/*-
* Copyright (c) 2007, 2008, 2009, 2019, 2020 The NetBSD Foundation, Inc.
* All rights reserved.
*
* This code is derived from software contributed to The NetBSD Foundation
* by Andrew Doran, and by Wasabi Systems, Inc.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
* ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
* TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
* BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
/*-
* Copyright (c) 1982, 1986, 1989, 1993
* The Regents of the University of California. All rights reserved.
* (c) UNIX System Laboratories, Inc.
* All or some portions of this file are derived from material licensed
* to the University of California by American Telephone and Telegraph
* Co. or Unix System Laboratories, Inc. and are reproduced herein with
* the permission of UNIX System Laboratories, Inc.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. Neither the name of the University nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
* @(#)vfs_bio.c 8.6 (Berkeley) 1/11/94
*/
/*-
* Copyright (c) 1994 Christopher G. Demetriou
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. All advertising materials mentioning features or use of this software
* must display the following acknowledgement:
* This product includes software developed by the University of
* California, Berkeley and its contributors.
* 4. Neither the name of the University nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
* @(#)vfs_bio.c 8.6 (Berkeley) 1/11/94
*/
/*
* The buffer cache subsystem.
*
* Some references:
* Bach: The Design of the UNIX Operating System (Prentice Hall, 1986)
* Leffler, et al.: The Design and Implementation of the 4.3BSD
* UNIX Operating System (Addison Welley, 1989)
*
* Locking
*
* There are three locks:
* - bufcache_lock: protects global buffer cache state.
* - BC_BUSY: a long term per-buffer lock.
* - buf_t::b_objlock: lock on completion (biowait vs biodone).
*
* For buffers associated with vnodes (a most common case) b_objlock points
* to the vnode_t::v_interlock. Otherwise, it points to generic buffer_lock.
*
* Lock order:
* bufcache_lock ->
* buf_t::b_objlock
*/
#ifdef BUFCACHE
# if (BUFCACHE < 5) || (BUFCACHE > 95)
# error BUFCACHE is not between 5 and 95
# endif
#else
# define BUFCACHE 15
#endif
u_int nbuf; /* desired number of buffer headers */
u_int bufpages = BUFPAGES; /* optional hardwired count */
u_int bufcache = BUFCACHE; /* max % of RAM to use for buffer cache */
/*
* Definitions for the buffer free lists.
*/
#define BQUEUES 3 /* number of free buffer queues */
/*
* note that for some ports this is used by pmap bootstrap code to
* determine kva size.
*/
u_long
buf_memcalc(void)
{
u_long n;
vsize_t mapsz = 0;
/*
* Determine the upper bound of memory to use for buffers.
*
* - If bufpages is specified, use that as the number
* pages.
*
* - Otherwise, use bufcache as the percentage of
* physical memory.
*/
if (bufpages != 0) {
n = bufpages;
} else {
if (bufcache < 5) {
printf("forcing bufcache %d -> 5", bufcache);
bufcache = 5;
}
if (bufcache > 95) {
printf("forcing bufcache %d -> 95", bufcache);
bufcache = 95;
}
if (buf_map != NULL)
mapsz = vm_map_max(buf_map) - vm_map_min(buf_map);
n = calc_cache_size(mapsz, bufcache,
(buf_map != kernel_map) ? 100 : BUFCACHE_VA_MAXPCT)
/ PAGE_SIZE;
}
n <<= PAGE_SHIFT;
if (bufmem_valimit != 0 && n > bufmem_valimit)
n = bufmem_valimit;
return n;
}
/*
* Initialize buffers and hash links for buffers.
*/
void
bufinit(void)
{
struct bqueue *dp;
int use_std;
u_int i;
/* Initialize the buffer queues */
for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) {
TAILQ_INIT(&dp->bq_queue);
dp->bq_bytes = 0;
}
/*
* Estimate hash table size based on the amount of memory we
* intend to use for the buffer cache. The average buffer
* size is dependent on our clients (i.e. filesystems).
*
* For now, use an empirical 3K per buffer.
