Home | History | Annotate | Line # | Download | only in uvm
uvm_pglist.c revision 1.24
      1 /*	$NetBSD: uvm_pglist.c,v 1.24 2002/06/27 18:05:29 drochner Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 1997 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
      9  * NASA Ames Research Center.
     10  *
     11  * Redistribution and use in source and binary forms, with or without
     12  * modification, are permitted provided that the following conditions
     13  * are met:
     14  * 1. Redistributions of source code must retain the above copyright
     15  *    notice, this list of conditions and the following disclaimer.
     16  * 2. Redistributions in binary form must reproduce the above copyright
     17  *    notice, this list of conditions and the following disclaimer in the
     18  *    documentation and/or other materials provided with the distribution.
     19  * 3. All advertising materials mentioning features or use of this software
     20  *    must display the following acknowledgement:
     21  *      This product includes software developed by the NetBSD
     22  *      Foundation, Inc. and its contributors.
     23  * 4. Neither the name of The NetBSD Foundation nor the names of its
     24  *    contributors may be used to endorse or promote products derived
     25  *    from this software without specific prior written permission.
     26  *
     27  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     28  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     29  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     30  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     31  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     32  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     33  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     34  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     35  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     36  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     37  * POSSIBILITY OF SUCH DAMAGE.
     38  */
     39 
     40 /*
     41  * uvm_pglist.c: pglist functions
     42  */
     43 
     44 #include <sys/cdefs.h>
     45 __KERNEL_RCSID(0, "$NetBSD: uvm_pglist.c,v 1.24 2002/06/27 18:05:29 drochner Exp $");
     46 
     47 #include <sys/param.h>
     48 #include <sys/systm.h>
     49 #include <sys/malloc.h>
     50 #include <sys/proc.h>
     51 
     52 #include <uvm/uvm.h>
     53 
     54 #ifdef VM_PAGE_ALLOC_MEMORY_STATS
     55 #define	STAT_INCR(v)	(v)++
     56 #define	STAT_DECR(v)	do { \
     57 		if ((v) == 0) \
     58 			printf("%s:%d -- Already 0!\n", __FILE__, __LINE__); \
     59 		else \
     60 			(v)--; \
     61 	} while (0)
     62 u_long	uvm_pglistalloc_npages;
     63 #else
     64 #define	STAT_INCR(v)
     65 #define	STAT_DECR(v)
     66 #endif
     67 
     68 /*
     69  * uvm_pglistalloc: allocate a list of pages
     70  *
     71  * => allocated pages are placed at the tail of rlist.  rlist is
     72  *    assumed to be properly initialized by caller.
     73  * => returns 0 on success or errno on failure
     74  * => XXX: implementation allocates only a single segment, also
     75  *	might be able to better advantage of vm_physeg[].
     76  * => doesn't take into account clean non-busy pages on inactive list
     77  *	that could be used(?)
     78  * => params:
     79  *	size		the size of the allocation, rounded to page size.
     80  *	low		the low address of the allowed allocation range.
     81  *	high		the high address of the allowed allocation range.
     82  *	alignment	memory must be aligned to this power-of-two boundary.
     83  *	boundary	no segment in the allocation may cross this
     84  *			power-of-two boundary (relative to zero).
