[hcoop/debian/openafs.git] / src / afs / FBSD / osi_vm.c

/*
 * Copyright 2000, International Business Machines Corporation and others.
 * All Rights Reserved.
 * 
 * This software has been released under the terms of the IBM Public
 * License.  For details, see the LICENSE file in the top-level source
 * directory or online at http://www.openafs.org/dl/license10.html
 */


/* osi_vm.c implements:
 *
 * osi_VM_FlushVCache(avc)
 * osi_ubc_flush_dirty_and_wait(vp, flags)
 * osi_VM_StoreAllSegments(avc)
 * osi_VM_TryToSmush(avc, acred, sync)
 * osi_VM_FlushPages(avc, credp)
 * osi_VM_Truncate(avc, alen, acred)
 */

#include <afsconfig.h>
#include "afs/param.h"
#include <sys/param.h>
#include <sys/vnode.h>


#include "afs/sysincludes.h"	/* Standard vendor system headers */
#include "afsincludes.h"	/* Afs-based standard headers */
#include "afs/afs_stats.h"	/* statistics */
#include <vm/vm_object.h>
#include <vm/vm_map.h>
#include <sys/limits.h>
#if __FreeBSD_version >= 1000030
#include <sys/rwlock.h>
#endif

/*
 * FreeBSD implementation notes:
 * Most of these operations require us to frob vm_objects.  Most
 * functions require that the object be locked (with VM_OBJECT_*LOCK)
 * on entry and leave it locked on exit.  The locking protocol
 * requires that we access vp->v_object with the heavy vnode lock
 * held and the vnode interlock unlocked.
 *
 * The locking protocol for vnodes is defined in
 * kern/vnode_if.src and sys/vnode.h; unfortunately, it is not *quite*
 * constant from version to version so to be properly correct we must
 * check the VCS history of those files.
 */

#if defined(AFS_FBSD80_ENV)
#define	lock_vnode(v, f)	vn_lock((v), (f))
#define ilock_vnode(v)	vn_lock((v), LK_INTERLOCK|LK_EXCLUSIVE|LK_RETRY)
#define unlock_vnode(v)	VOP_UNLOCK((v), 0)
#define islocked_vnode(v)	VOP_ISLOCKED((v))
#else
#define	lock_vnode(v, f)	vn_lock((v), (f), curthread)
#define ilock_vnode(v)	vn_lock((v), LK_INTERLOCK|LK_EXCLUSIVE|LK_RETRY, curthread)
#define unlock_vnode(v)	VOP_UNLOCK((v), 0, curthread)
#define islocked_vnode(v)	VOP_ISLOCKED((v), curthread)
#endif

#if __FreeBSD_version >= 1000030
#define AFS_VM_OBJECT_WLOCK(o)	VM_OBJECT_WLOCK(o)
#define AFS_VM_OBJECT_WUNLOCK(o)	VM_OBJECT_WUNLOCK(o)
#else
#define AFS_VM_OBJECT_WLOCK(o)	VM_OBJECT_LOCK(o)
#define AFS_VM_OBJECT_WUNLOCK(o)	VM_OBJECT_UNLOCK(o)
#endif

/* Try to discard pages, in order to recycle a vcache entry.
 *
 * We also make some sanity checks:  ref count, open count, held locks.
 *
 * We also do some non-VM-related chores, such as releasing the cred pointer
 * (for AIX and Solaris) and releasing the gnode (for AIX).
 *
 * Locking:  afs_xvcache lock is held. It must not be dropped.
 *
 */
int
osi_VM_FlushVCache(struct vcache *avc)
{
    struct vnode *vp;
    int code;

    vp = AFSTOV(avc);

    if (!VI_TRYLOCK(vp))
	return EBUSY;
    code = osi_fbsd_checkinuse(avc);
    if (code) {
	VI_UNLOCK(vp);
	return code;
    }

