making it ready to be used by the system.
1145 */
1146 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
1147 }
1148 }
1149
1150 /* pmap_init_b_tables() INTERNAL
1151 **
1152 * Initializes all B table managers, their MMU B tables, and
1153 * inserts them into the B manager pool for use by the system.
1154 */
1155 void
1156 pmap_init_b_tables(void)
1157 {
1158 int i, j;
1159 b_tmgr_t *b_tbl;
1160
1161 for (i = 0; i < NUM_B_TABLES; i++) {
1162 /* Select the next available B manager from the pool */
1163 b_tbl = &Btmgrbase[i];
1164
1165 b_tbl->bt_parent = NULL; /* clear its parent, */
1166 b_tbl->bt_pidx = 0; /* parent index, */
1167 b_tbl->bt_wcnt = 0; /* wired entry count, */
1168 b_tbl->bt_ecnt = 0; /* valid entry count. */
1169
1170 /* Assign it the next available MMU B table from the pool */
1171 b_tbl->bt_dtbl = &mmuBbase[i * MMU_B_TBL_SIZE];
1172
1173 /* Invalidate every descriptor in the table */
1174 for (j = 0; j < MMU_B_TBL_SIZE; j++)
1175 b_tbl->bt_dtbl[j].attr.raw = MMU_DT_INVALID;
1176
1177 /* Insert the manager into the B pool */
1178 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
1179 }
1180 }
1181
1182 /* pmap_init_c_tables() INTERNAL
1183 **
1184 * Initializes all C table managers, their MMU C tables, and
1185 * inserts them into the C manager pool for use by the system.
1186 */
1187 void
1188 pmap_init_c_tables(void)
1189 {
1190 int i, j;
1191 c_tmgr_t *c_tbl;
1192
1193 for (i = 0; i < NUM_C_TABLES; i++) {
1194 /* Select the next available C manager from the pool */
1195 c_tbl = &Ctmgrbase[i];
1196
1197 c_tbl->ct_parent = NULL; /* clear its parent, */
1198 c_tbl->ct_pidx = 0; /* parent index, */
1199 c_tbl->ct_wcnt = 0; /* wired entry count, */
1200 c_tbl->ct_ecnt = 0; /* valid entry count, */
1201 c_tbl->ct_pmap = NULL; /* parent pmap, */
1202 c_tbl->ct_va = 0; /* base of managed range */
1203
1204 /* Assign it the next available MMU C table from the pool */
1205 c_tbl->ct_dtbl = &mmuCbase[i * MMU_C_TBL_SIZE];
1206
1207 for (j = 0; j < MMU_C_TBL_SIZE; j++)
1208 c_tbl->ct_dtbl[j].attr.raw = MMU_DT_INVALID;
1209
1210 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
1211 }
1212 }
1213
1214 /* pmap_init_pv() INTERNAL
1215 **
1216 * Initializes the Physical to Virtual mapping system.
1217 */
1218 void
1219 pmap_init_pv(void)
1220 {
1221 int i;
1222
1223 /* Initialize every PV head. */
1224 for (i = 0; i < m68k_btop(total_phys_mem); i++) {
1225 pvbase[i].pv_idx = PVE_EOL; /* Indicate no mappings */
1226 pvbase[i].pv_flags = 0; /* Zero out page flags */
1227 }
1228 }
1229
1230 /* is_managed INTERNAL
1231 **
1232 * Determine if the given physical address is managed by the PV system.
1233 * Note that this logic assumes that no one will ask for the status of
1234 * addresses which lie in-between the memory banks on the 3/80. If they
1235 * do so, it will falsely report that it is managed.
1236 *
1237 * Note: A "managed" address is one that was reported to the VM system as
1238 * a "usable page" during system startup. As such, the VM system expects the
1239 * pmap module to keep an accurate track of the useage of those pages.
1240 * Any page not given to the VM system at startup does not exist (as far as
1241 * the VM system is concerned) and is therefore "unmanaged." Examples are
1242 * those pages which belong to the ROM monitor and the memory allocated before
1243 * the VM system was started.
1244 */
1245 static INLINE bool
1246 is_managed(paddr_t pa)
1247 {
1248 if (pa >= avail_start && pa < avail_end)
1249 return TRUE;
1250 else
1251 return FALSE;
1252 }
1253
1254 /* get_a_table INTERNAL
1255 **
1256 * Retrieve and return a level A table for use in a user map.
1257 */
1258 a_tmgr_t *
1259 get_a_table(void)
1260 {
1261 a_tmgr_t *tbl;
1262 pmap_t pmap;
1263
1264 /* Get the top A table in the pool */
1265 tbl = TAILQ_FIRST(&a_pool);
1266 if (tbl == NULL) {
1267 /*
1268 * XXX - Instead of panicking here and in other get_x_table
1269 * functions, we do have the option of sleeping on the head of
1270 * the table pool. Any function which updates the table pool
1271 * would then issue a wakeup() on the head, thus waking up any
1272 * processes waiting for a table.
1273 *
1274 * Actually, the place to sleep would be when some process
1275 * asks for a "wired" mapping that would run us short of
1276 * mapping resources. This design DEPENDS on always having
1277 * some mapping resources in the pool for stealing, so we
1278 * must make sure we NEVER let the pool become empty. -gwr
1279 */
1280 panic("get_a_table: out of A tables.");
1281 }
1282
1283 TAILQ_REMOVE(&a_pool, tbl, at_link);
1284 /*
1285 * If the table has a non-null parent pointer then it is in use.
1286 * Forcibly abduct it from its parent and clear its entries.
1287 * No re-entrancy worries here. This table would not be in the
1288 * table pool unless it was available for use.
1289 *
1290 * Note that the second argument to free_a_table() is FALSE. This
1291 * indicates that the table should not be relinked into the A table
1292 * pool. That is a job for the function that called us.
1293 */
1294 if (tbl->at_parent) {
1295 KASSERT(tbl->at_wcnt == 0);
1296 pmap = tbl->at_parent;
1297 free_a_table(tbl, FALSE);
1298 pmap->pm_a_tmgr = NULL;
1299 pmap->pm_a_phys = kernAphys;
1300 }
1301 return tbl;
1302 }
1303
1304 /* get_b_table INTERNAL
1305 **
1306 * Return a level B table for use.
1307 */
1308 b_tmgr_t *
1309 get_b_table(void)
1310 {
1311 b_tmgr_t *tbl;
1312
1313 /* See 'get_a_table' for comments. */
1314 tbl = TAILQ_FIRST(&b_pool);
1315 if (tbl == NULL)
1316 panic("get_b_table: out of B tables.");
1317 TAILQ_REMOVE(&b_pool, tbl, bt_link);
1318 if (tbl->bt_parent) {
1319 KASSERT(tbl->bt_wcnt == 0);
1320 tbl->bt_parent->at_dtbl[tbl->bt_pidx].attr.raw = MMU_DT_INVALID;
1321 tbl->bt_parent->at_ecnt--;
1322 free_b_table(tbl, FALSE);
1323 }
1324 return tbl;
1325 }
1326
1327 /* get_c_table INTERNAL
1328 **
1329 * Return a level C table for use.
1330 */
1331 c_tmgr_t *
1332 get_c_table(void)
1333 {
1334 c_tmgr_t *tbl;
1335
1336 /* See 'get_a_table' for comments */
1337 tbl = TAILQ_FIRST(&c_pool);
1338 if (tbl == NULL)
1339 panic("get_c_table: out of C tables.");
1340 TAILQ_REMOVE(&c_pool, tbl, ct_link);
1341 if (tbl->ct_parent) {
1342 KASSERT(tbl->ct_wcnt == 0);
1343 tbl->ct_parent->bt_dtbl[tbl->ct_pidx].attr.raw = MMU_DT_INVALID;
1344 tbl->ct_parent->bt_ecnt--;
1345 free_c_table(tbl, FALSE);
1346 }
1347 return tbl;
1348 }
1349
1350 /*
1351 * The following 'free_table' and 'steal_table' functions are called to
1352 * detach tables from their current obligations (parents and children) and
1353 * prepare them for reuse in another mapping.
1354 *
1355 * Free_table is used when the calling function will handle the fate
1356 * of the parent table, such as returning it to the free pool when it has
1357 * no valid entries. Functions that do not want to handle this should
1358 * call steal_table, in which the parent table's descriptors and entry
1359 * count are automatically modified when this table is removed.
1360 */
1361
1362 /* free_a_table INTERNAL
1363 **
1364 * Unmaps the given A table and all child tables from their current
1365 * mappings. Returns the number of pages that were invalidated.
1366 * If 'relink' is true, the function will return the table to the head
1367 * of the available table pool.
1368 *
1369 * Cache note: The MC68851 will automatically flush all
1370 * descriptors derived from a given A table from its
1371 * Automatic Translation Cache (ATC) if we issue a
1372 * 'PFLUSHR' instruction with the base address of the
1373 * table. This function should do, and does so.
1374 * Note note: We are using an MC68030 - there is no
1375 * PFLUSHR.
1376 */
1377 int
1378 free_a_table(a_tmgr_t *a_tbl, bool relink)
1379 {
1380 int i, removed_cnt;
1381 mmu_long_dte_t *dte;
1382 mmu_short_dte_t *dtbl;
1383 b_tmgr_t *b_tbl;
1384 uint8_t at_wired, bt_wired;
1385
1386 /*
1387 * Flush the ATC cache of all cached descriptors derived
1388 * from this table.
1389 * Sun3x does not use 68851's cached table feature
1390 * flush_atc_crp(mmu_vtop(a_tbl->dte));
1391 */
1392
1393 /*
1394 * Remove any pending cache flushes that were designated
1395 * for the pmap this A table belongs to.
1396 * a_tbl->parent->atc_flushq[0] = 0;
1397 * Not implemented in sun3x.
1398 */
1399
1400 /*
1401 * All A tables in the system should retain a map for the
1402 * kernel. If the table contains any valid descriptors
1403 * (other than those for the kernel area), invalidate them all,
1404 * stopping short of the kernel's entries.
1405 */
1406 removed_cnt = 0;
1407 at_wired = a_tbl->at_wcnt;
1408 if (a_tbl->at_ecnt) {
1409 dte = a_tbl->at_dtbl;
1410 for (i = 0; i < MMU_TIA(KERNBASE); i++) {
1411 /*
1412 * If a table entry points to a valid B table, free
1413 * it and its children.
1414 */
1415 if (MMU_VALID_DT(dte[i])) {
1416 /*
1417 * The following block does several things,
1418 * from innermost expression to the
1419 * outermost:
1420 * 1) It extracts the base (cc 1996)
1421 * address of the B table pointed
1422 * to in the A table entry dte[i].
1423 * 2) It converts this base address into
1424 * the virtual address it can be
1425 * accessed with. (all MMU tables point
1426 * to physical addresses.)
1427 * 3) It finds the corresponding manager
1428 * structure which manages this MMU table.
1429 * 4) It frees the manager structure.
1430 * (This frees the MMU table and all
1431 * child tables. See 'free_b_table' for
1432 * details.)
1433 */
1434 dtbl = mmu_ptov(dte[i].addr.raw);
1435 b_tbl = mmuB2tmgr(dtbl);
1436 bt_wired = b_tbl->bt_wcnt;
1437 removed_cnt += free_b_table(b_tbl, TRUE);
1438 if (bt_wired)
1439 a_tbl->at_wcnt--;
1440 dte[i].attr.raw = MMU_DT_INVALID;
1441 }
1442 }
1443 a_tbl->at_ecnt = 0;
1444 }
1445 KASSERT(a_tbl->at_wcnt == 0);
1446
1447 if (relink) {
1448 a_tbl->at_parent = NULL;
1449 if (!at_wired)
1450 TAILQ_REMOVE(&a_pool, a_tbl, at_link);
1451 TAILQ_INSERT_HEAD(&a_pool, a_tbl, at_link);
1452 }
1453 return removed_cnt;
1454 }
1455
1456 /* free_b_table INTERNAL
1457 **
1458 * Unmaps the given B table and all its children from their current
1459 * mappings. Returns the number of pages that were invalidated.
1460 * (For comments, see 'free_a_table()').