     85  */
     86 
     87 static void uvm_pglist_add(struct vm_page *, struct pglist *);
     88 static int uvm_pglistalloc_c_ps(struct vm_physseg *, int, paddr_t, paddr_t,
     89 				paddr_t, paddr_t, struct pglist *);
     90 static int uvm_pglistalloc_contig(int, paddr_t, paddr_t, paddr_t, paddr_t,
     91 				  struct pglist *);
     92 static int uvm_pglistalloc_s_ps(struct vm_physseg *, int, paddr_t, paddr_t,
     93 				struct pglist *);
     94 static int uvm_pglistalloc_simple(int, paddr_t, paddr_t,
     95 				  struct pglist *, int);
     96 
     97 static void
     98 uvm_pglist_add(pg, rlist)
     99 	struct vm_page *pg;
    100 	struct pglist *rlist;
    101 {
    102 	int free_list, color, pgflidx;
    103 #ifdef DEBUG
    104 	struct vm_page *tp;
    105 #endif
    106 
    107 #if PGFL_NQUEUES != 2
    108 #error uvm_pglistalloc needs to be updated
    109 #endif
    110 
    111 	free_list = uvm_page_lookup_freelist(pg);
    112 	color = VM_PGCOLOR_BUCKET(pg);
    113 	pgflidx = (pg->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN;
    114 #ifdef DEBUG
    115 	for (tp = TAILQ_FIRST(&uvm.page_free[
    116 		free_list].pgfl_buckets[color].pgfl_queues[pgflidx]);
    117 	     tp != NULL;
    118 	     tp = TAILQ_NEXT(tp, pageq)) {
    119 		if (tp == pg)
    120 			break;
    121 	}
    122 	if (tp == NULL)
    123 		panic("uvm_pglistalloc: page not on freelist");
    124 #endif
    125 	TAILQ_REMOVE(&uvm.page_free[free_list].pgfl_buckets[
    126 			color].pgfl_queues[pgflidx], pg, pageq);
    127 	uvmexp.free--;
    128 	if (pg->flags & PG_ZERO)
    129 		uvmexp.zeropages--;
    130 	pg->flags = PG_CLEAN;
    131 	pg->pqflags = 0;
    132 	pg->uobject = NULL;
    133 	pg->uanon = NULL;
    134 	TAILQ_INSERT_TAIL(rlist, pg, pageq);
    135 	STAT_INCR(uvm_pglistalloc_npages);
    136 }
    137 
    138 static int
    139 uvm_pglistalloc_c_ps(ps, num, low, high, alignment, boundary, rlist)
    140 	struct vm_physseg *ps;
    141 	int num;
    142 	paddr_t low, high, alignment, boundary;
    143 	struct pglist *rlist;
    144 {
    145 	int try, limit, tryidx, end, idx;
    146 	struct vm_page *pgs;
    147 	int pagemask;
    148 #ifdef DEBUG
    149 	paddr_t idxpa, lastidxpa;
    150 	int cidx;
    151 #endif
    152 #ifdef PGALLOC_VERBOSE
    153 	printf("pgalloc: contig %d pgs from psi %d\n", num, ps - vm_physmem);
    154 #endif
    155 
    156 	try = roundup(max(atop(low), ps->avail_start), atop(alignment));
    157 	limit = min(atop(high), ps->avail_end);
    158 	pagemask = ~((boundary >> PAGE_SHIFT) - 1);
    159 
    160 	for (;;) {
    161 		if (try + num > limit) {
    162 			/*
    163 			 * We've run past the allowable range.
    164 			 */
    165 			return (0); /* FAIL */
    166 		}
    167 		if (boundary != 0 &&
    168 		    ((try ^ (try + num - 1)) & pagemask) != 0) {
    169 			/*
    170 			 * Region crosses boundary. Jump to the boundary
    171 			 * just crossed and ensure alignment.
    172 			 */
    173 			try = (try + num - 1) & pagemask;
    174 			try = roundup(try, atop(alignment));
    175 			continue;
    176 		}
    177 #ifdef DEBUG
    178 		/*
    179 		 * Make sure this is a managed physical page.
    180 		 */
    181 
    182 		if (vm_physseg_find(try, &cidx) != ps - vm_physmem)
    183 			panic("pgalloc contig: botch1");
    184 		if (cidx != try - ps->start)
    185 			panic("pgalloc contig: botch2");
    186 		if (vm_physseg_find(try + num - 1, &cidx) != ps - vm_physmem)
    187 			panic("pgalloc contig: botch3");
    188 		if (cidx != try - ps->start + num - 1)
    189 			panic("pgalloc contig: botch4");
    190 #endif
    191 		tryidx = try - ps->start;
    192 		end = tryidx + num;
    193 		pgs = ps->pgs;
    194 
    195 		/*
    196 		 * Found a suitable starting page.  See if the range is free.
    197 		 */
    198 		for (idx = tryidx; idx < end; idx++) {
    199 			if (VM_PAGE_IS_FREE(&pgs[idx]) == 0)
    200 				break;
    201 
    202 #ifdef DEBUG
    203 			idxpa = VM_PAGE_TO_PHYS(&pgs[idx]);
    204 			if (idx > tryidx) {
    205 				lastidxpa = VM_PAGE_TO_PHYS(&pgs[idx - 1]);
    206 				if ((lastidxpa + PAGE_SIZE) != idxpa) {
    207 					/*
    208 					 * Region not contiguous.
    209 					 */
    210 					panic("pgalloc contig: botch5");
    211 				}
    212 				if (boundary != 0 &&
    213 				    ((lastidxpa ^ idxpa) & ~(boundary - 1))
    214 				    != 0) {
    215 					/*
    216 					 * Region crosses boundary.
    217 					 */
    218 					panic("pgalloc contig: botch6");
    219 				}
    220 			}
    221 #endif
    222 		}
    223 		if (idx == end)
    224 			break;
    225 
    226 		try += atop(alignment);
    227 	}
    228 
    229 	/*
    230 	 * we have a chunk of memory that conforms to the requested constraints.