*/
nbuf = (bufmem_hiwater / 1024) / 3;
bufhashtbl = hashinit(nbuf, HASH_LIST, true, &bufhash);
/* Always allocate if less than the low water mark. */
if (bufmem < bufmem_lowater)
return 1;
/* Never allocate if greater than the high water mark. */
if (bufmem > bufmem_hiwater)
return 0;
/* If there's anything on the AGE list, it should be eaten. */
if (TAILQ_FIRST(&bufqueues[BQ_AGE].bq_queue) != NULL)
return 0;
/*
* The probabily of getting a new allocation is inversely
* proportional to the current size of the cache above
* the low water mark. Divide the total first to avoid overflows
* in the product.
*/
guess = cprng_fast32() % 16;
/*
* Return estimate of bytes we think need to be
* released to help resolve low memory conditions.
*
* => called with bufcache_lock held.
*/
static int
buf_canrelease(void)
{
int pagedemand, ninvalid = 0;
KASSERT(mutex_owned(&bufcache_lock));
if (bufmem < bufmem_lowater)
return 0;
if (bufmem > bufmem_hiwater)
return bufmem - bufmem_hiwater;
while (1) {
addr = pool_get(&bmempools[n], PR_NOWAIT);
if (addr != NULL)
break;
/* No memory, see if we can free some. If so, try again */
mutex_enter(&bufcache_lock);
if (buf_drain(1) > 0) {
mutex_exit(&bufcache_lock);
continue;
}
if (curlwp == uvm.pagedaemon_lwp) {
mutex_exit(&bufcache_lock);
return NULL;
}
/* Wait for buffers to arrive on the LRU queue */
cv_timedwait(&needbuffer_cv, &bufcache_lock, hz / 4);
mutex_exit(&bufcache_lock);
}
/*
* bread()/breadn() helper.
*/
static buf_t *
bio_doread(struct vnode *vp, daddr_t blkno, int size, int async)
{
buf_t *bp;
struct mount *mp;
bp = getblk(vp, blkno, size, 0, 0);
/*
* getblk() may return NULL if we are the pagedaemon.
*/
if (bp == NULL) {
KASSERT(curlwp == uvm.pagedaemon_lwp);
return NULL;
}
/*
* If buffer does not have data valid, start a read.
* Note that if buffer is BC_INVAL, getblk() won't return it.
* Therefore, it's valid if its I/O has completed or been delayed.
*/
if (!ISSET(bp->b_oflags, (BO_DONE | BO_DELWRI))) {
/* Start I/O for the buffer. */
SET(bp->b_flags, B_READ | async);
if (async)
BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
else
BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
VOP_STRATEGY(vp, bp);
/* Pay for the read. */
curlwp->l_ru.ru_inblock++;
} else if (async)
brelse(bp, 0);
/*
* Collect statistics on synchronous and asynchronous reads.
* Reads from block devices are charged to their associated
* filesystem (if any).
*/
if (mp != NULL) {
if (async == 0)
mp->mnt_stat.f_syncreads++;
else
mp->mnt_stat.f_asyncreads++;
}
return bp;
}
/*
* Read a disk block.
* This algorithm described in Bach (p.54).
*/
int
bread(struct vnode *vp, daddr_t blkno, int size, int flags, buf_t **bpp)
{
buf_t *bp;
int error;
BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
/* Get buffer for block. */
bp = *bpp = bio_doread(vp, blkno, size, 0);
if (bp == NULL)
return SET_ERROR(ENOMEM);
/* Wait for the read to complete, and return result. */
error = biowait(bp);
if (error == 0 && (flags & B_MODIFY) != 0)
error = fscow_run(bp, true);
if (error) {
brelse(bp, 0);
*bpp = NULL;
}
return error;
}
/*
* Read-ahead multiple disk blocks. The first is sync, the rest async.
* Trivial modification to the breada algorithm presented in Bach (p.55).
*/
int
breadn(struct vnode *vp, daddr_t blkno, int size, daddr_t *rablks,
int *rasizes, int nrablks, int flags, buf_t **bpp)
{
buf_t *bp;
int error, i;
BIOHIST_FUNC(__func__); BIOHIST_CALLED(biohist);
bp = *bpp = bio_doread(vp, blkno, size, 0);
if (bp == NULL)
return SET_ERROR(ENOMEM);
/*
* For each of the read-ahead blocks, start a read, if necessary.