    /* must hold the vnode before calling cache_purge()
     * This code largely copied from vfs_subr.c:vlrureclaim() */
    vholdl(vp);
    VI_UNLOCK(vp);

    AFS_GUNLOCK();
    cache_purge(vp);
    AFS_GLOCK();

    vdrop(vp);

    return 0;
}

/* Try to store pages to cache, in order to store a file back to the server.
 *
 * Locking:  the vcache entry's lock is held.  It will usually be dropped and
 * re-obtained.
 */
void
osi_VM_StoreAllSegments(struct vcache *avc)
{
    struct vnode *vp;
    struct vm_object *obj;
    int anyio, tries;

    ReleaseWriteLock(&avc->lock);
    AFS_GUNLOCK();
    tries = 5;
    vp = AFSTOV(avc);

    /*
     * I don't understand this.  Why not just call vm_object_page_clean()
     * and be done with it?  I particularly don't understand why we're calling
     * vget() here.  Is there some reason to believe that the vnode might
     * be being recycled at this point?  I don't think there's any need for
     * this loop, either -- if we keep the vnode locked all the time,
     * that and the object lock will prevent any new pages from appearing.
     * The loop is what causes the race condition.  -GAW
     */
    do {
	anyio = 0;
	
	obj = vp->v_object;
	if (obj != NULL && obj->flags & OBJ_MIGHTBEDIRTY) {
	    if (!vget(vp, LK_EXCLUSIVE | LK_RETRY, curthread)) {
		    obj = vp->v_object;
		    if (obj != NULL) {
			AFS_VM_OBJECT_WLOCK(obj);
			vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
			AFS_VM_OBJECT_WUNLOCK(obj);
			anyio = 1;
		    }
		    vput(vp);
		}
	    }
    } while (anyio && (--tries > 0));
    AFS_GLOCK();
    ObtainWriteLock(&avc->lock, 94);
}

/* Try to invalidate pages, for "fs flush" or "fs flushv"; or
 * try to free pages, when deleting a file.
 *
 * Locking:  the vcache entry's lock is held.  It may be dropped and 
 * re-obtained.
 *
 * Since we drop and re-obtain the lock, we can't guarantee that there won't
 * be some pages around when we return, newly created by concurrent activity.
 */
void
osi_VM_TryToSmush(struct vcache *avc, afs_ucred_t *acred, int sync)
{
    struct vnode *vp;
    int tries, code;
    int islocked;

    vp = AFSTOV(avc);

    VI_LOCK(vp);
    if (vp->v_iflag & VI_DOOMED) {
	VI_UNLOCK(vp);
	return;
    }
    VI_UNLOCK(vp);

    islocked = islocked_vnode(vp);
    if (islocked == LK_EXCLOTHER)
	panic("Trying to Smush over someone else's lock");
    else if (islocked == LK_SHARED) {
	afs_warn("Trying to Smush with a shared lock");
	lock_vnode(vp, LK_UPGRADE);
    } else if (!islocked)
	lock_vnode(vp, LK_EXCLUSIVE);

    if (vp->v_bufobj.bo_object != NULL) {
	AFS_VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
	/*
	 * Do we really want OBJPC_SYNC?  OBJPC_INVAL would be
	 * faster, if invalidation is really what we are being
	 * asked to do.  (It would make more sense, too, since
	 * otherwise this function is practically identical to
	 * osi_VM_StoreAllSegments().)  -GAW
	 */

	/*
	 * Dunno.  We no longer resemble osi_VM_StoreAllSegments,
	 * though maybe that's wrong, now.  And OBJPC_SYNC is the
	 * common thing in 70 file systems, it seems.  Matt.
	 */

	vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
	AFS_VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
    }

    tries = 5;
    code = osi_vinvalbuf(vp, V_SAVE, PCATCH, 0);
    while (code && (tries > 0)) {
	afs_warn("TryToSmush retrying vinvalbuf");
	code = osi_vinvalbuf(vp, V_SAVE, PCATCH, 0);
	--tries;
    }
    if (islocked == LK_SHARED)
	lock_vnode(vp, LK_DOWNGRADE);
    else if (!islocked)
	unlock_vnode(vp);
}

