Home | History | Annotate | Line # | Download | only in uvm
uvm_device.c revision 1.55
      1 /*	$NetBSD: uvm_device.c,v 1.55 2008/12/17 20:51:39 cegger Exp $	*/
      2 
      3 /*
      4  *
      5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
      6  * All rights reserved.
      7  *
      8  * Redistribution and use in source and binary forms, with or without
      9  * modification, are permitted provided that the following conditions
     10  * are met:
     11  * 1. Redistributions of source code must retain the above copyright
     12  *    notice, this list of conditions and the following disclaimer.
     13  * 2. Redistributions in binary form must reproduce the above copyright
     14  *    notice, this list of conditions and the following disclaimer in the
     15  *    documentation and/or other materials provided with the distribution.
     16  * 3. All advertising materials mentioning features or use of this software
     17  *    must display the following acknowledgement:
     18  *      This product includes software developed by Charles D. Cranor and
     19  *      Washington University.
     20  * 4. The name of the author may not be used to endorse or promote products
     21  *    derived from this software without specific prior written permission.
     22  *
     23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
     24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
     25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
     26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
     27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
     28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
     29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
     30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
     31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
     32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     33  *
     34  * from: Id: uvm_device.c,v 1.1.2.9 1998/02/06 05:11:47 chs Exp
     35  */
     36 
     37 /*
     38  * uvm_device.c: the device pager.
     39  */
     40 
     41 #include <sys/cdefs.h>
     42 __KERNEL_RCSID(0, "$NetBSD: uvm_device.c,v 1.55 2008/12/17 20:51:39 cegger Exp $");
     43 
     44 #include "opt_uvmhist.h"
     45 
     46 #include <sys/param.h>
     47 #include <sys/systm.h>
     48 #include <sys/conf.h>
     49 #include <sys/proc.h>
     50 #include <sys/malloc.h>
     51 #include <sys/vnode.h>
     52 
     53 #include <uvm/uvm.h>
     54 #include <uvm/uvm_device.h>
     55 
     56 /*
     57  * private global data structure
     58  *
     59  * we keep a list of active device objects in the system.
     60  */
     61 
     62 LIST_HEAD(udv_list_struct, uvm_device);
     63 static struct udv_list_struct udv_list;
     64 static kmutex_t udv_lock;
     65 
     66 /*
     67  * functions
     68  */
     69 
     70 static void	udv_init(void);
     71 static void	udv_reference(struct uvm_object *);
     72 static void	udv_detach(struct uvm_object *);
     73 static int	udv_fault(struct uvm_faultinfo *, vaddr_t,
     74 			  struct vm_page **, int, int, vm_prot_t,
     75 			  int);
     76 
     77 /*
     78  * master pager structure
     79  */
     80 
     81 const struct uvm_pagerops uvm_deviceops = {
     82 	.pgo_init = udv_init,
     83 	.pgo_reference = udv_reference,
     84 	.pgo_detach = udv_detach,
     85 	.pgo_fault = udv_fault,
     86 };
     87 
     88 /*
     89  * the ops!
     90  */
     91 
     92 /*
     93  * udv_init
     94  *
     95  * init pager private data structures.
     96  */
     97 
     98 static void
     99 udv_init(void)
    100 {
    101 	LIST_INIT(&udv_list);
    102 	mutex_init(&udv_lock, MUTEX_DEFAULT, IPL_NONE);
    103 }
    104 
    105 /*
    106  * udv_attach
    107  *
    108  * get a VM object that is associated with a device.   allocate a new
    109  * one if needed.
    110  *
    111  * => caller must _not_ already be holding the lock on the uvm_object.
    112  * => in fact, nothing should be locked so that we can sleep here.