*/
mutex_enter(&bufcache_lock);
for (i = 0; i < nrablks; i++) {
/* If it's in the cache, just go on to next one. */
if (incore(vp, rablks[i]))
continue;
/* Get a buffer for the read-ahead block */
mutex_exit(&bufcache_lock);
(void) bio_doread(vp, rablks[i], rasizes[i], B_ASYNC);
mutex_enter(&bufcache_lock);
}
mutex_exit(&bufcache_lock);
/* Otherwise, we had to start a read for it; wait until it's valid. */
error = biowait(bp);
if (error == 0 && (flags & B_MODIFY) != 0)
error = fscow_run(bp, true);
if (error) {
brelse(bp, 0);
*bpp = NULL;
}
return error;
}
/*
* Block write. Described in Bach (p.56)
*/
int
bwrite(buf_t *bp)
{
int rv, sync, wasdelayed;
struct vnode *vp;
struct mount *mp;
/*
* dholland 20160728 AFAICT vp==NULL must be impossible as it
* will crash upon reaching VOP_STRATEGY below... see further
* analysis on tech-kern.
*/
KASSERTMSG(vp != NULL, "bwrite given buffer with null vnode");
if (mp && mp->mnt_wapbl) {
if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
bdwrite(bp);
return 0;
}
}
/*
* Remember buffer type, to switch on it later. If the write was
* synchronous, but the file system was mounted with MNT_ASYNC,
* convert it to a delayed write.
* XXX note that this relies on delayed tape writes being converted
* to async, not sync writes (which is safe, but ugly).
*/
sync = !ISSET(bp->b_flags, B_ASYNC);
if (sync && mp != NULL && ISSET(mp->mnt_flag, MNT_ASYNC)) {
bdwrite(bp);
return 0;
}
/*
* Collect statistics on synchronous and asynchronous writes.
* Writes to block devices are charged to their associated
* filesystem (if any).
*/
if (mp != NULL) {
if (sync)
mp->mnt_stat.f_syncwrites++;
else
mp->mnt_stat.f_asyncwrites++;
}
/*
* Pay for the I/O operation and make sure the buf is on the correct
* vnode queue.
*/
bp->b_error = 0;
wasdelayed = ISSET(bp->b_oflags, BO_DELWRI);
CLR(bp->b_flags, B_READ);
if (wasdelayed) {
mutex_enter(&bufcache_lock);
mutex_enter(bp->b_objlock);
CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
reassignbuf(bp, bp->b_vp);
/* Wake anyone trying to busy the buffer via vnode's lists. */
cv_broadcast(&bp->b_busy);
mutex_exit(&bufcache_lock);
} else {
curlwp->l_ru.ru_oublock++;
mutex_enter(bp->b_objlock);
CLR(bp->b_oflags, BO_DONE | BO_DELWRI);
}
if (vp != NULL)
vp->v_numoutput++;
mutex_exit(bp->b_objlock);
/* Initiate disk write. */
if (sync)
BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
else
BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
VOP_STRATEGY(vp, bp);
if (sync) {
/* If I/O was synchronous, wait for it to complete. */
rv = biowait(bp);
/* Release the buffer. */
brelse(bp, 0);
return rv;
} else {
return 0;
}
}
int
vn_bwrite(void *v)
{
struct vop_bwrite_args *ap = v;
return bwrite(ap->a_bp);
}
/*
* Delayed write.
*
* The buffer is marked dirty, but is not queued for I/O.
* This routine should be used when the buffer is expected
* to be modified again soon, typically a small write that
* partially fills a buffer.
*
* NB: magnetic tapes cannot be delayed; they must be
* written in the order that the writes are requested.
*
* Described in Leffler, et al. (pp. 208-213).
*/
void
bdwrite(buf_t *bp)
{
/* If this is a tape block, write the block now. */
if (bdev_type(bp->b_dev) == D_TAPE) {
bawrite(bp);
return;
}
if (wapbl_vphaswapbl(bp->b_vp)) {
struct mount *mp = wapbl_vptomp(bp->b_vp);
if (bp->b_iodone != mp->mnt_wapbl_op->wo_wapbl_biodone) {
WAPBL_ADD_BUF(mp, bp);
}
}
/*
* If the block hasn't been seen before:
* (1) Mark it as having been seen,
* (2) Charge for the write,
* (3) Make sure it's on its vnode's correct block list.
