2 * Copyright (c) 2011-2014 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
6 * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 #include <sys/param.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/nlookup.h>
39 #include <sys/vnode.h>
40 #include <sys/mount.h>
41 #include <sys/fcntl.h>
44 #include <sys/vfsops.h>
45 #include <sys/sysctl.h>
46 #include <sys/socket.h>
47 #include <sys/objcache.h>
50 #include <sys/namei.h>
51 #include <sys/mountctl.h>
52 #include <sys/dirent.h>
55 #include <sys/mutex.h>
56 #include <sys/mutex2.h>
59 #include "hammer2_disk.h"
60 #include "hammer2_mount.h"
63 #include "hammer2_lz4.h"
65 #include "zlib/hammer2_zlib.h"
67 #define REPORT_REFS_ERRORS 1 /* XXX remove me */
69 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
71 struct hammer2_sync_info {
72 hammer2_trans_t trans;
77 TAILQ_HEAD(hammer2_mntlist, hammer2_mount);
78 TAILQ_HEAD(hammer2_pfslist, hammer2_pfsmount);
79 static struct hammer2_mntlist hammer2_mntlist;
80 static struct hammer2_pfslist hammer2_pfslist;
81 static struct lock hammer2_mntlk;
84 int hammer2_cluster_enable = 1;
85 int hammer2_hardlink_enable = 1;
86 int hammer2_flush_pipe = 100;
87 int hammer2_synchronous_flush = 1;
88 int hammer2_dio_count;
89 long hammer2_limit_dirty_chains;
90 long hammer2_iod_file_read;
91 long hammer2_iod_meta_read;
92 long hammer2_iod_indr_read;
93 long hammer2_iod_fmap_read;
94 long hammer2_iod_volu_read;
95 long hammer2_iod_file_write;
96 long hammer2_iod_meta_write;
97 long hammer2_iod_indr_write;
98 long hammer2_iod_fmap_write;
99 long hammer2_iod_volu_write;
100 long hammer2_ioa_file_read;
101 long hammer2_ioa_meta_read;
102 long hammer2_ioa_indr_read;
103 long hammer2_ioa_fmap_read;
104 long hammer2_ioa_volu_read;
105 long hammer2_ioa_fmap_write;
106 long hammer2_ioa_file_write;
107 long hammer2_ioa_meta_write;
108 long hammer2_ioa_indr_write;
109 long hammer2_ioa_volu_write;
111 MALLOC_DECLARE(C_BUFFER);
112 MALLOC_DEFINE(C_BUFFER, "compbuffer", "Buffer used for compression.");
114 MALLOC_DECLARE(D_BUFFER);
115 MALLOC_DEFINE(D_BUFFER, "decompbuffer", "Buffer used for decompression.");
117 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
120 &hammer2_debug, 0, "");
121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_enable, CTLFLAG_RW,
122 &hammer2_cluster_enable, 0, "");
123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, hardlink_enable, CTLFLAG_RW,
124 &hammer2_hardlink_enable, 0, "");
125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
126 &hammer2_flush_pipe, 0, "");
127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
128 &hammer2_synchronous_flush, 0, "");
129 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
130 &hammer2_limit_dirty_chains, 0, "");
131 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
132 &hammer2_dio_count, 0, "");
134 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
135 &hammer2_iod_file_read, 0, "");
136 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
137 &hammer2_iod_meta_read, 0, "");
138 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
139 &hammer2_iod_indr_read, 0, "");
140 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
141 &hammer2_iod_fmap_read, 0, "");
142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
143 &hammer2_iod_volu_read, 0, "");
145 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
146 &hammer2_iod_file_write, 0, "");
147 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
148 &hammer2_iod_meta_write, 0, "");
149 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
150 &hammer2_iod_indr_write, 0, "");
151 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
152 &hammer2_iod_fmap_write, 0, "");
153 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
154 &hammer2_iod_volu_write, 0, "");
156 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_read, CTLFLAG_RW,
157 &hammer2_ioa_file_read, 0, "");
158 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_read, CTLFLAG_RW,
159 &hammer2_ioa_meta_read, 0, "");
160 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_read, CTLFLAG_RW,
161 &hammer2_ioa_indr_read, 0, "");
162 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_read, CTLFLAG_RW,
163 &hammer2_ioa_fmap_read, 0, "");
164 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_read, CTLFLAG_RW,
165 &hammer2_ioa_volu_read, 0, "");
167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_file_write, CTLFLAG_RW,
168 &hammer2_ioa_file_write, 0, "");
169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_meta_write, CTLFLAG_RW,
170 &hammer2_ioa_meta_write, 0, "");
171 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_indr_write, CTLFLAG_RW,
172 &hammer2_ioa_indr_write, 0, "");
173 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_fmap_write, CTLFLAG_RW,
174 &hammer2_ioa_fmap_write, 0, "");
175 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, ioa_volu_write, CTLFLAG_RW,
176 &hammer2_ioa_volu_write, 0, "");
178 static int hammer2_vfs_init(struct vfsconf *conf);
179 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
180 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
182 static int hammer2_remount(hammer2_mount_t *, struct mount *, char *,
183 struct vnode *, struct ucred *);
184 static int hammer2_recovery(hammer2_mount_t *hmp);
185 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
186 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
187 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
189 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
191 static int hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
192 ino_t ino, struct vnode **vpp);
193 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
194 struct fid *fhp, struct vnode **vpp);
195 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
196 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
197 int *exflagsp, struct ucred **credanonp);
199 static int hammer2_install_volume_header(hammer2_mount_t *hmp);
200 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
202 static void hammer2_write_thread(void *arg);
204 static void hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp);
205 static void hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp);
208 * Functions for compression in threads,
209 * from hammer2_vnops.c
211 static void hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
213 hammer2_inode_data_t *ipdata,
214 hammer2_cluster_t *cparent,
215 hammer2_key_t lbase, int ioflag, int pblksize,
217 static void hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
219 const hammer2_inode_data_t *ipdata,
220 hammer2_cluster_t *cparent,
221 hammer2_key_t lbase, int ioflag,
222 int pblksize, int *errorp, int comp_algo);
223 static void hammer2_zero_check_and_write(struct buf *bp,
224 hammer2_trans_t *trans, hammer2_inode_t *ip,
225 const hammer2_inode_data_t *ipdata,
226 hammer2_cluster_t *cparent,
228 int ioflag, int pblksize, int *errorp);
229 static int test_block_zeros(const char *buf, size_t bytes);
230 static void zero_write(struct buf *bp, hammer2_trans_t *trans,
232 const hammer2_inode_data_t *ipdata,
233 hammer2_cluster_t *cparent,
236 static void hammer2_write_bp(hammer2_cluster_t *cluster, struct buf *bp,
237 int ioflag, int pblksize, int *errorp);
239 static int hammer2_rcvdmsg(kdmsg_msg_t *msg);
240 static void hammer2_autodmsg(kdmsg_msg_t *msg);
241 static int hammer2_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg);
245 * HAMMER2 vfs operations.
247 static struct vfsops hammer2_vfsops = {
248 .vfs_init = hammer2_vfs_init,
249 .vfs_uninit = hammer2_vfs_uninit,
250 .vfs_sync = hammer2_vfs_sync,
251 .vfs_mount = hammer2_vfs_mount,
252 .vfs_unmount = hammer2_vfs_unmount,
253 .vfs_root = hammer2_vfs_root,
254 .vfs_statfs = hammer2_vfs_statfs,
255 .vfs_statvfs = hammer2_vfs_statvfs,
256 .vfs_vget = hammer2_vfs_vget,
257 .vfs_vptofh = hammer2_vfs_vptofh,
258 .vfs_fhtovp = hammer2_vfs_fhtovp,
259 .vfs_checkexp = hammer2_vfs_checkexp
262 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
264 VFS_SET(hammer2_vfsops, hammer2, 0);
265 MODULE_VERSION(hammer2, 1);
269 hammer2_vfs_init(struct vfsconf *conf)
271 static struct objcache_malloc_args margs_read;
272 static struct objcache_malloc_args margs_write;
278 if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
280 if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
282 if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
286 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
288 margs_read.objsize = 65536;
289 margs_read.mtype = D_BUFFER;
291 margs_write.objsize = 32768;
292 margs_write.mtype = C_BUFFER;
294 cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
295 0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
296 objcache_malloc_free, &margs_read);
297 cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
298 0, 1, NULL, NULL, NULL, objcache_malloc_alloc,
299 objcache_malloc_free, &margs_write);
301 lockinit(&hammer2_mntlk, "mntlk", 0, 0);
302 TAILQ_INIT(&hammer2_mntlist);
303 TAILQ_INIT(&hammer2_pfslist);
305 hammer2_limit_dirty_chains = desiredvnodes / 10;
307 hammer2_trans_manage_init();
314 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
316 objcache_destroy(cache_buffer_read);
317 objcache_destroy(cache_buffer_write);
322 * Core PFS allocator. Used to allocate the pmp structure for PFS cluster
323 * mounts and the spmp structure for media (hmp) structures.