    113  */
    114 
    115 struct uvm_object *
    116 udv_attach(void *arg, vm_prot_t accessprot,
    117     voff_t off,		/* used only for access check */
    118     vsize_t size	/* used only for access check */)
    119 {
    120 	dev_t device = *((dev_t *)arg);
    121 	struct uvm_device *udv, *lcv;
    122 	const struct cdevsw *cdev;
    123 	dev_type_mmap((*mapfn));
    124 
    125 	UVMHIST_FUNC("udv_attach"); UVMHIST_CALLED(maphist);
    126 
    127 	UVMHIST_LOG(maphist, "(device=0x%x)", device,0,0,0);
    128 
    129 	/*
    130 	 * before we do anything, ensure this device supports mmap
    131 	 */
    132 
    133 	cdev = cdevsw_lookup(device);
    134 	if (cdev == NULL) {
    135 		return (NULL);
    136 	}
    137 	mapfn = cdev->d_mmap;
    138 	if (mapfn == NULL || mapfn == nommap || mapfn == nullmmap) {
    139 		return(NULL);
    140 	}
    141 
    142 	/*
    143 	 * Negative offsets on the object are not allowed.
    144 	 */
    145 
    146 	if (off != UVM_UNKNOWN_OFFSET && off < 0)
    147 		return(NULL);
    148 
    149 	/*
    150 	 * Check that the specified range of the device allows the
    151 	 * desired protection.
    152 	 *
    153 	 * XXX assumes VM_PROT_* == PROT_*
    154 	 * XXX clobbers off and size, but nothing else here needs them.
    155 	 */
    156 
    157 	while (size != 0) {
    158 		if (cdev_mmap(device, off, accessprot) == -1) {
    159 			return (NULL);
    160 		}
    161 		off += PAGE_SIZE; size -= PAGE_SIZE;
    162 	}
    163 
    164 	/*
    165 	 * keep looping until we get it
    166 	 */
    167 
    168 	for (;;) {
    169 
    170 		/*
    171 		 * first, attempt to find it on the main list
    172 		 */
    173 
    174 		mutex_enter(&udv_lock);
    175 		LIST_FOREACH(lcv, &udv_list, u_list) {
    176 			if (device == lcv->u_device)
    177 				break;
    178 		}
    179 
    180 		/*
    181 		 * got it on main list.  put a hold on it and unlock udv_lock.
    182 		 */
    183 
    184 		if (lcv) {
    185 
    186 			/*
    187 			 * if someone else has a hold on it, sleep and start
    188 			 * over again.
    189 			 */
    190 
    191 			if (lcv->u_flags & UVM_DEVICE_HOLD) {
    192 				lcv->u_flags |= UVM_DEVICE_WANTED;
    193 				UVM_UNLOCK_AND_WAIT(lcv, &udv_lock, false,
    194 				    "udv_attach",0);
    195 				continue;
    196 			}
    197 
    198 			/* we are now holding it */
    199 			lcv->u_flags |= UVM_DEVICE_HOLD;
    200 			mutex_exit(&udv_lock);
    201 
    202 			/*
    203 			 * bump reference count, unhold, return.
    204 			 */
    205 
    206 			mutex_enter(&lcv->u_obj.vmobjlock);
    207 			lcv->u_obj.uo_refs++;
    208 			mutex_exit(&lcv->u_obj.vmobjlock);
    209 
    210 			mutex_enter(&udv_lock);
    211 			if (lcv->u_flags & UVM_DEVICE_WANTED)
    212 				wakeup(lcv);
    213 			lcv->u_flags &= ~(UVM_DEVICE_WANTED|UVM_DEVICE_HOLD);
    214 			mutex_exit(&udv_lock);
    215 			return(&lcv->u_obj);
    216 		}
    217 
    218 		/*
    219 		 * did not find it on main list.   need to malloc a new one.
    220 		 */
    221 
    222 		mutex_exit(&udv_lock);
    223 		/* NOTE: we could sleep in the following malloc() */
    224 		udv = malloc(sizeof(*udv), M_TEMP, M_WAITOK);
    225 		mutex_enter(&udv_lock);
    226 
    227 		/*
    228 		 * now we have to double check to make sure no one added it
    229 		 * to the list while we were sleeping...
    230 		 */
    231 
    232 		LIST_FOREACH(lcv, &udv_list, u_list) {
    233 			if (device == lcv->u_device)
    234 				break;
    235 		}
    236 
    237 		/*
    238 		 * did we lose a race to someone else?
    239 		 * free our memory and retry.
    240 		 */
    241 
    242 		if (lcv) {
    243 			mutex_exit(&udv_lock);
    244 			free(udv, M_TEMP);
    245 			continue;
    246 		}
    247 
    248 		/*
    249 		 * we have it!   init the data structures, add to list
    250 		 * and return.