*/
KASSERT(bp->b_vp == NULL || bp->b_objlock == bp->b_vp->v_interlock);
if (!ISSET(bp->b_oflags, BO_DELWRI)) {
mutex_enter(&bufcache_lock);
mutex_enter(bp->b_objlock);
SET(bp->b_oflags, BO_DELWRI);
curlwp->l_ru.ru_oublock++;
reassignbuf(bp, bp->b_vp);
/* Wake anyone trying to busy the buffer via vnode's lists. */
cv_broadcast(&bp->b_busy);
mutex_exit(&bufcache_lock);
} else {
mutex_enter(bp->b_objlock);
}
/* Otherwise, the "write" is done, so mark and release the buffer. */
CLR(bp->b_oflags, BO_DONE);
mutex_exit(bp->b_objlock);
brelse(bp, 0);
}
/*
* Asynchronous block write; just an asynchronous bwrite().
*/
void
bawrite(buf_t *bp)
{
/* Wake up any processes waiting for any buffer to become free. */
cv_signal(&needbuffer_cv);
/* Wake up any proceeses waiting for _this_ buffer to become free */
if (ISSET(bp->b_cflags, BC_WANTED))
CLR(bp->b_cflags, BC_WANTED|BC_AGE);
/* If it's clean clear the copy-on-write flag. */
if (ISSET(bp->b_flags, B_COWDONE)) {
mutex_enter(bp->b_objlock);
if (!ISSET(bp->b_oflags, BO_DELWRI))
CLR(bp->b_flags, B_COWDONE);
mutex_exit(bp->b_objlock);
}
/*
* Determine which queue the buffer should be on, then put it there.
*/
/* If it's locked, don't report an error; try again later. */
if (ISSET(bp->b_flags, B_LOCKED))
bp->b_error = 0;
/* If it's not cacheable, or an error, mark it invalid. */
if (ISSET(bp->b_cflags, BC_NOCACHE) || bp->b_error != 0)
SET(bp->b_cflags, BC_INVAL);
if (ISSET(bp->b_cflags, BC_VFLUSH)) {
/*
* This is a delayed write buffer that was just flushed to
* disk. It is still on the LRU queue. If it's become
* invalid, then we need to move it to a different queue;
* otherwise leave it in its current position.
*/
CLR(bp->b_cflags, BC_VFLUSH);
if (!ISSET(bp->b_cflags, BC_INVAL|BC_AGE) &&
!ISSET(bp->b_flags, B_LOCKED) && bp->b_error == 0) {
KDASSERT(checkfreelist(bp, &bufqueues[BQ_LRU], 1));
goto already_queued;
} else {
bremfree(bp);
}
}
if ((bp->b_bufsize <= 0) || ISSET(bp->b_cflags, BC_INVAL)) {
/*
* If it's invalid or empty, dissociate it from its vnode
* and put on the head of the appropriate queue.
*/
if (ISSET(bp->b_flags, B_LOCKED)) {
if (wapbl_vphaswapbl(vp = bp->b_vp)) {
struct mount *mp = wapbl_vptomp(vp);
mutex_enter(bp->b_objlock);
CLR(bp->b_oflags, BO_DONE|BO_DELWRI);
if ((vp = bp->b_vp) != NULL) {
KASSERT(bp->b_objlock == vp->v_interlock);
reassignbuf(bp, bp->b_vp);
brelvp(bp);
mutex_exit(vp->v_interlock);
} else {
KASSERT(bp->b_objlock == &buffer_lock);
mutex_exit(bp->b_objlock);
}
/* We want to dispose of the buffer, so wake everybody. */
cv_broadcast(&bp->b_busy);
if (bp->b_bufsize <= 0)
/* no data */
goto already_queued;
else
/* invalid data */
bufq = &bufqueues[BQ_AGE];
binsheadfree(bp, bufq);
} else {
/*
* It has valid data. Put it on the end of the appropriate
* queue, so that it'll stick around for as long as possible.
