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 * From hammer2_vnops.c.
1095 * The core write function which determines which path to take
1096 * depending on compression settings.
1100 hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
1101 hammer2_inode_t *ip, hammer2_inode_data_t *ipdata,
1102 hammer2_cluster_t *cparent,
1103 hammer2_key_t lbase, int ioflag, int pblksize,
1106 hammer2_cluster_t *cluster;
1108 switch(HAMMER2_DEC_COMP(ipdata->comp_algo)) {
1109 case HAMMER2_COMP_NONE:
1111 * We have to assign physical storage to the buffer
1112 * we intend to dirty or write now to avoid deadlocks
1113 * in the strategy code later.
1115 * This can return NOOFFSET for inode-embedded data.
1116 * The strategy code will take care of it in that case.
1118 cluster = hammer2_assign_physical(trans, ip, cparent,
1121 hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp);
1123 hammer2_cluster_unlock(cluster);
1125 case HAMMER2_COMP_AUTOZERO:
1127 * Check for zero-fill only
1129 hammer2_zero_check_and_write(bp, trans, ip,
1130 ipdata, cparent, lbase,
1131 ioflag, pblksize, errorp);
1133 case HAMMER2_COMP_LZ4:
1134 case HAMMER2_COMP_ZLIB:
1137 * Check for zero-fill and attempt compression.
1139 hammer2_compress_and_write(bp, trans, ip,
1149 * Generic function that will perform the compression in compression
1150 * write path. The compression algorithm is determined by the settings
1151 * obtained from inode.
1155 hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
1156 hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1157 hammer2_cluster_t *cparent,
1158 hammer2_key_t lbase, int ioflag, int pblksize,
1159 int *errorp, int comp_algo)
1161 hammer2_cluster_t *cluster;
1162 hammer2_chain_t *chain;
1164 int comp_block_size;
1168 if (test_block_zeros(bp->b_data, pblksize)) {
1169 zero_write(bp, trans, ip, ipdata, cparent, lbase, errorp);
1176 KKASSERT(pblksize / 2 <= 32768);
1178 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
1179 z_stream strm_compress;
1183 switch(HAMMER2_DEC_COMP(comp_algo)) {
1184 case HAMMER2_COMP_LZ4:
1185 comp_buffer = objcache_get(cache_buffer_write,
1187 comp_size = LZ4_compress_limitedOutput(
1189 &comp_buffer[sizeof(int)],
1191 pblksize / 2 - sizeof(int));
1193 * We need to prefix with the size, LZ4
1194 * doesn't do it for us. Add the related
1197 *(int *)comp_buffer = comp_size;
1199 comp_size += sizeof(int);
1201 case HAMMER2_COMP_ZLIB:
1202 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
1203 if (comp_level == 0)
1204 comp_level = 6; /* default zlib compression */
1205 else if (comp_level < 6)
1207 else if (comp_level > 9)
1209 ret = deflateInit(&strm_compress, comp_level);
1211 kprintf("HAMMER2 ZLIB: fatal error "
1212 "on deflateInit.\n");
1215 comp_buffer = objcache_get(cache_buffer_write,
1217 strm_compress.next_in = bp->b_data;
1218 strm_compress.avail_in = pblksize;
1219 strm_compress.next_out = comp_buffer;
1220 strm_compress.avail_out = pblksize / 2;
1221 ret = deflate(&strm_compress, Z_FINISH);
1222 if (ret == Z_STREAM_END) {
1223 comp_size = pblksize / 2 -
1224 strm_compress.avail_out;
1228 ret = deflateEnd(&strm_compress);
1231 kprintf("Error: Unknown compression method.\n");
1232 kprintf("Comp_method = %d.\n", comp_algo);
1237 if (comp_size == 0) {
1239 * compression failed or turned off
1241 comp_block_size = pblksize; /* safety */
1242 if (++ip->comp_heuristic > 128)
1243 ip->comp_heuristic = 8;
1246 * compression succeeded
1248 ip->comp_heuristic = 0;
1249 if (comp_size <= 1024) {
1250 comp_block_size = 1024;
1251 } else if (comp_size <= 2048) {
1252 comp_block_size = 2048;
1253 } else if (comp_size <= 4096) {
1254 comp_block_size = 4096;
1255 } else if (comp_size <= 8192) {
1256 comp_block_size = 8192;
1257 } else if (comp_size <= 16384) {
1258 comp_block_size = 16384;
1259 } else if (comp_size <= 32768) {
1260 comp_block_size = 32768;
1262 panic("hammer2: WRITE PATH: "
1263 "Weird comp_size value.");
1265 comp_block_size = pblksize;
1269 cluster = hammer2_assign_physical(trans, ip, cparent,
1270 lbase, comp_block_size,
1272 ipdata = &hammer2_cluster_data(cparent)->ipdata;
1275 kprintf("WRITE PATH: An error occurred while "
1276 "assigning physical space.\n");
1277 KKASSERT(cluster == NULL);
1281 for (i = 0; i < cluster->nchains; ++i) {
1286 chain = cluster->array[i];
1287 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1289 switch(chain->bref.type) {
1290 case HAMMER2_BREF_TYPE_INODE:
1291 KKASSERT(chain->data->ipdata.op_flags &
1292 HAMMER2_OPFLAG_DIRECTDATA);
1293 KKASSERT(bp->b_loffset == 0);
1294 bcopy(bp->b_data, chain->data->ipdata.u.data,
1295 HAMMER2_EMBEDDED_BYTES);
1297 case HAMMER2_BREF_TYPE_DATA:
1298 temp_check = HAMMER2_DEC_CHECK(chain->bref.methods);
1301 * Optimize out the read-before-write
1304 *errorp = hammer2_io_newnz(chain->hmp,
1305 chain->bref.data_off,
1309 hammer2_io_brelse(&dio);
1310 kprintf("hammer2: WRITE PATH: "
1311 "dbp bread error\n");
1314 bdata = hammer2_io_data(dio, chain->bref.data_off);
1317 * When loading the block make sure we don't
1318 * leave garbage after the compressed data.