    251 		 */
    252 
    253 		UVM_OBJ_INIT(&udv->u_obj, &uvm_deviceops, 1);
    254 		udv->u_flags = 0;
    255 		udv->u_device = device;
    256 		LIST_INSERT_HEAD(&udv_list, udv, u_list);
    257 		mutex_exit(&udv_lock);
    258 		return(&udv->u_obj);
    259 	}
    260 	/*NOTREACHED*/
    261 }
    262 
    263 /*
    264  * udv_reference
    265  *
    266  * add a reference to a VM object.   Note that the reference count must
    267  * already be one (the passed in reference) so there is no chance of the
    268  * udv being released or locked out here.
    269  *
    270  * => caller must call with object unlocked.
    271  */
    272 
    273 static void
    274 udv_reference(struct uvm_object *uobj)
    275 {
    276 	UVMHIST_FUNC("udv_reference"); UVMHIST_CALLED(maphist);
    277 
    278 	mutex_enter(&uobj->vmobjlock);
    279 	uobj->uo_refs++;
    280 	UVMHIST_LOG(maphist, "<- done (uobj=0x%x, ref = %d)",
    281 		    uobj, uobj->uo_refs,0,0);
    282 	mutex_exit(&uobj->vmobjlock);
    283 }
    284 
    285 /*
    286  * udv_detach
    287  *
    288  * remove a reference to a VM object.
    289  *
    290  * => caller must call with object unlocked and map locked.
    291  */
    292 
    293 static void
    294 udv_detach(struct uvm_object *uobj)
    295 {
    296 	struct uvm_device *udv = (struct uvm_device *)uobj;
    297 	UVMHIST_FUNC("udv_detach"); UVMHIST_CALLED(maphist);
    298 
    299 	/*
    300 	 * loop until done
    301 	 */
    302 again:
    303 	mutex_enter(&uobj->vmobjlock);
    304 	if (uobj->uo_refs > 1) {
    305 		uobj->uo_refs--;
    306 		mutex_exit(&uobj->vmobjlock);
    307 		UVMHIST_LOG(maphist," <- done, uobj=0x%x, ref=%d",
    308 			  uobj,uobj->uo_refs,0,0);
    309 		return;
    310 	}
    311 
    312 	/*
    313 	 * is it being held?   if so, wait until others are done.
    314 	 */
    315 
    316 	mutex_enter(&udv_lock);
    317 	if (udv->u_flags & UVM_DEVICE_HOLD) {
    318 		udv->u_flags |= UVM_DEVICE_WANTED;
    319 		mutex_exit(&uobj->vmobjlock);
    320 		UVM_UNLOCK_AND_WAIT(udv, &udv_lock, false, "udv_detach",0);
    321 		goto again;
    322 	}
    323 
    324 	/*
    325 	 * got it!   nuke it now.
    326 	 */
    327 
    328 	LIST_REMOVE(udv, u_list);
    329 	if (udv->u_flags & UVM_DEVICE_WANTED)
    330 		wakeup(udv);
    331 	mutex_exit(&udv_lock);
    332 	mutex_exit(&uobj->vmobjlock);
    333 	UVM_OBJ_DESTROY(uobj);
    334 	free(udv, M_TEMP);
    335 	UVMHIST_LOG(maphist," <- done, freed uobj=0x%x", uobj,0,0,0);
    336 }
    337 
    338 /*
    339  * udv_fault: non-standard fault routine for device "pages"
    340  *
    341  * => rather than having a "get" function, we have a fault routine
    342  *	since we don't return vm_pages we need full control over the
    343  *	pmap_enter map in
    344  * => all the usual fault data structured are locked by the caller
    345  *	(i.e. maps(read), amap (if any), uobj)
    346  * => on return, we unlock all fault data structures
    347  * => flags: PGO_ALLPAGES: get all of the pages
    348  *	     PGO_LOCKED: fault data structures are locked
    349  *    XXX: currently PGO_LOCKED is always required ... consider removing
    350  *	it as a flag
    351  * => NOTE: vaddr is the VA of pps[0] in ufi->entry, _NOT_ pps[centeridx]
    352  */
    353 
    354 static int
    355 udv_fault(struct uvm_faultinfo *ufi, vaddr_t vaddr, struct vm_page **pps,
    356     int npages, int centeridx, vm_prot_t access_type,
    357     int flags)
    358 {
    359 	struct vm_map_entry *entry = ufi->entry;
    360 	struct uvm_object *uobj = entry->object.uvm_obj;
    361 	struct uvm_device *udv = (struct uvm_device *)uobj;
    362 	vaddr_t curr_va;
    363 	off_t curr_offset;
    364 	paddr_t paddr, mdpgno;
    365 	int lcv, retval;
    366 	dev_t device;
    367 	vm_prot_t mapprot;
    368 	UVMHIST_FUNC("udv_fault"); UVMHIST_CALLED(maphist);
    369 	UVMHIST_LOG(maphist,"  flags=%d", flags,0,0,0);
    370 
    371 	/*
    372 	 * we do not allow device mappings to be mapped copy-on-write
    373 	 * so we kill any attempt to do so here.