* If buf is AGE, but has dependencies, must put it on last
* bufqueue to be scanned, ie LRU. This protects against the
* livelock where BQ_AGE only has buffers with dependencies,
* and we thus never get to the dependent buffers in BQ_LRU.
*/
if (ISSET(bp->b_flags, B_LOCKED)) {
/* locked in core */
bufq = &bufqueues[BQ_LOCKED];
} else if (!ISSET(bp->b_cflags, BC_AGE)) {
/* valid data */
bufq = &bufqueues[BQ_LRU];
} else {
/* stale but valid data */
bufq = &bufqueues[BQ_AGE];
}
binstailfree(bp, bufq);
}
already_queued:
/* Unlock the buffer. */
CLR(bp->b_cflags, BC_AGE|BC_BUSY|BC_NOCACHE);
CLR(bp->b_flags, B_ASYNC);
/*
* Wake only the highest priority waiter on the lock, in order to
* prevent a thundering herd: many LWPs simultaneously awakening and
* competing for the buffer's lock. Testing in 2019 revealed this
* to reduce contention on bufcache_lock tenfold during a kernel
* compile. Here and elsewhere, when the buffer is changing
* identity, being disposed of, or moving from one list to another,
* we wake all lock requestors.
*/
if (bp->b_bufsize <= 0) {
cv_broadcast(&bp->b_busy);
buf_destroy(bp);
#ifdef DEBUG
memset((char *)bp, 0, sizeof(*bp));
#endif
pool_cache_put(buf_cache, bp);
} else
cv_signal(&bp->b_busy);
}
/*
* Determine if a block is in the cache.
* Just look on what would be its hash chain. If it's there, return
* a pointer to it, unless it's marked invalid. If it's marked invalid,
* we normally don't return the buffer, unless the caller explicitly
* wants us to.
*/
buf_t *
incore(struct vnode *vp, daddr_t blkno)
{
buf_t *bp;
/*
* Get a block of requested size that is associated with
* a given vnode and block offset. If it is found in the
* block cache, mark it as having been found, make it busy
* and return it. Otherwise, return an empty block of the
* correct size. It is up to the caller to insure that the
* cached blocks be of the correct size.
*/
buf_t *
getblk(struct vnode *vp, daddr_t blkno, int size, int slpflag, int slptimeo)
{
int err, preserve;
buf_t *bp;
/*
* Expand or contract the actual memory allocated to a buffer.
*
* If the buffer shrinks, data is lost, so it's up to the
* caller to have written it out *first*; this routine will not
* start a write. If the buffer grows, it's the callers
* responsibility to fill out the buffer's additional contents.
*/
int
allocbuf(buf_t *bp, int size, int preserve)
{
void *addr;
vsize_t oldsize, desired_size;
int oldcount;
int delta;
desired_size = buf_roundsize(size);
if (desired_size > MAXBSIZE)
printf("allocbuf: buffer larger than MAXBSIZE requested");
oldcount = bp->b_bcount;
bp->b_bcount = size;
oldsize = bp->b_bufsize;
if (oldsize == desired_size) {
/*
* Do not short cut the WAPBL resize, as the buffer length
* could still have changed and this would corrupt the
* tracking of the transaction length.
*/
goto out;
}
/*
* If we want a buffer of a different size, re-allocate the
* buffer's memory; copy old content only if needed.
*/
addr = buf_alloc(desired_size);
if (addr == NULL)
return SET_ERROR(ENOMEM);
if (preserve)
memcpy(addr, bp->b_data, MIN(oldsize,desired_size));
if (bp->b_data != NULL)
buf_mrelease(bp->b_data, oldsize);
bp->b_data = addr;
bp->b_bufsize = desired_size;
mutex_enter(&bufcache_lock);
if ((bufmem += delta) > bufmem_hiwater) {
/*
* Need to trim overall memory usage.
*/
while (buf_canrelease()) {
if (preempt_needed()) {
mutex_exit(&bufcache_lock);
preempt();
mutex_enter(&bufcache_lock);
}
if (buf_trim() == 0)
break;
}
}
mutex_exit(&bufcache_lock);
out:
if (wapbl_vphaswapbl(bp->b_vp)) {
WAPBL_RESIZE_BUF(wapbl_vptomp(bp->b_vp), bp,
oldsize, oldcount);
}
return 0;
}
/*
* Find a buffer which is available for use.
