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->unlinkq_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);
529 lockinit(&hmp->vollk, "h2vol", 0, 0);
532 * vchain setup. vchain.data is embedded.
533 * vchain.refs is initialized and will never drop to 0.
535 * NOTE! voldata is not yet loaded.
537 hmp->vchain.hmp = hmp;
538 hmp->vchain.refs = 1;
539 hmp->vchain.data = (void *)&hmp->voldata;
540 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
541 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
542 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
543 hmp->vchain.delete_xid = HAMMER2_XID_MAX;
545 hammer2_chain_core_alloc(NULL, &hmp->vchain, NULL);
546 /* hmp->vchain.u.xxx is left NULL */
549 * fchain setup. fchain.data is embedded.
550 * fchain.refs is initialized and will never drop to 0.
552 * The data is not used but needs to be initialized to
553 * pass assertion muster. We use this chain primarily
554 * as a placeholder for the freemap's top-level RBTREE
555 * so it does not interfere with the volume's topology
558 hmp->fchain.hmp = hmp;
559 hmp->fchain.refs = 1;
560 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
561 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
562 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
563 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
564 hmp->fchain.bref.methods =
565 HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
566 HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
567 hmp->fchain.delete_xid = HAMMER2_XID_MAX;
569 hammer2_chain_core_alloc(NULL, &hmp->fchain, NULL);
570 /* hmp->fchain.u.xxx is left NULL */
573 * Install the volume header and initialize fields from
576 error = hammer2_install_volume_header(hmp);
579 hammer2_vfs_unmount_hmp1(mp, hmp);
580 hammer2_vfs_unmount_hmp2(mp, hmp);
581 hammer2_vfs_unmount(mp, MNT_FORCE);
586 * Really important to get these right or flush will get
589 hmp->spmp = hammer2_pfsalloc(NULL, hmp->voldata.mirror_tid);
590 kprintf("alloc spmp %p tid %016jx\n",
591 hmp->spmp, hmp->voldata.mirror_tid);
596 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
597 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
598 hmp->vchain.modify_xid = xid;
599 hmp->vchain.update_xlo = xid;
600 hmp->vchain.update_xhi = xid;
601 hmp->vchain.pmp = spmp;
602 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
603 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
604 hmp->fchain.modify_xid = xid;
605 hmp->fchain.update_xlo = xid;
606 hmp->fchain.update_xhi = xid;
607 hmp->fchain.pmp = spmp;
610 * First locate the super-root inode, which is key 0
611 * relative to the volume header's blockset.
613 * Then locate the root inode by scanning the directory keyspace
614 * represented by the label.
616 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
617 schain = hammer2_chain_lookup(&parent, &key_dummy,
618 HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
619 &cache_index, 0, &ddflag);
620 hammer2_chain_lookup_done(parent);
621 if (schain == NULL) {
622 kprintf("hammer2_mount: invalid super-root\n");
624 hammer2_vfs_unmount_hmp1(mp, hmp);
625 hammer2_vfs_unmount_hmp2(mp, hmp);
626 hammer2_vfs_unmount(mp, MNT_FORCE);
631 * Sanity-check schain's pmp, finish initializing spmp.
633 KKASSERT(schain->pmp == spmp);
634 spmp->pfs_clid = schain->data->ipdata.pfs_clid;
637 * NOTE: The CHAIN_PFSROOT is not set on the super-root inode.
638 * NOTE: inode_get sucks up schain's lock.
640 cluster = hammer2_cluster_from_chain(schain);
641 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster);
642 spmp->spmp_hmp = hmp;
643 hammer2_inode_ref(spmp->iroot);
644 hammer2_inode_unlock_ex(spmp->iroot, cluster);
646 /* leave spmp->iroot with one ref */
648 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
649 error = hammer2_recovery(hmp);
650 /* XXX do something with error */
655 * XXX RDONLY stuff is totally broken FIXME XXX
658 * Automatic handling of received LNK_SPAN
659 * Automatic handling of received LNK_CIRC
660 * No automatic LNK_SPAN generation - we do this ourselves
661 * No automatic LNK_CIRC generation - we do this ourselves
663 kdmsg_iocom_init(&hmp->iocom, hmp,
664 KDMSG_IOCOMF_AUTOCONN |
665 KDMSG_IOCOMF_AUTORXSPAN |
666 KDMSG_IOCOMF_AUTORXCIRC,
667 hmp->mchain, hammer2_rcvdmsg);
670 * Ref the cluster management messaging descriptor. The mount
671 * program deals with the other end of the communications pipe.
673 fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
675 hammer2_cluster_reconnect(hmp, fp);
677 kprintf("hammer2_mount: bad cluster_fd!\n");
685 * Lookup mount point under the media-localized super-root.
