Home | History | Annotate | Line # | Download | only in uvm
uvm_bio.c revision 1.56
      1  1.56   thorpej /*	$NetBSD: uvm_bio.c,v 1.56 2007/02/22 06:05:00 thorpej Exp $	*/
      2   1.2       chs 
      3  1.13       chs /*
      4   1.2       chs  * Copyright (c) 1998 Chuck Silvers.
      5   1.2       chs  * All rights reserved.
      6   1.2       chs  *
      7   1.2       chs  * Redistribution and use in source and binary forms, with or without
      8   1.2       chs  * modification, are permitted provided that the following conditions
      9   1.2       chs  * are met:
     10   1.2       chs  * 1. Redistributions of source code must retain the above copyright
     11   1.2       chs  *    notice, this list of conditions and the following disclaimer.
     12   1.2       chs  * 2. Redistributions in binary form must reproduce the above copyright
     13   1.2       chs  *    notice, this list of conditions and the following disclaimer in the
     14   1.2       chs  *    documentation and/or other materials provided with the distribution.
     15   1.2       chs  * 3. The name of the author may not be used to endorse or promote products
     16   1.2       chs  *    derived from this software without specific prior written permission.
     17   1.2       chs  *
     18   1.2       chs  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
     19   1.2       chs  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
     20   1.2       chs  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
     21   1.2       chs  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
     22   1.2       chs  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
     23   1.2       chs  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
     24   1.2       chs  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
     25   1.2       chs  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
     26   1.2       chs  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     27   1.2       chs  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     28   1.2       chs  * SUCH DAMAGE.
     29   1.2       chs  *
     30   1.2       chs  */
     31   1.2       chs 
     32   1.2       chs /*
     33  1.33       chs  * uvm_bio.c: buffered i/o object mapping cache
     34   1.2       chs  */
     35   1.2       chs 
     36  1.21     lukem #include <sys/cdefs.h>
     37  1.56   thorpej __KERNEL_RCSID(0, "$NetBSD: uvm_bio.c,v 1.56 2007/02/22 06:05:00 thorpej Exp $");
     38  1.21     lukem 
     39  1.21     lukem #include "opt_uvmhist.h"
     40  1.50      yamt #include "opt_ubc.h"
     41   1.2       chs 
     42   1.2       chs #include <sys/param.h>
     43   1.2       chs #include <sys/systm.h>
     44   1.2       chs #include <sys/malloc.h>
     45   1.2       chs #include <sys/kernel.h>
     46   1.2       chs 
     47   1.2       chs #include <uvm/uvm.h>
     48   1.2       chs 
     49   1.2       chs /*
     50   1.2       chs  * global data structures
     51   1.2       chs  */
     52   1.2       chs 
     53   1.2       chs /*
     54   1.2       chs  * local functions
     55   1.2       chs  */
     56   1.2       chs 
     57  1.39   thorpej static int	ubc_fault(struct uvm_faultinfo *, vaddr_t, struct vm_page **,
     58  1.44  drochner 			  int, int, vm_prot_t, int);
     59  1.39   thorpej static struct ubc_map *ubc_find_mapping(struct uvm_object *, voff_t);
     60   1.2       chs 
     61   1.2       chs /*
     62   1.2       chs  * local data structues
     63   1.2       chs  */
     64   1.2       chs 
     65  1.18       chs #define UBC_HASH(uobj, offset) 						\
     66  1.18       chs 	(((((u_long)(uobj)) >> 8) + (((u_long)(offset)) >> PAGE_SHIFT)) & \
     67   1.2       chs 				ubc_object.hashmask)
     68   1.2       chs 
     69  1.18       chs #define UBC_QUEUE(offset)						\
     70  1.18       chs 	(&ubc_object.inactive[(((u_long)(offset)) >> ubc_winshift) &	\
     71  1.18       chs 			     (UBC_NQUEUES - 1)])
     72  1.18       chs 
     73  1.18       chs #define UBC_UMAP_ADDR(u)						\
     74  1.18       chs 	(vaddr_t)(ubc_object.kva + (((u) - ubc_object.umap) << ubc_winshift))