* Select something from a free list.
* Preference is to AGE list, then LRU list.
*
* Called with the buffer queues locked.
* Return buffer locked.
*/
static buf_t *
getnewbuf(int slpflag, int slptimeo, int from_bufq)
{
buf_t *bp;
struct vnode *vp;
struct mount *transmp = NULL;
SDT_PROBE0(io, kernel, , getnewbuf__start);
start:
KASSERT(mutex_owned(&bufcache_lock));
/*
* Get a new buffer from the pool.
*/
if (!from_bufq && buf_lotsfree()) {
mutex_exit(&bufcache_lock);
bp = pool_cache_get(buf_cache, PR_NOWAIT);
if (bp != NULL) {
memset((char *)bp, 0, sizeof(*bp));
buf_init(bp);
SET(bp->b_cflags, BC_BUSY); /* mark buffer busy */
mutex_enter(&bufcache_lock);
#if defined(DIAGNOSTIC)
bp->b_freelistindex = -1;
#endif /* defined(DIAGNOSTIC) */
SDT_PROBE1(io, kernel, , getnewbuf__done, bp);
return bp;
}
mutex_enter(&bufcache_lock);
}
/* Buffer is no longer on free lists. */
SET(bp->b_cflags, BC_BUSY);
/* Wake anyone trying to lock the old identity. */
cv_broadcast(&bp->b_busy);
} else {
/*
* XXX: !from_bufq should be removed.
*/
if (!from_bufq || curlwp != uvm.pagedaemon_lwp) {
/* wait for a free buffer of any kind */
if ((slpflag & PCATCH) != 0)
(void)cv_timedwait_sig(&needbuffer_cv,
&bufcache_lock, slptimeo);
else
(void)cv_timedwait(&needbuffer_cv,
&bufcache_lock, slptimeo);
}
SDT_PROBE1(io, kernel, , getnewbuf__done, NULL);
return NULL;
}
#ifdef DIAGNOSTIC
if (bp->b_bufsize <= 0)
panic("buffer %p: on queue but empty", bp);
#endif
if (ISSET(bp->b_cflags, BC_VFLUSH)) {
/*
* This is a delayed write buffer being flushed to disk. Make
* sure it gets aged out of the queue when it's finished, and
* leave it off the LRU queue.
*/
CLR(bp->b_cflags, BC_VFLUSH);
SET(bp->b_cflags, BC_AGE);
goto start;
}
/*
* If buffer was a delayed write, start it and return NULL
* (since we might sleep while starting the write).
*/
if (ISSET(bp->b_oflags, BO_DELWRI)) {
/*
* This buffer has gone through the LRU, so make sure it gets
* reused ASAP.
*/
SET(bp->b_cflags, BC_AGE);
mutex_exit(&bufcache_lock);
bawrite(bp);
KASSERT(transmp != NULL);
fstrans_done(transmp);
mutex_enter(&bufcache_lock);
SDT_PROBE1(io, kernel, , getnewbuf__done, NULL);
return NULL;
}
/*
* Attempt to free an aged buffer off the queues.
* Called with queue lock held.
* Returns the amount of buffer memory freed.
*/
static int
buf_trim(void)
{
buf_t *bp;
long size;
KASSERT(mutex_owned(&bufcache_lock));
/* Instruct getnewbuf() to get buffers off the queues */
if ((bp = getnewbuf(PCATCH, 1, 1)) == NULL)
return 0;
KASSERT((bp->b_cflags & BC_WANTED) == 0);
size = bp->b_bufsize;
bufmem -= size;
if (size > 0) {
buf_mrelease(bp->b_data, size);
bp->b_bcount = bp->b_bufsize = 0;
}
/* brelse() will return the buffer to the global buffer pool */
brelsel(bp, 0);
return size;
}
int
buf_drain(int n)
{
int size = 0, sz;
KASSERT(mutex_owned(&bufcache_lock));
while (size < n && bufmem > bufmem_lowater) {
sz = buf_trim();
if (sz <= 0)
break;
size += sz;
}
return size;
}
/*
* Wait for operations on the buffer to complete.