687 * cluster->pmp will incorrectly point to spmp and must be fixed
690 cparent = hammer2_inode_lock_ex(spmp->iroot);
691 lhc = hammer2_dirhash(label, strlen(label));
692 cluster = hammer2_cluster_lookup(cparent, &key_next,
693 lhc, lhc + HAMMER2_DIRHASH_LOMASK,
696 if (hammer2_cluster_type(cluster) == HAMMER2_BREF_TYPE_INODE &&
698 hammer2_cluster_data(cluster)->ipdata.filename) == 0) {
701 cluster = hammer2_cluster_next(cparent, cluster, &key_next,
703 lhc + HAMMER2_DIRHASH_LOMASK, 0);
705 hammer2_inode_unlock_ex(spmp->iroot, cparent);
707 if (cluster == NULL) {
708 kprintf("hammer2_mount: PFS label not found\n");
709 hammer2_vfs_unmount_hmp1(mp, hmp);
710 hammer2_vfs_unmount_hmp2(mp, hmp);
711 lockmgr(&hammer2_mntlk, LK_RELEASE);
712 hammer2_vfs_unmount(mp, MNT_FORCE);
716 for (i = 0; i < cluster->nchains; ++i) {
717 rchain = cluster->array[i];
718 KKASSERT(rchain->pmp == NULL);
719 if (rchain->flags & HAMMER2_CHAIN_MOUNTED) {
720 kprintf("hammer2_mount: PFS label already mounted!\n");
721 hammer2_cluster_unlock(cluster);
722 hammer2_vfs_unmount_hmp1(mp, hmp);
723 hammer2_vfs_unmount_hmp2(mp, hmp);
724 lockmgr(&hammer2_mntlk, LK_RELEASE);
725 hammer2_vfs_unmount(mp, MNT_FORCE);
729 if (rchain->flags & HAMMER2_CHAIN_RECYCLE) {
730 kprintf("hammer2_mount: PFS label is recycling\n");
731 hammer2_cluster_unlock(cluster);
732 hammer2_vfs_unmount_hmp1(mp, hmp);
733 hammer2_vfs_unmount_hmp2(mp, hmp);
734 lockmgr(&hammer2_mntlk, LK_RELEASE);
735 hammer2_vfs_unmount(mp, MNT_FORCE);
742 * Check to see if the cluster id is already mounted at the mount
743 * point. If it is, add us to the cluster.
745 ipdata = &hammer2_cluster_data(cluster)->ipdata;
746 hammer2_cluster_bref(cluster, &bref);
747 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
748 if (pmp->spmp_hmp == NULL &&
749 bcmp(&pmp->pfs_clid, &ipdata->pfs_clid,
750 sizeof(pmp->pfs_clid)) == 0) {
759 hammer2_inode_ref(pmp->iroot);
760 ccms_thread_lock(&pmp->iroot->topo_cst, CCMS_STATE_EXCLUSIVE);
762 if (pmp->iroot->cluster.nchains + cluster->nchains >
763 HAMMER2_MAXCLUSTER) {
764 kprintf("hammer2_mount: cluster full!\n");
766 ccms_thread_unlock(&pmp->iroot->topo_cst);
767 hammer2_inode_drop(pmp->iroot);
769 hammer2_cluster_unlock(cluster);
770 hammer2_vfs_unmount_hmp1(mp, hmp);
771 hammer2_vfs_unmount_hmp2(mp, hmp);
772 lockmgr(&hammer2_mntlk, LK_RELEASE);
773 hammer2_vfs_unmount(mp, MNT_FORCE);
776 kprintf("hammer2_vfs_mount: Adding pfs to existing cluster\n");
777 j = pmp->iroot->cluster.nchains;
778 for (i = 0; i < cluster->nchains; ++i) {
779 rchain = cluster->array[i];
780 KKASSERT(rchain->pmp == NULL);
782 hammer2_chain_ref(cluster->array[i]);
783 pmp->iroot->cluster.array[j] = cluster->array[i];
786 pmp->iroot->cluster.nchains = j;
787 ccms_thread_unlock(&pmp->iroot->topo_cst);
788 hammer2_inode_drop(pmp->iroot);
789 hammer2_cluster_unlock(cluster);
790 lockmgr(&hammer2_mntlk, LK_RELEASE);
793 hammer2_inode_install_hidden(pmp);
799 * Block device opened successfully, finish initializing the
802 * From this point on we have to call hammer2_unmount() on failure.
804 pmp = hammer2_pfsalloc(ipdata, bref.mirror_tid);
805 kprintf("PMP mirror_tid is %016jx\n", bref.mirror_tid);
806 for (i = 0; i < cluster->nchains; ++i) {
807 rchain = cluster->array[i];
808 KKASSERT(rchain->pmp == NULL);
810 atomic_set_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
814 ccms_domain_init(&pmp->ccms_dom);
815 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
816 lockmgr(&hammer2_mntlk, LK_RELEASE);
818 kprintf("hammer2_mount hmp=%p pmp=%p pmpcnt=%d\n",
819 hmp, pmp, hmp->pmp_count);
821 mp->mnt_flag = MNT_LOCAL;
822 mp->mnt_kern_flag |= MNTK_ALL_MPSAFE; /* all entry pts are SMP */
823 mp->mnt_kern_flag |= MNTK_THR_SYNC; /* new vsyncscan semantics */
826 * required mount structure initializations
828 mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
829 mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
831 mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
832 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
837 mp->mnt_iosize_max = MAXPHYS;
838 mp->mnt_data = (qaddr_t)pmp;
842 * After this point hammer2_vfs_unmount() has visibility on hmp
843 * and manual hmp1/hmp2 calls are not needed on fatal errors.
845 pmp->iroot = hammer2_inode_get(pmp, NULL, cluster);
846 hammer2_inode_ref(pmp->iroot); /* ref for pmp->iroot */
847 hammer2_inode_unlock_ex(pmp->iroot, cluster);
850 * The logical file buffer bio write thread handles things
851 * like physical block assignment and compression.