     75  1.18       chs 
     76  1.18       chs 
     77  1.18       chs #define UMAP_PAGES_LOCKED	0x0001
     78  1.18       chs #define UMAP_MAPPING_CACHED	0x0002
     79   1.2       chs 
     80   1.2       chs struct ubc_map
     81   1.2       chs {
     82   1.2       chs 	struct uvm_object *	uobj;		/* mapped object */
     83   1.2       chs 	voff_t			offset;		/* offset into uobj */
     84  1.33       chs 	voff_t			writeoff;	/* write offset */
     85  1.33       chs 	vsize_t			writelen;	/* write len */
     86  1.18       chs 	int			refcount;	/* refcount on mapping */
     87  1.18       chs 	int			flags;		/* extra state */
     88  1.42      yamt 	int			advice;
     89   1.2       chs 
     90   1.2       chs 	LIST_ENTRY(ubc_map)	hash;		/* hash table */
     91   1.2       chs 	TAILQ_ENTRY(ubc_map)	inactive;	/* inactive queue */
     92   1.2       chs };
     93   1.2       chs 
     94   1.2       chs static struct ubc_object
     95   1.2       chs {
     96   1.2       chs 	struct uvm_object uobj;		/* glue for uvm_map() */
     97   1.2       chs 	char *kva;			/* where ubc_object is mapped */
     98   1.2       chs 	struct ubc_map *umap;		/* array of ubc_map's */
     99   1.2       chs 
    100   1.2       chs 	LIST_HEAD(, ubc_map) *hash;	/* hashtable for cached ubc_map's */
    101   1.2       chs 	u_long hashmask;		/* mask for hashtable */
    102   1.2       chs 
    103   1.2       chs 	TAILQ_HEAD(ubc_inactive_head, ubc_map) *inactive;
    104   1.2       chs 					/* inactive queues for ubc_map's */
    105   1.2       chs 
    106   1.2       chs } ubc_object;
    107   1.2       chs 
    108   1.2       chs struct uvm_pagerops ubc_pager =
    109   1.2       chs {
    110  1.48  christos 	.pgo_fault = ubc_fault,
    111   1.2       chs 	/* ... rest are NULL */
    112   1.2       chs };
    113   1.2       chs 
    114   1.2       chs int ubc_nwins = UBC_NWINS;
    115  1.11       chs int ubc_winshift = UBC_WINSHIFT;
    116  1.11       chs int ubc_winsize;
    117  1.27   thorpej #if defined(PMAP_PREFER)
    118   1.2       chs int ubc_nqueues;
    119   1.2       chs #define UBC_NQUEUES ubc_nqueues
    120   1.2       chs #else
    121   1.2       chs #define UBC_NQUEUES 1
    122   1.2       chs #endif
    123   1.2       chs 
    124  1.50      yamt #if defined(UBC_STATS)
    125  1.50      yamt 
    126  1.50      yamt #define	UBC_EVCNT_DEFINE(name) \
    127  1.50      yamt struct evcnt ubc_evcnt_##name = \
    128  1.50      yamt EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "ubc", #name); \
    129  1.50      yamt EVCNT_ATTACH_STATIC(ubc_evcnt_##name);
    130  1.50      yamt #define	UBC_EVCNT_INCR(name) ubc_evcnt_##name.ev_count++
    131  1.50      yamt 
    132  1.50      yamt #else /* defined(UBC_STATS) */
    133  1.50      yamt 
    134  1.50      yamt #define	UBC_EVCNT_DEFINE(name)	/* nothing */
    135  1.50      yamt #define	UBC_EVCNT_INCR(name)	/* nothing */
    136  1.50      yamt 
    137  1.50      yamt #endif /* defined(UBC_STATS) */
    138  1.50      yamt 
    139  1.50      yamt UBC_EVCNT_DEFINE(wincachehit)
    140  1.50      yamt UBC_EVCNT_DEFINE(wincachemiss)
    141  1.50      yamt 
    142   1.2       chs /*
    143   1.2       chs  * ubc_init
    144   1.2       chs  *
    145   1.2       chs  * init pager private data structures.
    146   1.2       chs  */
    147   1.2       chs 
    148   1.2       chs void
    149   1.2       chs ubc_init(void)
    150   1.2       chs {
    151   1.2       chs 	struct ubc_map *umap;
    152   1.2       chs 	vaddr_t va;
    153   1.2       chs 	int i;
    154  1.15    simonb 
    155  1.15    simonb 	/*
    156  1.15    simonb 	 * Make sure ubc_winshift is sane.
    157  1.15    simonb 	 */
    158  1.15    simonb 	if (ubc_winshift < PAGE_SHIFT)
    159  1.15    simonb 		ubc_winshift = PAGE_SHIFT;
    160   1.2       chs 
    161   1.2       chs 	/*
    162   1.2       chs 	 * init ubc_object.
    163   1.2       chs 	 * alloc and init ubc_map's.
    164   1.2       chs 	 * init inactive queues.
    165   1.2       chs 	 * alloc and init hashtable.
    166   1.2       chs 	 * map in ubc_object.