    231 	 */
    232 	idx = tryidx;
    233 	while (idx < end)
    234 		uvm_pglist_add(&pgs[idx++], rlist);
    235 
    236 #ifdef PGALLOC_VERBOSE
    237 	printf("got %d pgs\n", num);
    238 #endif
    239 	return (num); /* number of pages allocated */
    240 }
    241 
    242 static int
    243 uvm_pglistalloc_contig(num, low, high, alignment, boundary, rlist)
    244 	int num;
    245 	paddr_t low, high, alignment, boundary;
    246 	struct pglist *rlist;
    247 {
    248 	int fl, psi;
    249 	struct vm_physseg *ps;
    250 	int s, error;
    251 
    252 	/* Default to "lose". */
    253 	error = ENOMEM;
    254 
    255 	/*
    256 	 * Block all memory allocation and lock the free list.
    257 	 */
    258 	s = uvm_lock_fpageq();
    259 
    260 	/* Are there even any free pages? */
    261 	if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
    262 		goto out;
    263 
    264 	for (fl = 0; fl < VM_NFREELIST; fl++) {
    265 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
    266 		for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
    267 #else
    268 		for (psi = 0 ; psi < vm_nphysseg ; psi++)
    269 #endif
    270 		{
    271 			ps = &vm_physmem[psi];
    272 
    273 			if (ps->free_list != fl)
    274 				continue;
    275 
    276 			num -= uvm_pglistalloc_c_ps(ps, num, low, high,
    277 						    alignment, boundary, rlist);
    278 			if (num == 0) {
    279 #ifdef PGALLOC_VERBOSE
    280 				printf("pgalloc: %lx-%lx\n",
    281 				       TAILQ_FIRST(rlist)->phys_addr,
    282 				       TAILQ_LAST(rlist, pglist)->phys_addr);
    283 #endif
    284 				error = 0;
    285 				goto out;
    286 			}
    287 		}
    288 	}
    289 
    290 out:
    291 	/*
    292 	 * check to see if we need to generate some free pages waking
    293 	 * the pagedaemon.
    294 	 */
    295 
    296 	UVM_KICK_PDAEMON();
    297 	uvm_unlock_fpageq(s);
    298 	return (error);
    299 }
    300 
    301 static int
    302 uvm_pglistalloc_s_ps(ps, num, low, high, rlist)
    303 	struct vm_physseg *ps;
    304 	int num;
    305 	paddr_t low, high;
    306 	struct pglist *rlist;
    307 {
    308 	int todo, limit, try;
    309 	struct vm_page *pg;
    310 #ifdef DEBUG
    311 	int cidx;
    312 #endif
    313 #ifdef PGALLOC_VERBOSE
    314 	printf("pgalloc: simple %d pgs from psi %d\n", num, ps - vm_physmem);
    315 #endif
    316 
    317 	todo = num;
    318 	limit = min(atop(high), ps->avail_end);
    319 
    320 	for (try = max(atop(low), ps->avail_start);
    321 	     try < limit; try ++) {
    322 #ifdef DEBUG
    323 		if (vm_physseg_find(try, &cidx) != ps - vm_physmem)
    324 			panic("pgalloc simple: botch1");
    325 		if (cidx != (try - ps->start))
    326 			panic("pgalloc simple: botch2");
    327 #endif
    328 		pg = &ps->pgs[try - ps->start];
    329 		if (VM_PAGE_IS_FREE(pg) == 0)
    330 			continue;
    331 
    332 		uvm_pglist_add(pg, rlist);
    333 		if (--todo == 0)
    334 			break;
    335 	}
    336 
    337 #ifdef PGALLOC_VERBOSE
    338 	printf("got %d pgs\n", num - todo);
    339 #endif
    340 	return (num - todo); /* number of pages allocated */
    341 }
    342 
    343 static int
    344 uvm_pglistalloc_simple(num, low, high, rlist, waitok)
    345 	int num;
    346 	paddr_t low, high;
    347 	struct pglist *rlist;
    348 	int waitok;
    349 {
    350 	int fl, psi, s, error;
    351 	struct vm_physseg *ps;
    352 
    353 	/* Default to "lose". */
    354 	error = ENOMEM;
    355 
    356 again:
    357 	/*
    358 	 * Block all memory allocation and lock the free list.