1321 chain->bref.methods =
1322 HAMMER2_ENC_COMP(comp_algo) +
1323 HAMMER2_ENC_CHECK(temp_check);
1324 bcopy(comp_buffer, bdata, comp_size);
1325 if (comp_size != comp_block_size) {
1326 bzero(bdata + comp_size,
1327 comp_block_size - comp_size);
1330 chain->bref.methods =
1332 HAMMER2_COMP_NONE) +
1333 HAMMER2_ENC_CHECK(temp_check);
1334 bcopy(bp->b_data, bdata, pblksize);
1338 * Device buffer is now valid, chain is no longer in
1339 * the initial state.
1341 * (No blockref table worries with file data)
1343 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1345 /* Now write the related bdp. */
1346 if (ioflag & IO_SYNC) {
1348 * Synchronous I/O requested.
1350 hammer2_io_bwrite(&dio);
1352 } else if ((ioflag & IO_DIRECT) &&
1353 loff + n == pblksize) {
1354 hammer2_io_bdwrite(&dio);
1356 } else if (ioflag & IO_ASYNC) {
1357 hammer2_io_bawrite(&dio);
1359 hammer2_io_bdwrite(&dio);
1363 panic("hammer2_write_bp: bad chain type %d\n",
1371 hammer2_cluster_unlock(cluster);
1373 objcache_put(cache_buffer_write, comp_buffer);
1377 * Function that performs zero-checking and writing without compression,
1378 * it corresponds to default zero-checking path.
1382 hammer2_zero_check_and_write(struct buf *bp, hammer2_trans_t *trans,
1383 hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1384 hammer2_cluster_t *cparent,
1385 hammer2_key_t lbase, int ioflag, int pblksize, int *errorp)
1387 hammer2_cluster_t *cluster;
1389 if (test_block_zeros(bp->b_data, pblksize)) {
1390 zero_write(bp, trans, ip, ipdata, cparent, lbase, errorp);
1392 cluster = hammer2_assign_physical(trans, ip, cparent,
1393 lbase, pblksize, errorp);
1394 hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp);
1396 hammer2_cluster_unlock(cluster);
1401 * A function to test whether a block of data contains only zeros,
1402 * returns TRUE (non-zero) if the block is all zeros.
1406 test_block_zeros(const char *buf, size_t bytes)
1410 for (i = 0; i < bytes; i += sizeof(long)) {
1411 if (*(const long *)(buf + i) != 0)
1418 * Function to "write" a block that contains only zeros.
1422 zero_write(struct buf *bp, hammer2_trans_t *trans,
1423 hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1424 hammer2_cluster_t *cparent,
1425 hammer2_key_t lbase, int *errorp __unused)
1427 hammer2_cluster_t *cluster;
1428 hammer2_media_data_t *data;
1429 hammer2_key_t key_dummy;
1432 cparent = hammer2_cluster_lookup_init(cparent, 0);
1433 cluster = hammer2_cluster_lookup(cparent, &key_dummy, lbase, lbase,
1434 HAMMER2_LOOKUP_NODATA, &ddflag);
1436 data = hammer2_cluster_wdata(cluster);
1439 KKASSERT(cluster->focus->flags &
1440 HAMMER2_CHAIN_MODIFIED);
1441 bzero(data->ipdata.u.data, HAMMER2_EMBEDDED_BYTES);
1442 hammer2_cluster_modsync(cluster);
1444 hammer2_cluster_delete(trans, cparent, cluster,
1445 HAMMER2_DELETE_PERMANENT);
1447 hammer2_cluster_unlock(cluster);
1449 hammer2_cluster_lookup_done(cparent);
1453 * Function to write the data as it is, without performing any sort of
1454 * compression. This function is used in path without compression and
1455 * default zero-checking path.
1459 hammer2_write_bp(hammer2_cluster_t *cluster, struct buf *bp, int ioflag,
1460 int pblksize, int *errorp)
1462 hammer2_chain_t *chain;
1469 error = 0; /* XXX TODO below */
1471 for (i = 0; i < cluster->nchains; ++i) {
1472 chain = cluster->array[i];
1474 temp_check = HAMMER2_DEC_CHECK(chain->bref.methods);
1476 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1478 switch(chain->bref.type) {
1479 case HAMMER2_BREF_TYPE_INODE:
1480 KKASSERT(chain->data->ipdata.op_flags &
1481 HAMMER2_OPFLAG_DIRECTDATA);
1482 KKASSERT(bp->b_loffset == 0);
1483 bcopy(bp->b_data, chain->data->ipdata.u.data,
1484 HAMMER2_EMBEDDED_BYTES);
1487 case HAMMER2_BREF_TYPE_DATA:
1488 error = hammer2_io_newnz(chain->hmp,
1489 chain->bref.data_off,
1490 chain->bytes, &dio);
1492 hammer2_io_bqrelse(&dio);
1493 kprintf("hammer2: WRITE PATH: "
1494 "dbp bread error\n");
1497 bdata = hammer2_io_data(dio, chain->bref.data_off);
1499 chain->bref.methods = HAMMER2_ENC_COMP(
1500 HAMMER2_COMP_NONE) +
1501 HAMMER2_ENC_CHECK(temp_check);
1502 bcopy(bp->b_data, bdata, chain->bytes);
1505 * Device buffer is now valid, chain is no longer in
1506 * the initial state.
1508 * (No blockref table worries with file data)
1510 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1512 if (ioflag & IO_SYNC) {
1514 * Synchronous I/O requested.