    167   1.2       chs 	 */
    168   1.2       chs 
    169  1.38      yamt 	UVM_OBJ_INIT(&ubc_object.uobj, &ubc_pager, UVM_OBJ_KERN);
    170   1.2       chs 
    171   1.2       chs 	ubc_object.umap = malloc(ubc_nwins * sizeof(struct ubc_map),
    172   1.2       chs 				 M_TEMP, M_NOWAIT);
    173   1.7     enami 	if (ubc_object.umap == NULL)
    174   1.7     enami 		panic("ubc_init: failed to allocate ubc_map");
    175  1.16   thorpej 	memset(ubc_object.umap, 0, ubc_nwins * sizeof(struct ubc_map));
    176   1.2       chs 
    177  1.18       chs 	if (ubc_winshift < PAGE_SHIFT) {
    178  1.18       chs 		ubc_winshift = PAGE_SHIFT;
    179  1.18       chs 	}
    180   1.2       chs 	va = (vaddr_t)1L;
    181   1.2       chs #ifdef PMAP_PREFER
    182  1.36    atatat 	PMAP_PREFER(0, &va, 0, 0);	/* kernel is never topdown */
    183  1.11       chs 	ubc_nqueues = va >> ubc_winshift;
    184  1.11       chs 	if (ubc_nqueues == 0) {
    185  1.11       chs 		ubc_nqueues = 1;
    186   1.2       chs 	}
    187   1.2       chs #endif
    188  1.11       chs 	ubc_winsize = 1 << ubc_winshift;
    189   1.2       chs 	ubc_object.inactive = malloc(UBC_NQUEUES *
    190  1.18       chs 	    sizeof(struct ubc_inactive_head), M_TEMP, M_NOWAIT);
    191   1.7     enami 	if (ubc_object.inactive == NULL)
    192   1.7     enami 		panic("ubc_init: failed to allocate inactive queue heads");
    193   1.2       chs 	for (i = 0; i < UBC_NQUEUES; i++) {
    194   1.2       chs 		TAILQ_INIT(&ubc_object.inactive[i]);
    195   1.2       chs 	}
    196   1.2       chs 	for (i = 0; i < ubc_nwins; i++) {
    197   1.2       chs 		umap = &ubc_object.umap[i];
    198   1.2       chs 		TAILQ_INSERT_TAIL(&ubc_object.inactive[i & (UBC_NQUEUES - 1)],
    199   1.2       chs 				  umap, inactive);
    200   1.2       chs 	}
    201   1.2       chs 
    202   1.2       chs 	ubc_object.hash = hashinit(ubc_nwins, HASH_LIST, M_TEMP, M_NOWAIT,
    203   1.2       chs 				   &ubc_object.hashmask);
    204   1.2       chs 	for (i = 0; i <= ubc_object.hashmask; i++) {
    205   1.2       chs 		LIST_INIT(&ubc_object.hash[i]);
    206   1.2       chs 	}
    207   1.2       chs 
    208   1.2       chs 	if (uvm_map(kernel_map, (vaddr_t *)&ubc_object.kva,
    209  1.11       chs 		    ubc_nwins << ubc_winshift, &ubc_object.uobj, 0, (vsize_t)va,
    210   1.2       chs 		    UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, UVM_INH_NONE,
    211   1.9       chs 				UVM_ADV_RANDOM, UVM_FLAG_NOMERGE)) != 0) {
    212  1.26    provos 		panic("ubc_init: failed to map ubc_object");
    213   1.2       chs 	}
    214   1.2       chs 	UVMHIST_INIT(ubchist, 300);
    215   1.2       chs }
    216   1.2       chs 
    217   1.2       chs /*
    218   1.2       chs  * ubc_fault: fault routine for ubc mapping
    219   1.2       chs  */
    220  1.18       chs 
    221  1.39   thorpej static int
    222  1.54      yamt ubc_fault(struct uvm_faultinfo *ufi, vaddr_t ign1, struct vm_page **ign2,
    223  1.54      yamt     int ign3, int ign4, vm_prot_t access_type, int flags)
    224   1.2       chs {
    225   1.2       chs 	struct uvm_object *uobj;
    226   1.2       chs 	struct ubc_map *umap;
    227   1.2       chs 	vaddr_t va, eva, ubc_offset, slot_offset;
    228  1.18       chs 	int i, error, npages;
    229  1.18       chs 	struct vm_page *pgs[ubc_winsize >> PAGE_SHIFT], *pg;
    230  1.23       chs 	vm_prot_t prot;
    231   1.2       chs 	UVMHIST_FUNC("ubc_fault");  UVMHIST_CALLED(ubchist);
    232   1.2       chs 
    233   1.2       chs 	/*
    234   1.2       chs 	 * no need to try with PGO_LOCKED...
    235   1.2       chs 	 * we don't need to have the map locked since we know that
    236   1.2       chs 	 * no one will mess with it until our reference is released.
    237   1.2       chs 	 */
    238  1.18       chs 
    239   1.2       chs 	if (flags & PGO_LOCKED) {
    240   1.2       chs 		uvmfault_unlockall(ufi, NULL, &ubc_object.uobj, NULL);
    241   1.2       chs 		flags &= ~PGO_LOCKED;
    242   1.2       chs 	}
    243   1.2       chs 
    244   1.2       chs 	va = ufi->orig_rvaddr;
    245   1.2       chs 	ubc_offset = va - (vaddr_t)ubc_object.kva;
    246  1.11       chs 	umap = &ubc_object.umap[ubc_offset >> ubc_winshift];
    247   1.2       chs 	KASSERT(umap->refcount != 0);
    248  1.53      yamt 	KASSERT((umap->flags & UMAP_PAGES_LOCKED) == 0);
    249  1.18       chs 	slot_offset = ubc_offset & (ubc_winsize - 1);
    250   1.2       chs 
    251  1.34       chs 	/*
    252  1.34       chs 	 * some platforms cannot write to individual bytes atomically, so
    253  1.34       chs 	 * software has to do read/modify/write of larger quantities instead.
    254  1.34       chs 	 * this means that the access_type for "write" operations
    255  1.34       chs 	 * can be VM_PROT_READ, which confuses us mightily.
    256  1.37     perry 	 *
    257  1.34       chs 	 * deal with this by resetting access_type based on the info
    258  1.34       chs 	 * that ubc_alloc() stores for us.