325 static hammer2_pfsmount_t *
326 hammer2_pfsalloc(const hammer2_inode_data_t *ipdata, hammer2_tid_t alloc_tid)
328 hammer2_pfsmount_t *pmp;
330 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
331 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
332 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
333 lockinit(&pmp->lock, "pfslk", 0, 0);
334 spin_init(&pmp->inum_spin);
335 RB_INIT(&pmp->inum_tree);
336 TAILQ_INIT(&pmp->unlinkq);
337 spin_init(&pmp->list_spin);
339 pmp->alloc_tid = alloc_tid + 1; /* our first media transaction id */
340 pmp->flush_tid = pmp->alloc_tid;
342 pmp->inode_tid = ipdata->pfs_inum + 1;
343 pmp->pfs_clid = ipdata->pfs_clid;
345 mtx_init(&pmp->wthread_mtx);
346 bioq_init(&pmp->wthread_bioq);
352 * Mount or remount HAMMER2 fileystem from physical media
355 * mp mount point structure
361 * mp mount point structure
362 * path path to mount point
363 * data pointer to argument structure in user space
364 * volume volume path (device@LABEL form)
365 * hflags user mount flags
366 * cred user credentials
373 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
376 struct hammer2_mount_info info;
377 hammer2_pfsmount_t *pmp;
378 hammer2_pfsmount_t *spmp;
379 hammer2_mount_t *hmp;
380 hammer2_key_t key_next;
381 hammer2_key_t key_dummy;
384 struct nlookupdata nd;
385 hammer2_chain_t *parent;
386 hammer2_chain_t *rchain;
387 hammer2_cluster_t *cluster;
388 hammer2_cluster_t *cparent;
389 const hammer2_inode_data_t *ipdata;
390 hammer2_blockref_t bref;
392 char devstr[MNAMELEN];
410 kprintf("hammer2_mount\n");
416 bzero(&info, sizeof(info));
417 info.cluster_fd = -1;
421 * Non-root mount or updating a mount
423 error = copyin(data, &info, sizeof(info));
427 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
431 /* Extract device and label */
433 label = strchr(devstr, '@');
435 ((label + 1) - dev) > done) {
443 if (mp->mnt_flag & MNT_UPDATE) {
445 /* HAMMER2 implements NFS export via mountctl */
447 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
448 hmp = pmp->iroot->cluster.array[i]->hmp;
450 error = hammer2_remount(hmp, mp, path,
455 /*hammer2_inode_install_hidden(pmp);*/
464 * Lookup name and verify it refers to a block device.
466 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
468 error = nlookup(&nd);
470 error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
474 if (vn_isdisk(devvp, &error))
475 error = vfs_mountedon(devvp);
479 * Determine if the device has already been mounted. After this
480 * check hmp will be non-NULL if we are doing the second or more
481 * hammer2 mounts from the same device.
483 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
484 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
485 if (hmp->devvp == devvp)
490 * Open the device if this isn't a secondary mount and construct
491 * the H2 device mount (hmp).
494 hammer2_chain_t *schain;
497 if (error == 0 && vcount(devvp) > 0)
501 * Now open the device
504 ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
505 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
506 error = vinvalbuf(devvp, V_SAVE, 0, 0);
508 error = VOP_OPEN(devvp,
509 ronly ? FREAD : FREAD | FWRITE,
514 if (error && devvp) {
519 lockmgr(&hammer2_mntlk, LK_RELEASE);
522 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
525 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
526 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
527 RB_INIT(&hmp->iotree);
528 spin_init(&hmp->io_spin);
529 spin_init(&hmp->list_spin);
530 TAILQ_INIT(&hmp->flushq);
532 lockinit(&hmp->vollk, "h2vol", 0, 0);
535 * vchain setup. vchain.data is embedded.
536 * vchain.refs is initialized and will never drop to 0.
538 * NOTE! voldata is not yet loaded.
540 hmp->vchain.hmp = hmp;
541 hmp->vchain.refs = 1;
542 hmp->vchain.data = (void *)&hmp->voldata;
543 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
544 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
545 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
547 hammer2_chain_core_alloc(NULL, &hmp->vchain);
548 /* hmp->vchain.u.xxx is left NULL */
551 * fchain setup. fchain.data is embedded.
552 * fchain.refs is initialized and will never drop to 0.
554 * The data is not used but needs to be initialized to
555 * pass assertion muster. We use this chain primarily
556 * as a placeholder for the freemap's top-level RBTREE
557 * so it does not interfere with the volume's topology
560 hmp->fchain.hmp = hmp;
561 hmp->fchain.refs = 1;
562 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
563 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
564 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
565 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
566 hmp->fchain.bref.methods =
567 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
568 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
570 hammer2_chain_core_alloc(NULL, &hmp->fchain);
571 /* hmp->fchain.u.xxx is left NULL */
574 * Install the volume header and initialize fields from
577 error = hammer2_install_volume_header(hmp);
580 hammer2_vfs_unmount_hmp1(mp, hmp);
581 hammer2_vfs_unmount_hmp2(mp, hmp);
582 lockmgr(&hammer2_mntlk, LK_RELEASE);
583 hammer2_vfs_unmount(mp, MNT_FORCE);
588 * Really important to get these right or flush will get
591 hmp->spmp = hammer2_pfsalloc(NULL, hmp->voldata.mirror_tid);
592 kprintf("alloc spmp %p tid %016jx\n",
593 hmp->spmp, hmp->voldata.mirror_tid);
598 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
599 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
600 hmp->vchain.pmp = spmp;
601 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
602 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
603 hmp->fchain.pmp = spmp;
606 * First locate the super-root inode, which is key 0
607 * relative to the volume header's blockset.
609 * Then locate the root inode by scanning the directory keyspace
610 * represented by the label.
612 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
613 schain = hammer2_chain_lookup(&parent, &key_dummy,
614 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
615 &cache_index, 0, &ddflag);
616 hammer2_chain_lookup_done(parent);
617 if (schain == NULL) {
618 kprintf("hammer2_mount: invalid super-root\n");
620 hammer2_vfs_unmount_hmp1(mp, hmp);
621 hammer2_vfs_unmount_hmp2(mp, hmp);
622 lockmgr(&hammer2_mntlk, LK_RELEASE);
623 hammer2_vfs_unmount(mp, MNT_FORCE);
628 * Sanity-check schain's pmp, finish initializing spmp.
630 KKASSERT(schain->pmp == spmp);
631 spmp->pfs_clid = schain->data->ipdata.pfs_clid;
634 * NOTE: inode_get sucks up schain's lock.
636 cluster = hammer2_cluster_from_chain(schain);
637 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster);
638 spmp->spmp_hmp = hmp;
639 hammer2_inode_ref(spmp->iroot);
640 hammer2_inode_unlock_ex(spmp->iroot, cluster);
642 /* leave spmp->iroot with one ref */
644 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
645 error = hammer2_recovery(hmp);
646 /* XXX do something with error */
651 * XXX RDONLY stuff is totally broken FIXME XXX
654 * Automatic handling of received LNK_SPAN
655 * Automatic handling of received LNK_CIRC
656 * No automatic LNK_SPAN generation - we do this ourselves
657 * No automatic LNK_CIRC generation - we do this ourselves
659 kdmsg_iocom_init(&hmp->iocom, hmp,
660 KDMSG_IOCOMF_AUTOCONN |
661 KDMSG_IOCOMF_AUTORXSPAN,
662 hmp->mchain, hammer2_rcvdmsg);
665 * Ref the cluster management messaging descriptor. The mount
666 * program deals with the other end of the communications pipe.
668 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
670 hammer2_cluster_reconnect(hmp, fp);
672 kprintf("hammer2_mount: bad cluster_fd!\n");
680 * Lookup mount point under the media-localized super-root.
682 * cluster->pmp will incorrectly point to spmp and must be fixed
685 cparent = hammer2_inode_lock_ex(spmp->iroot);
686 lhc = hammer2_dirhash(label, strlen(label));
687 cluster = hammer2_cluster_lookup(cparent, &key_next,
688 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
691 if (hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE &&
693 hammer2_cluster_data(cluster)->ipdata.filename) == 0) {
696 cluster = hammer2_cluster_next(cparent, cluster, &key_next,
698 lhc + HAMMER2_DIRHASH_LOMASK, 0);
700 hammer2_inode_unlock_ex(spmp->iroot, cparent);
702 if (cluster == NULL) {
703 kprintf("hammer2_mount: PFS label not found\n");
704 hammer2_vfs_unmount_hmp1(mp, hmp);
705 hammer2_vfs_unmount_hmp2(mp, hmp);
706 lockmgr(&hammer2_mntlk, LK_RELEASE);
707 hammer2_vfs_unmount(mp, MNT_FORCE);
711 for (i = 0; i < cluster->nchains; ++i) {
712 rchain = cluster->array[i];
713 if (rchain->flags & HAMMER2_CHAIN_MOUNTED) {
714 kprintf("hammer2_mount: PFS label already mounted!\n");
715 hammer2_cluster_unlock(cluster);
716 hammer2_vfs_unmount_hmp1(mp, hmp);
717 hammer2_vfs_unmount_hmp2(mp, hmp);
718 lockmgr(&hammer2_mntlk, LK_RELEASE);
719 hammer2_vfs_unmount(mp, MNT_FORCE);
722 KKASSERT(rchain->pmp == NULL);
724 if (rchain->flags & HAMMER2_CHAIN_RECYCLE) {
725 kprintf("hammer2_mount: PFS label is recycling\n");
726 hammer2_cluster_unlock(cluster);
727 hammer2_vfs_unmount_hmp1(mp, hmp);
728 hammer2_vfs_unmount_hmp2(mp, hmp);
729 lockmgr(&hammer2_mntlk, LK_RELEASE);
730 hammer2_vfs_unmount(mp, MNT_FORCE);
737 * Check to see if the cluster id is already mounted at the mount
738 * point. If it is, add us to the cluster.