    374 	 */
    375 
    376 	if (UVM_ET_ISCOPYONWRITE(entry)) {
    377 		UVMHIST_LOG(maphist, "<- failed -- COW entry (etype=0x%x)",
    378 		entry->etype, 0,0,0);
    379 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
    380 		return(EIO);
    381 	}
    382 
    383 	/*
    384 	 * get device map function.
    385 	 */
    386 
    387 	device = udv->u_device;
    388 	if (cdevsw_lookup(device) == NULL) {
    389 		/* XXX This should not happen */
    390 		uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
    391 		return (EIO);
    392 	}
    393 
    394 	/*
    395 	 * now we must determine the offset in udv to use and the VA to
    396 	 * use for pmap_enter.  note that we always use orig_map's pmap
    397 	 * for pmap_enter (even if we have a submap).   since virtual
    398 	 * addresses in a submap must match the main map, this is ok.
    399 	 */
    400 
    401 	/* udv offset = (offset from start of entry) + entry's offset */
    402 	curr_offset = entry->offset + (vaddr - entry->start);
    403 	/* pmap va = vaddr (virtual address of pps[0]) */
    404 	curr_va = vaddr;
    405 
    406 	/*
    407 	 * loop over the page range entering in as needed
    408 	 */
    409 
    410 	retval = 0;
    411 	for (lcv = 0 ; lcv < npages ; lcv++, curr_offset += PAGE_SIZE,
    412 	    curr_va += PAGE_SIZE) {
    413 		if ((flags & PGO_ALLPAGES) == 0 && lcv != centeridx)
    414 			continue;
    415 
    416 		if (pps[lcv] == PGO_DONTCARE)
    417 			continue;
    418 
    419 		mdpgno = cdev_mmap(device, curr_offset, access_type);
    420 		if (mdpgno == -1) {
    421 			retval = EIO;
    422 			break;
    423 		}
    424 		paddr = pmap_phys_address(mdpgno);
    425 		mapprot = ufi->entry->protection;
    426 		UVMHIST_LOG(maphist,
    427 		    "  MAPPING: device: pm=0x%x, va=0x%x, pa=0x%lx, at=%d",
    428 		    ufi->orig_map->pmap, curr_va, paddr, mapprot);
    429 		if (pmap_enter(ufi->orig_map->pmap, curr_va, paddr,
    430 		    mapprot, PMAP_CANFAIL | mapprot) != 0) {
    431 			/*
    432 			 * pmap_enter() didn't have the resource to
    433 			 * enter this mapping.  Unlock everything,
    434 			 * wait for the pagedaemon to free up some
    435 			 * pages, and then tell uvm_fault() to start
    436 			 * the fault again.
    437 			 *
    438 			 * XXX Needs some rethinking for the PGO_ALLPAGES
    439 			 * XXX case.
    440 			 */
    441 			pmap_update(ufi->orig_map->pmap);	/* sync what we have so far */
    442 			uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap,
    443 			    uobj, NULL);
    444 			uvm_wait("udv_fault");
    445 			return (ERESTART);
    446 		}
    447 	}
    448 
    449 	pmap_update(ufi->orig_map->pmap);
    450 	uvmfault_unlockall(ufi, ufi->entry->aref.ar_amap, uobj, NULL);
    451 	return (retval);
    452 }
    453