/* Purge VM for a file when its callback is revoked.
 *
 * Locking:  No lock is held, not even the global lock.
 */
void
osi_VM_FlushPages(struct vcache *avc, afs_ucred_t *credp)
{
    struct vnode *vp;
    struct vm_object *obj;

    vp = AFSTOV(avc);
    ASSERT_VOP_LOCKED(vp, __func__);
    obj = vp->v_object;
    if (obj != NULL) {
	AFS_VM_OBJECT_WLOCK(obj);
	vm_object_page_remove(obj, 0, 0, FALSE);
	AFS_VM_OBJECT_WUNLOCK(obj);
    }
    osi_vinvalbuf(vp, 0, 0, 0);
}

/* Purge pages beyond end-of-file, when truncating a file.
 *
 * Locking:  no lock is held, not even the global lock.
 * activeV is raised.  This is supposed to block pageins, but at present
 * it only works on Solaris.
 */
void
osi_VM_Truncate(struct vcache *avc, int alen, afs_ucred_t *acred)
{
    vnode_pager_setsize(AFSTOV(avc), alen);
}
Commit	Line	Data
805e021f CE	1	/*
	2	* Copyright 2000, International Business Machines Corporation and others.
	3	* All Rights Reserved.
	4	*
	5	* This software has been released under the terms of the IBM Public
	6	* License. For details, see the LICENSE file in the top-level source
	7	* directory or online at http://www.openafs.org/dl/license10.html
	8	*/
	9
	10
	11	/* osi_vm.c implements:
	12	*
	13	* osi_VM_FlushVCache(avc)
	14	* osi_ubc_flush_dirty_and_wait(vp, flags)
	15	* osi_VM_StoreAllSegments(avc)
	16	* osi_VM_TryToSmush(avc, acred, sync)
	17	* osi_VM_FlushPages(avc, credp)
	18	* osi_VM_Truncate(avc, alen, acred)
	19	*/
	20
	21	#include <afsconfig.h>
	22	#include "afs/param.h"
	23	#include <sys/param.h>
	24	#include <sys/vnode.h>
	25
	26
	27	#include "afs/sysincludes.h" /* Standard vendor system headers */
	28	#include "afsincludes.h" /* Afs-based standard headers */
	29	#include "afs/afs_stats.h" /* statistics */
	30	#include <vm/vm_object.h>
	31	#include <vm/vm_map.h>
	32	#include <sys/limits.h>
	33	#if __FreeBSD_version >= 1000030
	34	#include <sys/rwlock.h>
	35	#endif
	36
	37	/*
	38	* FreeBSD implementation notes:
	39	* Most of these operations require us to frob vm_objects. Most
	40	* functions require that the object be locked (with VM_OBJECT_*LOCK)
	41	* on entry and leave it locked on exit. The locking protocol
	42	* requires that we access vp->v_object with the heavy vnode lock
	43	* held and the vnode interlock unlocked.
	44	*
	45	* The locking protocol for vnodes is defined in
	46	* kern/vnode_if.src and sys/vnode.h; unfortunately, it is not quite
	47	* constant from version to version so to be properly correct we must
	48	* check the VCS history of those files.
	49	*/
	50
	51	#if defined(AFS_FBSD80_ENV)
	52	#define lock_vnode(v, f) vn_lock((v), (f))
	53	#define ilock_vnode(v) vn_lock((v), LK_INTERLOCK\|LK_EXCLUSIVE\|LK_RETRY)
	54	#define unlock_vnode(v) VOP_UNLOCK((v), 0)
	55	#define islocked_vnode(v) VOP_ISLOCKED((v))
	56	#else
	57	#define lock_vnode(v, f) vn_lock((v), (f), curthread)