740 ipdata = &hammer2_cluster_data(cluster)->ipdata;
741 hammer2_cluster_bref(cluster, &bref);
742 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
743 if (pmp->spmp_hmp == NULL &&
744 bcmp(&pmp->pfs_clid, &ipdata->pfs_clid,
745 sizeof(pmp->pfs_clid)) == 0) {
754 hammer2_inode_ref(pmp->iroot);
755 ccms_thread_lock(&pmp->iroot->topo_cst, CCMS_STATE_EXCLUSIVE);
757 if (pmp->iroot->cluster.nchains + cluster->nchains >
758 HAMMER2_MAXCLUSTER) {
759 kprintf("hammer2_mount: cluster full!\n");
761 ccms_thread_unlock(&pmp->iroot->topo_cst);
762 hammer2_inode_drop(pmp->iroot);
764 hammer2_cluster_unlock(cluster);
765 hammer2_vfs_unmount_hmp1(mp, hmp);
766 hammer2_vfs_unmount_hmp2(mp, hmp);
767 lockmgr(&hammer2_mntlk, LK_RELEASE);
768 hammer2_vfs_unmount(mp, MNT_FORCE);
771 kprintf("hammer2_vfs_mount: Adding pfs to existing cluster\n");
772 j = pmp->iroot->cluster.nchains;
773 for (i = 0; i < cluster->nchains; ++i) {
774 rchain = cluster->array[i];
775 KKASSERT(rchain->pmp == NULL);
777 hammer2_chain_ref(cluster->array[i]);
778 pmp->iroot->cluster.array[j] = cluster->array[i];
781 pmp->iroot->cluster.nchains = j;
782 ccms_thread_unlock(&pmp->iroot->topo_cst);
783 hammer2_inode_drop(pmp->iroot);
784 hammer2_cluster_unlock(cluster);
785 lockmgr(&hammer2_mntlk, LK_RELEASE);
788 hammer2_inode_install_hidden(pmp);
794 * Block device opened successfully, finish initializing the
797 * From this point on we have to call hammer2_unmount() on failure.
799 pmp = hammer2_pfsalloc(ipdata, bref.mirror_tid);
800 kprintf("PMP mirror_tid is %016jx\n", bref.mirror_tid);
801 for (i = 0; i < cluster->nchains; ++i) {
802 rchain = cluster->array[i];
803 KKASSERT(rchain->pmp == NULL);
805 atomic_set_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
809 ccms_domain_init(&pmp->ccms_dom);
810 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
811 lockmgr(&hammer2_mntlk, LK_RELEASE);
813 kprintf("hammer2_mount hmp=%p pmp=%p pmpcnt=%d\n",
814 hmp, pmp, hmp->pmp_count);
816 mp->mnt_flag = MNT_LOCAL;
817 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */
818 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */
821 * required mount structure initializations
823 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
824 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
826 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
827 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
832 mp->mnt_iosize_max = MAXPHYS;
833 mp->mnt_data = (qaddr_t)pmp;
837 * After this point hammer2_vfs_unmount() has visibility on hmp
838 * and manual hmp1/hmp2 calls are not needed on fatal errors.
840 pmp->iroot = hammer2_inode_get(pmp, NULL, cluster);
841 hammer2_inode_ref(pmp->iroot); /* ref for pmp->iroot */
842 hammer2_inode_unlock_ex(pmp->iroot, cluster);
845 * The logical file buffer bio write thread handles things
846 * like physical block assignment and compression.
848 * (only applicable to pfs mounts, not applicable to spmp)
850 pmp->wthread_destroy = 0;
851 lwkt_create(hammer2_write_thread, pmp,
852 &pmp->wthread_td, NULL, 0, -1, "hwrite-%s", label);
855 * With the cluster operational install ihidden.
856 * (only applicable to pfs mounts, not applicable to spmp)
858 hammer2_inode_install_hidden(pmp);
864 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
865 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
866 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
868 copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
869 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
870 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
871 copyinstr(path, mp->mnt_stat.f_mntonname,
872 sizeof(mp->mnt_stat.f_mntonname) - 1,
876 * Initial statfs to prime mnt_stat.
878 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
884 * Handle bioq for strategy write
888 hammer2_write_thread(void *arg)
890 hammer2_pfsmount_t *pmp;
893 hammer2_trans_t trans;
896 hammer2_cluster_t *cparent;
897 hammer2_inode_data_t *wipdata;
905 mtx_lock(&pmp->wthread_mtx);
906 while (pmp->wthread_destroy == 0) {
907 if (bioq_first(&pmp->wthread_bioq) == NULL) {
908 mtxsleep(&pmp->wthread_bioq, &pmp->wthread_mtx,
913 hammer2_trans_init(&trans, pmp, HAMMER2_TRANS_BUFCACHE);
915 while ((bio = bioq_takefirst(&pmp->wthread_bioq)) != NULL) {
917 * dummy bio for synchronization. The transaction
918 * must be reinitialized.
920 if (bio->bio_buf == NULL) {
921 bio->bio_flags |= BIO_DONE;
923 hammer2_trans_done(&trans);
924 hammer2_trans_init(&trans, pmp,
925 HAMMER2_TRANS_BUFCACHE);
930 * else normal bio processing
932 mtx_unlock(&pmp->wthread_mtx);
934 hammer2_lwinprog_drop(pmp);
942 * Inode is modified, flush size and mtime changes
943 * to ensure that the file size remains consistent
944 * with the buffers being flushed.
946 * NOTE: The inode_fsync() call only flushes the
947 * inode's meta-data state, it doesn't try
948 * to flush underlying buffers or chains.
950 cparent = hammer2_inode_lock_ex(ip);
951 if (ip->flags & (HAMMER2_INODE_RESIZED |
952 HAMMER2_INODE_MTIME)) {
953 hammer2_inode_fsync(&trans, ip, cparent);
955 wipdata = hammer2_cluster_modify_ip(&trans, ip,
957 lblksize = hammer2_calc_logical(ip, bio->bio_offset,
959 pblksize = hammer2_calc_physical(ip, wipdata, lbase);
960 hammer2_write_file_core(bp, &trans, ip, wipdata,
964 hammer2_cluster_modsync(cparent);
965 hammer2_inode_unlock_ex(ip, cparent);
967 kprintf("hammer2: error in buffer write\n");
968 bp->b_flags |= B_ERROR;
972 mtx_lock(&pmp->wthread_mtx);
974 hammer2_trans_done(&trans);
976 pmp->wthread_destroy = -1;
977 wakeup(&pmp->wthread_destroy);
979 mtx_unlock(&pmp->wthread_mtx);
983 hammer2_bioq_sync(hammer2_pfsmount_t *pmp)
987 bzero(&sync_bio, sizeof(sync_bio)); /* dummy with no bio_buf */
988 mtx_lock(&pmp->wthread_mtx);
989 if (pmp->wthread_destroy == 0 &&
990 TAILQ_FIRST(&pmp->wthread_bioq.queue)) {
991 bioq_insert_tail(&pmp->wthread_bioq, &sync_bio);
992 while ((sync_bio.bio_flags & BIO_DONE) == 0)
993 mtxsleep(&sync_bio, &pmp->wthread_mtx, 0, "h2bioq", 0);
995 mtx_unlock(&pmp->wthread_mtx);
999 * Return a chain suitable for I/O, creating the chain if necessary
1000 * and assigning its physical block.
1004 hammer2_assign_physical(hammer2_trans_t *trans,
1005 hammer2_inode_t *ip, hammer2_cluster_t *cparent,
1006 hammer2_key_t lbase, int pblksize, int *errorp)
1008 hammer2_cluster_t *cluster;
1009 hammer2_cluster_t *dparent;
1010 hammer2_key_t key_dummy;
1011 int pradix = hammer2_getradix(pblksize);
1015 * Locate the chain associated with lbase, return a locked chain.
1016 * However, do not instantiate any data reference (which utilizes a
1017 * device buffer) because we will be using direct IO via the
1018 * logical buffer cache buffer.
1021 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
1023 dparent = hammer2_cluster_lookup_init(cparent, 0);
1024 cluster = hammer2_cluster_lookup(dparent, &key_dummy,
1026 HAMMER2_LOOKUP_NODATA, &ddflag);
1028 if (cluster == NULL) {
1030 * We found a hole, create a new chain entry.
1032 * NOTE: DATA chains are created without device backing
1033 * store (nor do we want any).
1035 *errorp = hammer2_cluster_create(trans, dparent, &cluster,
1036 lbase, HAMMER2_PBUFRADIX,
1037 HAMMER2_BREF_TYPE_DATA,
1039 if (cluster == NULL) {
1040 hammer2_cluster_lookup_done(dparent);
1041 panic("hammer2_cluster_create: par=%p error=%d\n",
1042 dparent->focus, *errorp);
1045 /*ip->delta_dcount += pblksize;*/
1047 switch (hammer2_cluster_type(cluster)) {
1048 case HAMMER2_BREF_TYPE_INODE:
1050 * The data is embedded in the inode. The
1051 * caller is responsible for marking the inode
1052 * modified and copying the data to the embedded
1056 case HAMMER2_BREF_TYPE_DATA:
1057 if (hammer2_cluster_need_resize(cluster, pblksize)) {
1058 hammer2_cluster_resize(trans, ip,
1061 HAMMER2_MODIFY_OPTDATA);
1065 * DATA buffers must be marked modified whether the
1066 * data is in a logical buffer or not. We also have
1067 * to make this call to fixup the chain data pointers
1068 * after resizing in case this is an encrypted or
1069 * compressed buffer.