1516 hammer2_io_bwrite(&dio);
1518 } else if ((ioflag & IO_DIRECT) &&
1519 loff + n == pblksize) {
1520 hammer2_io_bdwrite(&dio);
1522 } else if (ioflag & IO_ASYNC) {
1523 hammer2_io_bawrite(&dio);
1525 hammer2_io_bdwrite(&dio);
1529 panic("hammer2_write_bp: bad chain type %d\n",
1535 KKASSERT(error == 0); /* XXX TODO */
1542 hammer2_remount(hammer2_mount_t *hmp, struct mount *mp, char *path,
1543 struct vnode *devvp, struct ucred *cred)
1547 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1548 error = hammer2_recovery(hmp);
1557 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1559 hammer2_pfsmount_t *pmp;
1560 hammer2_mount_t *hmp;
1561 hammer2_chain_t *rchain;
1562 hammer2_cluster_t *cluster;
1572 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1573 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
1576 * If mount initialization proceeded far enough we must flush
1579 if (mntflags & MNT_FORCE)
1584 error = vflush(mp, 0, flags);
1589 ccms_domain_uninit(&pmp->ccms_dom);
1591 if (pmp->wthread_td) {
1592 mtx_lock(&pmp->wthread_mtx);
1593 pmp->wthread_destroy = 1;
1594 wakeup(&pmp->wthread_bioq);
1595 while (pmp->wthread_destroy != -1) {
1596 mtxsleep(&pmp->wthread_destroy,
1597 &pmp->wthread_mtx, 0,
1600 mtx_unlock(&pmp->wthread_mtx);
1601 pmp->wthread_td = NULL;
1605 * Cleanup our reference on ihidden.
1608 hammer2_inode_drop(pmp->ihidden);
1609 pmp->ihidden = NULL;
1613 * Cleanup our reference on iroot. iroot is (should) not be needed
1614 * by the flush code.
1617 cluster = &pmp->iroot->cluster;
1618 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1619 rchain = pmp->iroot->cluster.array[i];
1623 hammer2_vfs_unmount_hmp1(mp, hmp);
1625 atomic_clear_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
1626 #if REPORT_REFS_ERRORS
1627 if (rchain->refs != 1)
1628 kprintf("PMP->RCHAIN %p REFS WRONG %d\n",
1629 rchain, rchain->refs);
1631 KKASSERT(rchain->refs == 1);
1633 hammer2_chain_drop(rchain);
1634 cluster->array[i] = NULL;
1635 hammer2_vfs_unmount_hmp2(mp, hmp);
1637 cluster->focus = NULL;
1639 #if REPORT_REFS_ERRORS
1640 if (pmp->iroot->refs != 1)
1641 kprintf("PMP->IROOT %p REFS WRONG %d\n",
1642 pmp->iroot, pmp->iroot->refs);
1644 KKASSERT(pmp->iroot->refs == 1);
1646 /* ref for pmp->iroot */
1647 hammer2_inode_drop(pmp->iroot);
1652 mp->mnt_data = NULL;
1654 kmalloc_destroy(&pmp->mmsg);
1655 kmalloc_destroy(&pmp->minode);
1657 kfree(pmp, M_HAMMER2);
1661 lockmgr(&hammer2_mntlk, LK_RELEASE);
1668 hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp)
1670 hammer2_mount_exlock(hmp);
1673 kprintf("hammer2_unmount hmp=%p pmpcnt=%d\n", hmp, hmp->pmp_count);
1675 kdmsg_iocom_uninit(&hmp->iocom); /* XXX chain depend deadlck? */
1678 * Cycle the volume data lock as a safety (probably not needed any
1679 * more). To ensure everything is out we need to flush at least
1680 * three times. (1) The running of the unlinkq can dirty the
1681 * filesystem, (2) A normal flush can dirty the freemap, and
1682 * (3) ensure that the freemap is fully synchronized.
1684 * The next mount's recovery scan can clean everything up but we want
1685 * to leave the filesystem in a 100% clean state on a normal unmount.
1687 hammer2_voldata_lock(hmp);
1688 hammer2_voldata_unlock(hmp);
1690 hammer2_vfs_sync(mp, MNT_WAIT);
1691 hammer2_vfs_sync(mp, MNT_WAIT);
1692 hammer2_vfs_sync(mp, MNT_WAIT);
1695 if (hmp->pmp_count == 0) {
1696 if ((hmp->vchain.flags | hmp->fchain.flags) &
1697 HAMMER2_CHAIN_FLUSH_MASK) {
1698 kprintf("hammer2_unmount: chains left over "
1699 "after final sync\n");
1700 kprintf(" vchain %08x\n", hmp->vchain.flags);
1701 kprintf(" fchain %08x\n", hmp->fchain.flags);
1703 if (hammer2_debug & 0x0010)
1704 Debugger("entered debugger");
1711 hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp)
1713 hammer2_pfsmount_t *spmp;
1714 struct vnode *devvp;
1716 int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1719 * If no PFS's left drop the master hammer2_mount for the
1722 if (hmp->pmp_count == 0) {
1724 * Clean up SPMP and the super-root inode
1729 hammer2_inode_drop(spmp->iroot);
1733 kmalloc_destroy(&spmp->mmsg);
1734 kmalloc_destroy(&spmp->minode);
1735 kfree(spmp, M_HAMMER2);
1739 * Finish up with the device vnode
1741 if ((devvp = hmp->devvp) != NULL) {
1742 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1743 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1745 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1752 * Clear vchain/fchain flags that might prevent final cleanup
1755 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1756 atomic_clear_int(&hmp->vchain.flags,
1757 HAMMER2_CHAIN_MODIFIED);
1758 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1759 hammer2_chain_drop(&hmp->vchain);
1761 if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1762 atomic_clear_int(&hmp->vchain.flags,
1763 HAMMER2_CHAIN_UPDATE);
1764 hammer2_chain_drop(&hmp->vchain);
1767 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1768 atomic_clear_int(&hmp->fchain.flags,