    359 	 */
    360 	s = uvm_lock_fpageq();
    361 
    362 	/* Are there even any free pages? */
    363 	if (uvmexp.free <= (uvmexp.reserve_pagedaemon + uvmexp.reserve_kernel))
    364 		goto out;
    365 
    366 	for (fl = 0; fl < VM_NFREELIST; fl++) {
    367 #if (VM_PHYSSEG_STRAT == VM_PSTRAT_BIGFIRST)
    368 		for (psi = vm_nphysseg - 1 ; psi >= 0 ; psi--)
    369 #else
    370 		for (psi = 0 ; psi < vm_nphysseg ; psi++)
    371 #endif
    372 		{
    373 			ps = &vm_physmem[psi];
    374 
    375 			if (ps->free_list != fl)
    376 				continue;
    377 
    378 			num -= uvm_pglistalloc_s_ps(ps, num, low, high, rlist);
    379 			if (num == 0) {
    380 				error = 0;
    381 				goto out;
    382 			}
    383 		}
    384 
    385 	}
    386 
    387 out:
    388 	/*
    389 	 * check to see if we need to generate some free pages waking
    390 	 * the pagedaemon.
    391 	 */
    392 
    393 	UVM_KICK_PDAEMON();
    394 	uvm_unlock_fpageq(s);
    395 	if (error) {
    396 		if (waitok) {
    397 			/* XXX perhaps some time limitation? */
    398 #ifdef DEBUG
    399 			printf("pglistalloc waiting\n");
    400 #endif
    401 			uvm_wait("pglalloc");
    402 			goto again;
    403 		} else
    404 			uvm_pglistfree(rlist);
    405 	}
    406 #ifdef PGALLOC_VERBOSE
    407 	if (!error)
    408 		printf("pgalloc: %lx..%lx\n",
    409 		       TAILQ_FIRST(rlist)->phys_addr,
    410 		       TAILQ_LAST(rlist, pglist)->phys_addr);
    411 #endif
    412 	return (error);
    413 }
    414 
    415 int
    416 uvm_pglistalloc(size, low, high, alignment, boundary, rlist, nsegs, waitok)
    417 	psize_t size;
    418 	paddr_t low, high, alignment, boundary;
    419 	struct pglist *rlist;
    420 	int nsegs, waitok;
    421 {
    422 	int num, res;
    423 
    424 	KASSERT((alignment & (alignment - 1)) == 0);
    425 	KASSERT((boundary & (boundary - 1)) == 0);
    426 
    427 	/*
    428 	 * Our allocations are always page granularity, so our alignment
    429 	 * must be, too.
    430 	 */
    431 	if (alignment < PAGE_SIZE)
    432 		alignment = PAGE_SIZE;
    433 	if (boundary != 0 && boundary < size)
    434 		return (EINVAL);
    435 	num = atop(round_page(size));
    436 	low = roundup(low, alignment);
    437 
    438 	TAILQ_INIT(rlist);
    439 
    440 	if ((nsegs < size >> PAGE_SHIFT) || (alignment != PAGE_SIZE) ||
    441 	    (boundary != 0))
    442 		res = uvm_pglistalloc_contig(num, low, high, alignment,
    443 					     boundary, rlist);
    444 	else
    445 		res = uvm_pglistalloc_simple(num, low, high, rlist, waitok);
    446 
    447 	return (res);
    448 }
    449 
    450 /*
    451  * uvm_pglistfree: free a list of pages
    452  *
    453  * => pages should already be unmapped
    454  */
    455 
    456 void
    457 uvm_pglistfree(list)
    458 	struct pglist *list;
    459 {
    460 	struct vm_page *pg;
    461 	int s;
    462 
    463 	/*
    464 	 * Lock the free list and free each page.
    465 	 */
    466 
    467 	s = uvm_lock_fpageq();
    468 	while ((pg = TAILQ_FIRST(list)) != NULL) {
    469 		KASSERT((pg->pqflags & (PQ_ACTIVE|PQ_INACTIVE)) == 0);
    470 		TAILQ_REMOVE(list, pg, pageq);
    471 		pg->pqflags = PQ_FREE;
    472 		TAILQ_INSERT_TAIL(&uvm.page_free[uvm_page_lookup_freelist(pg)].
    473 		    pgfl_buckets[VM_PGCOLOR_BUCKET(pg)].
    474 		    pgfl_queues[PGFL_UNKNOWN], pg, pageq);
    475 		uvmexp.free++;
    476 		if (uvmexp.zeropages < UVM_PAGEZERO_TARGET)
    477 			uvm.page_idle_zero = vm_page_zero_enable;
    478 		STAT_DECR(uvm_pglistalloc_npages);
    479 	}
    480 	uvm_unlock_fpageq(s);
    481 }
    482