1461 */
1462 int
1463 free_b_table(b_tmgr_t *b_tbl, bool relink)
1464 {
1465 int i, removed_cnt;
1466 mmu_short_dte_t *dte;
1467 mmu_short_pte_t *dtbl;
1468 c_tmgr_t *c_tbl;
1469 uint8_t bt_wired, ct_wired;
1470
1471 removed_cnt = 0;
1472 bt_wired = b_tbl->bt_wcnt;
1473 if (b_tbl->bt_ecnt) {
1474 dte = b_tbl->bt_dtbl;
1475 for (i = 0; i < MMU_B_TBL_SIZE; i++) {
1476 if (MMU_VALID_DT(dte[i])) {
1477 dtbl = mmu_ptov(MMU_DTE_PA(dte[i]));
1478 c_tbl = mmuC2tmgr(dtbl);
1479 ct_wired = c_tbl->ct_wcnt;
1480 removed_cnt += free_c_table(c_tbl, TRUE);
1481 if (ct_wired)
1482 b_tbl->bt_wcnt--;
1483 dte[i].attr.raw = MMU_DT_INVALID;
1484 }
1485 }
1486 b_tbl->bt_ecnt = 0;
1487 }
1488 KASSERT(b_tbl->bt_wcnt == 0);
1489
1490 if (relink) {
1491 b_tbl->bt_parent = NULL;
1492 if (!bt_wired)
1493 TAILQ_REMOVE(&b_pool, b_tbl, bt_link);
1494 TAILQ_INSERT_HEAD(&b_pool, b_tbl, bt_link);
1495 }
1496 return removed_cnt;
1497 }
1498
1499 /* free_c_table INTERNAL
1500 **
1501 * Unmaps the given C table from use and returns it to the pool for
1502 * re-use. Returns the number of pages that were invalidated.
1503 *
1504 * This function preserves any physical page modification information
1505 * contained in the page descriptors within the C table by calling
1506 * 'pmap_remove_pte().'
1507 */
1508 int
1509 free_c_table(c_tmgr_t *c_tbl, bool relink)
1510 {
1511 mmu_short_pte_t *c_pte;
1512 int i, removed_cnt;
1513 uint8_t ct_wired;
1514
1515 removed_cnt = 0;
1516 ct_wired = c_tbl->ct_wcnt;
1517 if (c_tbl->ct_ecnt) {
1518 for (i = 0; i < MMU_C_TBL_SIZE; i++) {
1519 c_pte = &c_tbl->ct_dtbl[i];
1520 if (MMU_VALID_DT(*c_pte)) {
1521 if (c_pte->attr.raw & MMU_SHORT_PTE_WIRED)
1522 c_tbl->ct_wcnt--;
1523 pmap_remove_pte(c_pte);
1524 removed_cnt++;
1525 }
1526 }
1527 c_tbl->ct_ecnt = 0;
1528 }
1529 KASSERT(c_tbl->ct_wcnt == 0);
1530
1531 if (relink) {
1532 c_tbl->ct_parent = NULL;
1533 if (!ct_wired)
1534 TAILQ_REMOVE(&c_pool, c_tbl, ct_link);
1535 TAILQ_INSERT_HEAD(&c_pool, c_tbl, ct_link);
1536 }
1537 return removed_cnt;
1538 }
1539
1540
1541 /* pmap_remove_pte INTERNAL
1542 **
1543 * Unmap the given pte and preserve any page modification
1544 * information by transfering it to the pv head of the
1545 * physical page it maps to. This function does not update
1546 * any reference counts because it is assumed that the calling
1547 * function will do so.
1548 */
1549 void
1550 pmap_remove_pte(mmu_short_pte_t *pte)
1551 {
1552 u_short pv_idx, targ_idx;
1553 paddr_t pa;
1554 pv_t *pv;
1555
1556 pa = MMU_PTE_PA(*pte);
1557 if (is_managed(pa)) {
1558 pv = pa2pv(pa);
1559 targ_idx = pteidx(pte); /* Index of PTE being removed */
1560
1561 /*
1562 * If the PTE being removed is the first (or only) PTE in
1563 * the list of PTEs currently mapped to this page, remove the
1564 * PTE by changing the index found on the PV head. Otherwise
1565 * a linear search through the list will have to be executed
1566 * in order to find the PVE which points to the PTE being
1567 * removed, so that it may be modified to point to its new
1568 * neighbor.
1569 */
1570
1571 pv_idx = pv->pv_idx; /* Index of first PTE in PV list */
1572 if (pv_idx == targ_idx) {
1573 pv->pv_idx = pvebase[targ_idx].pve_next;
1574 } else {
1575
1576 /*
1577 * Find the PV element pointing to the target
1578 * element. Note: may have pv_idx==PVE_EOL
1579 */
1580
1581 for (;;) {
1582 if (pv_idx == PVE_EOL) {
1583 goto pv_not_found;
1584 }
1585 if (pvebase[pv_idx].pve_next == targ_idx)
1586 break;
1587 pv_idx = pvebase[pv_idx].pve_next;
1588 }
1589
1590 /*
1591 * At this point, pv_idx is the index of the PV
1592 * element just before the target element in the list.
1593 * Unlink the target.
1594 */
1595
1596 pvebase[pv_idx].pve_next = pvebase[targ_idx].pve_next;
1597 }
1598
1599 /*
1600 * Save the mod/ref bits of the pte by simply
1601 * ORing the entire pte onto the pv_flags member
1602 * of the pv structure.
1603 * There is no need to use a separate bit pattern
1604 * for usage information on the pv head than that
1605 * which is used on the MMU ptes.
1606 */
1607
1608 pv_not_found:
1609 pv->pv_flags |= (u_short) pte->attr.raw;
1610 }
1611 pte->attr.raw = MMU_DT_INVALID;
1612 }
1613
1614 /* pmap_stroll INTERNAL
1615 **
1616 * Retrieve the addresses of all table managers involved in the mapping of
1617 * the given virtual address. If the table walk completed successfully,
1618 * return TRUE. If it was only partially successful, return FALSE.
1619 * The table walk performed by this function is important to many other
1620 * functions in this module.
1621 *
1622 * Note: This function ought to be easier to read.
1623 */
1624 bool
1625 pmap_stroll(pmap_t pmap, vaddr_t va, a_tmgr_t **a_tbl, b_tmgr_t **b_tbl,
1626 c_tmgr_t **c_tbl, mmu_short_pte_t **pte, int *a_idx, int *b_idx,
1627 int *pte_idx)
1628 {
1629 mmu_long_dte_t *a_dte; /* A: long descriptor table */
1630 mmu_short_dte_t *b_dte; /* B: short descriptor table */
1631
1632 if (pmap == pmap_kernel())
1633 return FALSE;
1634
1635 /* Does the given pmap have its own A table? */
1636 *a_tbl = pmap->pm_a_tmgr;
1637 if (*a_tbl == NULL)
1638 return FALSE; /* No. Return unknown. */
1639 /* Does the A table have a valid B table
1640 * under the corresponding table entry?
1641 */
1642 *a_idx = MMU_TIA(va);
1643 a_dte = &((*a_tbl)->at_dtbl[*a_idx]);
1644 if (!MMU_VALID_DT(*a_dte))
1645 return FALSE; /* No. Return unknown. */
1646 /* Yes. Extract B table from the A table. */
1647 *b_tbl = mmuB2tmgr(mmu_ptov(a_dte->addr.raw));
1648 /*
1649 * Does the B table have a valid C table
1650 * under the corresponding table entry?
1651 */
1652 *b_idx = MMU_TIB(va);
1653 b_dte = &((*b_tbl)->bt_dtbl[*b_idx]);
1654 if (!MMU_VALID_DT(*b_dte))
1655 return FALSE; /* No. Return unknown. */
1656 /* Yes. Extract C table from the B table. */
1657 *c_tbl = mmuC2tmgr(mmu_ptov(MMU_DTE_PA(*b_dte)));
1658 *pte_idx = MMU_TIC(va);
1659 *pte = &((*c_tbl)->ct_dtbl[*pte_idx]);
1660
1661 return TRUE;
1662 }
1663
1664 /* pmap_enter INTERFACE
1665 **
1666 * Called by the kernel to map a virtual address
1667 * to a physical address in the given process map.
1668 *
1669 * Note: this function should apply an exclusive lock
1670 * on the pmap system for its duration. (it certainly
1671 * would save my hair!!)
1672 * This function ought to be easier to read.
1673 */
1674 int
1675 pmap_enter(pmap_t pmap, vaddr_t va, paddr_t pa, vm_prot_t prot, int flags)
1676 {
1677 bool insert, managed; /* Marks the need for PV insertion.*/
1678 u_short nidx; /* PV list index */
1679 int mapflags; /* Flags for the mapping (see NOTE1) */
1680 u_int a_idx, b_idx, pte_idx; /* table indices */
1681 a_tmgr_t *a_tbl; /* A: long descriptor table manager */
1682 b_tmgr_t *b_tbl; /* B: short descriptor table manager */
1683 c_tmgr_t *c_tbl; /* C: short page table manager */
1684 mmu_long_dte_t *a_dte; /* A: long descriptor table */
1685 mmu_short_dte_t *b_dte; /* B: short descriptor table */
1686 mmu_short_pte_t *c_pte; /* C: short page descriptor table */
1687 pv_t *pv; /* pv list head */
1688 bool wired; /* is the mapping to be wired? */
1689 enum {NONE, NEWA, NEWB, NEWC} llevel; /* used at end */
1690
1691 if (pmap == pmap_kernel()) {
1692 pmap_enter_kernel(va, pa, prot);
1693 return 0;
1694 }
1695
1696 /*
1697 * Determine if the mapping should be wired.
1698 */
1699 wired = ((flags & PMAP_WIRED) != 0);
1700
1701 /*
1702 * NOTE1:
1703 *
1704 * On November 13, 1999, someone changed the pmap_enter() API such
1705 * that it now accepts a 'flags' argument. This new argument
1706 * contains bit-flags for the architecture-independent (UVM) system to
1707 * use in signalling certain mapping requirements to the architecture-
1708 * dependent (pmap) system. The argument it replaces, 'wired', is now
1709 * one of the flags within it.
1710 *
1711 * In addition to flags signaled by the architecture-independent
1712 * system, parts of the architecture-dependent section of the sun3x
1713 * kernel pass their own flags in the lower, unused bits of the
1714 * physical address supplied to this function. These flags are
1715 * extracted and stored in the temporary variable 'mapflags'.
1716 *
1717 * Extract sun3x specific flags from the physical address.
1718 */
1719 mapflags = (pa & ~MMU_PAGE_MASK);
1720 pa &= MMU_PAGE_MASK;
1721
1722 /*
1723 * Determine if the physical address being mapped is on-board RAM.
1724 * Any other area of the address space is likely to belong to a
1725 * device and hence it would be disasterous to cache its contents.
1726 */
1727 if ((managed = is_managed(pa)) == FALSE)
1728 mapflags |= PMAP_NC;
1729
1730 /*
1731 * For user mappings we walk along the MMU tables of the given
1732 * pmap, reaching a PTE which describes the virtual page being
1733 * mapped or changed. If any level of the walk ends in an invalid
1734 * entry, a table must be allocated and the entry must be updated
1735 * to point to it.
1736 * There is a bit of confusion as to whether this code must be
1737 * re-entrant. For now we will assume it is. To support
1738 * re-entrancy we must unlink tables from the table pool before
1739 * we assume we may use them. Tables are re-linked into the pool
1740 * when we are finished with them at the end of the function.
1741 * But I don't feel like doing that until we have proof that this
1742 * needs to be re-entrant.
1743 * 'llevel' records which tables need to be relinked.
1744 */
1745 llevel = NONE;
1746
1747 /*
1748 * Step 1 - Retrieve the A table from the pmap. If it has no
1749 * A table, allocate a new one from the available pool.
1750 */
1751
1752 a_tbl = pmap->pm_a_tmgr;
1753 if (a_tbl == NULL) {
1754 /*
1755 * This pmap does not currently have an A table. Allocate
1756 * a new one.