    259  1.34       chs 	 */
    260  1.34       chs 
    261  1.34       chs 	access_type = umap->writelen ? VM_PROT_WRITE : VM_PROT_READ;
    262  1.34       chs 	UVMHIST_LOG(ubchist, "va 0x%lx ubc_offset 0x%lx access_type %d",
    263  1.34       chs 	    va, ubc_offset, access_type, 0);
    264  1.34       chs 
    265  1.33       chs #ifdef DIAGNOSTIC
    266  1.33       chs 	if ((access_type & VM_PROT_WRITE) != 0) {
    267  1.33       chs 		if (slot_offset < trunc_page(umap->writeoff) ||
    268  1.33       chs 		    umap->writeoff + umap->writelen <= slot_offset) {
    269  1.33       chs 			panic("ubc_fault: out of range write");
    270  1.33       chs 		}
    271  1.33       chs 	}
    272  1.33       chs #endif
    273  1.33       chs 
    274   1.2       chs 	/* no umap locking needed since we have a ref on the umap */
    275   1.2       chs 	uobj = umap->uobj;
    276   1.2       chs 
    277  1.33       chs 	if ((access_type & VM_PROT_WRITE) == 0) {
    278  1.33       chs 		npages = (ubc_winsize - slot_offset) >> PAGE_SHIFT;
    279  1.33       chs 	} else {
    280  1.33       chs 		npages = (round_page(umap->offset + umap->writeoff +
    281  1.33       chs 		    umap->writelen) - (umap->offset + slot_offset))
    282  1.33       chs 		    >> PAGE_SHIFT;
    283  1.33       chs 		flags |= PGO_PASTEOF;
    284  1.33       chs 	}
    285   1.2       chs 
    286   1.2       chs again:
    287   1.2       chs 	memset(pgs, 0, sizeof (pgs));
    288   1.2       chs 	simple_lock(&uobj->vmobjlock);
    289   1.2       chs 
    290  1.33       chs 	UVMHIST_LOG(ubchist, "slot_offset 0x%x writeoff 0x%x writelen 0x%x ",
    291  1.33       chs 	    slot_offset, umap->writeoff, umap->writelen, 0);
    292  1.33       chs 	UVMHIST_LOG(ubchist, "getpages uobj %p offset 0x%x npages %d",
    293  1.18       chs 	    uobj, umap->offset + slot_offset, npages, 0);
    294   1.2       chs 
    295  1.33       chs 	error = (*uobj->pgops->pgo_get)(uobj, umap->offset + slot_offset, pgs,
    296  1.42      yamt 	    &npages, 0, access_type, umap->advice, flags | PGO_NOBLOCKALLOC |
    297  1.41      yamt 	    PGO_NOTIMESTAMP);
    298  1.24    simonb 	UVMHIST_LOG(ubchist, "getpages error %d npages %d", error, npages, 0,
    299  1.24    simonb 	    0);
    300   1.2       chs 
    301   1.5       chs 	if (error == EAGAIN) {
    302   1.2       chs 		tsleep(&lbolt, PVM, "ubc_fault", 0);
    303   1.2       chs 		goto again;
    304   1.2       chs 	}
    305   1.5       chs 	if (error) {
    306  1.10       chs 		return error;
    307   1.5       chs 	}
    308   1.2       chs 
    309   1.2       chs 	va = ufi->orig_rvaddr;
    310   1.2       chs 	eva = ufi->orig_rvaddr + (npages << PAGE_SHIFT);
    311   1.2       chs 
    312  1.28      yamt 	UVMHIST_LOG(ubchist, "va 0x%lx eva 0x%lx", va, eva, 0, 0);
    313  1.28      yamt 	for (i = 0; va < eva; i++, va += PAGE_SIZE) {
    314  1.55   thorpej 		bool rdonly;
    315  1.40      yamt 		vm_prot_t mask;
    316  1.34       chs 
    317  1.28      yamt 		/*
    318  1.28      yamt 		 * for virtually-indexed, virtually-tagged caches we should
    319  1.28      yamt 		 * avoid creating writable mappings when we don't absolutely
    320  1.28      yamt 		 * need them, since the "compatible alias" trick doesn't work
    321  1.28      yamt 		 * on such caches.  otherwise, we can always map the pages
    322  1.28      yamt 		 * writable.
    323  1.28      yamt 		 */
    324  1.23       chs 
    325  1.23       chs #ifdef PMAP_CACHE_VIVT
    326  1.28      yamt 		prot = VM_PROT_READ | access_type;
    327  1.23       chs #else
    328  1.28      yamt 		prot = VM_PROT_READ | VM_PROT_WRITE;
    329  1.23       chs #endif
    330  1.24    simonb 		UVMHIST_LOG(ubchist, "pgs[%d] = %p", i, pgs[i], 0, 0);
    331   1.2       chs 		pg = pgs[i];
    332   1.2       chs 
    333   1.2       chs 		if (pg == NULL || pg == PGO_DONTCARE) {
    334   1.2       chs 			continue;
    335   1.2       chs 		}
    336  1.43      yamt 
    337  1.43      yamt 		uobj = pg->uobject;
    338  1.43      yamt 		simple_lock(&uobj->vmobjlock);
    339   1.2       chs 		if (pg->flags & PG_WANTED) {
    340   1.2       chs 			wakeup(pg);
    341   1.2       chs 		}
    342   1.2       chs 		KASSERT((pg->flags & PG_FAKE) == 0);
    343   1.2       chs 		if (pg->flags & PG_RELEASED) {
    344  1.43      yamt 			uvm_lock_pageq();
    345  1.18       chs 			uvm_pagefree(pg);
    346  1.43      yamt 			uvm_unlock_pageq();
    347  1.43      yamt 			simple_unlock(&uobj->vmobjlock);
    348   1.2       chs 			continue;
    349   1.2       chs 		}
    350  1.28      yamt 		if (pg->loan_count != 0) {
    351  1.34       chs 
    352  1.28      yamt 			/*
    353  1.28      yamt 			 * avoid unneeded loan break if possible.