* When they do, extract and return the I/O's error value.
*/
int
biowait(buf_t *bp)
{
/*
* Mark I/O complete on a buffer.
*
* If a callback has been requested, e.g. the pageout
* daemon, do so. Otherwise, awaken waiting processes.
*
* [ Leffler, et al., says on p.247:
* "This routine wakes up the blocked process, frees the buffer
* for an asynchronous write, or, for a request by the pagedaemon
* process, invokes a procedure specified in the buffer structure" ]
*
* In real life, the pagedaemon (or other system processes) wants
* to do async stuff too, and doesn't want the buffer brelse()'d.
* (for swap pager, that puts swap buffers on the free lists (!!!),
* for the vn device, that puts allocated buffers on the free lists!)
*/
void
biodone(buf_t *bp)
{
int s;
BIOHIST_FUNC(__func__);
KASSERT(!ISSET(bp->b_oflags, BO_DONE));
if (cpu_intr_p()) {
/* From interrupt mode: defer to a soft interrupt. */
s = splvm();
TAILQ_INSERT_TAIL(&curcpu()->ci_data.cpu_biodone, bp, b_actq);
BIOHIST_CALLARGS(biohist, "bp=%#jx, softint scheduled",
(uintptr_t)bp, 0, 0, 0);
softint_schedule(biodone_sih);
splx(s);
} else {
/* Process now - the buffer may be freed soon. */
biodone2(bp);
}
}
mutex_enter(bp->b_objlock);
/* Note that the transfer is done. */
if (ISSET(bp->b_oflags, BO_DONE))
panic("biodone2 already");
CLR(bp->b_flags, B_COWDONE);
SET(bp->b_oflags, BO_DONE);
BIO_SETPRIO(bp, BPRIO_DEFAULT);
/* Wake up waiting writers. */
if (!ISSET(bp->b_flags, B_READ))
vwakeup(bp);
/*
* at the moment, these are just "placeholders" to make the
* API for retrieving kern.buf data more extensible in the
* future.
*
* XXX kern.buf currently has "netbsd32" issues. hopefully
* these will be resolved at a later point.
*/
if (op != KERN_BUF_ALL || arg != KERN_BUF_ALL ||
elem_size < 1 || elem_count < 0)
return SET_ERROR(EINVAL);
if (oldp == NULL) {
/* count only, don't run through the buffer queues */
needed = pool_cache_nget(buf_cache) -
pool_cache_nput(buf_cache);
*oldlenp = (needed + KERN_BUFSLOP) * elem_size;
return 0;
}
error = 0;
needed = 0;
sysctl_unlock();
mutex_enter(&bufcache_lock);
for (i = 0; i < BQUEUES; i++) {
bq = &bufqueues[i];
TAILQ_FOREACH(bp, &bq->bq_queue, b_freelist) {
bq->bq_marker = bp;
if (len >= elem_size && elem_count > 0) {
sysctl_fillbuf(bp, &bs);
mutex_exit(&bufcache_lock);
error = copyout(&bs, dp, out_size);
mutex_enter(&bufcache_lock);
if (error)
break;
if (bq->bq_marker != bp) {
/*
* This sysctl node is only for
* statistics. Retry; if the
* queue keeps changing, then
* bail out.