853 * (only applicable to pfs mounts, not applicable to spmp)
855 pmp->wthread_destroy = 0;
856 lwkt_create(hammer2_write_thread, pmp,
857 &pmp->wthread_td, NULL, 0, -1, "hwrite-%s", label);
860 * With the cluster operational install ihidden.
861 * (only applicable to pfs mounts, not applicable to spmp)
863 hammer2_inode_install_hidden(pmp);
869 vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
870 vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
871 vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
873 copyinstr(info.volume, mp->mnt_stat.f_mntfromname, MNAMELEN - 1, &size);
874 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
875 bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
876 copyinstr(path, mp->mnt_stat.f_mntonname,
877 sizeof(mp->mnt_stat.f_mntonname) - 1,
881 * Initial statfs to prime mnt_stat.
883 hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
889 * Handle bioq for strategy write
893 hammer2_write_thread(void *arg)
895 hammer2_pfsmount_t *pmp;
898 hammer2_trans_t trans;
901 hammer2_cluster_t *cparent;
902 hammer2_inode_data_t *wipdata;
910 mtx_lock(&pmp->wthread_mtx);
911 while (pmp->wthread_destroy == 0) {
912 if (bioq_first(&pmp->wthread_bioq) == NULL) {
913 mtxsleep(&pmp->wthread_bioq, &pmp->wthread_mtx,
918 hammer2_trans_init(&trans, pmp, HAMMER2_TRANS_BUFCACHE);
920 while ((bio = bioq_takefirst(&pmp->wthread_bioq)) != NULL) {
922 * dummy bio for synchronization. The transaction
923 * must be reinitialized.
925 if (bio->bio_buf == NULL) {
926 bio->bio_flags |= BIO_DONE;
928 hammer2_trans_done(&trans);
929 hammer2_trans_init(&trans, pmp,
930 HAMMER2_TRANS_BUFCACHE);
935 * else normal bio processing
937 mtx_unlock(&pmp->wthread_mtx);
939 hammer2_lwinprog_drop(pmp);
947 * Inode is modified, flush size and mtime changes
948 * to ensure that the file size remains consistent
949 * with the buffers being flushed.
951 * NOTE: The inode_fsync() call only flushes the
952 * inode's meta-data state, it doesn't try
953 * to flush underlying buffers or chains.
955 cparent = hammer2_inode_lock_ex(ip);
956 if (ip->flags & (HAMMER2_INODE_RESIZED |
957 HAMMER2_INODE_MTIME)) {
958 hammer2_inode_fsync(&trans, ip, cparent);
960 wipdata = hammer2_cluster_modify_ip(&trans, ip,
962 lblksize = hammer2_calc_logical(ip, bio->bio_offset,
964 pblksize = hammer2_calc_physical(ip, wipdata, lbase);
965 hammer2_write_file_core(bp, &trans, ip, wipdata,
969 hammer2_cluster_modsync(cparent);
970 hammer2_inode_unlock_ex(ip, cparent);
972 kprintf("hammer2: error in buffer write\n");
973 bp->b_flags |= B_ERROR;
977 mtx_lock(&pmp->wthread_mtx);
979 hammer2_trans_done(&trans);
981 pmp->wthread_destroy = -1;
982 wakeup(&pmp->wthread_destroy);
984 mtx_unlock(&pmp->wthread_mtx);
988 hammer2_bioq_sync(hammer2_pfsmount_t *pmp)
992 bzero(&sync_bio, sizeof(sync_bio)); /* dummy with no bio_buf */
993 mtx_lock(&pmp->wthread_mtx);
994 if (pmp->wthread_destroy == 0 &&
995 TAILQ_FIRST(&pmp->wthread_bioq.queue)) {
996 bioq_insert_tail(&pmp->wthread_bioq, &sync_bio);
997 while ((sync_bio.bio_flags & BIO_DONE) == 0)
998 mtxsleep(&sync_bio, &pmp->wthread_mtx, 0, "h2bioq", 0);
1000 mtx_unlock(&pmp->wthread_mtx);
1004 * Return a chain suitable for I/O, creating the chain if necessary
1005 * and assigning its physical block.
1009 hammer2_assign_physical(hammer2_trans_t *trans,
1010 hammer2_inode_t *ip, hammer2_cluster_t *cparent,
1011 hammer2_key_t lbase, int pblksize, int *errorp)
1013 hammer2_cluster_t *cluster;
1014 hammer2_cluster_t *dparent;
1015 hammer2_key_t key_dummy;
1016 int pradix = hammer2_getradix(pblksize);
1020 * Locate the chain associated with lbase, return a locked chain.
1021 * However, do not instantiate any data reference (which utilizes a
1022 * device buffer) because we will be using direct IO via the
1023 * logical buffer cache buffer.
1026 KKASSERT(pblksize >= HAMMER2_ALLOC_MIN);
1028 dparent = hammer2_cluster_lookup_init(cparent, 0);
1029 cluster = hammer2_cluster_lookup(dparent, &key_dummy,
1031 HAMMER2_LOOKUP_NODATA, &ddflag);
1033 if (cluster == NULL) {
1035 * We found a hole, create a new chain entry.
1037 * NOTE: DATA chains are created without device backing
1038 * store (nor do we want any).