1757 */
1758 a_tbl = get_a_table();
1759 a_tbl->at_parent = pmap;
1760
1761 /*
1762 * Assign this new A table to the pmap, and calculate its
1763 * physical address so that loadcrp() can be used to make
1764 * the table active.
1765 */
1766 pmap->pm_a_tmgr = a_tbl;
1767 pmap->pm_a_phys = mmu_vtop(a_tbl->at_dtbl);
1768
1769 /*
1770 * If the process receiving a new A table is the current
1771 * process, we are responsible for setting the MMU so that
1772 * it becomes the current address space. This only adds
1773 * new mappings, so no need to flush anything.
1774 */
1775 if (pmap == current_pmap()) {
1776 kernel_crp.rp_addr = pmap->pm_a_phys;
1777 loadcrp(&kernel_crp);
1778 }
1779
1780 if (!wired)
1781 llevel = NEWA;
1782 } else {
1783 /*
1784 * Use the A table already allocated for this pmap.
1785 * Unlink it from the A table pool if necessary.
1786 */
1787 if (wired && !a_tbl->at_wcnt)
1788 TAILQ_REMOVE(&a_pool, a_tbl, at_link);
1789 }
1790
1791 /*
1792 * Step 2 - Walk into the B table. If there is no valid B table,
1793 * allocate one.
1794 */
1795
1796 a_idx = MMU_TIA(va); /* Calculate the TIA of the VA. */
1797 a_dte = &a_tbl->at_dtbl[a_idx]; /* Retrieve descriptor from table */
1798 if (MMU_VALID_DT(*a_dte)) { /* Is the descriptor valid? */
1799 /* The descriptor is valid. Use the B table it points to. */
1800 /*************************************
1801 * a_idx *
1802 * v *
1803 * a_tbl -> +-+-+-+-+-+-+-+-+-+-+-+- *
1804 * | | | | | | | | | | | | *
1805 * +-+-+-+-+-+-+-+-+-+-+-+- *
1806 * | *
1807 * \- b_tbl -> +-+- *
1808 * | | *
1809 * +-+- *
1810 *************************************/
1811 b_dte = mmu_ptov(a_dte->addr.raw);
1812 b_tbl = mmuB2tmgr(b_dte);
1813
1814 /*
1815 * If the requested mapping must be wired, but this table
1816 * being used to map it is not, the table must be removed
1817 * from the available pool and its wired entry count
1818 * incremented.
1819 */
1820 if (wired && !b_tbl->bt_wcnt) {
1821 TAILQ_REMOVE(&b_pool, b_tbl, bt_link);
1822 a_tbl->at_wcnt++;
1823 }
1824 } else {
1825 /* The descriptor is invalid. Allocate a new B table. */
1826 b_tbl = get_b_table();
1827
1828 /* Point the parent A table descriptor to this new B table. */
1829 a_dte->addr.raw = mmu_vtop(b_tbl->bt_dtbl);
1830 a_dte->attr.raw = MMU_LONG_DTE_LU | MMU_DT_SHORT;
1831 a_tbl->at_ecnt++; /* Update parent's valid entry count */
1832
1833 /* Create the necessary back references to the parent table */
1834 b_tbl->bt_parent = a_tbl;
1835 b_tbl->bt_pidx = a_idx;
1836
1837 /*
1838 * If this table is to be wired, make sure the parent A table
1839 * wired count is updated to reflect that it has another wired
1840 * entry.
1841 */
1842 if (wired)
1843 a_tbl->at_wcnt++;
1844 else if (llevel == NONE)
1845 llevel = NEWB;
1846 }
1847
1848 /*
1849 * Step 3 - Walk into the C table, if there is no valid C table,
1850 * allocate one.
1851 */
1852
1853 b_idx = MMU_TIB(va); /* Calculate the TIB of the VA */
1854 b_dte = &b_tbl->bt_dtbl[b_idx]; /* Retrieve descriptor from table */
1855 if (MMU_VALID_DT(*b_dte)) { /* Is the descriptor valid? */
1856 /* The descriptor is valid. Use the C table it points to. */
1857 /**************************************
1858 * c_idx *
1859 * | v *
1860 * \- b_tbl -> +-+-+-+-+-+-+-+-+-+-+- *
1861 * | | | | | | | | | | | *
1862 * +-+-+-+-+-+-+-+-+-+-+- *
1863 * | *
1864 * \- c_tbl -> +-+-- *
1865 * | | | *
1866 * +-+-- *
1867 **************************************/
1868 c_pte = mmu_ptov(MMU_PTE_PA(*b_dte));
1869 c_tbl = mmuC2tmgr(c_pte);
1870
1871 /* If mapping is wired and table is not */
1872 if (wired && !c_tbl->ct_wcnt) {
1873 TAILQ_REMOVE(&c_pool, c_tbl, ct_link);
1874 b_tbl->bt_wcnt++;
1875 }
1876 } else {
1877 /* The descriptor is invalid. Allocate a new C table. */
1878 c_tbl = get_c_table();
1879
1880 /* Point the parent B table descriptor to this new C table. */
1881 b_dte->attr.raw = mmu_vtop(c_tbl->ct_dtbl);
1882 b_dte->attr.raw |= MMU_DT_SHORT;
1883 b_tbl->bt_ecnt++; /* Update parent's valid entry count */
1884
1885 /* Create the necessary back references to the parent table */
1886 c_tbl->ct_parent = b_tbl;
1887 c_tbl->ct_pidx = b_idx;
1888 /*
1889 * Store the pmap and base virtual managed address for faster
1890 * retrieval in the PV functions.
1891 */
1892 c_tbl->ct_pmap = pmap;
1893 c_tbl->ct_va = (va & (MMU_TIA_MASK|MMU_TIB_MASK));
1894
1895 /*
1896 * If this table is to be wired, make sure the parent B table
1897 * wired count is updated to reflect that it has another wired
1898 * entry.
1899 */
1900 if (wired)
1901 b_tbl->bt_wcnt++;
1902 else if (llevel == NONE)
1903 llevel = NEWC;
1904 }
1905
1906 /*
1907 * Step 4 - Deposit a page descriptor (PTE) into the appropriate
1908 * slot of the C table, describing the PA to which the VA is mapped.
1909 */
1910
1911 pte_idx = MMU_TIC(va);
1912 c_pte = &c_tbl->ct_dtbl[pte_idx];
1913 if (MMU_VALID_DT(*c_pte)) { /* Is the entry currently valid? */
1914 /*
1915 * The PTE is currently valid. This particular call
1916 * is just a synonym for one (or more) of the following
1917 * operations:
1918 * change protection of a page
1919 * change wiring status of a page
1920 * remove the mapping of a page
1921 */
1922
1923 /* First check if this is a wiring operation. */
1924 if (c_pte->attr.raw & MMU_SHORT_PTE_WIRED) {
1925 /*
1926 * The existing mapping is wired, so adjust wired
1927 * entry count here. If new mapping is still wired,
1928 * wired entry count will be incremented again later.
1929 */
1930 c_tbl->ct_wcnt--;
1931 if (!wired) {
1932 /*
1933 * The mapping of this PTE is being changed
1934 * from wired to unwired.
1935 * Adjust wired entry counts in each table and
1936 * set llevel flag to put unwired tables back
1937 * into the active pool.
1938 */
1939 if (c_tbl->ct_wcnt == 0) {
1940 llevel = NEWC;
1941 if (--b_tbl->bt_wcnt == 0) {
1942 llevel = NEWB;
1943 if (--a_tbl->at_wcnt == 0) {
1944 llevel = NEWA;
1945 }
1946 }
1947 }
1948 }
1949 }
1950
1951 /* Is the new address the same as the old? */
1952 if (MMU_PTE_PA(*c_pte) == pa) {
1953 /*
1954 * Yes, mark that it does not need to be reinserted
1955 * into the PV list.
1956 */
1957 insert = FALSE;
1958
1959 /*
1960 * Clear all but the modified, referenced and wired
1961 * bits on the PTE.
1962 */
1963 c_pte->attr.raw &= (MMU_SHORT_PTE_M
1964 | MMU_SHORT_PTE_USED | MMU_SHORT_PTE_WIRED);
1965 } else {
1966 /* No, remove the old entry */
1967 pmap_remove_pte(c_pte);
1968 insert = TRUE;
1969 }
1970
1971 /*
1972 * TLB flush is only necessary if modifying current map.
1973 * However, in pmap_enter(), the pmap almost always IS
1974 * the current pmap, so don't even bother to check.
1975 */
1976 TBIS(va);
1977 } else {
1978 /*
1979 * The PTE is invalid. Increment the valid entry count in
1980 * the C table manager to reflect the addition of a new entry.
1981 */
1982 c_tbl->ct_ecnt++;
1983
1984 /* XXX - temporarily make sure the PTE is cleared. */
1985 c_pte->attr.raw = 0;
1986
1987 /* It will also need to be inserted into the PV list. */
1988 insert = TRUE;
1989 }
1990
1991 /*
1992 * If page is changing from unwired to wired status, set an unused bit
1993 * within the PTE to indicate that it is wired. Also increment the
1994 * wired entry count in the C table manager.
1995 */
1996 if (wired) {
1997 c_pte->attr.raw |= MMU_SHORT_PTE_WIRED;
1998 c_tbl->ct_wcnt++;
1999 }
2000
2001 /*
2002 * Map the page, being careful to preserve modify/reference/wired
2003 * bits. At this point it is assumed that the PTE either has no bits
2004 * set, or if there are set bits, they are only modified, reference or
2005 * wired bits. If not, the following statement will cause erratic
2006 * behavior.
2007 */
2008 #ifdef PMAP_DEBUG
2009 if (c_pte->attr.raw & ~(MMU_SHORT_PTE_M |
2010 MMU_SHORT_PTE_USED | MMU_SHORT_PTE_WIRED)) {
2011 printf("pmap_enter: junk left in PTE at %p\n", c_pte);
2012 Debugger();
2013 }
2014 #endif
2015 c_pte->attr.raw |= ((u_long) pa | MMU_DT_PAGE);
2016
2017 /*
2018 * If the mapping should be read-only, set the write protect
2019 * bit in the PTE.
2020 */
2021 if (!(prot & VM_PROT_WRITE))
2022 c_pte->attr.raw |= MMU_SHORT_PTE_WP;
2023
2024 /*
2025 * Mark the PTE as used and/or modified as specified by the flags arg.
2026 */
2027 if (flags & VM_PROT_ALL) {
2028 c_pte->attr.raw |= MMU_SHORT_PTE_USED;
2029 if (flags & VM_PROT_WRITE) {
2030 c_pte->attr.raw |= MMU_SHORT_PTE_M;
2031 }
2032 }
2033
2034 /*
2035 * If the mapping should be cache inhibited (indicated by the flag
2036 * bits found on the lower order of the physical address.)
2037 * mark the PTE as a cache inhibited page.
2038 */
2039 if (mapflags & PMAP_NC)
2040 c_pte->attr.raw |= MMU_SHORT_PTE_CI;
2041
2042 /*
2043 * If the physical address being mapped is managed by the PV
2044 * system then link the pte into the list of pages mapped to that
2045 * address.
2046 */
2047 if (insert && managed) {
2048 pv = pa2pv(pa);
2049 nidx = pteidx(c_pte);
2050
2051 pvebase[nidx].pve_next = pv->pv_idx;
2052 pv->pv_idx = nidx;
2053 }
2054
2055 /* Move any allocated or unwired tables back into the active pool. */
2056
2057 switch (llevel) {
2058 case NEWA:
2059 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
2060 /* FALLTHROUGH */
2061 case NEWB:
2062 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
2063 /* FALLTHROUGH */
2064 case NEWC:
2065 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
2066 /* FALLTHROUGH */
2067 default:
2068 break;
2069 }
2070
2071 return 0;
2072 }
2073
2074 /* pmap_enter_kernel INTERNAL
2075 **
2076 * Map the given virtual address to the given physical address within the
2077 * kernel address space. This function exists because the kernel map does
2078 * not do dynamic table allocation. It consists of a contiguous array of ptes
2079 * and can be edited directly without the need to walk through any tables.
2080 *
2081 * XXX: "Danger, Will Robinson!"