1071 hammer2_cluster_modify(trans, cluster,
1072 HAMMER2_MODIFY_OPTDATA);
1075 panic("hammer2_assign_physical: bad type");
1082 * Cleanup. If cluster wound up being the inode itself, i.e.
1083 * the DIRECTDATA case for offset 0, then we need to update cparent.
1084 * The caller expects cparent to not become stale.
1086 hammer2_cluster_lookup_done(dparent);
1087 /* dparent = NULL; safety */
1088 if (cluster && ddflag)
1089 hammer2_cluster_replace_locked(cparent, cluster);
1094 * bio queued from hammer2_vnops.c.
1096 * The core write function which determines which path to take
1097 * depending on compression settings. We also have to locate the
1098 * related clusters so we can calculate and set the check data for
1103 hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
1104 hammer2_inode_t *ip, hammer2_inode_data_t *ipdata,
1105 hammer2_cluster_t *cparent,
1106 hammer2_key_t lbase, int ioflag, int pblksize,
1109 hammer2_cluster_t *cluster;
1111 switch(HAMMER2_DEC_COMP(ipdata->comp_algo)) {
1112 case HAMMER2_COMP_NONE:
1114 * We have to assign physical storage to the buffer
1115 * we intend to dirty or write now to avoid deadlocks
1116 * in the strategy code later.
1118 * This can return NOOFFSET for inode-embedded data.
1119 * The strategy code will take care of it in that case.
1121 cluster = hammer2_assign_physical(trans, ip, cparent,
1124 hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp);
1126 hammer2_cluster_unlock(cluster);
1128 case HAMMER2_COMP_AUTOZERO:
1130 * Check for zero-fill only
1132 hammer2_zero_check_and_write(bp, trans, ip,
1133 ipdata, cparent, lbase,
1134 ioflag, pblksize, errorp);
1136 case HAMMER2_COMP_LZ4:
1137 case HAMMER2_COMP_ZLIB:
1140 * Check for zero-fill and attempt compression.
1142 hammer2_compress_and_write(bp, trans, ip,
1152 * Generic function that will perform the compression in compression
1153 * write path. The compression algorithm is determined by the settings
1154 * obtained from inode.
1158 hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
1159 hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1160 hammer2_cluster_t *cparent,
1161 hammer2_key_t lbase, int ioflag, int pblksize,
1162 int *errorp, int comp_algo)
1164 hammer2_cluster_t *cluster;
1165 hammer2_chain_t *chain;
1167 int comp_block_size;
1171 if (test_block_zeros(bp->b_data, pblksize)) {
1172 zero_write(bp, trans, ip, ipdata, cparent, lbase, errorp);
1179 KKASSERT(pblksize / 2 <= 32768);
1181 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
1182 z_stream strm_compress;
1186 switch(HAMMER2_DEC_COMP(comp_algo)) {
1187 case HAMMER2_COMP_LZ4:
1188 comp_buffer = objcache_get(cache_buffer_write,
1190 comp_size = LZ4_compress_limitedOutput(
1192 &comp_buffer[sizeof(int)],
1194 pblksize / 2 - sizeof(int));
1196 * We need to prefix with the size, LZ4
1197 * doesn't do it for us. Add the related
1200 *(int *)comp_buffer = comp_size;
1202 comp_size += sizeof(int);
1204 case HAMMER2_COMP_ZLIB:
1205 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
1206 if (comp_level == 0)
1207 comp_level = 6; /* default zlib compression */
1208 else if (comp_level < 6)
1210 else if (comp_level > 9)
1212 ret = deflateInit(&strm_compress, comp_level);
1214 kprintf("HAMMER2 ZLIB: fatal error "
1215 "on deflateInit.\n");
1218 comp_buffer = objcache_get(cache_buffer_write,
1220 strm_compress.next_in = bp->b_data;
1221 strm_compress.avail_in = pblksize;
1222 strm_compress.next_out = comp_buffer;
1223 strm_compress.avail_out = pblksize / 2;
1224 ret = deflate(&strm_compress, Z_FINISH);
1225 if (ret == Z_STREAM_END) {
1226 comp_size = pblksize / 2 -
1227 strm_compress.avail_out;
1231 ret = deflateEnd(&strm_compress);
1234 kprintf("Error: Unknown compression method.\n");
1235 kprintf("Comp_method = %d.\n", comp_algo);
1240 if (comp_size == 0) {
1242 * compression failed or turned off
1244 comp_block_size = pblksize; /* safety */
1245 if (++ip->comp_heuristic > 128)
1246 ip->comp_heuristic = 8;
1249 * compression succeeded
1251 ip->comp_heuristic = 0;
1252 if (comp_size <= 1024) {
1253 comp_block_size = 1024;
1254 } else if (comp_size <= 2048) {
1255 comp_block_size = 2048;
1256 } else if (comp_size <= 4096) {
1257 comp_block_size = 4096;
1258 } else if (comp_size <= 8192) {
1259 comp_block_size = 8192;
1260 } else if (comp_size <= 16384) {
1261 comp_block_size = 16384;
1262 } else if (comp_size <= 32768) {
1263 comp_block_size = 32768;
1265 panic("hammer2: WRITE PATH: "
1266 "Weird comp_size value.");
1268 comp_block_size = pblksize;
1272 cluster = hammer2_assign_physical(trans, ip, cparent,
1273 lbase, comp_block_size,
1275 ipdata = &hammer2_cluster_data(cparent)->ipdata;
1278 kprintf("WRITE PATH: An error occurred while "
1279 "assigning physical space.\n");
1280 KKASSERT(cluster == NULL);
1284 for (i = 0; i < cluster->nchains; ++i) {
1289 chain = cluster->array[i];
1290 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1292 switch(chain->bref.type) {
1293 case HAMMER2_BREF_TYPE_INODE:
1294 KKASSERT(chain->data->ipdata.op_flags &
1295 HAMMER2_OPFLAG_DIRECTDATA);
1296 KKASSERT(bp->b_loffset == 0);
1297 bcopy(bp->b_data, chain->data->ipdata.u.data,
1298 HAMMER2_EMBEDDED_BYTES);
1300 case HAMMER2_BREF_TYPE_DATA:
1301 temp_check = HAMMER2_DEC_CHECK(chain->bref.methods);
1304 * Optimize out the read-before-write
1307 *errorp = hammer2_io_newnz(chain->hmp,
1308 chain->bref.data_off,
1312 hammer2_io_brelse(&dio);
1313 kprintf("hammer2: WRITE PATH: "
1314 "dbp bread error\n");
1317 bdata = hammer2_io_data(dio, chain->bref.data_off);
1320 * When loading the block make sure we don't
1321 * leave garbage after the compressed data.
1324 chain->bref.methods =
1325 HAMMER2_ENC_COMP(comp_algo) +
1326 HAMMER2_ENC_CHECK(temp_check);
1327 bcopy(comp_buffer, bdata, comp_size);
1328 if (comp_size != comp_block_size) {
1329 bzero(bdata + comp_size,
1330 comp_block_size - comp_size);
1333 chain->bref.methods =
1335 HAMMER2_COMP_NONE) +
1336 HAMMER2_ENC_CHECK(temp_check);
1337 bcopy(bp->b_data, bdata, pblksize);
1341 * The flush code doesn't calculate check codes for
1342 * file data (doing so can result in excessive I/O),
1345 hammer2_chain_setcheck(chain, bdata);
1348 * Device buffer is now valid, chain is no longer in
1349 * the initial state.
1351 * (No blockref table worries with file data)
1353 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1355 /* Now write the related bdp. */
1356 if (ioflag & IO_SYNC) {
1358 * Synchronous I/O requested.
1360 hammer2_io_bwrite(&dio);
1362 } else if ((ioflag & IO_DIRECT) &&
1363 loff + n == pblksize) {
1364 hammer2_io_bdwrite(&dio);
1366 } else if (ioflag & IO_ASYNC) {
1367 hammer2_io_bawrite(&dio);
1369 hammer2_io_bdwrite(&dio);
1373 panic("hammer2_write_bp: bad chain type %d\n",
1381 hammer2_cluster_unlock(cluster);
1383 objcache_put(cache_buffer_write, comp_buffer);
1387 * Function that performs zero-checking and writing without compression,
1388 * it corresponds to default zero-checking path.
1392 hammer2_zero_check_and_write(struct buf *bp, hammer2_trans_t *trans,
1393 hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1394 hammer2_cluster_t *cparent,
1395 hammer2_key_t lbase, int ioflag, int pblksize, int *errorp)
1397 hammer2_cluster_t *cluster;
1399 if (test_block_zeros(bp->b_data, pblksize)) {
1400 zero_write(bp, trans, ip, ipdata, cparent, lbase, errorp);
1402 cluster = hammer2_assign_physical(trans, ip, cparent,
1403 lbase, pblksize, errorp);
1404 hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp);
1406 hammer2_cluster_unlock(cluster);
1411 * A function to test whether a block of data contains only zeros,
1412 * returns TRUE (non-zero) if the block is all zeros.
1416 test_block_zeros(const char *buf, size_t bytes)
1420 for (i = 0; i < bytes; i += sizeof(long)) {
1421 if (*(const long *)(buf + i) != 0)
1428 * Function to "write" a block that contains only zeros.