1040 *errorp = hammer2_cluster_create(trans, dparent, &cluster,
1041 lbase, HAMMER2_PBUFRADIX,
1042 HAMMER2_BREF_TYPE_DATA,
1044 if (cluster == NULL) {
1045 hammer2_cluster_lookup_done(dparent);
1046 panic("hammer2_cluster_create: par=%p error=%d\n",
1047 dparent->focus, *errorp);
1050 /*ip->delta_dcount += pblksize;*/
1052 switch (hammer2_cluster_type(cluster)) {
1053 case HAMMER2_BREF_TYPE_INODE:
1055 * The data is embedded in the inode. The
1056 * caller is responsible for marking the inode
1057 * modified and copying the data to the embedded
1061 case HAMMER2_BREF_TYPE_DATA:
1062 if (hammer2_cluster_bytes(cluster) != pblksize) {
1063 hammer2_cluster_resize(trans, ip,
1066 HAMMER2_MODIFY_OPTDATA);
1068 hammer2_cluster_modify(trans, cluster,
1069 HAMMER2_MODIFY_OPTDATA);
1072 panic("hammer2_assign_physical: bad type");
1079 * Cleanup. If cluster wound up being the inode itself, i.e.
1080 * the DIRECTDATA case for offset 0, then we need to update cparent.
1081 * The caller expects cparent to not become stale.
1083 hammer2_cluster_lookup_done(dparent);
1084 /* dparent = NULL; safety */
1085 if (cluster && ddflag)
1086 hammer2_cluster_replace_locked(cparent, cluster);
1091 * From hammer2_vnops.c.
1092 * The core write function which determines which path to take
1093 * depending on compression settings.
1097 hammer2_write_file_core(struct buf *bp, hammer2_trans_t *trans,
1098 hammer2_inode_t *ip, hammer2_inode_data_t *ipdata,
1099 hammer2_cluster_t *cparent,
1100 hammer2_key_t lbase, int ioflag, int pblksize,
1103 hammer2_cluster_t *cluster;
1105 switch(HAMMER2_DEC_COMP(ipdata->comp_algo)) {
1106 case HAMMER2_COMP_NONE:
1108 * We have to assign physical storage to the buffer
1109 * we intend to dirty or write now to avoid deadlocks
1110 * in the strategy code later.
1112 * This can return NOOFFSET for inode-embedded data.
1113 * The strategy code will take care of it in that case.
1115 cluster = hammer2_assign_physical(trans, ip, cparent,
1118 hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp);
1120 hammer2_cluster_unlock(cluster);
1122 case HAMMER2_COMP_AUTOZERO:
1124 * Check for zero-fill only
1126 hammer2_zero_check_and_write(bp, trans, ip,
1127 ipdata, cparent, lbase,
1128 ioflag, pblksize, errorp);
1130 case HAMMER2_COMP_LZ4:
1131 case HAMMER2_COMP_ZLIB:
1134 * Check for zero-fill and attempt compression.
1136 hammer2_compress_and_write(bp, trans, ip,
1146 * Generic function that will perform the compression in compression
1147 * write path. The compression algorithm is determined by the settings
1148 * obtained from inode.
1152 hammer2_compress_and_write(struct buf *bp, hammer2_trans_t *trans,
1153 hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1154 hammer2_cluster_t *cparent,
1155 hammer2_key_t lbase, int ioflag, int pblksize,
1156 int *errorp, int comp_algo)
1158 hammer2_cluster_t *cluster;
1159 hammer2_chain_t *chain;
1161 int comp_block_size;
1165 if (test_block_zeros(bp->b_data, pblksize)) {
1166 zero_write(bp, trans, ip, ipdata, cparent, lbase, errorp);
1173 KKASSERT(pblksize / 2 <= 32768);
1175 if (ip->comp_heuristic < 8 || (ip->comp_heuristic & 7) == 0) {
1176 z_stream strm_compress;
1180 switch(HAMMER2_DEC_COMP(comp_algo)) {
1181 case HAMMER2_COMP_LZ4:
1182 comp_buffer = objcache_get(cache_buffer_write,
1184 comp_size = LZ4_compress_limitedOutput(
1186 &comp_buffer[sizeof(int)],
1188 pblksize / 2 - sizeof(int));
1190 * We need to prefix with the size, LZ4
1191 * doesn't do it for us. Add the related
1194 *(int *)comp_buffer = comp_size;
1196 comp_size += sizeof(int);
1198 case HAMMER2_COMP_ZLIB:
1199 comp_level = HAMMER2_DEC_LEVEL(comp_algo);
1200 if (comp_level == 0)
1201 comp_level = 6; /* default zlib compression */
1202 else if (comp_level < 6)
1204 else if (comp_level > 9)
1206 ret = deflateInit(&strm_compress, comp_level);
1208 kprintf("HAMMER2 ZLIB: fatal error "
1209 "on deflateInit.\n");
1212 comp_buffer = objcache_get(cache_buffer_write,
1214 strm_compress.next_in = bp->b_data;
1215 strm_compress.avail_in = pblksize;
1216 strm_compress.next_out = comp_buffer;
1217 strm_compress.avail_out = pblksize / 2;
1218 ret = deflate(&strm_compress, Z_FINISH);
1219 if (ret == Z_STREAM_END) {
1220 comp_size = pblksize / 2 -
1221 strm_compress.avail_out;
1225 ret = deflateEnd(&strm_compress);
1228 kprintf("Error: Unknown compression method.\n");
1229 kprintf("Comp_method = %d.\n", comp_algo);
1234 if (comp_size == 0) {
1236 * compression failed or turned off
1238 comp_block_size = pblksize; /* safety */
1239 if (++ip->comp_heuristic > 128)
1240 ip->comp_heuristic = 8;
1243 * compression succeeded
1245 ip->comp_heuristic = 0;