    354  1.28      yamt 			 */
    355  1.34       chs 
    356  1.28      yamt 			if ((access_type & VM_PROT_WRITE) == 0)
    357  1.28      yamt 				prot &= ~VM_PROT_WRITE;
    358  1.28      yamt 
    359  1.28      yamt 			if (prot & VM_PROT_WRITE) {
    360  1.47      yamt 				struct vm_page *newpg;
    361  1.47      yamt 
    362  1.47      yamt 				newpg = uvm_loanbreak(pg);
    363  1.47      yamt 				if (newpg == NULL) {
    364  1.47      yamt 					uvm_page_unbusy(&pg, 1);
    365  1.47      yamt 					simple_unlock(&uobj->vmobjlock);
    366  1.47      yamt 					uvm_wait("ubc_loanbrk");
    367  1.28      yamt 					continue; /* will re-fault */
    368  1.47      yamt 				}
    369  1.47      yamt 				pg = newpg;
    370  1.28      yamt 			}
    371  1.28      yamt 		}
    372  1.40      yamt 
    373  1.40      yamt 		/*
    374  1.40      yamt 		 * note that a page whose backing store is partially allocated
    375  1.40      yamt 		 * is marked as PG_RDONLY.
    376  1.40      yamt 		 */
    377  1.40      yamt 
    378  1.49      yamt 		rdonly = ((access_type & VM_PROT_WRITE) == 0 &&
    379  1.49      yamt 		    (pg->flags & PG_RDONLY) != 0) ||
    380  1.49      yamt 		    UVM_OBJ_NEEDS_WRITEFAULT(uobj);
    381  1.40      yamt 		KASSERT((pg->flags & PG_RDONLY) == 0 ||
    382  1.40      yamt 		    (access_type & VM_PROT_WRITE) == 0 ||
    383  1.40      yamt 		    pg->offset < umap->writeoff ||
    384  1.40      yamt 		    pg->offset + PAGE_SIZE > umap->writeoff + umap->writelen);
    385  1.40      yamt 		mask = rdonly ? ~VM_PROT_WRITE : VM_PROT_ALL;
    386  1.46      yamt 		error = pmap_enter(ufi->orig_map->pmap, va, VM_PAGE_TO_PHYS(pg),
    387  1.46      yamt 		    prot & mask, PMAP_CANFAIL | (access_type & mask));
    388  1.43      yamt 		uvm_lock_pageq();
    389   1.2       chs 		uvm_pageactivate(pg);
    390  1.43      yamt 		uvm_unlock_pageq();
    391  1.45      yamt 		pg->flags &= ~(PG_BUSY|PG_WANTED);
    392   1.2       chs 		UVM_PAGE_OWN(pg, NULL);
    393  1.43      yamt 		simple_unlock(&uobj->vmobjlock);
    394  1.46      yamt 		if (error) {
    395  1.46      yamt 			UVMHIST_LOG(ubchist, "pmap_enter fail %d",
    396  1.46      yamt 			    error, 0, 0, 0);
    397  1.46      yamt 			uvm_wait("ubc_pmfail");
    398  1.46      yamt 			/* will refault */
    399  1.46      yamt 		}
    400   1.2       chs 	}
    401  1.17     chris 	pmap_update(ufi->orig_map->pmap);
    402   1.8       chs 	return 0;
    403   1.2       chs }
    404   1.2       chs 
    405   1.2       chs /*
    406   1.2       chs  * local functions
    407   1.2       chs  */
    408   1.2       chs 
    409  1.39   thorpej static struct ubc_map *
    410  1.39   thorpej ubc_find_mapping(struct uvm_object *uobj, voff_t offset)
    411   1.2       chs {
    412   1.2       chs 	struct ubc_map *umap;
    413   1.2       chs 
    414   1.2       chs 	LIST_FOREACH(umap, &ubc_object.hash[UBC_HASH(uobj, offset)], hash) {
    415   1.2       chs 		if (umap->uobj == uobj && umap->offset == offset) {
    416   1.2       chs 			return umap;
    417   1.2       chs 		}
    418   1.2       chs 	}
    419   1.2       chs 	return NULL;
    420   1.2       chs }
    421   1.2       chs 
    422   1.2       chs 
    423   1.2       chs /*
    424   1.2       chs  * ubc interface functions
    425   1.2       chs  */
    426   1.2       chs 
    427   1.2       chs /*
    428  1.18       chs  * ubc_alloc:  allocate a file mapping window
    429   1.2       chs  */
    430  1.18       chs 
    431   1.2       chs void *
    432  1.42      yamt ubc_alloc(struct uvm_object *uobj, voff_t offset, vsize_t *lenp, int advice,
    433  1.42      yamt     int flags)
    434   1.2       chs {
    435   1.6       chs 	vaddr_t slot_offset, va;
    436   1.2       chs 	struct ubc_map *umap;
    437   1.6       chs 	voff_t umap_offset;
    438  1.18       chs 	int error;
    439   1.2       chs 	UVMHIST_FUNC("ubc_alloc"); UVMHIST_CALLED(ubchist);
    440   1.2       chs 
    441  1.33       chs 	UVMHIST_LOG(ubchist, "uobj %p offset 0x%lx len 0x%lx",
    442  1.33       chs 	    uobj, offset, *lenp, 0);
    443   1.2       chs 
    444  1.34       chs 	KASSERT(*lenp > 0);
    445   1.6       chs 	umap_offset = (offset & ~((voff_t)ubc_winsize - 1));
    446   1.4     enami 	slot_offset = (vaddr_t)(offset & ((voff_t)ubc_winsize - 1));
    447  1.18       chs 	*lenp = MIN(*lenp, ubc_winsize - slot_offset);
    448   1.2       chs 
    449   1.2       chs 	/*
    450  1.33       chs 	 * the object is always locked here, so we don't need to add a ref.