1432 zero_write(struct buf *bp, hammer2_trans_t *trans,
1433 hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1434 hammer2_cluster_t *cparent,
1435 hammer2_key_t lbase, int *errorp __unused)
1437 hammer2_cluster_t *cluster;
1438 hammer2_media_data_t *data;
1439 hammer2_key_t key_dummy;
1442 cparent = hammer2_cluster_lookup_init(cparent, 0);
1443 cluster = hammer2_cluster_lookup(cparent, &key_dummy, lbase, lbase,
1444 HAMMER2_LOOKUP_NODATA, &ddflag);
1446 data = hammer2_cluster_wdata(cluster);
1449 KKASSERT(cluster->focus->flags &
1450 HAMMER2_CHAIN_MODIFIED);
1451 bzero(data->ipdata.u.data, HAMMER2_EMBEDDED_BYTES);
1452 hammer2_cluster_modsync(cluster);
1454 hammer2_cluster_delete(trans, cparent, cluster,
1455 HAMMER2_DELETE_PERMANENT);
1457 hammer2_cluster_unlock(cluster);
1459 hammer2_cluster_lookup_done(cparent);
1463 * Function to write the data as it is, without performing any sort of
1464 * compression. This function is used in path without compression and
1465 * default zero-checking path.
1469 hammer2_write_bp(hammer2_cluster_t *cluster, struct buf *bp, int ioflag,
1470 int pblksize, int *errorp)
1472 hammer2_chain_t *chain;
1479 error = 0; /* XXX TODO below */
1481 for (i = 0; i < cluster->nchains; ++i) {
1482 chain = cluster->array[i];
1484 temp_check = HAMMER2_DEC_CHECK(chain->bref.methods);
1486 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1488 switch(chain->bref.type) {
1489 case HAMMER2_BREF_TYPE_INODE:
1490 KKASSERT(chain->data->ipdata.op_flags &
1491 HAMMER2_OPFLAG_DIRECTDATA);
1492 KKASSERT(bp->b_loffset == 0);
1493 bcopy(bp->b_data, chain->data->ipdata.u.data,
1494 HAMMER2_EMBEDDED_BYTES);
1497 case HAMMER2_BREF_TYPE_DATA:
1498 error = hammer2_io_newnz(chain->hmp,
1499 chain->bref.data_off,
1500 chain->bytes, &dio);
1502 hammer2_io_bqrelse(&dio);
1503 kprintf("hammer2: WRITE PATH: "
1504 "dbp bread error\n");
1507 bdata = hammer2_io_data(dio, chain->bref.data_off);
1509 chain->bref.methods = HAMMER2_ENC_COMP(
1510 HAMMER2_COMP_NONE) +
1511 HAMMER2_ENC_CHECK(temp_check);
1512 bcopy(bp->b_data, bdata, chain->bytes);
1515 * The flush code doesn't calculate check codes for
1516 * file data (doing so can result in excessive I/O),
1519 hammer2_chain_setcheck(chain, bdata);
1522 * Device buffer is now valid, chain is no longer in
1523 * the initial state.
1525 * (No blockref table worries with file data)
1527 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1529 if (ioflag & IO_SYNC) {
1531 * Synchronous I/O requested.
1533 hammer2_io_bwrite(&dio);
1535 } else if ((ioflag & IO_DIRECT) &&
1536 loff + n == pblksize) {
1537 hammer2_io_bdwrite(&dio);
1539 } else if (ioflag & IO_ASYNC) {
1540 hammer2_io_bawrite(&dio);
1542 hammer2_io_bdwrite(&dio);
1546 panic("hammer2_write_bp: bad chain type %d\n",
1552 KKASSERT(error == 0); /* XXX TODO */
1559 hammer2_remount(hammer2_mount_t *hmp, struct mount *mp, char *path,
1560 struct vnode *devvp, struct ucred *cred)
1564 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1565 error = hammer2_recovery(hmp);
1574 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1576 hammer2_pfsmount_t *pmp;
1577 hammer2_mount_t *hmp;
1578 hammer2_chain_t *rchain;
1579 hammer2_cluster_t *cluster;
1589 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1590 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
1593 * If mount initialization proceeded far enough we must flush
1596 if (mntflags & MNT_FORCE)
1601 error = vflush(mp, 0, flags);
1606 ccms_domain_uninit(&pmp->ccms_dom);
1608 if (pmp->wthread_td) {
1609 mtx_lock(&pmp->wthread_mtx);
1610 pmp->wthread_destroy = 1;
1611 wakeup(&pmp->wthread_bioq);
1612 while (pmp->wthread_destroy != -1) {
1613 mtxsleep(&pmp->wthread_destroy,
1614 &pmp->wthread_mtx, 0,
1617 mtx_unlock(&pmp->wthread_mtx);
1618 pmp->wthread_td = NULL;
1622 * Cleanup our reference on ihidden.
1625 hammer2_inode_drop(pmp->ihidden);
1626 pmp->ihidden = NULL;
1630 * Cleanup our reference on iroot. iroot is (should) not be needed
1631 * by the flush code.
1634 cluster = &pmp->iroot->cluster;
1635 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1636 rchain = pmp->iroot->cluster.array[i];
1640 hammer2_vfs_unmount_hmp1(mp, hmp);
1642 atomic_clear_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
1643 #if REPORT_REFS_ERRORS
1644 if (rchain->refs != 1)
1645 kprintf("PMP->RCHAIN %p REFS WRONG %d\n",
1646 rchain, rchain->refs);
1648 KKASSERT(rchain->refs == 1);
1650 hammer2_chain_drop(rchain);
1651 cluster->array[i] = NULL;
1652 hammer2_vfs_unmount_hmp2(mp, hmp);
1654 cluster->focus = NULL;
1656 #if REPORT_REFS_ERRORS
1657 if (pmp->iroot->refs != 1)
1658 kprintf("PMP->IROOT %p REFS WRONG %d\n",
1659 pmp->iroot, pmp->iroot->refs);
1661 KKASSERT(pmp->iroot->refs == 1);
1663 /* ref for pmp->iroot */
1664 hammer2_inode_drop(pmp->iroot);
1669 mp->mnt_data = NULL;
1671 kmalloc_destroy(&pmp->mmsg);
1672 kmalloc_destroy(&pmp->minode);
1674 kfree(pmp, M_HAMMER2);
1678 lockmgr(&hammer2_mntlk, LK_RELEASE);
1685 hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp)
1687 hammer2_mount_exlock(hmp);
1690 kprintf("hammer2_unmount hmp=%p pmpcnt=%d\n", hmp, hmp->pmp_count);
1692 kdmsg_iocom_uninit(&hmp->iocom); /* XXX chain depend deadlck? */
1695 * Cycle the volume data lock as a safety (probably not needed any
1696 * more). To ensure everything is out we need to flush at least
1697 * three times. (1) The running of the unlinkq can dirty the
1698 * filesystem, (2) A normal flush can dirty the freemap, and
1699 * (3) ensure that the freemap is fully synchronized.
1701 * The next mount's recovery scan can clean everything up but we want
1702 * to leave the filesystem in a 100% clean state on a normal unmount.