1769 HAMMER2_CHAIN_MODIFIED);
1770 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1771 hammer2_chain_drop(&hmp->fchain);
1773 if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1774 atomic_clear_int(&hmp->fchain.flags,
1775 HAMMER2_CHAIN_UPDATE);
1776 hammer2_chain_drop(&hmp->fchain);
1780 * Final drop of embedded freemap root chain to
1781 * clean up fchain.core (fchain structure is not
1782 * flagged ALLOCATED so it is cleaned out and then
1785 hammer2_chain_drop(&hmp->fchain);
1788 * Final drop of embedded volume root chain to clean
1789 * up vchain.core (vchain structure is not flagged
1790 * ALLOCATED so it is cleaned out and then left to
1794 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1796 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1797 hammer2_mount_unlock(hmp);
1798 hammer2_chain_drop(&hmp->vchain);
1800 hammer2_io_cleanup(hmp, &hmp->iotree);
1801 if (hmp->iofree_count) {
1802 kprintf("io_cleanup: %d I/O's left hanging\n",
1806 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1807 kmalloc_destroy(&hmp->mchain);
1808 kfree(hmp, M_HAMMER2);
1810 hammer2_mount_unlock(hmp);
1816 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1817 ino_t ino, struct vnode **vpp)
1819 kprintf("hammer2_vget\n");
1820 return (EOPNOTSUPP);
1825 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1827 hammer2_pfsmount_t *pmp;
1828 hammer2_cluster_t *cparent;
1833 if (pmp->iroot == NULL) {
1837 cparent = hammer2_inode_lock_sh(pmp->iroot);
1838 vp = hammer2_igetv(pmp->iroot, cparent, &error);
1839 hammer2_inode_unlock_sh(pmp->iroot, cparent);
1842 kprintf("vnodefail\n");
1851 * XXX incorporate ipdata->inode_quota and data_quota
1855 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1857 hammer2_pfsmount_t *pmp;
1858 hammer2_mount_t *hmp;
1861 KKASSERT(pmp->iroot->cluster.nchains >= 1);
1862 hmp = pmp->iroot->cluster.focus->hmp; /* XXX */
1864 mp->mnt_stat.f_files = pmp->inode_count;
1865 mp->mnt_stat.f_ffree = 0;
1866 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1867 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1868 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1870 *sbp = mp->mnt_stat;
1876 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1878 hammer2_pfsmount_t *pmp;
1879 hammer2_mount_t *hmp;
1882 KKASSERT(pmp->iroot->cluster.nchains >= 1);
1883 hmp = pmp->iroot->cluster.focus->hmp; /* XXX */
1885 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1886 mp->mnt_vstat.f_files = pmp->inode_count;
1887 mp->mnt_vstat.f_ffree = 0;
1888 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1889 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1890 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1892 *sbp = mp->mnt_vstat;
1897 * Mount-time recovery (RW mounts)
1899 * Updates to the free block table are allowed to lag flushes by one
1900 * transaction. In case of a crash, then on a fresh mount we must do an
1901 * incremental scan of the last committed transaction id and make sure that
1902 * all related blocks have been marked allocated.
1904 * The super-root topology and each PFS has its own transaction id domain,
1905 * so we must track PFS boundary transitions.
1907 struct hammer2_recovery_elm {
1908 TAILQ_ENTRY(hammer2_recovery_elm) entry;
1909 hammer2_chain_t *chain;
1910 hammer2_tid_t sync_tid;
1913 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1915 struct hammer2_recovery_info {
1916 struct hammer2_recovery_list list;
1920 static int hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1921 hammer2_chain_t *parent,
1922 struct hammer2_recovery_info *info,
1923 hammer2_tid_t sync_tid);
1925 #define HAMMER2_RECOVERY_MAXDEPTH 10
1929 hammer2_recovery(hammer2_mount_t *hmp)
1931 hammer2_trans_t trans;
1932 struct hammer2_recovery_info info;
1933 struct hammer2_recovery_elm *elm;
1934 hammer2_chain_t *parent;
1935 hammer2_tid_t sync_tid;
1937 int cumulative_error = 0;
1939 hammer2_trans_init(&trans, hmp->spmp, 0);
1942 TAILQ_INIT(&info.list);
1944 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1945 cumulative_error = hammer2_recovery_scan(&trans, hmp, parent,
1947 hammer2_chain_lookup_done(parent);
1949 while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1950 TAILQ_REMOVE(&info.list, elm, entry);
1951 parent = elm->chain;
1952 sync_tid = elm->sync_tid;
1953 kfree(elm, M_HAMMER2);
1955 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
1956 HAMMER2_RESOLVE_NOREF);
1957 error = hammer2_recovery_scan(&trans, hmp, parent,
1959 hammer2_chain_unlock(parent);
1961 cumulative_error = error;
1963 hammer2_trans_done(&trans);
1965 return cumulative_error;
1970 hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1971 hammer2_chain_t *parent,
1972 struct hammer2_recovery_info *info,
1973 hammer2_tid_t sync_tid)
1975 hammer2_chain_t *chain;
1977 int cumulative_error = 0;
1978 int pfs_boundary = 0;
1982 * Adjust freemap to ensure that the block(s) are marked allocated.