1246 if (comp_size <= 1024) {
1247 comp_block_size = 1024;
1248 } else if (comp_size <= 2048) {
1249 comp_block_size = 2048;
1250 } else if (comp_size <= 4096) {
1251 comp_block_size = 4096;
1252 } else if (comp_size <= 8192) {
1253 comp_block_size = 8192;
1254 } else if (comp_size <= 16384) {
1255 comp_block_size = 16384;
1256 } else if (comp_size <= 32768) {
1257 comp_block_size = 32768;
1259 panic("hammer2: WRITE PATH: "
1260 "Weird comp_size value.");
1262 comp_block_size = pblksize;
1266 cluster = hammer2_assign_physical(trans, ip, cparent,
1267 lbase, comp_block_size,
1269 ipdata = &hammer2_cluster_data(cparent)->ipdata;
1272 kprintf("WRITE PATH: An error occurred while "
1273 "assigning physical space.\n");
1274 KKASSERT(cluster == NULL);
1278 for (i = 0; i < cluster->nchains; ++i) {
1283 chain = cluster->array[i];
1284 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1286 switch(chain->bref.type) {
1287 case HAMMER2_BREF_TYPE_INODE:
1288 KKASSERT(chain->data->ipdata.op_flags &
1289 HAMMER2_OPFLAG_DIRECTDATA);
1290 KKASSERT(bp->b_loffset == 0);
1291 bcopy(bp->b_data, chain->data->ipdata.u.data,
1292 HAMMER2_EMBEDDED_BYTES);
1294 case HAMMER2_BREF_TYPE_DATA:
1295 temp_check = HAMMER2_DEC_CHECK(chain->bref.methods);
1298 * Optimize out the read-before-write
1301 *errorp = hammer2_io_newnz(chain->hmp,
1302 chain->bref.data_off,
1306 hammer2_io_brelse(&dio);
1307 kprintf("hammer2: WRITE PATH: "
1308 "dbp bread error\n");
1311 bdata = hammer2_io_data(dio, chain->bref.data_off);
1314 * When loading the block make sure we don't
1315 * leave garbage after the compressed data.
1318 chain->bref.methods =
1319 HAMMER2_ENC_COMP(comp_algo) +
1320 HAMMER2_ENC_CHECK(temp_check);
1321 bcopy(comp_buffer, bdata, comp_size);
1322 if (comp_size != comp_block_size) {
1323 bzero(bdata + comp_size,
1324 comp_block_size - comp_size);
1327 chain->bref.methods =
1329 HAMMER2_COMP_NONE) +
1330 HAMMER2_ENC_CHECK(temp_check);
1331 bcopy(bp->b_data, bdata, pblksize);
1335 * Device buffer is now valid, chain is no
1336 * longer in the initial state.
1338 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1340 /* Now write the related bdp. */
1341 if (ioflag & IO_SYNC) {
1343 * Synchronous I/O requested.
1345 hammer2_io_bwrite(&dio);
1347 } else if ((ioflag & IO_DIRECT) &&
1348 loff + n == pblksize) {
1349 hammer2_io_bdwrite(&dio);
1351 } else if (ioflag & IO_ASYNC) {
1352 hammer2_io_bawrite(&dio);
1354 hammer2_io_bdwrite(&dio);
1358 panic("hammer2_write_bp: bad chain type %d\n",
1366 hammer2_cluster_unlock(cluster);
1368 objcache_put(cache_buffer_write, comp_buffer);
1372 * Function that performs zero-checking and writing without compression,
1373 * it corresponds to default zero-checking path.
1377 hammer2_zero_check_and_write(struct buf *bp, hammer2_trans_t *trans,
1378 hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1379 hammer2_cluster_t *cparent,
1380 hammer2_key_t lbase, int ioflag, int pblksize, int *errorp)
1382 hammer2_cluster_t *cluster;
1384 if (test_block_zeros(bp->b_data, pblksize)) {
1385 zero_write(bp, trans, ip, ipdata, cparent, lbase, errorp);
1387 cluster = hammer2_assign_physical(trans, ip, cparent,
1388 lbase, pblksize, errorp);
1389 hammer2_write_bp(cluster, bp, ioflag, pblksize, errorp);
1391 hammer2_cluster_unlock(cluster);
1396 * A function to test whether a block of data contains only zeros,
1397 * returns TRUE (non-zero) if the block is all zeros.
1401 test_block_zeros(const char *buf, size_t bytes)
1405 for (i = 0; i < bytes; i += sizeof(long)) {
1406 if (*(const long *)(buf + i) != 0)
1413 * Function to "write" a block that contains only zeros.
1417 zero_write(struct buf *bp, hammer2_trans_t *trans,
1418 hammer2_inode_t *ip, const hammer2_inode_data_t *ipdata,
1419 hammer2_cluster_t *cparent,
1420 hammer2_key_t lbase, int *errorp __unused)
1422 hammer2_cluster_t *cluster;
1423 hammer2_media_data_t *data;
1424 hammer2_key_t key_dummy;
1427 cparent = hammer2_cluster_lookup_init(cparent, 0);
1428 cluster = hammer2_cluster_lookup(cparent, &key_dummy, lbase, lbase,
1429 HAMMER2_LOOKUP_NODATA, &ddflag);
1431 data = hammer2_cluster_wdata(cluster);
1434 KKASSERT(cluster->focus->flags &
1435 HAMMER2_CHAIN_MODIFIED);
1436 bzero(data->ipdata.u.data, HAMMER2_EMBEDDED_BYTES);
1437 hammer2_cluster_modsync(cluster);
1439 hammer2_cluster_delete(trans, cluster, 0);
1441 hammer2_cluster_unlock(cluster);
1443 hammer2_cluster_lookup_done(cparent);
1447 * Function to write the data as it is, without performing any sort of
1448 * compression. This function is used in path without compression and
1449 * default zero-checking path.