1704 hammer2_voldata_lock(hmp);
1705 hammer2_voldata_unlock(hmp);
1707 hammer2_vfs_sync(mp, MNT_WAIT);
1708 hammer2_vfs_sync(mp, MNT_WAIT);
1709 hammer2_vfs_sync(mp, MNT_WAIT);
1712 if (hmp->pmp_count == 0) {
1713 if ((hmp->vchain.flags | hmp->fchain.flags) &
1714 HAMMER2_CHAIN_FLUSH_MASK) {
1715 kprintf("hammer2_unmount: chains left over "
1716 "after final sync\n");
1717 kprintf(" vchain %08x\n", hmp->vchain.flags);
1718 kprintf(" fchain %08x\n", hmp->fchain.flags);
1720 if (hammer2_debug & 0x0010)
1721 Debugger("entered debugger");
1728 hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp)
1730 hammer2_pfsmount_t *spmp;
1731 struct vnode *devvp;
1733 int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1736 * If no PFS's left drop the master hammer2_mount for the
1739 if (hmp->pmp_count == 0) {
1741 * Clean up SPMP and the super-root inode
1746 hammer2_inode_drop(spmp->iroot);
1750 kmalloc_destroy(&spmp->mmsg);
1751 kmalloc_destroy(&spmp->minode);
1752 kfree(spmp, M_HAMMER2);
1756 * Finish up with the device vnode
1758 if ((devvp = hmp->devvp) != NULL) {
1759 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1760 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1762 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1769 * Clear vchain/fchain flags that might prevent final cleanup
1772 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1773 atomic_clear_int(&hmp->vchain.flags,
1774 HAMMER2_CHAIN_MODIFIED);
1775 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1776 hammer2_chain_drop(&hmp->vchain);
1778 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1779 atomic_clear_int(&hmp->vchain.flags,
1780 HAMMER2_CHAIN_UPDATE);
1781 hammer2_chain_drop(&hmp->vchain);
1784 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1785 atomic_clear_int(&hmp->fchain.flags,
1786 HAMMER2_CHAIN_MODIFIED);
1787 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1788 hammer2_chain_drop(&hmp->fchain);
1790 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1791 atomic_clear_int(&hmp->fchain.flags,
1792 HAMMER2_CHAIN_UPDATE);
1793 hammer2_chain_drop(&hmp->fchain);
1797 * Final drop of embedded freemap root chain to
1798 * clean up fchain.core (fchain structure is not
1799 * flagged ALLOCATED so it is cleaned out and then
1802 hammer2_chain_drop(&hmp->fchain);
1805 * Final drop of embedded volume root chain to clean
1806 * up vchain.core (vchain structure is not flagged
1807 * ALLOCATED so it is cleaned out and then left to
1811 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1813 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1814 hammer2_mount_unlock(hmp);
1815 hammer2_chain_drop(&hmp->vchain);
1817 hammer2_io_cleanup(hmp, &hmp->iotree);
1818 if (hmp->iofree_count) {
1819 kprintf("io_cleanup: %d I/O's left hanging\n",
1823 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1824 kmalloc_destroy(&hmp->mchain);
1825 kfree(hmp, M_HAMMER2);
1827 hammer2_mount_unlock(hmp);
1833 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1834 ino_t ino, struct vnode **vpp)
1836 kprintf("hammer2_vget\n");
1837 return (EOPNOTSUPP);
1842 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1844 hammer2_pfsmount_t *pmp;
1845 hammer2_cluster_t *cparent;
1850 if (pmp->iroot == NULL) {
1854 cparent = hammer2_inode_lock_sh(pmp->iroot);
1855 vp = hammer2_igetv(pmp->iroot, cparent, &error);
1856 hammer2_inode_unlock_sh(pmp->iroot, cparent);
1859 kprintf("vnodefail\n");
1868 * XXX incorporate ipdata->inode_quota and data_quota
1872 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1874 hammer2_pfsmount_t *pmp;
1875 hammer2_mount_t *hmp;
1878 KKASSERT(pmp->iroot->cluster.nchains >= 1);
1879 hmp = pmp->iroot->cluster.focus->hmp; /* XXX */
1881 mp->mnt_stat.f_files = pmp->inode_count;
1882 mp->mnt_stat.f_ffree = 0;
1883 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1884 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1885 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1887 *sbp = mp->mnt_stat;
1893 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1895 hammer2_pfsmount_t *pmp;
1896 hammer2_mount_t *hmp;
1899 KKASSERT(pmp->iroot->cluster.nchains >= 1);
1900 hmp = pmp->iroot->cluster.focus->hmp; /* XXX */
1902 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1903 mp->mnt_vstat.f_files = pmp->inode_count;
1904 mp->mnt_vstat.f_ffree = 0;
1905 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1906 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1907 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1909 *sbp = mp->mnt_vstat;
1914 * Mount-time recovery (RW mounts)
1916 * Updates to the free block table are allowed to lag flushes by one
1917 * transaction. In case of a crash, then on a fresh mount we must do an
1918 * incremental scan of the last committed transaction id and make sure that
1919 * all related blocks have been marked allocated.
1921 * The super-root topology and each PFS has its own transaction id domain,
1922 * so we must track PFS boundary transitions.
1924 struct hammer2_recovery_elm {
1925 TAILQ_ENTRY(hammer2_recovery_elm) entry;
1926 hammer2_chain_t *chain;
1927 hammer2_tid_t sync_tid;
1930 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1932 struct hammer2_recovery_info {
1933 struct hammer2_recovery_list list;
1937 static int hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1938 hammer2_chain_t *parent,
1939 struct hammer2_recovery_info *info,
1940 hammer2_tid_t sync_tid);
1942 #define HAMMER2_RECOVERY_MAXDEPTH 10
1946 hammer2_recovery(hammer2_mount_t *hmp)
1948 hammer2_trans_t trans;
1949 struct hammer2_recovery_info info;
1950 struct hammer2_recovery_elm *elm;
1951 hammer2_chain_t *parent;
1952 hammer2_tid_t sync_tid;
1954 int cumulative_error = 0;
1956 hammer2_trans_init(&trans, hmp->spmp, 0);
1959 TAILQ_INIT(&info.list);
1961 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1962 cumulative_error = hammer2_recovery_scan(&trans, hmp, parent,
1964 hammer2_chain_lookup_done(parent);
1966 while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1967 TAILQ_REMOVE(&info.list, elm, entry);
1968 parent = elm->chain;
1969 sync_tid = elm->sync_tid;
1970 kfree(elm, M_HAMMER2);
1972 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
1973 HAMMER2_RESOLVE_NOREF);
1974 error = hammer2_recovery_scan(&trans, hmp, parent,
1976 hammer2_chain_unlock(parent);
1978 cumulative_error = error;
1980 hammer2_trans_done(&trans);
1982 return cumulative_error;
1987 hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1988 hammer2_chain_t *parent,
1989 struct hammer2_recovery_info *info,
1990 hammer2_tid_t sync_tid)
1992 hammer2_chain_t *chain;
1994 int cumulative_error = 0;
1995 int pfs_boundary = 0;
1999 * Adjust freemap to ensure that the block(s) are marked allocated.
2001 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2002 hammer2_freemap_adjust(trans, hmp, &parent->bref,
2003 HAMMER2_FREEMAP_DORECOVER);
2007 * Check type for recursive scan
2009 switch(parent->bref.type) {
2010 case HAMMER2_BREF_TYPE_VOLUME:
2011 /* data already instantiated */
2013 case HAMMER2_BREF_TYPE_INODE:
2015 * Must instantiate data for DIRECTDATA test and also
2018 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2019 if (parent->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2020 /* not applicable to recovery scan */
2021 hammer2_chain_unlock(parent);
2024 if ((parent->data->ipdata.op_flags & HAMMER2_OPFLAG_PFSROOT) &&
2027 sync_tid = parent->bref.mirror_tid - 1;
2029 hammer2_chain_unlock(parent);
2031 case HAMMER2_BREF_TYPE_INDIRECT:
2033 * Must instantiate data for recursion
2035 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2036 hammer2_chain_unlock(parent);
2038 case HAMMER2_BREF_TYPE_DATA:
2039 case HAMMER2_BREF_TYPE_FREEMAP:
2040 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2041 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2042 /* not applicable to recovery scan */
2050 * Defer operation if depth limit reached or if we are crossing a
2053 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH || pfs_boundary) {
2054 struct hammer2_recovery_elm *elm;
2056 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2057 elm->chain = parent;
2058 elm->sync_tid = sync_tid;
2059 hammer2_chain_ref(parent);
2060 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2061 /* unlocked by caller */
2068 * Recursive scan of the last flushed transaction only. We are
2069 * doing this without pmp assignments so don't leave the chains
2070 * hanging around after we are done with them.
2073 chain = hammer2_chain_scan(parent, NULL, &cache_index,
2074 HAMMER2_LOOKUP_NODATA);
2076 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2077 if (chain->bref.mirror_tid >= sync_tid) {
2079 error = hammer2_recovery_scan(trans, hmp, chain,
2083 cumulative_error = error;
2085 chain = hammer2_chain_scan(parent, chain, &cache_index,
2086 HAMMER2_LOOKUP_NODATA);
2089 return cumulative_error;
2093 * Sync the entire filesystem; this is called from the filesystem syncer
2094 * process periodically and whenever a user calls sync(1) on the hammer
2097 * Currently is actually called from the syncer! \o/
2099 * This task will have to snapshot the state of the dirty inode chain.
2100 * From that, it will have to make sure all of the inodes on the dirty
2101 * chain have IO initiated. We make sure that io is initiated for the root
2104 * If waitfor is set, we wait for media to acknowledge the new rootblock.
2106 * THINKS: side A vs side B, to have sync not stall all I/O?
2109 hammer2_vfs_sync(struct mount *mp, int waitfor)
2111 struct hammer2_sync_info info;
2112 hammer2_inode_t *iroot;
2113 hammer2_chain_t *chain;
2114 hammer2_chain_t *parent;
2115 hammer2_pfsmount_t *pmp;
2116 hammer2_mount_t *hmp;
2127 KKASSERT(iroot->pmp == pmp);
2130 * We can't acquire locks on existing vnodes while in a transaction
2131 * without risking a deadlock. This assumes that vfsync() can be
2132 * called without the vnode locked (which it can in DragonFly).
2133 * Otherwise we'd have to implement a multi-pass or flag the lock
2134 * failures and retry.
2136 * The reclamation code interlocks with the sync list's token
2137 * (by removing the vnode from the scan list) before unlocking
2138 * the inode, giving us time to ref the inode.
2140 /*flags = VMSC_GETVP;*/
2142 if (waitfor & MNT_LAZY)
2143 flags |= VMSC_ONEPASS;
2146 * Start our flush transaction. This does not return until all
2147 * concurrent transactions have completed and will prevent any
2148 * new transactions from running concurrently, except for the
2149 * buffer cache transactions.
2151 * For efficiency do an async pass before making sure with a
2152 * synchronous pass on all related buffer cache buffers. It
2153 * should theoretically not be possible for any new file buffers
2154 * to be instantiated during this sequence.