1984 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1985 hammer2_freemap_adjust(trans, hmp, &parent->bref,
1986 HAMMER2_FREEMAP_DORECOVER);
1990 * Check type for recursive scan
1992 switch(parent->bref.type) {
1993 case HAMMER2_BREF_TYPE_VOLUME:
1994 /* data already instantiated */
1996 case HAMMER2_BREF_TYPE_INODE:
1998 * Must instantiate data for DIRECTDATA test and also
2001 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2002 if (parent->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2003 /* not applicable to recovery scan */
2004 hammer2_chain_unlock(parent);
2007 if ((parent->data->ipdata.op_flags & HAMMER2_OPFLAG_PFSROOT) &&
2010 sync_tid = parent->bref.mirror_tid - 1;
2012 hammer2_chain_unlock(parent);
2014 case HAMMER2_BREF_TYPE_INDIRECT:
2016 * Must instantiate data for recursion
2018 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2019 hammer2_chain_unlock(parent);
2021 case HAMMER2_BREF_TYPE_DATA:
2022 case HAMMER2_BREF_TYPE_FREEMAP:
2023 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2024 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2025 /* not applicable to recovery scan */
2033 * Defer operation if depth limit reached or if we are crossing a
2036 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH || pfs_boundary) {
2037 struct hammer2_recovery_elm *elm;
2039 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2040 elm->chain = parent;
2041 elm->sync_tid = sync_tid;
2042 hammer2_chain_ref(parent);
2043 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2044 /* unlocked by caller */
2051 * Recursive scan of the last flushed transaction only. We are
2052 * doing this without pmp assignments so don't leave the chains
2053 * hanging around after we are done with them.
2056 chain = hammer2_chain_scan(parent, NULL, &cache_index,
2057 HAMMER2_LOOKUP_NODATA);
2059 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2060 if (chain->bref.mirror_tid >= sync_tid) {
2062 error = hammer2_recovery_scan(trans, hmp, chain,
2066 cumulative_error = error;
2068 chain = hammer2_chain_scan(parent, chain, &cache_index,
2069 HAMMER2_LOOKUP_NODATA);
2072 return cumulative_error;
2076 * Sync the entire filesystem; this is called from the filesystem syncer
2077 * process periodically and whenever a user calls sync(1) on the hammer
2080 * Currently is actually called from the syncer! \o/
2082 * This task will have to snapshot the state of the dirty inode chain.
2083 * From that, it will have to make sure all of the inodes on the dirty
2084 * chain have IO initiated. We make sure that io is initiated for the root
2087 * If waitfor is set, we wait for media to acknowledge the new rootblock.
2089 * THINKS: side A vs side B, to have sync not stall all I/O?
2092 hammer2_vfs_sync(struct mount *mp, int waitfor)
2094 struct hammer2_sync_info info;
2095 hammer2_inode_t *iroot;
2096 hammer2_chain_t *chain;
2097 hammer2_chain_t *parent;
2098 hammer2_pfsmount_t *pmp;
2099 hammer2_mount_t *hmp;
2110 KKASSERT(iroot->pmp == pmp);
2113 * We can't acquire locks on existing vnodes while in a transaction
2114 * without risking a deadlock. This assumes that vfsync() can be
2115 * called without the vnode locked (which it can in DragonFly).
2116 * Otherwise we'd have to implement a multi-pass or flag the lock
2117 * failures and retry.
2119 * The reclamation code interlocks with the sync list's token
2120 * (by removing the vnode from the scan list) before unlocking
2121 * the inode, giving us time to ref the inode.
2123 /*flags = VMSC_GETVP;*/
2125 if (waitfor & MNT_LAZY)
2126 flags |= VMSC_ONEPASS;
2129 * Start our flush transaction. This does not return until all
2130 * concurrent transactions have completed and will prevent any
2131 * new transactions from running concurrently, except for the
2132 * buffer cache transactions.
2134 * For efficiency do an async pass before making sure with a
2135 * synchronous pass on all related buffer cache buffers. It
2136 * should theoretically not be possible for any new file buffers
2137 * to be instantiated during this sequence.
2139 hammer2_trans_init(&info.trans, pmp, HAMMER2_TRANS_ISFLUSH |
2140 HAMMER2_TRANS_PREFLUSH);
2141 hammer2_run_unlinkq(&info.trans, pmp);
2144 info.waitfor = MNT_NOWAIT;
2145 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2146 info.waitfor = MNT_WAIT;
2147 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2150 * Clear PREFLUSH. This prevents (or asserts on) any new logical
2151 * buffer cache flushes which occur during the flush. Device buffers
2156 if (info.error == 0 && (waitfor & MNT_WAIT)) {
2157 info.waitfor = waitfor;
2158 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2162 hammer2_bioq_sync(info.trans.pmp);
2163 atomic_clear_int(&info.trans.flags, HAMMER2_TRANS_PREFLUSH);
2168 * Flush all storage elements making up the cluster
2170 * We must also flush any deleted siblings because the super-root
2171 * flush won't do it for us. They all must be staged or the
2172 * super-root flush will not be able to update its block table
2175 * XXX currently done serially instead of concurrently
2177 for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2178 chain = iroot->cluster.array[i];
2180 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
2181 hammer2_flush(&info.trans, chain);
2182 hammer2_chain_unlock(chain);
2186 hammer2_trans_done(&info.trans);
2190 * Flush all volume roots to synchronize PFS flushes with the
2191 * storage media. Use a super-root transaction for each one.
2193 * The flush code will detect super-root -> pfs-root chain
2194 * transitions using the last pfs-root flush.
2196 for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2197 chain = iroot->cluster.array[i];
2204 * We only have to flush each hmp once
2206 for (j = i - 1; j >= 0; --j) {
2207 if (iroot->cluster.array[j] &&
2208 iroot->cluster.array[j]->hmp == hmp)
2213 hammer2_trans_spmp(&info.trans, hmp->spmp);
2216 * Force an update of the XID from the PFS root to the
2217 * topology root. We couldn't do this from the PFS
2218 * transaction because a SPMP transaction is needed.
2219 * This does not modify blocks, instead what it does is
2220 * allow the flush code to find the transition point and
2221 * then update on the way back up.
2223 parent = chain->parent;
2224 KKASSERT(chain->pmp != parent->pmp);
2225 hammer2_chain_setflush(&info.trans, parent);
2228 * Media mounts have two 'roots', vchain for the topology
2229 * and fchain for the free block table. Flush both.