2082 * Note that the kernel should never take a fault on any page
2083 * between [ KERNBASE .. virtual_avail ] and this is checked in
2084 * trap.c for kernel-mode MMU faults. This means that mappings
2085 * created in that range must be implicily wired. -gwr
2086 */
2087 void
2088 pmap_enter_kernel(vaddr_t va, paddr_t pa, vm_prot_t prot)
2089 {
2090 bool was_valid, insert;
2091 u_short pte_idx;
2092 int flags;
2093 mmu_short_pte_t *pte;
2094 pv_t *pv;
2095 paddr_t old_pa;
2096
2097 flags = (pa & ~MMU_PAGE_MASK);
2098 pa &= MMU_PAGE_MASK;
2099
2100 if (is_managed(pa))
2101 insert = TRUE;
2102 else
2103 insert = FALSE;
2104
2105 /*
2106 * Calculate the index of the PTE being modified.
2107 */
2108 pte_idx = (u_long)m68k_btop(va - KERNBASE);
2109
2110 /* This array is traditionally named "Sysmap" */
2111 pte = &kernCbase[pte_idx];
2112
2113 if (MMU_VALID_DT(*pte)) {
2114 was_valid = TRUE;
2115 /*
2116 * If the PTE already maps a different
2117 * physical address, umap and pv_unlink.
2118 */
2119 old_pa = MMU_PTE_PA(*pte);
2120 if (pa != old_pa)
2121 pmap_remove_pte(pte);
2122 else {
2123 /*
2124 * Old PA and new PA are the same. No need to
2125 * relink the mapping within the PV list.
2126 */
2127 insert = FALSE;
2128
2129 /*
2130 * Save any mod/ref bits on the PTE.
2131 */
2132 pte->attr.raw &= (MMU_SHORT_PTE_USED|MMU_SHORT_PTE_M);
2133 }
2134 } else {
2135 pte->attr.raw = MMU_DT_INVALID;
2136 was_valid = FALSE;
2137 }
2138
2139 /*
2140 * Map the page. Being careful to preserve modified/referenced bits
2141 * on the PTE.
2142 */
2143 pte->attr.raw |= (pa | MMU_DT_PAGE);
2144
2145 if (!(prot & VM_PROT_WRITE)) /* If access should be read-only */
2146 pte->attr.raw |= MMU_SHORT_PTE_WP;
2147 if (flags & PMAP_NC)
2148 pte->attr.raw |= MMU_SHORT_PTE_CI;
2149 if (was_valid)
2150 TBIS(va);
2151
2152 /*
2153 * Insert the PTE into the PV system, if need be.
2154 */
2155 if (insert) {
2156 pv = pa2pv(pa);
2157 pvebase[pte_idx].pve_next = pv->pv_idx;
2158 pv->pv_idx = pte_idx;
2159 }
2160 }
2161
2162 void
2163 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot)
2164 {
2165 mmu_short_pte_t *pte;
2166
2167 /* This array is traditionally named "Sysmap" */
2168 pte = &kernCbase[(u_long)m68k_btop(va - KERNBASE)];
2169
2170 KASSERT(!MMU_VALID_DT(*pte));
2171 pte->attr.raw = MMU_DT_INVALID | MMU_DT_PAGE | (pa & MMU_PAGE_MASK);
2172 if (!(prot & VM_PROT_WRITE))
2173 pte->attr.raw |= MMU_SHORT_PTE_WP;
2174 }
2175
2176 void
2177 pmap_kremove(vaddr_t va, vsize_t len)
2178 {
2179 int idx, eidx;
2180
2181 #ifdef PMAP_DEBUG
2182 if ((va & PGOFSET) || (len & PGOFSET))
2183 panic("pmap_kremove: alignment");
2184 #endif
2185
2186 idx = m68k_btop(va - KERNBASE);
2187 eidx = m68k_btop(va + len - KERNBASE);
2188
2189 while (idx < eidx) {
2190 kernCbase[idx++].attr.raw = MMU_DT_INVALID;
2191 TBIS(va);
2192 va += PAGE_SIZE;
2193 }
2194 }
2195
2196 /* pmap_map INTERNAL
2197 **
2198 * Map a contiguous range of physical memory into a contiguous range of
2199 * the kernel virtual address space.
2200 *
2201 * Used for device mappings and early mapping of the kernel text/data/bss.
2202 * Returns the first virtual address beyond the end of the range.
2203 */
2204 vaddr_t
2205 pmap_map(vaddr_t va, paddr_t pa, paddr_t endpa, int prot)
2206 {
2207 int sz;
2208
2209 sz = endpa - pa;
2210 do {
2211 pmap_enter_kernel(va, pa, prot);
2212 va += PAGE_SIZE;
2213 pa += PAGE_SIZE;
2214 sz -= PAGE_SIZE;
2215 } while (sz > 0);
2216 pmap_update(pmap_kernel());
2217 return va;
2218 }
2219
2220 /* pmap_protect_kernel INTERNAL
2221 **
2222 * Apply the given protection code to a kernel address range.
2223 */
2224 static INLINE void
2225 pmap_protect_kernel(vaddr_t startva, vaddr_t endva, vm_prot_t prot)
2226 {
2227 vaddr_t va;
2228 mmu_short_pte_t *pte;
2229
2230 pte = &kernCbase[(unsigned long) m68k_btop(startva - KERNBASE)];
2231 for (va = startva; va < endva; va += PAGE_SIZE, pte++) {
2232 if (MMU_VALID_DT(*pte)) {
2233 switch (prot) {
2234 case VM_PROT_ALL:
2235 break;
2236 case VM_PROT_EXECUTE:
2237 case VM_PROT_READ:
2238 case VM_PROT_READ|VM_PROT_EXECUTE:
2239 pte->attr.raw |= MMU_SHORT_PTE_WP;
2240 break;
2241 case VM_PROT_NONE:
2242 /* this is an alias for 'pmap_remove_kernel' */
2243 pmap_remove_pte(pte);
2244 break;
2245 default:
2246 break;
2247 }
2248 /*
2249 * since this is the kernel, immediately flush any cached
2250 * descriptors for this address.
2251 */
2252 TBIS(va);
2253 }
2254 }
2255 }
2256
2257 /* pmap_protect INTERFACE
2258 **
2259 * Apply the given protection to the given virtual address range within
2260 * the given map.
2261 *
2262 * It is ok for the protection applied to be stronger than what is
2263 * specified. We use this to our advantage when the given map has no
2264 * mapping for the virtual address. By skipping a page when this
2265 * is discovered, we are effectively applying a protection of VM_PROT_NONE,
2266 * and therefore do not need to map the page just to apply a protection
2267 * code. Only pmap_enter() needs to create new mappings if they do not exist.
2268 *
2269 * XXX - This function could be speeded up by using pmap_stroll() for inital
2270 * setup, and then manual scrolling in the for() loop.
2271 */
2272 void
2273 pmap_protect(pmap_t pmap, vaddr_t startva, vaddr_t endva, vm_prot_t prot)
2274 {
2275 bool iscurpmap;
2276 int a_idx, b_idx, c_idx;
2277 a_tmgr_t *a_tbl;
2278 b_tmgr_t *b_tbl;
2279 c_tmgr_t *c_tbl;
2280 mmu_short_pte_t *pte;
2281
2282 if (pmap == pmap_kernel()) {
2283 pmap_protect_kernel(startva, endva, prot);
2284 return;
2285 }
2286
2287 /*
2288 * In this particular pmap implementation, there are only three
2289 * types of memory protection: 'all' (read/write/execute),
2290 * 'read-only' (read/execute) and 'none' (no mapping.)
2291 * It is not possible for us to treat 'executable' as a separate
2292 * protection type. Therefore, protection requests that seek to
2293 * remove execute permission while retaining read or write, and those
2294 * that make little sense (write-only for example) are ignored.
2295 */
2296 switch (prot) {
2297 case VM_PROT_NONE:
2298 /*
2299 * A request to apply the protection code of
2300 * 'VM_PROT_NONE' is a synonym for pmap_remove().
2301 */
2302 pmap_remove(pmap, startva, endva);
2303 return;
2304 case VM_PROT_EXECUTE:
2305 case VM_PROT_READ:
2306 case VM_PROT_READ|VM_PROT_EXECUTE:
2307 /* continue */
2308 break;
2309 case VM_PROT_WRITE:
2310 case VM_PROT_WRITE|VM_PROT_READ:
2311 case VM_PROT_WRITE|VM_PROT_EXECUTE:
2312 case VM_PROT_ALL:
2313 /* None of these should happen in a sane system. */
2314 return;
2315 }
2316
2317 /*
2318 * If the pmap has no A table, it has no mappings and therefore
2319 * there is nothing to protect.
2320 */
2321 if ((a_tbl = pmap->pm_a_tmgr) == NULL)
2322 return;
2323
2324 a_idx = MMU_TIA(startva);
2325 b_idx = MMU_TIB(startva);
2326 c_idx = MMU_TIC(startva);
2327 b_tbl = NULL;
2328 c_tbl = NULL;
2329
2330 iscurpmap = (pmap == current_pmap());
2331 while (startva < endva) {
2332 if (b_tbl || MMU_VALID_DT(a_tbl->at_dtbl[a_idx])) {
2333 if (b_tbl == NULL) {
2334 b_tbl = (b_tmgr_t *) a_tbl->at_dtbl[a_idx].addr.raw;
2335 b_tbl = mmu_ptov((vaddr_t)b_tbl);
2336 b_tbl = mmuB2tmgr((mmu_short_dte_t *)b_tbl);
2337 }
2338 if (c_tbl || MMU_VALID_DT(b_tbl->bt_dtbl[b_idx])) {
2339 if (c_tbl == NULL) {
2340 c_tbl = (c_tmgr_t *) MMU_DTE_PA(b_tbl->bt_dtbl[b_idx]);
2341 c_tbl = mmu_ptov((vaddr_t)c_tbl);
2342 c_tbl = mmuC2tmgr((mmu_short_pte_t *)c_tbl);
2343 }
2344 if (MMU_VALID_DT(c_tbl->ct_dtbl[c_idx])) {
2345 pte = &c_tbl->ct_dtbl[c_idx];
2346 /* make the mapping read-only */
2347 pte->attr.raw |= MMU_SHORT_PTE_WP;
2348 /*
2349 * If we just modified the current address space,
2350 * flush any translations for the modified page from
2351 * the translation cache and any data from it in the
2352 * data cache.
2353 */
2354 if (iscurpmap)
2355 TBIS(startva);
2356 }
2357 startva += PAGE_SIZE;
2358
2359 if (++c_idx >= MMU_C_TBL_SIZE) { /* exceeded C table? */
2360 c_tbl = NULL;
2361 c_idx = 0;
2362 if (++b_idx >= MMU_B_TBL_SIZE) { /* exceeded B table? */
2363 b_tbl = NULL;
2364 b_idx = 0;
2365 }
2366 }
2367 } else { /* C table wasn't valid */
2368 c_tbl = NULL;
2369 c_idx = 0;
2370 startva += MMU_TIB_RANGE;
2371 if (++b_idx >= MMU_B_TBL_SIZE) { /* exceeded B table? */
2372 b_tbl = NULL;
2373 b_idx = 0;
2374 }
2375 } /* C table */
2376 } else { /* B table wasn't valid */
2377 b_tbl = NULL;
2378 b_idx = 0;
2379 startva += MMU_TIA_RANGE;
2380 a_idx++;
2381 } /* B table */
2382 }
2383 }
2384
2385 /* pmap_unwire INTERFACE
2386 **
2387 * Clear the wired attribute of the specified page.
2388 *
2389 * This function is called from vm_fault.c to unwire
2390 * a mapping.
2391 */
2392 void
2393 pmap_unwire(pmap_t pmap, vaddr_t va)
2394 {
2395 int a_idx, b_idx, c_idx;
2396 a_tmgr_t *a_tbl;
2397 b_tmgr_t *b_tbl;
2398 c_tmgr_t *c_tbl;
2399 mmu_short_pte_t *pte;
2400
2401 /* Kernel mappings always remain wired. */
2402 if (pmap == pmap_kernel())
2403 return;
2404
2405 /*
2406 * Walk through the tables. If the walk terminates without
2407 * a valid PTE then the address wasn't wired in the first place.