2156 hammer2_trans_init(&info.trans, pmp, HAMMER2_TRANS_ISFLUSH |
2157 HAMMER2_TRANS_PREFLUSH);
2158 hammer2_run_unlinkq(&info.trans, pmp);
2161 info.waitfor = MNT_NOWAIT;
2162 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2163 info.waitfor = MNT_WAIT;
2164 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2167 * Clear PREFLUSH. This prevents (or asserts on) any new logical
2168 * buffer cache flushes which occur during the flush. Device buffers
2173 if (info.error == 0 && (waitfor & MNT_WAIT)) {
2174 info.waitfor = waitfor;
2175 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2179 hammer2_bioq_sync(info.trans.pmp);
2180 atomic_clear_int(&info.trans.flags, HAMMER2_TRANS_PREFLUSH);
2185 * Flush all storage elements making up the cluster
2187 * We must also flush any deleted siblings because the super-root
2188 * flush won't do it for us. They all must be staged or the
2189 * super-root flush will not be able to update its block table
2192 * XXX currently done serially instead of concurrently
2194 for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2195 chain = iroot->cluster.array[i];
2197 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
2198 hammer2_flush(&info.trans, chain);
2199 hammer2_chain_unlock(chain);
2203 hammer2_trans_done(&info.trans);
2207 * Flush all volume roots to synchronize PFS flushes with the
2208 * storage media. Use a super-root transaction for each one.
2210 * The flush code will detect super-root -> pfs-root chain
2211 * transitions using the last pfs-root flush.
2213 for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2214 chain = iroot->cluster.array[i];
2221 * We only have to flush each hmp once
2223 for (j = i - 1; j >= 0; --j) {
2224 if (iroot->cluster.array[j] &&
2225 iroot->cluster.array[j]->hmp == hmp)
2230 hammer2_trans_spmp(&info.trans, hmp->spmp);
2233 * Force an update of the XID from the PFS root to the
2234 * topology root. We couldn't do this from the PFS
2235 * transaction because a SPMP transaction is needed.
2236 * This does not modify blocks, instead what it does is
2237 * allow the flush code to find the transition point and
2238 * then update on the way back up.
2240 parent = chain->parent;
2241 KKASSERT(chain->pmp != parent->pmp);
2242 hammer2_chain_setflush(&info.trans, parent);
2245 * Media mounts have two 'roots', vchain for the topology
2246 * and fchain for the free block table. Flush both.
2248 * Note that the topology and free block table are handled
2249 * independently, so the free block table can wind up being
2250 * ahead of the topology. We depend on the bulk free scan
2251 * code to deal with any loose ends.
2253 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2254 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2255 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2257 * This will also modify vchain as a side effect,
2258 * mark vchain as modified now.
2260 hammer2_voldata_modify(hmp);
2261 chain = &hmp->fchain;
2262 hammer2_flush(&info.trans, chain);
2263 KKASSERT(chain == &hmp->fchain);
2265 hammer2_chain_unlock(&hmp->fchain);
2266 hammer2_chain_unlock(&hmp->vchain);
2268 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2269 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2270 chain = &hmp->vchain;
2271 hammer2_flush(&info.trans, chain);
2272 KKASSERT(chain == &hmp->vchain);
2277 hammer2_chain_unlock(&hmp->vchain);
2280 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2281 if ((hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) ||
2283 /* this will also modify vchain as a side effect */
2284 chain = &hmp->fchain;
2285 hammer2_flush(&info.trans, chain);
2286 KKASSERT(chain == &hmp->fchain);
2288 hammer2_chain_unlock(&hmp->fchain);
2294 * We can't safely flush the volume header until we have
2295 * flushed any device buffers which have built up.
2297 * XXX this isn't being incremental
2299 vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
2300 error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
2301 vn_unlock(hmp->devvp);
2304 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
2305 * volume header needs synchronization via hmp->volsync.
2307 * XXX synchronize the flag & data with only this flush XXX
2310 (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
2314 * Synchronize the disk before flushing the volume
2318 bp->b_bio1.bio_offset = 0;
2321 bp->b_cmd = BUF_CMD_FLUSH;
2322 bp->b_bio1.bio_done = biodone_sync;
2323 bp->b_bio1.bio_flags |= BIO_SYNC;
2324 vn_strategy(hmp->devvp, &bp->b_bio1);
2325 biowait(&bp->b_bio1, "h2vol");
2329 * Then we can safely flush the version of the
2330 * volume header synchronized by the flush code.
2332 i = hmp->volhdrno + 1;
2333 if (i >= HAMMER2_NUM_VOLHDRS)
2335 if (i * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
2336 hmp->volsync.volu_size) {
2339 kprintf("sync volhdr %d %jd\n",
2340 i, (intmax_t)hmp->volsync.volu_size);
2341 bp = getblk(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2342 HAMMER2_PBUFSIZE, 0, 0);
2343 atomic_clear_int(&hmp->vchain.flags,
2344 HAMMER2_CHAIN_VOLUMESYNC);
2345 bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
2350 total_error = error;
2353 hammer2_trans_done(&info.trans);
2356 hammer2_trans_done(&info.trans);
2358 return (total_error);
2365 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2367 struct hammer2_sync_info *info = data;
2368 hammer2_inode_t *ip;
2377 if (vp->v_type == VNON || vp->v_type == VBAD) {
2381 if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2382 RB_EMPTY(&vp->v_rbdirty_tree)) {
2388 * VOP_FSYNC will start a new transaction so replicate some code
2389 * here to do it inline (see hammer2_vop_fsync()).
2391 * WARNING: The vfsync interacts with the buffer cache and might
2392 * block, we can't hold the inode lock at that time.
2393 * However, we MUST ref ip before blocking to ensure that
2394 * it isn't ripped out from under us (since we do not
2395 * hold a lock on the vnode).
2397 hammer2_inode_ref(ip);
2398 atomic_clear_int(&ip->flags, HAMMER2_INODE_MODIFIED);
2400 vfsync(vp, MNT_NOWAIT, 1, NULL, NULL);
2402 hammer2_inode_drop(ip);
2406 info->error = error;
2413 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2420 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2421 struct fid *fhp, struct vnode **vpp)
2428 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2429 int *exflagsp, struct ucred **credanonp)
2435 * Support code for hammer2_mount(). Read, verify, and install the volume
2436 * header into the HMP
2438 * XXX read four volhdrs and use the one with the highest TID whos CRC
2443 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2444 * nonexistant locations.
2446 * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2450 hammer2_install_volume_header(hammer2_mount_t *hmp)
2452 hammer2_volume_data_t *vd;
2454 hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2466 * There are up to 4 copies of the volume header (syncs iterate
2467 * between them so there is no single master). We don't trust the
2468 * volu_size field so we don't know precisely how large the filesystem
2469 * is, so depend on the OS to return an error if we go beyond the
2470 * block device's EOF.
2472 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2473 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2474 HAMMER2_VOLUME_BYTES, &bp);
2481 vd = (struct hammer2_volume_data *) bp->b_data;
2482 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2483 (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2489 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2490 /* XXX: Reversed-endianness filesystem */
2491 kprintf("hammer2: reverse-endian filesystem detected");
2497 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2498 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2499 HAMMER2_VOLUME_ICRC0_SIZE);
2500 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2501 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2502 HAMMER2_VOLUME_ICRC1_SIZE);
2503 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2504 kprintf("hammer2 volume header crc "
2505 "mismatch copy #%d %08x/%08x\n",
2512 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2521 hmp->volsync = hmp->voldata;
2523 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2524 kprintf("hammer2: using volume header #%d\n",
2529 kprintf("hammer2: no valid volume headers found!\n");
2535 * Reconnect using the passed file pointer. The caller must ref the
2539 hammer2_cluster_reconnect(hammer2_mount_t *hmp, struct file *fp)
2542 const char *name = "disk-volume";
2545 * Closes old comm descriptor, kills threads, cleans up
2546 * states, then installs the new descriptor and creates
2549 kdmsg_iocom_reconnect(&hmp->iocom, fp, "hammer2");
2552 * Setup LNK_CONN fields for autoinitiated state machine. We
2553 * will use SPANs to advertise multiple PFSs so only pass the
2554 * fsid and HAMMER2_PFSTYPE_SUPROOT for the AUTOCONN.
2556 * We are not initiating a LNK_SPAN so we do not have to set-up
2557 * iocom.auto_lnk_span.
2559 bzero(&hmp->iocom.auto_lnk_conn.pfs_clid,
2560 sizeof(hmp->iocom.auto_lnk_conn.pfs_clid));
2561 hmp->iocom.auto_lnk_conn.pfs_fsid = hmp->voldata.fsid;
2562 hmp->iocom.auto_lnk_conn.pfs_type = HAMMER2_PFSTYPE_SUPROOT;
2563 hmp->iocom.auto_lnk_conn.proto_version = DMSG_SPAN_PROTO_1;
2565 hmp->iocom.auto_lnk_conn.peer_type = hmp->voldata.peer_type;
2567 hmp->iocom.auto_lnk_conn.peer_type = DMSG_PEER_HAMMER2;
2570 * Filter adjustment. Clients do not need visibility into other
2571 * clients (otherwise millions of clients would present a serious
2572 * problem). The fs_label also serves to restrict the namespace.