2231 * Note that the topology and free block table are handled
2232 * independently, so the free block table can wind up being
2233 * ahead of the topology. We depend on the bulk free scan
2234 * code to deal with any loose ends.
2236 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2237 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2238 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2240 * This will also modify vchain as a side effect,
2241 * mark vchain as modified now.
2243 hammer2_voldata_modify(hmp);
2244 chain = &hmp->fchain;
2245 hammer2_flush(&info.trans, chain);
2246 KKASSERT(chain == &hmp->fchain);
2248 hammer2_chain_unlock(&hmp->fchain);
2249 hammer2_chain_unlock(&hmp->vchain);
2251 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2252 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_MASK) {
2253 chain = &hmp->vchain;
2254 hammer2_flush(&info.trans, chain);
2255 KKASSERT(chain == &hmp->vchain);
2260 hammer2_chain_unlock(&hmp->vchain);
2263 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2264 if ((hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_MASK) ||
2266 /* this will also modify vchain as a side effect */
2267 chain = &hmp->fchain;
2268 hammer2_flush(&info.trans, chain);
2269 KKASSERT(chain == &hmp->fchain);
2271 hammer2_chain_unlock(&hmp->fchain);
2277 * We can't safely flush the volume header until we have
2278 * flushed any device buffers which have built up.
2280 * XXX this isn't being incremental
2282 vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
2283 error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
2284 vn_unlock(hmp->devvp);
2287 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
2288 * volume header needs synchronization via hmp->volsync.
2290 * XXX synchronize the flag & data with only this flush XXX
2293 (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
2297 * Synchronize the disk before flushing the volume
2301 bp->b_bio1.bio_offset = 0;
2304 bp->b_cmd = BUF_CMD_FLUSH;
2305 bp->b_bio1.bio_done = biodone_sync;
2306 bp->b_bio1.bio_flags |= BIO_SYNC;
2307 vn_strategy(hmp->devvp, &bp->b_bio1);
2308 biowait(&bp->b_bio1, "h2vol");
2312 * Then we can safely flush the version of the
2313 * volume header synchronized by the flush code.
2315 i = hmp->volhdrno + 1;
2316 if (i >= HAMMER2_NUM_VOLHDRS)
2318 if (i * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
2319 hmp->volsync.volu_size) {
2322 kprintf("sync volhdr %d %jd\n",
2323 i, (intmax_t)hmp->volsync.volu_size);
2324 bp = getblk(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2325 HAMMER2_PBUFSIZE, 0, 0);
2326 atomic_clear_int(&hmp->vchain.flags,
2327 HAMMER2_CHAIN_VOLUMESYNC);
2328 bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
2333 total_error = error;
2336 hammer2_trans_done(&info.trans);
2339 hammer2_trans_done(&info.trans);
2341 return (total_error);
2348 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2350 struct hammer2_sync_info *info = data;
2351 hammer2_inode_t *ip;
2360 if (vp->v_type == VNON || vp->v_type == VBAD) {
2364 if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2365 RB_EMPTY(&vp->v_rbdirty_tree)) {
2371 * VOP_FSYNC will start a new transaction so replicate some code
2372 * here to do it inline (see hammer2_vop_fsync()).
2374 * WARNING: The vfsync interacts with the buffer cache and might
2375 * block, we can't hold the inode lock at that time.
2376 * However, we MUST ref ip before blocking to ensure that
2377 * it isn't ripped out from under us (since we do not
2378 * hold a lock on the vnode).
2380 hammer2_inode_ref(ip);
2381 atomic_clear_int(&ip->flags, HAMMER2_INODE_MODIFIED);
2383 vfsync(vp, MNT_NOWAIT, 1, NULL, NULL);
2385 hammer2_inode_drop(ip);
2389 info->error = error;
2396 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2403 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2404 struct fid *fhp, struct vnode **vpp)
2411 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2412 int *exflagsp, struct ucred **credanonp)
2418 * Support code for hammer2_mount(). Read, verify, and install the volume
2419 * header into the HMP
2421 * XXX read four volhdrs and use the one with the highest TID whos CRC
2426 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2427 * nonexistant locations.
2429 * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2433 hammer2_install_volume_header(hammer2_mount_t *hmp)
2435 hammer2_volume_data_t *vd;
2437 hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2449 * There are up to 4 copies of the volume header (syncs iterate
2450 * between them so there is no single master). We don't trust the
2451 * volu_size field so we don't know precisely how large the filesystem
2452 * is, so depend on the OS to return an error if we go beyond the
2453 * block device's EOF.
2455 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2456 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2457 HAMMER2_VOLUME_BYTES, &bp);
2464 vd = (struct hammer2_volume_data *) bp->b_data;
2465 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2466 (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2472 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2473 /* XXX: Reversed-endianness filesystem */
2474 kprintf("hammer2: reverse-endian filesystem detected");
2480 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2481 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2482 HAMMER2_VOLUME_ICRC0_SIZE);
2483 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2484 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2485 HAMMER2_VOLUME_ICRC1_SIZE);
2486 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2487 kprintf("hammer2 volume header crc "
2488 "mismatch copy #%d %08x/%08x\n",
2495 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2504 hmp->volsync = hmp->voldata;
2506 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2507 kprintf("hammer2: using volume header #%d\n",
2512 kprintf("hammer2: no valid volume headers found!\n");
2518 * Reconnect using the passed file pointer. The caller must ref the
2522 hammer2_cluster_reconnect(hammer2_mount_t *hmp, struct file *fp)
2525 const char *name = "disk-volume";
2528 * Closes old comm descriptor, kills threads, cleans up
2529 * states, then installs the new descriptor and creates
2532 kdmsg_iocom_reconnect(&hmp->iocom, fp, "hammer2");
2535 * Setup LNK_CONN fields for autoinitiated state machine. We
2536 * will use SPANs to advertise multiple PFSs so only pass the
2537 * fsid and HAMMER2_PFSTYPE_SUPROOT for the AUTOCONN.