    451   1.2       chs 	 */
    452   1.2       chs 
    453   1.2       chs again:
    454   1.2       chs 	simple_lock(&ubc_object.uobj.vmobjlock);
    455   1.2       chs 	umap = ubc_find_mapping(uobj, umap_offset);
    456   1.2       chs 	if (umap == NULL) {
    457  1.50      yamt 		UBC_EVCNT_INCR(wincachemiss);
    458   1.2       chs 		umap = TAILQ_FIRST(UBC_QUEUE(offset));
    459   1.2       chs 		if (umap == NULL) {
    460   1.2       chs 			simple_unlock(&ubc_object.uobj.vmobjlock);
    461   1.2       chs 			tsleep(&lbolt, PVM, "ubc_alloc", 0);
    462   1.2       chs 			goto again;
    463   1.2       chs 		}
    464   1.2       chs 
    465   1.2       chs 		/*
    466  1.18       chs 		 * remove from old hash (if any), add to new hash.
    467   1.2       chs 		 */
    468   1.2       chs 
    469   1.2       chs 		if (umap->uobj != NULL) {
    470   1.2       chs 			LIST_REMOVE(umap, hash);
    471   1.2       chs 		}
    472   1.2       chs 		umap->uobj = uobj;
    473   1.2       chs 		umap->offset = umap_offset;
    474   1.2       chs 		LIST_INSERT_HEAD(&ubc_object.hash[UBC_HASH(uobj, umap_offset)],
    475  1.18       chs 		    umap, hash);
    476  1.18       chs 		va = UBC_UMAP_ADDR(umap);
    477  1.18       chs 		if (umap->flags & UMAP_MAPPING_CACHED) {
    478  1.18       chs 			umap->flags &= ~UMAP_MAPPING_CACHED;
    479  1.18       chs 			pmap_remove(pmap_kernel(), va, va + ubc_winsize);
    480  1.18       chs 			pmap_update(pmap_kernel());
    481  1.18       chs 		}
    482  1.18       chs 	} else {
    483  1.50      yamt 		UBC_EVCNT_INCR(wincachehit);
    484  1.18       chs 		va = UBC_UMAP_ADDR(umap);
    485   1.2       chs 	}
    486   1.2       chs 
    487   1.2       chs 	if (umap->refcount == 0) {
    488   1.2       chs 		TAILQ_REMOVE(UBC_QUEUE(offset), umap, inactive);
    489   1.2       chs 	}
    490   1.2       chs 
    491   1.2       chs #ifdef DIAGNOSTIC
    492  1.18       chs 	if ((flags & UBC_WRITE) && (umap->writeoff || umap->writelen)) {
    493  1.33       chs 		panic("ubc_alloc: concurrent writes uobj %p", uobj);
    494   1.2       chs 	}
    495   1.2       chs #endif
    496   1.2       chs 	if (flags & UBC_WRITE) {
    497   1.2       chs 		umap->writeoff = slot_offset;
    498   1.2       chs 		umap->writelen = *lenp;
    499   1.2       chs 	}
    500   1.2       chs 
    501   1.2       chs 	umap->refcount++;
    502  1.42      yamt 	umap->advice = advice;
    503   1.2       chs 	simple_unlock(&ubc_object.uobj.vmobjlock);
    504  1.18       chs 	UVMHIST_LOG(ubchist, "umap %p refs %d va %p flags 0x%x",
    505  1.18       chs 	    umap, umap->refcount, va, flags);
    506  1.18       chs 
    507  1.18       chs 	if (flags & UBC_FAULTBUSY) {
    508  1.18       chs 		int npages = (*lenp + PAGE_SIZE - 1) >> PAGE_SHIFT;
    509  1.18       chs 		struct vm_page *pgs[npages];
    510  1.40      yamt 		int gpflags =
    511  1.41      yamt 		    PGO_SYNCIO|PGO_OVERWRITE|PGO_PASTEOF|PGO_NOBLOCKALLOC|
    512  1.41      yamt 		    PGO_NOTIMESTAMP;
    513  1.18       chs 		int i;
    514  1.30       dbj 		KDASSERT(flags & UBC_WRITE);
    515   1.2       chs 
    516  1.18       chs 		if (umap->flags & UMAP_MAPPING_CACHED) {
    517  1.18       chs 			umap->flags &= ~UMAP_MAPPING_CACHED;
    518  1.18       chs 			pmap_remove(pmap_kernel(), va, va + ubc_winsize);
    519  1.18       chs 		}
    520  1.22     enami 		memset(pgs, 0, sizeof(pgs));
    521  1.18       chs 		simple_lock(&uobj->vmobjlock);
    522  1.33       chs 		error = (*uobj->pgops->pgo_get)(uobj, trunc_page(offset), pgs,
    523  1.42      yamt 		    &npages, 0, VM_PROT_READ | VM_PROT_WRITE, advice, gpflags);
    524  1.24    simonb 		UVMHIST_LOG(ubchist, "faultbusy getpages %d", error, 0, 0, 0);
    525  1.18       chs 		if (error) {