2408 * Return immediately.
2409 */
2410 if (pmap_stroll(pmap, va, &a_tbl, &b_tbl, &c_tbl, &pte, &a_idx,
2411 &b_idx, &c_idx) == FALSE)
2412 return;
2413
2414
2415 /* Is the PTE wired? If not, return. */
2416 if (!(pte->attr.raw & MMU_SHORT_PTE_WIRED))
2417 return;
2418
2419 /* Remove the wiring bit. */
2420 pte->attr.raw &= ~(MMU_SHORT_PTE_WIRED);
2421
2422 /*
2423 * Decrement the wired entry count in the C table.
2424 * If it reaches zero the following things happen:
2425 * 1. The table no longer has any wired entries and is considered
2426 * unwired.
2427 * 2. It is placed on the available queue.
2428 * 3. The parent table's wired entry count is decremented.
2429 * 4. If it reaches zero, this process repeats at step 1 and
2430 * stops at after reaching the A table.
2431 */
2432 if (--c_tbl->ct_wcnt == 0) {
2433 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
2434 if (--b_tbl->bt_wcnt == 0) {
2435 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
2436 if (--a_tbl->at_wcnt == 0) {
2437 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
2438 }
2439 }
2440 }
2441 }
2442
2443 /* pmap_copy INTERFACE
2444 **
2445 * Copy the mappings of a range of addresses in one pmap, into
2446 * the destination address of another.
2447 *
2448 * This routine is advisory. Should we one day decide that MMU tables
2449 * may be shared by more than one pmap, this function should be used to
2450 * link them together. Until that day however, we do nothing.
2451 */
2452 void
2453 pmap_copy(pmap_t pmap_a, pmap_t pmap_b, vaddr_t dst, vsize_t len, vaddr_t src)
2454 {
2455
2456 /* not implemented. */
2457 }
2458
2459 /* pmap_copy_page INTERFACE
2460 **
2461 * Copy the contents of one physical page into another.
2462 *
2463 * This function makes use of two virtual pages allocated in pmap_bootstrap()
2464 * to map the two specified physical pages into the kernel address space.
2465 *
2466 * Note: We could use the transparent translation registers to make the
2467 * mappings. If we do so, be sure to disable interrupts before using them.
2468 */
2469 void
2470 pmap_copy_page(paddr_t srcpa, paddr_t dstpa)
2471 {
2472 vaddr_t srcva, dstva;
2473 int s;
2474
2475 srcva = tmp_vpages[0];
2476 dstva = tmp_vpages[1];
2477
2478 s = splvm();
2479 #ifdef DIAGNOSTIC
2480 if (tmp_vpages_inuse++)
2481 panic("pmap_copy_page: temporary vpages are in use.");
2482 #endif
2483
2484 /* Map pages as non-cacheable to avoid cache polution? */
2485 pmap_kenter_pa(srcva, srcpa, VM_PROT_READ);
2486 pmap_kenter_pa(dstva, dstpa, VM_PROT_READ | VM_PROT_WRITE);
2487
2488 /* Hand-optimized version of bcopy(src, dst, PAGE_SIZE) */
2489 copypage((char *)srcva, (char *)dstva);
2490
2491 pmap_kremove(srcva, PAGE_SIZE);
2492 pmap_kremove(dstva, PAGE_SIZE);
2493
2494 #ifdef DIAGNOSTIC
2495 --tmp_vpages_inuse;
2496 #endif
2497 splx(s);
2498 }
2499
2500 /* pmap_zero_page INTERFACE
2501 **
2502 * Zero the contents of the specified physical page.
2503 *
2504 * Uses one of the virtual pages allocated in pmap_boostrap()
2505 * to map the specified page into the kernel address space.
2506 */
2507 void
2508 pmap_zero_page(paddr_t dstpa)
2509 {
2510 vaddr_t dstva;
2511 int s;
2512
2513 dstva = tmp_vpages[1];
2514 s = splvm();
2515 #ifdef DIAGNOSTIC
2516 if (tmp_vpages_inuse++)
2517 panic("pmap_zero_page: temporary vpages are in use.");
2518 #endif
2519
2520 /* The comments in pmap_copy_page() above apply here also. */
2521 pmap_kenter_pa(dstva, dstpa, VM_PROT_READ | VM_PROT_WRITE);
2522
2523 /* Hand-optimized version of bzero(ptr, PAGE_SIZE) */
2524 zeropage((char *)dstva);
2525
2526 pmap_kremove(dstva, PAGE_SIZE);
2527 #ifdef DIAGNOSTIC
2528 --tmp_vpages_inuse;
2529 #endif
2530 splx(s);
2531 }
2532
2533 /* pmap_collect INTERFACE
2534 **
2535 * Called from the VM system when we are about to swap out
2536 * the process using this pmap. This should give up any
2537 * resources held here, including all its MMU tables.
2538 */
2539 void
2540 pmap_collect(pmap_t pmap)
2541 {
2542
2543 /* XXX - todo... */
2544 }
2545
2546 /* pmap_pinit INTERNAL
2547 **
2548 * Initialize a pmap structure.
2549 */
2550 static INLINE void
2551 pmap_pinit(pmap_t pmap)
2552 {
2553
2554 memset(pmap, 0, sizeof(struct pmap));
2555 pmap->pm_a_tmgr = NULL;
2556 pmap->pm_a_phys = kernAphys;
2557 pmap->pm_refcount = 1;
2558 simple_lock_init(&pmap->pm_lock);
2559 }
2560
2561 /* pmap_create INTERFACE
2562 **
2563 * Create and return a pmap structure.
2564 */
2565 pmap_t
2566 pmap_create(void)
2567 {
2568 pmap_t pmap;
2569
2570 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK);
2571 pmap_pinit(pmap);
2572 return pmap;
2573 }
2574
2575 /* pmap_release INTERNAL
2576 **
2577 * Release any resources held by the given pmap.
2578 *
2579 * This is the reverse analog to pmap_pinit. It does not
2580 * necessarily mean for the pmap structure to be deallocated,
2581 * as in pmap_destroy.
2582 */
2583 static INLINE void
2584 pmap_release(pmap_t pmap)
2585 {
2586
2587 /*
2588 * As long as the pmap contains no mappings,
2589 * which always should be the case whenever
2590 * this function is called, there really should
2591 * be nothing to do.
2592 */
2593 #ifdef PMAP_DEBUG
2594 if (pmap == pmap_kernel())
2595 panic("pmap_release: kernel pmap");
2596 #endif
2597 /*
2598 * XXX - If this pmap has an A table, give it back.
2599 * The pmap SHOULD be empty by now, and pmap_remove
2600 * should have already given back the A table...
2601 * However, I see: pmap->pm_a_tmgr->at_ecnt == 1
2602 * at this point, which means some mapping was not
2603 * removed when it should have been. -gwr
2604 */
2605 if (pmap->pm_a_tmgr != NULL) {
2606 /* First make sure we are not using it! */
2607 if (kernel_crp.rp_addr == pmap->pm_a_phys) {
2608 kernel_crp.rp_addr = kernAphys;
2609 loadcrp(&kernel_crp);
2610 }
2611 #ifdef PMAP_DEBUG /* XXX - todo! */
2612 /* XXX - Now complain... */
2613 printf("pmap_release: still have table\n");
2614 Debugger();
2615 #endif
2616 free_a_table(pmap->pm_a_tmgr, TRUE);
2617 pmap->pm_a_tmgr = NULL;
2618 pmap->pm_a_phys = kernAphys;
2619 }
2620 }
2621
2622 /* pmap_reference INTERFACE
2623 **
2624 * Increment the reference count of a pmap.
2625 */
2626 void
2627 pmap_reference(pmap_t pmap)
2628 {
2629 pmap_lock(pmap);
2630 pmap_add_ref(pmap);
2631 pmap_unlock(pmap);
2632 }
2633
2634 /* pmap_dereference INTERNAL
2635 **
2636 * Decrease the reference count on the given pmap
2637 * by one and return the current count.
2638 */
2639 static INLINE int
2640 pmap_dereference(pmap_t pmap)
2641 {
2642 int rtn;
2643
2644 pmap_lock(pmap);
2645 rtn = pmap_del_ref(pmap);
2646 pmap_unlock(pmap);
2647
2648 return rtn;
2649 }
2650
2651 /* pmap_destroy INTERFACE
2652 **
2653 * Decrement a pmap's reference count and delete
2654 * the pmap if it becomes zero. Will be called
2655 * only after all mappings have been removed.
2656 */
2657 void
2658 pmap_destroy(pmap_t pmap)
2659 {
2660
2661 if (pmap_dereference(pmap) == 0) {
2662 pmap_release(pmap);
2663 pool_put(&pmap_pmap_pool, pmap);
2664 }
2665 }
2666
2667 /* pmap_is_referenced INTERFACE
2668 **
2669 * Determine if the given physical page has been
2670 * referenced (read from [or written to.])
2671 */
2672 bool
2673 pmap_is_referenced(struct vm_page *pg)
2674 {
2675 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2676 pv_t *pv;
2677 int idx;
2678
2679 /*
2680 * Check the flags on the pv head. If they are set,
2681 * return immediately. Otherwise a search must be done.
2682 */
2683
2684 pv = pa2pv(pa);
2685 if (pv->pv_flags & PV_FLAGS_USED)
2686 return TRUE;
2687
2688 /*
2689 * Search through all pv elements pointing
2690 * to this page and query their reference bits
2691 */
2692
2693 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) {
2694 if (MMU_PTE_USED(kernCbase[idx])) {
2695 return TRUE;
2696 }
2697 }
2698 return FALSE;
2699 }
2700
2701 /* pmap_is_modified INTERFACE
2702 **
2703 * Determine if the given physical page has been
2704 * modified (written to.)
2705 */
2706 bool
2707 pmap_is_modified(struct vm_page *pg)
2708 {
2709 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2710 pv_t *pv;
2711 int idx;
2712
2713 /* see comments in pmap_is_referenced() */
2714 pv = pa2pv(pa);
2715 if (pv->pv_flags & PV_FLAGS_MDFY)
2716 return TRUE;
2717
2718 for (idx = pv->pv_idx;
2719 idx != PVE_EOL;
2720 idx = pvebase[idx].pve_next) {
2721
2722 if (MMU_PTE_MODIFIED(kernCbase[idx])) {
2723 return TRUE;
2724 }
2725 }
2726
2727 return FALSE;
2728 }
2729
2730 /* pmap_page_protect INTERFACE
2731 **
2732 * Applies the given protection to all mappings to the given
2733 * physical page.
2734 */
2735 void
2736 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
2737 {
2738 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2739 pv_t *pv;
2740 int idx;
2741 vaddr_t va;
2742 struct mmu_short_pte_struct *pte;
2743 c_tmgr_t *c_tbl;
2744 pmap_t pmap, curpmap;
2745
2746 curpmap = current_pmap();
2747 pv = pa2pv(pa);
2748
2749 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) {
2750 pte = &kernCbase[idx];
2751 switch (prot) {
2752 case VM_PROT_ALL:
2753 /* do nothing */
2754 break;
2755 case VM_PROT_EXECUTE:
2756 case VM_PROT_READ:
2757 case VM_PROT_READ|VM_PROT_EXECUTE:
2758 /*
2759 * Determine the virtual address mapped by
2760 * the PTE and flush ATC entries if necessary.
2761 */
2762 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
2763 pte->attr.raw |= MMU_SHORT_PTE_WP;
2764 if (pmap == curpmap || pmap == pmap_kernel())
2765 TBIS(va);
2766 break;
2767 case VM_PROT_NONE:
2768 /* Save the mod/ref bits. */
2769 pv->pv_flags |= pte->attr.raw;
2770 /* Invalidate the PTE. */
2771 pte->attr.raw = MMU_DT_INVALID;
2772
2773 /*
2774 * Update table counts. And flush ATC entries
2775 * if necessary.
2776 */
2777 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
2778
2779 /*
2780 * If the PTE belongs to the kernel map,
2781 * be sure to flush the page it maps.