	58	#define ilock_vnode(v) vn_lock((v), LK_INTERLOCK\|LK_EXCLUSIVE\|LK_RETRY, curthread)
	59	#define unlock_vnode(v) VOP_UNLOCK((v), 0, curthread)
	60	#define islocked_vnode(v) VOP_ISLOCKED((v), curthread)
	61	#endif
	62
	63	#if __FreeBSD_version >= 1000030
	64	#define AFS_VM_OBJECT_WLOCK(o) VM_OBJECT_WLOCK(o)
65	#define AFS_VM_OBJECT_WUNLOCK(o) VM_OBJECT_WUNLOCK(o)
66	#else
67	#define AFS_VM_OBJECT_WLOCK(o) VM_OBJECT_LOCK(o)
68	#define AFS_VM_OBJECT_WUNLOCK(o) VM_OBJECT_UNLOCK(o)
69	#endif
70
71	/* Try to discard pages, in order to recycle a vcache entry.
72	*
73	* We also make some sanity checks: ref count, open count, held locks.
74	*
75	* We also do some non-VM-related chores, such as releasing the cred pointer
76	* (for AIX and Solaris) and releasing the gnode (for AIX).
77	*
78	* Locking: afs_xvcache lock is held. It must not be dropped.
79	*
80	*/
81	int
82	osi_VM_FlushVCache(struct vcache *avc)
83	{
84	struct vnode *vp;
85	int code;
86
87	vp = AFSTOV(avc);
88
89	if (!VI_TRYLOCK(vp))
90	return EBUSY;
91	code = osi_fbsd_checkinuse(avc);
92	if (code) {
93	VI_UNLOCK(vp);
94	return code;
95	}
96
97	/* must hold the vnode before calling cache_purge()
98	* This code largely copied from vfs_subr.c:vlrureclaim() */
99	vholdl(vp);
100	VI_UNLOCK(vp);
101
102	AFS_GUNLOCK();
103	cache_purge(vp);
104	AFS_GLOCK();
105
106	vdrop(vp);
107
108	return 0;
109	}
110
111	/* Try to store pages to cache, in order to store a file back to the server.
112	*
113	* Locking: the vcache entry's lock is held. It will usually be dropped and
114	* re-obtained.
115	*/
116	void
117	osi_VM_StoreAllSegments(struct vcache *avc)
118	{
119	struct vnode *vp;
120	struct vm_object *obj;
121	int anyio, tries;
122
123	ReleaseWriteLock(&avc->lock);
124	AFS_GUNLOCK();
125	tries = 5;
126	vp = AFSTOV(avc);
127
128	/*
129	* I don't understand this. Why not just call vm_object_page_clean()
130	* and be done with it? I particularly don't understand why we're calling
131	* vget() here. Is there some reason to believe that the vnode might
132	* be being recycled at this point? I don't think there's any need for
133	* this loop, either -- if we keep the vnode locked all the time,
134	* that and the object lock will prevent any new pages from appearing.
135	* The loop is what causes the race condition. -GAW
136	*/
137	do {
138	anyio = 0;
139
140	obj = vp->v_object;
141	if (obj != NULL && obj->flags & OBJ_MIGHTBEDIRTY) {
142	if (!vget(vp, LK_EXCLUSIVE \| LK_RETRY, curthread)) {
143	obj = vp->v_object;
144	if (obj != NULL) {
145	AFS_VM_OBJECT_WLOCK(obj);
146	vm_object_page_clean(obj, 0, 0, OBJPC_SYNC);
147	AFS_VM_OBJECT_WUNLOCK(obj);
148	anyio = 1;
149	}
150	vput(vp);
151	}
152	}
153	} while (anyio && (--tries > 0));
154	AFS_GLOCK();
155	ObtainWriteLock(&avc->lock, 94);
156	}
157
158	/* Try to invalidate pages, for "fs flush" or "fs flushv"; or