2539 * We are not initiating a LNK_SPAN so we do not have to set-up
2540 * iocom.auto_lnk_span.
2542 bzero(&hmp->iocom.auto_lnk_conn.pfs_clid,
2543 sizeof(hmp->iocom.auto_lnk_conn.pfs_clid));
2544 hmp->iocom.auto_lnk_conn.pfs_fsid = hmp->voldata.fsid;
2545 hmp->iocom.auto_lnk_conn.pfs_type = HAMMER2_PFSTYPE_SUPROOT;
2546 hmp->iocom.auto_lnk_conn.proto_version = DMSG_SPAN_PROTO_1;
2548 hmp->iocom.auto_lnk_conn.peer_type = hmp->voldata.peer_type;
2550 hmp->iocom.auto_lnk_conn.peer_type = DMSG_PEER_HAMMER2;
2553 * Filter adjustment. Clients do not need visibility into other
2554 * clients (otherwise millions of clients would present a serious
2555 * problem). The fs_label also serves to restrict the namespace.
2557 hmp->iocom.auto_lnk_conn.peer_mask = 1LLU << DMSG_PEER_HAMMER2;
2558 hmp->iocom.auto_lnk_conn.pfs_mask = (uint64_t)-1;
2561 switch (ipdata->pfs_type) {
2562 case DMSG_PFSTYPE_CLIENT:
2563 hmp->iocom.auto_lnk_conn.peer_mask &=
2564 ~(1LLU << DMSG_PFSTYPE_CLIENT);
2571 name_len = strlen(name);
2572 if (name_len >= sizeof(hmp->iocom.auto_lnk_conn.fs_label))
2573 name_len = sizeof(hmp->iocom.auto_lnk_conn.fs_label) - 1;
2574 bcopy(name, hmp->iocom.auto_lnk_conn.fs_label, name_len);
2575 hmp->iocom.auto_lnk_conn.fs_label[name_len] = 0;
2577 kdmsg_iocom_autoinitiate(&hmp->iocom, hammer2_autodmsg);
2581 hammer2_rcvdmsg(kdmsg_msg_t *msg)
2583 kprintf("RCVMSG %08x\n", msg->tcmd);
2586 case DMSG_DBG_SHELL:
2589 * Execute shell command (not supported atm)
2591 kdmsg_msg_result(msg, DMSG_ERR_NOSUPP);
2593 case DMSG_DBG_SHELL | DMSGF_REPLY:
2597 if (msg->aux_data) {
2598 msg->aux_data[msg->aux_size - 1] = 0;
2599 kprintf("HAMMER2 DBG: %s\n", msg->aux_data);
2604 * Unsupported message received. We only need to
2605 * reply if it's a transaction in order to close our end.
2606 * Ignore any one-way messages or any further messages
2607 * associated with the transaction.
2609 * NOTE: This case also includes DMSG_LNK_ERROR messages
2610 * which might be one-way, replying to those would
2611 * cause an infinite ping-pong.
2613 if (msg->any.head.cmd & DMSGF_CREATE)
2614 kdmsg_msg_reply(msg, DMSG_ERR_NOSUPP);
2621 * This function is called after KDMSG has automatically handled processing
2622 * of a LNK layer message (typically CONN, SPAN, or CIRC).
2624 * We tag off the LNK_CONN to trigger our LNK_VOLCONF messages which
2625 * advertises all available hammer2 super-root volumes.
2627 static void hammer2_update_spans(hammer2_mount_t *hmp, kdmsg_state_t *state);
2630 hammer2_autodmsg(kdmsg_msg_t *msg)
2632 hammer2_mount_t *hmp = msg->state->iocom->handle;
2635 kprintf("RCAMSG %08x\n", msg->tcmd);
2638 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_REPLY:
2639 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_DELETE | DMSGF_REPLY:
2640 if (msg->any.head.cmd & DMSGF_CREATE) {
2641 kprintf("HAMMER2: VOLDATA DUMP\n");
2644 * Dump the configuration stored in the volume header.
2645 * This will typically be import/export access rights,
2646 * master encryption keys (encrypted), etc.
2648 hammer2_voldata_lock(hmp);
2650 while (copyid < HAMMER2_COPYID_COUNT) {
2651 if (hmp->voldata.copyinfo[copyid].copyid)
2652 hammer2_volconf_update(hmp, copyid);
2655 hammer2_voldata_unlock(hmp);
2657 kprintf("HAMMER2: INITIATE SPANs\n");
2658 hammer2_update_spans(hmp, msg->state);
2660 if ((msg->any.head.cmd & DMSGF_DELETE) &&
2661 msg->state && (msg->state->txcmd & DMSGF_DELETE) == 0) {
2662 kprintf("HAMMER2: CONN WAS TERMINATED\n");
2671 * Update LNK_SPAN state
2674 hammer2_update_spans(hammer2_mount_t *hmp, kdmsg_state_t *state)
2676 const hammer2_inode_data_t *ipdata;
2677 hammer2_cluster_t *cparent;
2678 hammer2_cluster_t *cluster;
2679 hammer2_pfsmount_t *spmp;
2680 hammer2_key_t key_next;
2686 * Lookup mount point under the media-localized super-root.