2782 */
2783 if (pmap == pmap_kernel()) {
2784 TBIS(va);
2785 } else {
2786 /*
2787 * The PTE belongs to a user map.
2788 * update the entry count in the C
2789 * table to which it belongs and flush
2790 * the ATC if the mapping belongs to
2791 * the current pmap.
2792 */
2793 c_tbl->ct_ecnt--;
2794 if (pmap == curpmap)
2795 TBIS(va);
2796 }
2797 break;
2798 default:
2799 break;
2800 }
2801 }
2802
2803 /*
2804 * If the protection code indicates that all mappings to the page
2805 * be removed, truncate the PV list to zero entries.
2806 */
2807 if (prot == VM_PROT_NONE)
2808 pv->pv_idx = PVE_EOL;
2809 }
2810
2811 /* pmap_get_pteinfo INTERNAL
2812 **
2813 * Called internally to find the pmap and virtual address within that
2814 * map to which the pte at the given index maps. Also includes the PTE's C
2815 * table manager.
2816 *
2817 * Returns the pmap in the argument provided, and the virtual address
2818 * by return value.
2819 */
2820 vaddr_t
2821 pmap_get_pteinfo(u_int idx, pmap_t *pmap, c_tmgr_t **tbl)
2822 {
2823 vaddr_t va = 0;
2824
2825 /*
2826 * Determine if the PTE is a kernel PTE or a user PTE.
2827 */
2828 if (idx >= NUM_KERN_PTES) {
2829 /*
2830 * The PTE belongs to a user mapping.
2831 */
2832 /* XXX: Would like an inline for this to validate idx... */
2833 *tbl = &Ctmgrbase[(idx - NUM_KERN_PTES) / MMU_C_TBL_SIZE];
2834
2835 *pmap = (*tbl)->ct_pmap;
2836 /*
2837 * To find the va to which the PTE maps, we first take
2838 * the table's base virtual address mapping which is stored
2839 * in ct_va. We then increment this address by a page for
2840 * every slot skipped until we reach the PTE.
2841 */
2842 va = (*tbl)->ct_va;
2843 va += m68k_ptob(idx % MMU_C_TBL_SIZE);
2844 } else {
2845 /*
2846 * The PTE belongs to the kernel map.
2847 */
2848 *pmap = pmap_kernel();
2849
2850 va = m68k_ptob(idx);
2851 va += KERNBASE;
2852 }
2853
2854 return va;
2855 }
2856
2857 /* pmap_clear_modify INTERFACE
2858 **
2859 * Clear the modification bit on the page at the specified
2860 * physical address.
2861 *
2862 */
2863 bool
2864 pmap_clear_modify(struct vm_page *pg)
2865 {
2866 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2867 bool rv;
2868
2869 rv = pmap_is_modified(pg);
2870 pmap_clear_pv(pa, PV_FLAGS_MDFY);
2871 return rv;
2872 }
2873
2874 /* pmap_clear_reference INTERFACE
2875 **
2876 * Clear the referenced bit on the page at the specified
2877 * physical address.
2878 */
2879 bool
2880 pmap_clear_reference(struct vm_page *pg)
2881 {
2882 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2883 bool rv;
2884
2885 rv = pmap_is_referenced(pg);
2886 pmap_clear_pv(pa, PV_FLAGS_USED);
2887 return rv;
2888 }
2889
2890 /* pmap_clear_pv INTERNAL
2891 **
2892 * Clears the specified flag from the specified physical address.
2893 * (Used by pmap_clear_modify() and pmap_clear_reference().)
2894 *
2895 * Flag is one of:
2896 * PV_FLAGS_MDFY - Page modified bit.
2897 * PV_FLAGS_USED - Page used (referenced) bit.
2898 *
2899 * This routine must not only clear the flag on the pv list
2900 * head. It must also clear the bit on every pte in the pv
2901 * list associated with the address.
2902 */
2903 void
2904 pmap_clear_pv(paddr_t pa, int flag)
2905 {
2906 pv_t *pv;
2907 int idx;
2908 vaddr_t va;
2909 pmap_t pmap;
2910 mmu_short_pte_t *pte;
2911 c_tmgr_t *c_tbl;
2912
2913 pv = pa2pv(pa);
2914 pv->pv_flags &= ~(flag);
2915 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) {
2916 pte = &kernCbase[idx];
2917 pte->attr.raw &= ~(flag);
2918
2919 /*
2920 * The MC68030 MMU will not set the modified or
2921 * referenced bits on any MMU tables for which it has
2922 * a cached descriptor with its modify bit set. To insure
2923 * that it will modify these bits on the PTE during the next
2924 * time it is written to or read from, we must flush it from
2925 * the ATC.
2926 *
2927 * Ordinarily it is only necessary to flush the descriptor
2928 * if it is used in the current address space. But since I
2929 * am not sure that there will always be a notion of
2930 * 'the current address space' when this function is called,
2931 * I will skip the test and always flush the address. It
2932 * does no harm.
2933 */
2934
2935 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
2936 TBIS(va);
2937 }
2938 }
2939
2940 /* pmap_extract_kernel INTERNAL
2941 **
2942 * Extract a translation from the kernel address space.
2943 */
2944 static INLINE bool
2945 pmap_extract_kernel(vaddr_t va, paddr_t *pap)
2946 {
2947 mmu_short_pte_t *pte;
2948
2949 pte = &kernCbase[(u_int)m68k_btop(va - KERNBASE)];
2950 if (!MMU_VALID_DT(*pte))
2951 return FALSE;
2952 if (pap != NULL)
2953 *pap = MMU_PTE_PA(*pte);
2954 return TRUE;
2955 }
2956
2957 /* pmap_extract INTERFACE
2958 **
2959 * Return the physical address mapped by the virtual address
2960 * in the specified pmap.
2961 *
2962 * Note: this function should also apply an exclusive lock
2963 * on the pmap system during its duration.
2964 */
2965 bool
2966 pmap_extract(pmap_t pmap, vaddr_t va, paddr_t *pap)
2967 {
2968 int a_idx, b_idx, pte_idx;
2969 a_tmgr_t *a_tbl;
2970 b_tmgr_t *b_tbl;
2971 c_tmgr_t *c_tbl;
2972 mmu_short_pte_t *c_pte;
2973
2974 if (pmap == pmap_kernel())
2975 return pmap_extract_kernel(va, pap);
2976
2977 if (pmap_stroll(pmap, va, &a_tbl, &b_tbl, &c_tbl,
2978 &c_pte, &a_idx, &b_idx, &pte_idx) == FALSE)
2979 return FALSE;
2980
2981 if (!MMU_VALID_DT(*c_pte))
2982 return FALSE;
2983
2984 if (pap != NULL)
2985 *pap = MMU_PTE_PA(*c_pte);
2986 return TRUE;
2987 }
2988
2989 /* pmap_remove_kernel INTERNAL
2990 **
2991 * Remove the mapping of a range of virtual addresses from the kernel map.
2992 * The arguments are already page-aligned.
2993 */
2994 static INLINE void
2995 pmap_remove_kernel(vaddr_t sva, vaddr_t eva)
2996 {
2997 int idx, eidx;
2998
2999 #ifdef PMAP_DEBUG
3000 if ((sva & PGOFSET) || (eva & PGOFSET))
3001 panic("pmap_remove_kernel: alignment");
3002 #endif
3003
3004 idx = m68k_btop(sva - KERNBASE);
3005 eidx = m68k_btop(eva - KERNBASE);
3006
3007 while (idx < eidx) {
3008 pmap_remove_pte(&kernCbase[idx++]);
3009 TBIS(sva);
3010 sva += PAGE_SIZE;
3011 }
3012 }
3013
3014 /* pmap_remove INTERFACE
3015 **
3016 * Remove the mapping of a range of virtual addresses from the given pmap.
3017 *
3018 */
3019 void
3020 pmap_remove(pmap_t pmap, vaddr_t sva, vaddr_t eva)
3021 {
3022
3023 if (pmap == pmap_kernel()) {
3024 pmap_remove_kernel(sva, eva);
3025 return;
3026 }
3027
3028 /*
3029 * If the pmap doesn't have an A table of its own, it has no mappings
3030 * that can be removed.
3031 */
3032 if (pmap->pm_a_tmgr == NULL)
3033 return;
3034
3035 /*
3036 * Remove the specified range from the pmap. If the function
3037 * returns true, the operation removed all the valid mappings
3038 * in the pmap and freed its A table. If this happened to the
3039 * currently loaded pmap, the MMU root pointer must be reloaded
3040 * with the default 'kernel' map.
3041 */
3042 if (pmap_remove_a(pmap->pm_a_tmgr, sva, eva)) {
3043 if (kernel_crp.rp_addr == pmap->pm_a_phys) {
3044 kernel_crp.rp_addr = kernAphys;
3045 loadcrp(&kernel_crp);
3046 /* will do TLB flush below */
3047 }
3048 pmap->pm_a_tmgr = NULL;
3049 pmap->pm_a_phys = kernAphys;
3050 }
3051
3052 /*
3053 * If we just modified the current address space,
3054 * make sure to flush the MMU cache.
3055 *
3056 * XXX - this could be an unecessarily large flush.
3057 * XXX - Could decide, based on the size of the VA range
3058 * to be removed, whether to flush "by pages" or "all".
3059 */
3060 if (pmap == current_pmap())
3061 TBIAU();
3062 }
3063
3064 /* pmap_remove_a INTERNAL
3065 **
3066 * This is function number one in a set of three that removes a range
3067 * of memory in the most efficient manner by removing the highest possible
3068 * tables from the memory space. This particular function attempts to remove
3069 * as many B tables as it can, delegating the remaining fragmented ranges to
3070 * pmap_remove_b().
3071 *
3072 * If the removal operation results in an empty A table, the function returns
3073 * TRUE.
3074 *
3075 * It's ugly but will do for now.
3076 */
3077 bool
3078 pmap_remove_a(a_tmgr_t *a_tbl, vaddr_t sva, vaddr_t eva)
3079 {
3080 bool empty;
3081 int idx;
3082 vaddr_t nstart, nend;
3083 b_tmgr_t *b_tbl;
3084 mmu_long_dte_t *a_dte;
3085 mmu_short_dte_t *b_dte;
3086 uint8_t at_wired, bt_wired;
3087
3088 /*
3089 * The following code works with what I call a 'granularity
3090 * reduction algorithim'. A range of addresses will always have
3091 * the following properties, which are classified according to
3092 * how the range relates to the size of the current granularity
3093 * - an A table entry:
3094 *
3095 * 1 2 3 4
3096 * -+---+---+---+---+---+---+---+-
3097 * -+---+---+---+---+---+---+---+-
3098 *
3099 * A range will always start on a granularity boundary, illustrated
3100 * by '+' signs in the table above, or it will start at some point
3101 * inbetween a granularity boundary, as illustrated by point 1.
3102 * The first step in removing a range of addresses is to remove the
3103 * range between 1 and 2, the nearest granularity boundary. This
3104 * job is handled by the section of code governed by the
3105 * 'if (start < nstart)' statement.
3106 *
3107 * A range will always encompass zero or more intergral granules,
3108 * illustrated by points 2 and 3. Integral granules are easy to
3109 * remove. The removal of these granules is the second step, and
3110 * is handled by the code block 'if (nstart < nend)'.
3111 *
3112 * Lastly, a range will always end on a granularity boundary,
3113 * ill. by point 3, or it will fall just beyond one, ill. by point
3114 * 4. The last step involves removing this range and is handled by
3115 * the code block 'if (nend < end)'.
3116 */
3117 nstart = MMU_ROUND_UP_A(sva);
3118 nend = MMU_ROUND_A(eva);
3119
3120 at_wired = a_tbl->at_wcnt;
3121
3122 if (sva < nstart) {
3123 /*
3124 * This block is executed if the range starts between
3125 * a granularity boundary.
3126 *
3127 * First find the DTE which is responsible for mapping
3128 * the start of the range.
3129 */
3130 idx = MMU_TIA(sva);
3131 a_dte = &a_tbl->at_dtbl[idx];
3132
3133 /*
3134 * If the DTE is valid then delegate the removal of the sub
3135 * range to pmap_remove_b(), which can remove addresses at
3136 * a finer granularity.