159	* try to free pages, when deleting a file.
160	*
161	* Locking: the vcache entry's lock is held. It may be dropped and
162	* re-obtained.
163	*
164	* Since we drop and re-obtain the lock, we can't guarantee that there won't
165	* be some pages around when we return, newly created by concurrent activity.
166	*/
167	void
168	osi_VM_TryToSmush(struct vcache avc, afs_ucred_t acred, int sync)
169	{
170	struct vnode *vp;
171	int tries, code;
172	int islocked;
173
174	vp = AFSTOV(avc);
175
176	VI_LOCK(vp);
177	if (vp->v_iflag & VI_DOOMED) {
178	VI_UNLOCK(vp);
179	return;
180	}
181	VI_UNLOCK(vp);
182
183	islocked = islocked_vnode(vp);
184	if (islocked == LK_EXCLOTHER)
185	panic("Trying to Smush over someone else's lock");
186	else if (islocked == LK_SHARED) {
187	afs_warn("Trying to Smush with a shared lock");
188	lock_vnode(vp, LK_UPGRADE);
189	} else if (!islocked)
190	lock_vnode(vp, LK_EXCLUSIVE);
191
192	if (vp->v_bufobj.bo_object != NULL) {
193	AFS_VM_OBJECT_WLOCK(vp->v_bufobj.bo_object);
194	/*
195	* Do we really want OBJPC_SYNC? OBJPC_INVAL would be
196	* faster, if invalidation is really what we are being
197	* asked to do. (It would make more sense, too, since
198	* otherwise this function is practically identical to
199	* osi_VM_StoreAllSegments().) -GAW
200	*/
201
202	/*
203	* Dunno. We no longer resemble osi_VM_StoreAllSegments,
204	* though maybe that's wrong, now. And OBJPC_SYNC is the
205	* common thing in 70 file systems, it seems. Matt.
206	*/
207
208	vm_object_page_clean(vp->v_bufobj.bo_object, 0, 0, OBJPC_SYNC);
209	AFS_VM_OBJECT_WUNLOCK(vp->v_bufobj.bo_object);
210	}
211
212	tries = 5;
213	code = osi_vinvalbuf(vp, V_SAVE, PCATCH, 0);
214	while (code && (tries > 0)) {
215	afs_warn("TryToSmush retrying vinvalbuf");
216	code = osi_vinvalbuf(vp, V_SAVE, PCATCH, 0);
217	--tries;
218	}
219	if (islocked == LK_SHARED)
220	lock_vnode(vp, LK_DOWNGRADE);
221	else if (!islocked)
222	unlock_vnode(vp);
223	}
224
225	/* Purge VM for a file when its callback is revoked.
226	*
227	* Locking: No lock is held, not even the global lock.
228	*/
229	void
230	osi_VM_FlushPages(struct vcache avc, afs_ucred_t credp)
231	{
232	struct vnode *vp;
233	struct vm_object *obj;
234
235	vp = AFSTOV(avc);
236	ASSERT_VOP_LOCKED(vp, __func__);
237	obj = vp->v_object;
238	if (obj != NULL) {
239	AFS_VM_OBJECT_WLOCK(obj);
240	vm_object_page_remove(obj, 0, 0, FALSE);
241	AFS_VM_OBJECT_WUNLOCK(obj);
242	}
243	osi_vinvalbuf(vp, 0, 0, 0);
244	}
245
246	/* Purge pages beyond end-of-file, when truncating a file.
247	*
248	* Locking: no lock is held, not even the global lock.
249	* activeV is raised. This is supposed to block pageins, but at present
250	* it only works on Solaris.
251	*/
252	void
253	osi_VM_Truncate(struct vcache avc, int alen, afs_ucred_t acred)
254	{
255	vnode_pager_setsize(AFSTOV(avc), alen);
256	}