2688 * cluster->pmp will incorrectly point to spmp and must be fixed
2692 cparent = hammer2_inode_lock_ex(spmp->iroot);
2693 cluster = hammer2_cluster_lookup(cparent, &key_next,
2698 if (hammer2_cluster_type(cluster) != HAMMER2_BREF_TYPE_INODE)
2700 ipdata = &hammer2_cluster_data(cluster)->ipdata;
2701 kprintf("UPDATE SPANS: %s\n", ipdata->filename);
2703 rmsg = kdmsg_msg_alloc(state, DMSG_LNK_SPAN | DMSGF_CREATE,
2704 hammer2_lnk_span_reply, NULL);
2705 rmsg->any.lnk_span.pfs_clid = ipdata->pfs_clid;
2706 rmsg->any.lnk_span.pfs_fsid = ipdata->pfs_fsid;
2707 rmsg->any.lnk_span.pfs_type = ipdata->pfs_type;
2708 rmsg->any.lnk_span.peer_type = DMSG_PEER_HAMMER2;
2709 rmsg->any.lnk_span.proto_version = DMSG_SPAN_PROTO_1;
2710 name_len = ipdata->name_len;
2711 if (name_len >= sizeof(rmsg->any.lnk_span.fs_label))
2712 name_len = sizeof(rmsg->any.lnk_span.fs_label) - 1;
2713 bcopy(ipdata->filename, rmsg->any.lnk_span.fs_label, name_len);
2715 kdmsg_msg_write(rmsg);
2717 cluster = hammer2_cluster_next(cparent, cluster,
2723 hammer2_inode_unlock_ex(spmp->iroot, cparent);
2728 hammer2_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg)
2730 if ((state->txcmd & DMSGF_DELETE) == 0 &&
2731 (msg->any.head.cmd & DMSGF_DELETE)) {
2732 kdmsg_msg_reply(msg, 0);
2738 * Volume configuration updates are passed onto the userland service
2739 * daemon via the open LNK_CONN transaction.
2742 hammer2_volconf_update(hammer2_mount_t *hmp, int index)
2746 /* XXX interlock against connection state termination */
2747 kprintf("volconf update %p\n", hmp->iocom.conn_state);
2748 if (hmp->iocom.conn_state) {
2749 kprintf("TRANSMIT VOLCONF VIA OPEN CONN TRANSACTION\n");
2750 msg = kdmsg_msg_alloc(hmp->iocom.conn_state,
2751 DMSG_LNK_HAMMER2_VOLCONF,
2753 H2_LNK_VOLCONF(msg)->copy = hmp->voldata.copyinfo[index];
2754 H2_LNK_VOLCONF(msg)->mediaid = hmp->voldata.fsid;
2755 H2_LNK_VOLCONF(msg)->index = index;
2756 kdmsg_msg_write(msg);
2761 * This handles hysteresis on regular file flushes. Because the BIOs are
2762 * routed to a thread it is possible for an excessive number to build up
2763 * and cause long front-end stalls long before the runningbuffspace limit
2764 * is hit, so we implement hammer2_flush_pipe to control the
2767 * This is a particular problem when compression is used.
2770 hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp)
2772 atomic_add_int(&pmp->count_lwinprog, 1);
2776 hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp)
2780 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2781 if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2782 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2783 atomic_clear_int(&pmp->count_lwinprog,
2784 HAMMER2_LWINPROG_WAITING);
2785 wakeup(&pmp->count_lwinprog);
2790 hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp)
2795 lwinprog = pmp->count_lwinprog;
2797 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2799 tsleep_interlock(&pmp->count_lwinprog, 0);
2800 atomic_set_int(&pmp->count_lwinprog, HAMMER2_LWINPROG_WAITING);
2801 lwinprog = pmp->count_lwinprog;
2802 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2804 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2809 * Manage excessive memory resource use for chain and related
2813 hammer2_pfs_memory_wait(hammer2_pfsmount_t *pmp)
2823 * Atomic check condition and wait. Also do an early speedup of
2824 * the syncer to try to avoid hitting the wait.
2827 waiting = pmp->inmem_dirty_chains;
2829 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2831 limit = pmp->mp->mnt_nvnodelistsize / 10;
2832 if (limit < hammer2_limit_dirty_chains)
2833 limit = hammer2_limit_dirty_chains;
2838 if ((int)(ticks - zzticks) > hz) {
2840 kprintf("count %ld %ld\n", count, limit);
2845 * Block if there are too many dirty chains present, wait
2846 * for the flush to clean some out.
2848 if (count > limit) {
2849 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2850 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2852 waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2853 speedup_syncer(pmp->mp);
2854 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2857 continue; /* loop on success or fail */
2861 * Try to start an early flush before we are forced to block.
2863 if (count > limit * 7 / 10)
2864 speedup_syncer(pmp->mp);
2870 hammer2_pfs_memory_inc(hammer2_pfsmount_t *pmp)
2873 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2878 hammer2_pfs_memory_wakeup(hammer2_pfsmount_t *pmp)
2886 waiting = pmp->inmem_dirty_chains;
2888 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2891 ~HAMMER2_DIRTYCHAIN_WAITING)) {
2896 if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2897 wakeup(&pmp->inmem_dirty_chains);
2904 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2906 hammer2_chain_t *scan;
2907 hammer2_chain_t *parent;
2911 kprintf("%*.*s...\n", tab, tab, "");
2916 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2918 chain, chain->bref.type,
2919 chain->bref.key, chain->bref.keybits,
2920 chain->bref.mirror_tid);
2922 kprintf("%*.*s [%08x] (%s) refs=%d\n",
2925 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2926 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
2929 kprintf("%*.*s core [%08x]",
2933 parent = chain->parent;
2935 kprintf("\n%*.*s p=%p [pflags %08x prefs %d",
2937 parent, parent->flags, parent->refs);
2938 if (RB_EMPTY(&chain->core.rbtree)) {
2942 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2943 hammer2_dump_chain(scan, tab + 4, countp, 'a');
2944 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2945 kprintf("%*.*s}(%s)\n", tab, tab, "",
2946 chain->data->ipdata.filename);
2948 kprintf("%*.*s}\n", tab, tab, "");