3137 */
3138 if (MMU_VALID_DT(*a_dte)) {
3139 b_dte = mmu_ptov(a_dte->addr.raw);
3140 b_tbl = mmuB2tmgr(b_dte);
3141 bt_wired = b_tbl->bt_wcnt;
3142
3143 /*
3144 * The sub range to be removed starts at the start
3145 * of the full range we were asked to remove, and ends
3146 * at the greater of:
3147 * 1. The end of the full range, -or-
3148 * 2. The end of the full range, rounded down to the
3149 * nearest granularity boundary.
3150 */
3151 if (eva < nstart)
3152 empty = pmap_remove_b(b_tbl, sva, eva);
3153 else
3154 empty = pmap_remove_b(b_tbl, sva, nstart);
3155
3156 /*
3157 * If the child table no longer has wired entries,
3158 * decrement wired entry count.
3159 */
3160 if (bt_wired && b_tbl->bt_wcnt == 0)
3161 a_tbl->at_wcnt--;
3162
3163 /*
3164 * If the removal resulted in an empty B table,
3165 * invalidate the DTE that points to it and decrement
3166 * the valid entry count of the A table.
3167 */
3168 if (empty) {
3169 a_dte->attr.raw = MMU_DT_INVALID;
3170 a_tbl->at_ecnt--;
3171 }
3172 }
3173 /*
3174 * If the DTE is invalid, the address range is already non-
3175 * existent and can simply be skipped.
3176 */
3177 }
3178 if (nstart < nend) {
3179 /*
3180 * This block is executed if the range spans a whole number
3181 * multiple of granules (A table entries.)
3182 *
3183 * First find the DTE which is responsible for mapping
3184 * the start of the first granule involved.
3185 */
3186 idx = MMU_TIA(nstart);
3187 a_dte = &a_tbl->at_dtbl[idx];
3188
3189 /*
3190 * Remove entire sub-granules (B tables) one at a time,
3191 * until reaching the end of the range.
3192 */
3193 for (; nstart < nend; a_dte++, nstart += MMU_TIA_RANGE)
3194 if (MMU_VALID_DT(*a_dte)) {
3195 /*
3196 * Find the B table manager for the
3197 * entry and free it.
3198 */
3199 b_dte = mmu_ptov(a_dte->addr.raw);
3200 b_tbl = mmuB2tmgr(b_dte);
3201 bt_wired = b_tbl->bt_wcnt;
3202
3203 free_b_table(b_tbl, TRUE);
3204
3205 /*
3206 * All child entries has been removed.
3207 * If there were any wired entries in it,
3208 * decrement wired entry count.
3209 */
3210 if (bt_wired)
3211 a_tbl->at_wcnt--;
3212
3213 /*
3214 * Invalidate the DTE that points to the
3215 * B table and decrement the valid entry
3216 * count of the A table.
3217 */
3218 a_dte->attr.raw = MMU_DT_INVALID;
3219 a_tbl->at_ecnt--;
3220 }
3221 }
3222 if (nend < eva) {
3223 /*
3224 * This block is executed if the range ends beyond a
3225 * granularity boundary.
3226 *
3227 * First find the DTE which is responsible for mapping
3228 * the start of the nearest (rounded down) granularity
3229 * boundary.
3230 */
3231 idx = MMU_TIA(nend);
3232 a_dte = &a_tbl->at_dtbl[idx];
3233
3234 /*
3235 * If the DTE is valid then delegate the removal of the sub
3236 * range to pmap_remove_b(), which can remove addresses at
3237 * a finer granularity.
3238 */
3239 if (MMU_VALID_DT(*a_dte)) {
3240 /*
3241 * Find the B table manager for the entry
3242 * and hand it to pmap_remove_b() along with
3243 * the sub range.
3244 */
3245 b_dte = mmu_ptov(a_dte->addr.raw);
3246 b_tbl = mmuB2tmgr(b_dte);
3247 bt_wired = b_tbl->bt_wcnt;
3248
3249 empty = pmap_remove_b(b_tbl, nend, eva);
3250
3251 /*
3252 * If the child table no longer has wired entries,
3253 * decrement wired entry count.
3254 */
3255 if (bt_wired && b_tbl->bt_wcnt == 0)
3256 a_tbl->at_wcnt--;
3257 /*
3258 * If the removal resulted in an empty B table,
3259 * invalidate the DTE that points to it and decrement
3260 * the valid entry count of the A table.
3261 */
3262 if (empty) {
3263 a_dte->attr.raw = MMU_DT_INVALID;
3264 a_tbl->at_ecnt--;
3265 }
3266 }
3267 }
3268
3269 /*
3270 * If there are no more entries in the A table, release it
3271 * back to the available pool and return TRUE.
3272 */
3273 if (a_tbl->at_ecnt == 0) {
3274 KASSERT(a_tbl->at_wcnt == 0);
3275 a_tbl->at_parent = NULL;
3276 if (!at_wired)
3277 TAILQ_REMOVE(&a_pool, a_tbl, at_link);
3278 TAILQ_INSERT_HEAD(&a_pool, a_tbl, at_link);
3279 empty = TRUE;
3280 } else {
3281 /*
3282 * If the table doesn't have wired entries any longer
3283 * but still has unwired entries, put it back into
3284 * the available queue.
3285 */
3286 if (at_wired && a_tbl->at_wcnt == 0)
3287 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link);
3288 empty = FALSE;
3289 }
3290
3291 return empty;
3292 }
3293
3294 /* pmap_remove_b INTERNAL
3295 **
3296 * Remove a range of addresses from an address space, trying to remove entire
3297 * C tables if possible.
3298 *
3299 * If the operation results in an empty B table, the function returns TRUE.
3300 */
3301 bool
3302 pmap_remove_b(b_tmgr_t *b_tbl, vaddr_t sva, vaddr_t eva)
3303 {
3304 bool empty;
3305 int idx;
3306 vaddr_t nstart, nend, rstart;
3307 c_tmgr_t *c_tbl;
3308 mmu_short_dte_t *b_dte;
3309 mmu_short_pte_t *c_dte;
3310 uint8_t bt_wired, ct_wired;
3311
3312 nstart = MMU_ROUND_UP_B(sva);
3313 nend = MMU_ROUND_B(eva);
3314
3315 bt_wired = b_tbl->bt_wcnt;
3316
3317 if (sva < nstart) {
3318 idx = MMU_TIB(sva);
3319 b_dte = &b_tbl->bt_dtbl[idx];
3320 if (MMU_VALID_DT(*b_dte)) {
3321 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte));
3322 c_tbl = mmuC2tmgr(c_dte);
3323 ct_wired = c_tbl->ct_wcnt;
3324
3325 if (eva < nstart)
3326 empty = pmap_remove_c(c_tbl, sva, eva);
3327 else
3328 empty = pmap_remove_c(c_tbl, sva, nstart);
3329
3330 /*
3331 * If the child table no longer has wired entries,
3332 * decrement wired entry count.
3333 */
3334 if (ct_wired && c_tbl->ct_wcnt == 0)
3335 b_tbl->bt_wcnt--;
3336
3337 if (empty) {
3338 b_dte->attr.raw = MMU_DT_INVALID;
3339 b_tbl->bt_ecnt--;
3340 }
3341 }
3342 }
3343 if (nstart < nend) {
3344 idx = MMU_TIB(nstart);
3345 b_dte = &b_tbl->bt_dtbl[idx];
3346 rstart = nstart;
3347 while (rstart < nend) {
3348 if (MMU_VALID_DT(*b_dte)) {
3349 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte));
3350 c_tbl = mmuC2tmgr(c_dte);
3351 ct_wired = c_tbl->ct_wcnt;
3352
3353 free_c_table(c_tbl, TRUE);
3354
3355 /*
3356 * All child entries has been removed.
3357 * If there were any wired entries in it,
3358 * decrement wired entry count.
3359 */
3360 if (ct_wired)
3361 b_tbl->bt_wcnt--;
3362
3363 b_dte->attr.raw = MMU_DT_INVALID;
3364 b_tbl->bt_ecnt--;
3365 }
3366 b_dte++;
3367 rstart += MMU_TIB_RANGE;
3368 }
3369 }
3370 if (nend < eva) {
3371 idx = MMU_TIB(nend);
3372 b_dte = &b_tbl->bt_dtbl[idx];
3373 if (MMU_VALID_DT(*b_dte)) {
3374 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte));
3375 c_tbl = mmuC2tmgr(c_dte);
3376 ct_wired = c_tbl->ct_wcnt;
3377 empty = pmap_remove_c(c_tbl, nend, eva);
3378
3379 /*
3380 * If the child table no longer has wired entries,
3381 * decrement wired entry count.
3382 */
3383 if (ct_wired && c_tbl->ct_wcnt == 0)
3384 b_tbl->bt_wcnt--;
3385
3386 if (empty) {
3387 b_dte->attr.raw = MMU_DT_INVALID;
3388 b_tbl->bt_ecnt--;
3389 }
3390 }
3391 }
3392
3393 if (b_tbl->bt_ecnt == 0) {
3394 KASSERT(b_tbl->bt_wcnt == 0);
3395 b_tbl->bt_parent = NULL;
3396 if (!bt_wired)
3397 TAILQ_REMOVE(&b_pool, b_tbl, bt_link);
3398 TAILQ_INSERT_HEAD(&b_pool, b_tbl, bt_link);
3399 empty = TRUE;
3400 } else {
3401 /*
3402 * If the table doesn't have wired entries any longer
3403 * but still has unwired entries, put it back into
3404 * the available queue.
3405 */
3406 if (bt_wired && b_tbl->bt_wcnt == 0)
3407 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link);
3408
3409 empty = FALSE;
3410 }
3411
3412 return empty;
3413 }
3414
3415 /* pmap_remove_c INTERNAL
3416 **
3417 * Remove a range of addresses from the given C table.
3418 */
3419 bool
3420 pmap_remove_c(c_tmgr_t *c_tbl, vaddr_t sva, vaddr_t eva)
3421 {
3422 bool empty;
3423 int idx;
3424 mmu_short_pte_t *c_pte;
3425 uint8_t ct_wired;
3426
3427 ct_wired = c_tbl->ct_wcnt;
3428
3429 idx = MMU_TIC(sva);
3430 c_pte = &c_tbl->ct_dtbl[idx];
3431 for (; sva < eva; sva += MMU_PAGE_SIZE, c_pte++) {
3432 if (MMU_VALID_DT(*c_pte)) {
3433 if (c_pte->attr.raw & MMU_SHORT_PTE_WIRED)
3434 c_tbl->ct_wcnt--;
3435 pmap_remove_pte(c_pte);
3436 c_tbl->ct_ecnt--;
3437 }
3438 }
3439
3440 if (c_tbl->ct_ecnt == 0) {
3441 KASSERT(c_tbl->ct_wcnt == 0);
3442 c_tbl->ct_parent = NULL;
3443 if (!ct_wired)
3444 TAILQ_REMOVE(&c_pool, c_tbl, ct_link);
3445 TAILQ_INSERT_HEAD(&c_pool, c_tbl, ct_link);
3446 empty = TRUE;
3447 } else {
3448 /*
3449 * If the table doesn't have wired entries any longer
3450 * but still has unwired entries, put it back into
3451 * the available queue.
3452 */
3453 if (ct_wired && c_tbl->ct_wcnt == 0)
3454 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link);
3455 empty = FALSE;
3456 }
3457
3458 return empty;
3459 }
3460
3461 /* pmap_bootstrap_alloc INTERNAL
3462 **
3463 * Used internally for memory allocation at startup when malloc is not
3464 * available. This code will fail once it crosses the first memory
3465 * bank boundary on the 3/80. Hopefully by then however, the VM system
3466 * will be in charge of allocation.