    526  1.18       chs 			goto out;
    527  1.18       chs 		}
    528  1.18       chs 		for (i = 0; i < npages; i++) {
    529  1.18       chs 			pmap_kenter_pa(va + slot_offset + (i << PAGE_SHIFT),
    530  1.18       chs 			    VM_PAGE_TO_PHYS(pgs[i]),
    531  1.18       chs 			    VM_PROT_READ | VM_PROT_WRITE);
    532  1.18       chs 		}
    533  1.18       chs 		pmap_update(pmap_kernel());
    534  1.18       chs 		umap->flags |= UMAP_PAGES_LOCKED;
    535  1.18       chs 	}
    536  1.18       chs 
    537  1.18       chs out:
    538  1.18       chs 	return (void *)(va + slot_offset);
    539   1.2       chs }
    540   1.2       chs 
    541  1.18       chs /*
    542  1.18       chs  * ubc_release:  free a file mapping window.
    543  1.18       chs  */
    544   1.2       chs 
    545   1.2       chs void
    546  1.39   thorpej ubc_release(void *va, int flags)
    547   1.2       chs {
    548   1.2       chs 	struct ubc_map *umap;
    549   1.2       chs 	struct uvm_object *uobj;
    550  1.18       chs 	vaddr_t umapva;
    551  1.55   thorpej 	bool unmapped;
    552   1.2       chs 	UVMHIST_FUNC("ubc_release"); UVMHIST_CALLED(ubchist);
    553   1.2       chs 
    554  1.24    simonb 	UVMHIST_LOG(ubchist, "va %p", va, 0, 0, 0);
    555  1.11       chs 	umap = &ubc_object.umap[((char *)va - ubc_object.kva) >> ubc_winshift];
    556  1.18       chs 	umapva = UBC_UMAP_ADDR(umap);
    557   1.2       chs 	uobj = umap->uobj;
    558   1.2       chs 	KASSERT(uobj != NULL);
    559   1.2       chs 
    560  1.18       chs 	if (umap->flags & UMAP_PAGES_LOCKED) {
    561  1.18       chs 		int slot_offset = umap->writeoff;
    562  1.18       chs 		int endoff = umap->writeoff + umap->writelen;
    563  1.18       chs 		int zerolen = round_page(endoff) - endoff;
    564  1.18       chs 		int npages = (int)(round_page(umap->writeoff + umap->writelen)
    565  1.18       chs 				   - trunc_page(umap->writeoff)) >> PAGE_SHIFT;
    566  1.18       chs 		struct vm_page *pgs[npages];
    567  1.18       chs 		paddr_t pa;
    568  1.18       chs 		int i;
    569  1.55   thorpej 		bool rv;
    570  1.18       chs 
    571  1.18       chs 		if (zerolen) {
    572  1.18       chs 			memset((char *)umapva + endoff, 0, zerolen);
    573  1.18       chs 		}
    574  1.18       chs 		umap->flags &= ~UMAP_PAGES_LOCKED;
    575  1.18       chs 		uvm_lock_pageq();
    576  1.18       chs 		for (i = 0; i < npages; i++) {
    577  1.18       chs 			rv = pmap_extract(pmap_kernel(),
    578  1.18       chs 			    umapva + slot_offset + (i << PAGE_SHIFT), &pa);
    579  1.18       chs 			KASSERT(rv);
    580  1.18       chs 			pgs[i] = PHYS_TO_VM_PAGE(pa);
    581  1.18       chs 			pgs[i]->flags &= ~(PG_FAKE|PG_CLEAN);
    582  1.28      yamt 			KASSERT(pgs[i]->loan_count == 0);
    583  1.18       chs 			uvm_pageactivate(pgs[i]);
    584  1.18       chs 		}
    585  1.18       chs 		uvm_unlock_pageq();
    586  1.18       chs 		pmap_kremove(umapva, ubc_winsize);
    587  1.18       chs 		pmap_update(pmap_kernel());
    588  1.32      yamt 		simple_lock(&uobj->vmobjlock);
    589  1.18       chs 		uvm_page_unbusy(pgs, npages);
    590  1.32      yamt 		simple_unlock(&uobj->vmobjlock);
    591  1.56   thorpej 		unmapped = true;
    592  1.18       chs 	} else {
    593  1.56   thorpej 		unmapped = false;
    594  1.18       chs 	}
    595  1.18       chs 
    596  1.18       chs 	simple_lock(&ubc_object.uobj.vmobjlock);
    597   1.2       chs 	umap->writeoff = 0;
    598   1.2       chs 	umap->writelen = 0;
    599   1.2       chs 	umap->refcount--;
    600   1.2       chs 	if (umap->refcount == 0) {
    601  1.33       chs 		if (flags & UBC_UNMAP) {
    602   1.2       chs 
    603   1.2       chs 			/*
    604  1.33       chs 			 * Invalidate any cached mappings if requested.