1453 hammer2_write_bp(hammer2_cluster_t *cluster, struct buf *bp, int ioflag,
1454 int pblksize, int *errorp)
1456 hammer2_chain_t *chain;
1463 error = 0; /* XXX TODO below */
1465 for (i = 0; i < cluster->nchains; ++i) {
1466 chain = cluster->array[i];
1468 temp_check = HAMMER2_DEC_CHECK(chain->bref.methods);
1470 KKASSERT(chain->flags & HAMMER2_CHAIN_MODIFIED);
1472 switch(chain->bref.type) {
1473 case HAMMER2_BREF_TYPE_INODE:
1474 KKASSERT(chain->data->ipdata.op_flags &
1475 HAMMER2_OPFLAG_DIRECTDATA);
1476 KKASSERT(bp->b_loffset == 0);
1477 bcopy(bp->b_data, chain->data->ipdata.u.data,
1478 HAMMER2_EMBEDDED_BYTES);
1481 case HAMMER2_BREF_TYPE_DATA:
1482 error = hammer2_io_newnz(chain->hmp,
1483 chain->bref.data_off,
1484 chain->bytes, &dio);
1486 hammer2_io_bqrelse(&dio);
1487 kprintf("hammer2: WRITE PATH: "
1488 "dbp bread error\n");
1491 bdata = hammer2_io_data(dio, chain->bref.data_off);
1493 chain->bref.methods = HAMMER2_ENC_COMP(
1494 HAMMER2_COMP_NONE) +
1495 HAMMER2_ENC_CHECK(temp_check);
1496 bcopy(bp->b_data, bdata, chain->bytes);
1499 * Device buffer is now valid, chain is no
1500 * longer in the initial state.
1502 atomic_clear_int(&chain->flags, HAMMER2_CHAIN_INITIAL);
1504 if (ioflag & IO_SYNC) {
1506 * Synchronous I/O requested.
1508 hammer2_io_bwrite(&dio);
1510 } else if ((ioflag & IO_DIRECT) &&
1511 loff + n == pblksize) {
1512 hammer2_io_bdwrite(&dio);
1514 } else if (ioflag & IO_ASYNC) {
1515 hammer2_io_bawrite(&dio);
1517 hammer2_io_bdwrite(&dio);
1521 panic("hammer2_write_bp: bad chain type %d\n",
1527 KKASSERT(error == 0); /* XXX TODO */
1534 hammer2_remount(hammer2_mount_t *hmp, struct mount *mp, char *path,
1535 struct vnode *devvp, struct ucred *cred)
1539 if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1540 error = hammer2_recovery(hmp);
1549 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1551 hammer2_pfsmount_t *pmp;
1552 hammer2_mount_t *hmp;
1553 hammer2_chain_t *rchain;
1554 hammer2_cluster_t *cluster;
1564 lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1565 TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
1568 * If mount initialization proceeded far enough we must flush
1571 if (mntflags & MNT_FORCE)
1576 error = vflush(mp, 0, flags);
1581 ccms_domain_uninit(&pmp->ccms_dom);
1583 if (pmp->wthread_td) {
1584 mtx_lock(&pmp->wthread_mtx);
1585 pmp->wthread_destroy = 1;
1586 wakeup(&pmp->wthread_bioq);
1587 while (pmp->wthread_destroy != -1) {
1588 mtxsleep(&pmp->wthread_destroy,
1589 &pmp->wthread_mtx, 0,
1592 mtx_unlock(&pmp->wthread_mtx);
1593 pmp->wthread_td = NULL;
1597 * Cleanup our reference on ihidden.
1600 hammer2_inode_drop(pmp->ihidden);
1601 pmp->ihidden = NULL;
1605 * Cleanup our reference on iroot. iroot is (should) not be needed
1606 * by the flush code.
1609 cluster = &pmp->iroot->cluster;
1610 for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1611 rchain = pmp->iroot->cluster.array[i];
1615 hammer2_vfs_unmount_hmp1(mp, hmp);
1617 atomic_clear_int(&rchain->flags, HAMMER2_CHAIN_MOUNTED);
1618 #if REPORT_REFS_ERRORS
1619 if (rchain->refs != 1)
1620 kprintf("PMP->RCHAIN %p REFS WRONG %d\n",
1621 rchain, rchain->refs);
1623 KKASSERT(rchain->refs == 1);
1625 hammer2_chain_drop(rchain);
1626 cluster->array[i] = NULL;
1627 hammer2_vfs_unmount_hmp2(mp, hmp);
1629 cluster->focus = NULL;
1631 #if REPORT_REFS_ERRORS
1632 if (pmp->iroot->refs != 1)
1633 kprintf("PMP->IROOT %p REFS WRONG %d\n",
1634 pmp->iroot, pmp->iroot->refs);
1636 KKASSERT(pmp->iroot->refs == 1);
1638 /* ref for pmp->iroot */
1639 hammer2_inode_drop(pmp->iroot);
1644 mp->mnt_data = NULL;
1646 kmalloc_destroy(&pmp->mmsg);
1647 kmalloc_destroy(&pmp->minode);
1649 kfree(pmp, M_HAMMER2);
1653 lockmgr(&hammer2_mntlk, LK_RELEASE);
1660 hammer2_vfs_unmount_hmp1(struct mount *mp, hammer2_mount_t *hmp)
1662 hammer2_mount_exlock(hmp);
1665 kprintf("hammer2_unmount hmp=%p pmpcnt=%d\n", hmp, hmp->pmp_count);
1667 kdmsg_iocom_uninit(&hmp->iocom); /* XXX chain depend deadlck? */
1670 * Flush any left over chains. The voldata lock is only used
1671 * to synchronize against HAMMER2_CHAIN_MODIFIED_AUX.