3467 */
3468 void *
3469 pmap_bootstrap_alloc(int size)
3470 {
3471 void *rtn;
3472
3473 #ifdef PMAP_DEBUG
3474 if (bootstrap_alloc_enabled == FALSE) {
3475 mon_printf("pmap_bootstrap_alloc: disabled\n");
3476 sunmon_abort();
3477 }
3478 #endif
3479
3480 rtn = (void *) virtual_avail;
3481 virtual_avail += size;
3482
3483 #ifdef PMAP_DEBUG
3484 if (virtual_avail > virtual_contig_end) {
3485 mon_printf("pmap_bootstrap_alloc: out of mem\n");
3486 sunmon_abort();
3487 }
3488 #endif
3489
3490 return rtn;
3491 }
3492
3493 /* pmap_bootstap_aalign INTERNAL
3494 **
3495 * Used to insure that the next call to pmap_bootstrap_alloc() will
3496 * return a chunk of memory aligned to the specified size.
3497 *
3498 * Note: This function will only support alignment sizes that are powers
3499 * of two.
3500 */
3501 void
3502 pmap_bootstrap_aalign(int size)
3503 {
3504 int off;
3505
3506 off = virtual_avail & (size - 1);
3507 if (off) {
3508 (void)pmap_bootstrap_alloc(size - off);
3509 }
3510 }
3511
3512 /* pmap_pa_exists
3513 **
3514 * Used by the /dev/mem driver to see if a given PA is memory
3515 * that can be mapped. (The PA is not in a hole.)
3516 */
3517 int
3518 pmap_pa_exists(paddr_t pa)
3519 {
3520 int i;
3521
3522 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
3523 if ((pa >= avail_mem[i].pmem_start) &&
3524 (pa < avail_mem[i].pmem_end))
3525 return 1;
3526 if (avail_mem[i].pmem_next == NULL)
3527 break;
3528 }
3529 return 0;
3530 }
3531
3532 /* Called only from locore.s and pmap.c */
3533 void _pmap_switch(pmap_t pmap);
3534
3535 /*
3536 * _pmap_switch INTERNAL
3537 *
3538 * This is called by locore.s:cpu_switch() when it is
3539 * switching to a new process. Load new translations.
3540 * Note: done in-line by locore.s unless PMAP_DEBUG
3541 *
3542 * Note that we do NOT allocate a context here, but
3543 * share the "kernel only" context until we really
3544 * need our own context for user-space mappings in
3545 * pmap_enter_user(). [ s/context/mmu A table/ ]
3546 */
3547 void
3548 _pmap_switch(pmap_t pmap)
3549 {
3550 u_long rootpa;
3551
3552 /*
3553 * Only do reload/flush if we have to.
3554 * Note that if the old and new process
3555 * were BOTH using the "null" context,
3556 * then this will NOT flush the TLB.
3557 */
3558 rootpa = pmap->pm_a_phys;
3559 if (kernel_crp.rp_addr != rootpa) {
3560 DPRINT(("pmap_activate(%p)\n", pmap));
3561 kernel_crp.rp_addr = rootpa;
3562 loadcrp(&kernel_crp);
3563 TBIAU();
3564 }
3565 }
3566
3567 /*
3568 * Exported version of pmap_activate(). This is called from the
3569 * machine-independent VM code when a process is given a new pmap.
3570 * If (p == curlwp) do like cpu_switch would do; otherwise just
3571 * take this as notification that the process has a new pmap.
3572 */
3573 void
3574 pmap_activate(struct lwp *l)
3575 {
3576
3577 if (l->l_proc == curproc) {
3578 _pmap_switch(l->l_proc->p_vmspace->vm_map.pmap);
3579 }
3580 }
3581
3582 /*
3583 * pmap_deactivate INTERFACE
3584 **
3585 * This is called to deactivate the specified process's address space.
3586 */
3587 void
3588 pmap_deactivate(struct lwp *l)
3589 {
3590
3591 /* Nothing to do. */
3592 }
3593
3594 /*
3595 * Fill in the sun3x-specific part of the kernel core header
3596 * for dumpsys(). (See machdep.c for the rest.)
3597 */
3598 void
3599 pmap_kcore_hdr(struct sun3x_kcore_hdr *sh)
3600 {
3601 u_long spa, len;
3602 int i;
3603
3604 sh->pg_frame = MMU_SHORT_PTE_BASEADDR;
3605 sh->pg_valid = MMU_DT_PAGE;
3606 sh->contig_end = virtual_contig_end;
3607 sh->kernCbase = (u_long)kernCbase;
3608 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
3609 spa = avail_mem[i].pmem_start;
3610 spa = m68k_trunc_page(spa);
3611 len = avail_mem[i].pmem_end - spa;
3612 len = m68k_round_page(len);
3613 sh->ram_segs[i].start = spa;
3614 sh->ram_segs[i].size = len;
3615 }
3616 }
3617
3618
3619 /* pmap_virtual_space INTERFACE
3620 **
3621 * Return the current available range of virtual addresses in the
3622 * arguuments provided. Only really called once.
3623 */
3624 void
3625 pmap_virtual_space(vaddr_t *vstart, vaddr_t *vend)
3626 {
3627
3628 *vstart = virtual_avail;
3629 *vend = virtual_end;
3630 }
3631
3632 /*
3633 * Provide memory to the VM system.
3634 *
3635 * Assume avail_start is always in the
3636 * first segment as pmap_bootstrap does.
3637 */
3638 static void
3639 pmap_page_upload(void)
3640 {
3641 paddr_t a, b; /* memory range */
3642 int i;
3643
3644 /* Supply the memory in segments. */
3645 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) {
3646 a = atop(avail_mem[i].pmem_start);
3647 b = atop(avail_mem[i].pmem_end);
3648 if (i == 0)
3649 a = atop(avail_start);
3650 if (avail_mem[i].pmem_end > avail_end)
3651 b = atop(avail_end);
3652
3653 uvm_page_physload(a, b, a, b, VM_FREELIST_DEFAULT);
3654
3655 if (avail_mem[i].pmem_next == NULL)
3656 break;
3657 }
3658 }
3659
3660 /* pmap_count INTERFACE
3661 **
3662 * Return the number of resident (valid) pages in the given pmap.
3663 *
3664 * Note: If this function is handed the kernel map, it will report
3665 * that it has no mappings. Hopefully the VM system won't ask for kernel
3666 * map statistics.
3667 */
3668 segsz_t
3669 pmap_count(pmap_t pmap, int type)
3670 {
3671 u_int count;
3672 int a_idx, b_idx;
3673 a_tmgr_t *a_tbl;
3674 b_tmgr_t *b_tbl;
3675 c_tmgr_t *c_tbl;
3676
3677 /*
3678 * If the pmap does not have its own A table manager, it has no
3679 * valid entires.
3680 */
3681 if (pmap->pm_a_tmgr == NULL)
3682 return 0;
3683
3684 a_tbl = pmap->pm_a_tmgr;
3685
3686 count = 0;
3687 for (a_idx = 0; a_idx < MMU_TIA(KERNBASE); a_idx++) {
3688 if (MMU_VALID_DT(a_tbl->at_dtbl[a_idx])) {
3689 b_tbl = mmuB2tmgr(mmu_ptov(a_tbl->at_dtbl[a_idx].addr.raw));
3690 for (b_idx = 0; b_idx < MMU_B_TBL_SIZE; b_idx++) {
3691 if (MMU_VALID_DT(b_tbl->bt_dtbl[b_idx])) {
3692 c_tbl = mmuC2tmgr(
3693 mmu_ptov(MMU_DTE_PA(b_tbl->bt_dtbl[b_idx])));
3694 if (type == 0)
3695 /*
3696 * A resident entry count has been requested.
3697 */
3698 count += c_tbl->ct_ecnt;
3699 else
3700 /*
3701 * A wired entry count has been requested.
3702 */
3703 count += c_tbl->ct_wcnt;
3704 }
3705 }
3706 }
3707 }
3708
3709 return count;
3710 }
3711
3712 /************************ SUN3 COMPATIBILITY ROUTINES ********************
3713 * The following routines are only used by DDB for tricky kernel text *
3714 * text operations in db_memrw.c. They are provided for sun3 *
3715 * compatibility. *
3716 *************************************************************************/
3717 /* get_pte INTERNAL
3718 **
3719 * Return the page descriptor the describes the kernel mapping
3720 * of the given virtual address.
3721 */
3722 extern u_long ptest_addr(u_long); /* XXX: locore.s */
3723 u_int
3724 get_pte(vaddr_t va)
3725 {
3726 u_long pte_pa;
3727 mmu_short_pte_t *pte;
3728
3729 /* Get the physical address of the PTE */
3730 pte_pa = ptest_addr(va & ~PGOFSET);
3731
3732 /* Convert to a virtual address... */
3733 pte = (mmu_short_pte_t *) (KERNBASE + pte_pa);
3734
3735 /* Make sure it is in our level-C tables... */
3736 if ((pte < kernCbase) ||
3737 (pte >= &mmuCbase[NUM_USER_PTES]))
3738 return 0;
3739
3740 /* ... and just return its contents. */
3741 return (pte->attr.raw);
3742 }
3743
3744
3745 /* set_pte INTERNAL
3746 **
3747 * Set the page descriptor that describes the kernel mapping
3748 * of the given virtual address.
3749 */
3750 void
3751 set_pte(vaddr_t va, u_int pte)
3752 {
3753 u_long idx;
3754
3755 if (va < KERNBASE)
3756 return;
3757
3758 idx = (unsigned long) m68k_btop(va - KERNBASE);
3759 kernCbase[idx].attr.raw = pte;
3760 TBIS(va);
3761 }
3762
3763 /*
3764 * Routine: pmap_procwr
3765 *
3766 * Function:
3767 * Synchronize caches corresponding to [addr, addr+len) in p.
3768 */
3769 void
3770 pmap_procwr(struct proc *p, vaddr_t va, size_t len)
3771 {
3772
3773 (void)cachectl1(0x80000004, va, len, p);
3774 }
3775
3776
3777 #ifdef PMAP_DEBUG
3778 /************************** DEBUGGING ROUTINES **************************
3779 * The following routines are meant to be an aid to debugging the pmap *
3780 * system. They are callable from the DDB command line and should be *
3781 * prepared to be handed unstable or incomplete states of the system. *
3782 ************************************************************************/
3783
3784 /* pv_list
3785 **
3786 * List all pages found on the pv list for the given physical page.
3787 * To avoid endless loops, the listing will stop at the end of the list
3788 * or after 'n' entries - whichever comes first.
3789 */
3790 void
3791 pv_list(paddr_t pa, int n)
3792 {
3793 int idx;
3794 vaddr_t va;
3795 pv_t *pv;
3796 c_tmgr_t *c_tbl;
3797 pmap_t pmap;
3798
3799 pv = pa2pv(pa);
3800 idx = pv->pv_idx;
3801 for (; idx != PVE_EOL && n > 0; idx = pvebase[idx].pve_next, n--) {
3802 va = pmap_get_pteinfo(idx, &pmap, &c_tbl);
3803 printf("idx %d, pmap 0x%x, va 0x%x, c_tbl %x\n",
3804 idx, (u_int) pmap, (u_int) va, (u_int) c_tbl);
3805 }
3806 }
3807 #endif /* PMAP_DEBUG */
3808
3809 #ifdef NOT_YET
3810 /* and maybe not ever */
3811 /************************** LOW-LEVEL ROUTINES **************************
3812 * These routines will eventually be re-written into assembly and placed*
3813 * in locore.s. They are here now as stubs so that the pmap module can *
3814 * be linked as a standalone user program for testing. *
3815 ************************************************************************/
3816 /* flush_atc_crp INTERNAL
3817 **
3818 * Flush all page descriptors derived from the given CPU Root Pointer
3819 * (CRP), or 'A' table as it is known here, from the 68851's automatic
3820 * cache.
3821 */
3822 void
3823 flush_atc_crp(int a_tbl)
3824 {
3825 mmu_long_rp_t rp;
3826
3827 /* Create a temporary root table pointer that points to the
3828 * given A table.
3829 */
3830 rp.attr.raw = ~MMU_LONG_RP_LU;
3831 rp.addr.raw = (unsigned int) a_tbl;
3832
3833 mmu_pflushr(&rp);
3834 /* mmu_pflushr:
3835 * movel sp(4)@,a0
3836 * pflushr a0@
3837 * rts
3838 */
3839 }
3840 #endif /* NOT_YET */
3841