    605  1.33       chs 			 * This is typically used to avoid leaving
    606  1.33       chs 			 * incompatible cache aliases around indefinitely.
    607   1.2       chs 			 */
    608   1.2       chs 
    609  1.18       chs 			pmap_remove(pmap_kernel(), umapva,
    610  1.18       chs 				    umapva + ubc_winsize);
    611  1.18       chs 			umap->flags &= ~UMAP_MAPPING_CACHED;
    612  1.17     chris 			pmap_update(pmap_kernel());
    613   1.2       chs 			LIST_REMOVE(umap, hash);
    614   1.2       chs 			umap->uobj = NULL;
    615   1.2       chs 			TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap,
    616   1.2       chs 			    inactive);
    617   1.2       chs 		} else {
    618  1.18       chs 			if (!unmapped) {
    619  1.18       chs 				umap->flags |= UMAP_MAPPING_CACHED;
    620  1.18       chs 			}
    621   1.2       chs 			TAILQ_INSERT_TAIL(UBC_QUEUE(umap->offset), umap,
    622   1.2       chs 			    inactive);
    623   1.2       chs 		}
    624   1.2       chs 	}
    625  1.24    simonb 	UVMHIST_LOG(ubchist, "umap %p refs %d", umap, umap->refcount, 0, 0);
    626   1.2       chs 	simple_unlock(&ubc_object.uobj.vmobjlock);
    627   1.2       chs }
    628   1.2       chs 
    629   1.2       chs 
    630  1.29      yamt #if 0 /* notused */
    631   1.2       chs /*
    632   1.2       chs  * removing a range of mappings from the ubc mapping cache.
    633   1.2       chs  */
    634   1.2       chs 
    635   1.2       chs void
    636  1.39   thorpej ubc_flush(struct uvm_object *uobj, voff_t start, voff_t end)
    637   1.2       chs {
    638   1.2       chs 	struct ubc_map *umap;
    639   1.2       chs 	vaddr_t va;
    640   1.2       chs 	UVMHIST_FUNC("ubc_flush");  UVMHIST_CALLED(ubchist);
    641   1.2       chs 
    642   1.2       chs 	UVMHIST_LOG(ubchist, "uobj %p start 0x%lx end 0x%lx",
    643  1.24    simonb 		    uobj, start, end, 0);
    644   1.2       chs 
    645   1.2       chs 	simple_lock(&ubc_object.uobj.vmobjlock);
    646   1.2       chs 	for (umap = ubc_object.umap;
    647   1.2       chs 	     umap < &ubc_object.umap[ubc_nwins];
    648   1.2       chs 	     umap++) {
    649   1.2       chs 
    650  1.18       chs 		if (umap->uobj != uobj || umap->offset < start ||
    651   1.2       chs 		    (umap->offset >= end && end != 0) ||
    652   1.2       chs 		    umap->refcount > 0) {
    653   1.2       chs 			continue;
    654   1.2       chs 		}
    655   1.2       chs 
    656   1.2       chs 		/*
    657   1.2       chs 		 * remove from hash,
    658   1.2       chs 		 * move to head of inactive queue.
    659   1.2       chs 		 */
    660   1.2       chs 
    661   1.2       chs 		va = (vaddr_t)(ubc_object.kva +
    662  1.18       chs 		    ((umap - ubc_object.umap) << ubc_winshift));
    663   1.4     enami 		pmap_remove(pmap_kernel(), va, va + ubc_winsize);
    664   1.2       chs 
    665   1.2       chs 		LIST_REMOVE(umap, hash);
    666   1.2       chs 		umap->uobj = NULL;
    667   1.2       chs 		TAILQ_REMOVE(UBC_QUEUE(umap->offset), umap, inactive);
    668   1.2       chs 		TAILQ_INSERT_HEAD(UBC_QUEUE(umap->offset), umap, inactive);
    669   1.2       chs 	}
    670  1.18       chs 	pmap_update(pmap_kernel());
    671   1.2       chs 	simple_unlock(&ubc_object.uobj.vmobjlock);
    672   1.2       chs }
    673  1.29      yamt #endif /* notused */
    674