1673 * Flush twice to ensure that the freemap is completely
1674 * synchronized. If we only do it once the next mount's
1675 * recovery scan will have to do some fixups (which isn't
1676 * bad, but we don't want it to have to do it except when
1677 * recovering from a crash).
1679 hammer2_voldata_lock(hmp);
1680 if (((hmp->vchain.flags | hmp->fchain.flags) &
1681 HAMMER2_CHAIN_MODIFIED) ||
1682 hmp->vchain.update_xhi > hmp->vchain.update_xlo ||
1683 hmp->fchain.update_xhi > hmp->fchain.update_xlo) {
1684 hammer2_voldata_unlock(hmp);
1685 hammer2_vfs_sync(mp, MNT_WAIT);
1686 /*hammer2_vfs_sync(mp, MNT_WAIT);*/
1688 hammer2_voldata_unlock(hmp);
1690 if (hmp->pmp_count == 0) {
1691 if (((hmp->vchain.flags | hmp->fchain.flags) &
1692 HAMMER2_CHAIN_MODIFIED) ||
1693 hmp->vchain.update_xhi > hmp->vchain.update_xlo ||
1694 hmp->fchain.update_xhi > hmp->fchain.update_xlo) {
1695 kprintf("hammer2_unmount: chains left over "
1696 "after final sync\n");
1697 kprintf(" vchain %08x update_xlo/hi %08x/%08x\n",
1699 hmp->vchain.update_xlo,
1700 hmp->vchain.update_xhi);
1701 kprintf(" fchain %08x update_xhi/hi %08x/%08x\n",
1703 hmp->fchain.update_xlo,
1704 hmp->fchain.update_xhi);
1706 if (hammer2_debug & 0x0010)
1707 Debugger("entered debugger");
1714 hammer2_vfs_unmount_hmp2(struct mount *mp, hammer2_mount_t *hmp)
1716 hammer2_pfsmount_t *spmp;
1717 struct vnode *devvp;
1719 int ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1722 * If no PFS's left drop the master hammer2_mount for the
1725 if (hmp->pmp_count == 0) {
1727 * Clean up SPMP and the super-root inode
1732 hammer2_inode_drop(spmp->iroot);
1736 kmalloc_destroy(&spmp->mmsg);
1737 kmalloc_destroy(&spmp->minode);
1738 kfree(spmp, M_HAMMER2);
1742 * Finish up with the device vnode
1744 if ((devvp = hmp->devvp) != NULL) {
1745 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1746 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1748 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1755 * Clear vchain/fchain flags that might prevent final cleanup
1758 if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1759 atomic_clear_int(&hmp->vchain.flags,
1760 HAMMER2_CHAIN_MODIFIED);
1761 hammer2_chain_drop(&hmp->vchain);
1763 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_CREATE) {
1764 atomic_clear_int(&hmp->vchain.flags,
1765 HAMMER2_CHAIN_FLUSH_CREATE);
1766 hammer2_chain_drop(&hmp->vchain);
1768 if (hmp->vchain.flags & HAMMER2_CHAIN_FLUSH_DELETE) {
1769 atomic_clear_int(&hmp->vchain.flags,
1770 HAMMER2_CHAIN_FLUSH_DELETE);
1771 hammer2_chain_drop(&hmp->vchain);
1774 if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1775 atomic_clear_int(&hmp->fchain.flags,
1776 HAMMER2_CHAIN_MODIFIED);
1777 hammer2_chain_drop(&hmp->fchain);
1779 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_CREATE) {
1780 atomic_clear_int(&hmp->fchain.flags,
1781 HAMMER2_CHAIN_FLUSH_CREATE);
1782 hammer2_chain_drop(&hmp->fchain);
1784 if (hmp->fchain.flags & HAMMER2_CHAIN_FLUSH_DELETE) {
1785 atomic_clear_int(&hmp->fchain.flags,
1786 HAMMER2_CHAIN_FLUSH_DELETE);
1787 hammer2_chain_drop(&hmp->fchain);
1791 * Final drop of embedded freemap root chain to
1792 * clean up fchain.core (fchain structure is not
1793 * flagged ALLOCATED so it is cleaned out and then
1796 hammer2_chain_drop(&hmp->fchain);
1799 * Final drop of embedded volume root chain to clean
1800 * up vchain.core (vchain structure is not flagged
1801 * ALLOCATED so it is cleaned out and then left to
1805 hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1807 hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1808 hammer2_mount_unlock(hmp);
1809 hammer2_chain_drop(&hmp->vchain);
1811 hammer2_io_cleanup(hmp, &hmp->iotree);
1812 if (hmp->iofree_count) {
1813 kprintf("io_cleanup: %d I/O's left hanging\n",
1817 TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1818 kmalloc_destroy(&hmp->mchain);
1819 kfree(hmp, M_HAMMER2);
1821 hammer2_mount_unlock(hmp);
1827 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1828 ino_t ino, struct vnode **vpp)
1830 kprintf("hammer2_vget\n");
1831 return (EOPNOTSUPP);
1836 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1838 hammer2_pfsmount_t *pmp;
1839 hammer2_cluster_t *cparent;
1844 if (pmp->iroot == NULL) {
1848 cparent = hammer2_inode_lock_sh(pmp->iroot);
1849 vp = hammer2_igetv(pmp->iroot, cparent, &error);
1850 hammer2_inode_unlock_sh(pmp->iroot, cparent);