*/
if (retries-- == 0) {
error = SET_ERROR(EAGAIN);
break;
}
mutex_exit(&bufcache_lock);
sysctl_relock();
goto retry;
}
dp += elem_size;
len -= elem_size;
}
needed += elem_size;
if (elem_count > 0 && elem_count != INT_MAX)
elem_count--;
}
if (error != 0)
break;
}
mutex_exit(&bufcache_lock);
sysctl_relock();
*oldlenp = needed;
return error;
}
static int
sysctl_bufvm_update(SYSCTLFN_ARGS)
{
int error, rv;
struct sysctlnode node;
unsigned int temp_bufcache;
unsigned long temp_water;
/* Take a copy of the supplied node and its data */
node = *rnode;
if (node.sysctl_data == &bufcache) {
node.sysctl_data = &temp_bufcache;
temp_bufcache = *(unsigned int *)rnode->sysctl_data;
} else {
node.sysctl_data = &temp_water;
temp_water = *(unsigned long *)rnode->sysctl_data;
}
/* Update the copy */
error = sysctl_lookup(SYSCTLFN_CALL(&node));
if (error || newp == NULL)
return error;
if (rnode->sysctl_data == &bufcache) {
if (temp_bufcache > 100)
return SET_ERROR(EINVAL);
bufcache = temp_bufcache;
buf_setwm();
} else if (rnode->sysctl_data == &bufmem_lowater) {
if (bufmem_hiwater - temp_water < 16)
return SET_ERROR(EINVAL);
bufmem_lowater = temp_water;
} else if (rnode->sysctl_data == &bufmem_hiwater) {
if (temp_water - bufmem_lowater < 16)
return SET_ERROR(EINVAL);
bufmem_hiwater = temp_water;
} else
return SET_ERROR(EINVAL);
/* Drain until below new high water mark */
sysctl_unlock();
mutex_enter(&bufcache_lock);
while (bufmem > bufmem_hiwater) {
rv = buf_drain((bufmem - bufmem_hiwater) / (2 * 1024));
if (rv <= 0)
break;
}
mutex_exit(&bufcache_lock);
sysctl_relock();
if (chain > 0) {
hs->hash_used++;
hs->hash_items += chain;
if (chain > hs->hash_maxchain)
hs->hash_maxchain = chain;
}
preempt_point();
}
return 0;
}
#ifdef DEBUG
/*
* Print out statistics on the current allocation of the buffer pool.
* Can be enabled to print out on every ``sync'' by setting "syncprt"
* in vfs_syscalls.c using sysctl.
*/
void
vfs_bufstats(void)
{
int i, j, count;
buf_t *bp;
struct bqueue *dp;
int counts[MAXBSIZE / MIN_PAGE_SIZE + 1];
static const char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE" };
error = bp->b_error;
if (bp->b_error == 0 &&
(bp->b_bcount < bp->b_bufsize || bp->b_resid > 0)) {
/*
* Not all got transferred, raise an error. We have no way to
* propagate these conditions to mbp.
*/
error = SET_ERROR(EIO);
}
/*
* nestiobuf_setup: setup a "nested" buffer.
*
* => 'mbp' is a "master" buffer which is being divided into sub pieces.
* => 'bp' should be a buffer allocated by getiobuf.
* => 'offset' is a byte offset in the master buffer.
* => 'size' is a size in bytes of this nested buffer.
*/
/*
* nestiobuf_done: propagate completion to the master buffer.
*
* => 'donebytes' specifies how many bytes in the 'mbp' is completed.
* => 'error' is an errno(2) that 'donebytes' has been completed with.
*/
void
nestiobuf_done(buf_t *mbp, int donebytes, int error)
{
if (donebytes == 0) {
return;
}
mutex_enter(mbp->b_objlock);
KASSERT(mbp->b_resid >= donebytes);
mbp->b_resid -= donebytes;
if (error)
mbp->b_error = error;
if (mbp->b_resid == 0) {
if (mbp->b_error)
mbp->b_resid = mbp->b_bcount;
mutex_exit(mbp->b_objlock);
biodone(mbp);
} else
mutex_exit(mbp->b_objlock);
}
if ((bp->b_cflags & BC_BUSY) != 0) {
if (curlwp == uvm.pagedaemon_lwp) {
error = SET_ERROR(EDEADLK);
goto out;
}
bp->b_cflags |= BC_WANTED;
if (interlock != NULL)
mutex_exit(interlock);
if (intr) {
error = cv_timedwait_sig(&bp->b_busy, &bufcache_lock,
timo);
} else {
error = cv_timedwait(&bp->b_busy, &bufcache_lock,
timo);
}
/*
* At this point the buffer may be gone: don't touch it
* again. The caller needs to find it again and retry.
*/
if (interlock != NULL)
mutex_enter(interlock);
if (error == 0)
error = SET_ERROR(EPASSTHROUGH);
} else {
bp->b_cflags |= BC_BUSY;
error = 0;
}
/*
* Nothing outside this file should really need to know about nbuf,
* but a few things still want to read it, so give them a way to do that.
*/
u_int
buf_nbuf(void)
{