2574 hmp->iocom.auto_lnk_conn.peer_mask = 1LLU << DMSG_PEER_HAMMER2;
2575 hmp->iocom.auto_lnk_conn.pfs_mask = (uint64_t)-1;
2578 switch (ipdata->pfs_type) {
2579 case DMSG_PFSTYPE_CLIENT:
2580 hmp->iocom.auto_lnk_conn.peer_mask &=
2581 ~(1LLU << DMSG_PFSTYPE_CLIENT);
2588 name_len = strlen(name);
2589 if (name_len >= sizeof(hmp->iocom.auto_lnk_conn.fs_label))
2590 name_len = sizeof(hmp->iocom.auto_lnk_conn.fs_label) - 1;
2591 bcopy(name, hmp->iocom.auto_lnk_conn.fs_label, name_len);
2592 hmp->iocom.auto_lnk_conn.fs_label[name_len] = 0;
2594 kdmsg_iocom_autoinitiate(&hmp->iocom, hammer2_autodmsg);
2598 hammer2_rcvdmsg(kdmsg_msg_t *msg)
2600 kprintf("RCVMSG %08x\n", msg->tcmd);
2603 case DMSG_DBG_SHELL:
2606 * Execute shell command (not supported atm)
2608 kdmsg_msg_result(msg, DMSG_ERR_NOSUPP);
2610 case DMSG_DBG_SHELL | DMSGF_REPLY:
2614 if (msg->aux_data) {
2615 msg->aux_data[msg->aux_size - 1] = 0;
2616 kprintf("HAMMER2 DBG: %s\n", msg->aux_data);
2621 * Unsupported message received. We only need to
2622 * reply if it's a transaction in order to close our end.
2623 * Ignore any one-way messages or any further messages
2624 * associated with the transaction.
2626 * NOTE: This case also includes DMSG_LNK_ERROR messages
2627 * which might be one-way, replying to those would
2628 * cause an infinite ping-pong.
2630 if (msg->any.head.cmd & DMSGF_CREATE)
2631 kdmsg_msg_reply(msg, DMSG_ERR_NOSUPP);
2638 * This function is called after KDMSG has automatically handled processing
2639 * of a LNK layer message (typically CONN, SPAN, or CIRC).
2641 * We tag off the LNK_CONN to trigger our LNK_VOLCONF messages which
2642 * advertises all available hammer2 super-root volumes.
2644 static void hammer2_update_spans(hammer2_mount_t *hmp, kdmsg_state_t *state);
2647 hammer2_autodmsg(kdmsg_msg_t *msg)
2649 hammer2_mount_t *hmp = msg->state->iocom->handle;
2652 kprintf("RCAMSG %08x\n", msg->tcmd);
2655 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_REPLY:
2656 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_DELETE | DMSGF_REPLY:
2657 if (msg->any.head.cmd & DMSGF_CREATE) {
2658 kprintf("HAMMER2: VOLDATA DUMP\n");
2661 * Dump the configuration stored in the volume header.
2662 * This will typically be import/export access rights,
2663 * master encryption keys (encrypted), etc.
2665 hammer2_voldata_lock(hmp);
2667 while (copyid < HAMMER2_COPYID_COUNT) {
2668 if (hmp->voldata.copyinfo[copyid].copyid)
2669 hammer2_volconf_update(hmp, copyid);
2672 hammer2_voldata_unlock(hmp);
2674 kprintf("HAMMER2: INITIATE SPANs\n");
2675 hammer2_update_spans(hmp, msg->state);
2677 if ((msg->any.head.cmd & DMSGF_DELETE) &&
2678 msg->state && (msg->state->txcmd & DMSGF_DELETE) == 0) {
2679 kprintf("HAMMER2: CONN WAS TERMINATED\n");
2688 * Update LNK_SPAN state
2691 hammer2_update_spans(hammer2_mount_t *hmp, kdmsg_state_t *state)
2693 const hammer2_inode_data_t *ipdata;
2694 hammer2_cluster_t *cparent;
2695 hammer2_cluster_t *cluster;
2696 hammer2_pfsmount_t *spmp;
2697 hammer2_key_t key_next;
2703 * Lookup mount point under the media-localized super-root.
2705 * cluster->pmp will incorrectly point to spmp and must be fixed
2709 cparent = hammer2_inode_lock_ex(spmp->iroot);
2710 cluster = hammer2_cluster_lookup(cparent, &key_next,
2715 if (hammer2_cluster_type(cluster) != HAMMER2_BREF_TYPE_INODE)
2717 ipdata = &hammer2_cluster_data(cluster)->ipdata;
2718 kprintf("UPDATE SPANS: %s\n", ipdata->filename);
2720 rmsg = kdmsg_msg_alloc(state, DMSG_LNK_SPAN | DMSGF_CREATE,
2721 hammer2_lnk_span_reply, NULL);
2722 rmsg->any.lnk_span.pfs_clid = ipdata->pfs_clid;
2723 rmsg->any.lnk_span.pfs_fsid = ipdata->pfs_fsid;
2724 rmsg->any.lnk_span.pfs_type = ipdata->pfs_type;
2725 rmsg->any.lnk_span.peer_type = DMSG_PEER_HAMMER2;
2726 rmsg->any.lnk_span.proto_version = DMSG_SPAN_PROTO_1;
2727 name_len = ipdata->name_len;
2728 if (name_len >= sizeof(rmsg->any.lnk_span.fs_label))
2729 name_len = sizeof(rmsg->any.lnk_span.fs_label) - 1;
2730 bcopy(ipdata->filename, rmsg->any.lnk_span.fs_label, name_len);
2732 kdmsg_msg_write(rmsg);
2734 cluster = hammer2_cluster_next(cparent, cluster,
2740 hammer2_inode_unlock_ex(spmp->iroot, cparent);
2745 hammer2_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg)
2747 if ((state->txcmd & DMSGF_DELETE) == 0 &&
2748 (msg->any.head.cmd & DMSGF_DELETE)) {
2749 kdmsg_msg_reply(msg, 0);
2755 * Volume configuration updates are passed onto the userland service
2756 * daemon via the open LNK_CONN transaction.
2759 hammer2_volconf_update(hammer2_mount_t *hmp, int index)
2763 /* XXX interlock against connection state termination */
2764 kprintf("volconf update %p\n", hmp->iocom.conn_state);
2765 if (hmp->iocom.conn_state) {
2766 kprintf("TRANSMIT VOLCONF VIA OPEN CONN TRANSACTION\n");
2767 msg = kdmsg_msg_alloc(hmp->iocom.conn_state,
2768 DMSG_LNK_HAMMER2_VOLCONF,
2770 H2_LNK_VOLCONF(msg)->copy = hmp->voldata.copyinfo[index];
2771 H2_LNK_VOLCONF(msg)->mediaid = hmp->voldata.fsid;
2772 H2_LNK_VOLCONF(msg)->index = index;
2773 kdmsg_msg_write(msg);
2778 * This handles hysteresis on regular file flushes. Because the BIOs are
2779 * routed to a thread it is possible for an excessive number to build up
2780 * and cause long front-end stalls long before the runningbuffspace limit
2781 * is hit, so we implement hammer2_flush_pipe to control the
2784 * This is a particular problem when compression is used.
2787 hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp)
2789 atomic_add_int(&pmp->count_lwinprog, 1);
2793 hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp)
2797 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2798 if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2799 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2800 atomic_clear_int(&pmp->count_lwinprog,
2801 HAMMER2_LWINPROG_WAITING);
2802 wakeup(&pmp->count_lwinprog);
2807 hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp)
2812 lwinprog = pmp->count_lwinprog;
2814 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2816 tsleep_interlock(&pmp->count_lwinprog, 0);
2817 atomic_set_int(&pmp->count_lwinprog, HAMMER2_LWINPROG_WAITING);
2818 lwinprog = pmp->count_lwinprog;
2819 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2821 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2826 * Manage excessive memory resource use for chain and related
2830 hammer2_pfs_memory_wait(hammer2_pfsmount_t *pmp)
2840 * Atomic check condition and wait. Also do an early speedup of
2841 * the syncer to try to avoid hitting the wait.
2844 waiting = pmp->inmem_dirty_chains;
2846 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2848 limit = pmp->mp->mnt_nvnodelistsize / 10;
2849 if (limit < hammer2_limit_dirty_chains)
2850 limit = hammer2_limit_dirty_chains;
2855 if ((int)(ticks - zzticks) > hz) {
2857 kprintf("count %ld %ld\n", count, limit);
2862 * Block if there are too many dirty chains present, wait
2863 * for the flush to clean some out.
2865 if (count > limit) {
2866 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2867 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2869 waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2870 speedup_syncer(pmp->mp);
2871 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2874 continue; /* loop on success or fail */
2878 * Try to start an early flush before we are forced to block.
2880 if (count > limit * 7 / 10)
2881 speedup_syncer(pmp->mp);
2887 hammer2_pfs_memory_inc(hammer2_pfsmount_t *pmp)
2890 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2895 hammer2_pfs_memory_wakeup(hammer2_pfsmount_t *pmp)
2903 waiting = pmp->inmem_dirty_chains;
2905 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2908 ~HAMMER2_DIRTYCHAIN_WAITING)) {
2913 if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2914 wakeup(&pmp->inmem_dirty_chains);
2921 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2923 hammer2_chain_t *scan;
2924 hammer2_chain_t *parent;
2928 kprintf("%*.*s...\n", tab, tab, "");
2933 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2935 chain, chain->bref.type,
2936 chain->bref.key, chain->bref.keybits,
2937 chain->bref.mirror_tid);
2939 kprintf("%*.*s [%08x] (%s) refs=%d\n",
2942 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2943 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
2946 kprintf("%*.*s core [%08x]",
2950 parent = chain->parent;
2952 kprintf("\n%*.*s p=%p [pflags %08x prefs %d",
2954 parent, parent->flags, parent->refs);
2955 if (RB_EMPTY(&chain->core.rbtree)) {
2959 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2960 hammer2_dump_chain(scan, tab + 4, countp, 'a');
2961 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2962 kprintf("%*.*s}(%s)\n", tab, tab, "",
2963 chain->data->ipdata.filename);
2965 kprintf("%*.*s}\n", tab, tab, "");