1853 kprintf("vnodefail\n");
1862 * XXX incorporate ipdata->inode_quota and data_quota
1866 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1868 hammer2_pfsmount_t *pmp;
1869 hammer2_mount_t *hmp;
1872 KKASSERT(pmp->iroot->cluster.nchains >= 1);
1873 hmp = pmp->iroot->cluster.focus->hmp; /* XXX */
1875 mp->mnt_stat.f_files = pmp->inode_count;
1876 mp->mnt_stat.f_ffree = 0;
1877 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1878 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1879 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1881 *sbp = mp->mnt_stat;
1887 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1889 hammer2_pfsmount_t *pmp;
1890 hammer2_mount_t *hmp;
1893 KKASSERT(pmp->iroot->cluster.nchains >= 1);
1894 hmp = pmp->iroot->cluster.focus->hmp; /* XXX */
1896 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1897 mp->mnt_vstat.f_files = pmp->inode_count;
1898 mp->mnt_vstat.f_ffree = 0;
1899 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size / HAMMER2_PBUFSIZE;
1900 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free / HAMMER2_PBUFSIZE;
1901 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1903 *sbp = mp->mnt_vstat;
1908 * Mount-time recovery (RW mounts)
1910 * Updates to the free block table are allowed to lag flushes by one
1911 * transaction. In case of a crash, then on a fresh mount we must do an
1912 * incremental scan of the last committed transaction id and make sure that
1913 * all related blocks have been marked allocated.
1915 * The super-root topology and each PFS has its own transaction id domain,
1916 * so we must track PFS boundary transitions.
1918 struct hammer2_recovery_elm {
1919 TAILQ_ENTRY(hammer2_recovery_elm) entry;
1920 hammer2_chain_t *chain;
1921 hammer2_tid_t sync_tid;
1924 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1926 struct hammer2_recovery_info {
1927 struct hammer2_recovery_list list;
1931 static int hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1932 hammer2_chain_t *parent,
1933 struct hammer2_recovery_info *info,
1934 hammer2_tid_t sync_tid);
1936 #define HAMMER2_RECOVERY_MAXDEPTH 10
1940 hammer2_recovery(hammer2_mount_t *hmp)
1942 hammer2_trans_t trans;
1943 struct hammer2_recovery_info info;
1944 struct hammer2_recovery_elm *elm;
1945 hammer2_chain_t *parent;
1946 hammer2_tid_t sync_tid;
1948 int cumulative_error = 0;
1950 hammer2_trans_init(&trans, hmp->spmp, 0);
1953 TAILQ_INIT(&info.list);
1955 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1956 cumulative_error = hammer2_recovery_scan(&trans, hmp, parent,
1958 hammer2_chain_lookup_done(parent);
1960 while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1961 TAILQ_REMOVE(&info.list, elm, entry);
1962 parent = elm->chain;
1963 sync_tid = elm->sync_tid;
1964 kfree(elm, M_HAMMER2);
1966 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS |
1967 HAMMER2_RESOLVE_NOREF);
1968 error = hammer2_recovery_scan(&trans, hmp, parent,
1970 hammer2_chain_unlock(parent);
1972 cumulative_error = error;
1974 hammer2_trans_done(&trans);
1976 return cumulative_error;
1981 hammer2_recovery_scan(hammer2_trans_t *trans, hammer2_mount_t *hmp,
1982 hammer2_chain_t *parent,
1983 struct hammer2_recovery_info *info,
1984 hammer2_tid_t sync_tid)
1986 hammer2_chain_t *chain;
1988 int cumulative_error = 0;
1989 int pfs_boundary = 0;
1993 * Adjust freemap to ensure that the block(s) are marked allocated.
1995 if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
1996 hammer2_freemap_adjust(trans, hmp, &parent->bref,
1997 HAMMER2_FREEMAP_DORECOVER);
2001 * Check type for recursive scan
2003 switch(parent->bref.type) {
2004 case HAMMER2_BREF_TYPE_VOLUME:
2005 /* data already instantiated */
2007 case HAMMER2_BREF_TYPE_INODE:
2009 * Must instantiate data for DIRECTDATA test and also
2012 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2013 if (parent->data->ipdata.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2014 /* not applicable to recovery scan */
2015 hammer2_chain_unlock(parent);
2018 if ((parent->data->ipdata.op_flags & HAMMER2_OPFLAG_PFSROOT) &&
2021 sync_tid = parent->bref.mirror_tid - 1;
2023 hammer2_chain_unlock(parent);
2025 case HAMMER2_BREF_TYPE_INDIRECT:
2027 * Must instantiate data for recursion
2029 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2030 hammer2_chain_unlock(parent);
2032 case HAMMER2_BREF_TYPE_DATA:
2033 case HAMMER2_BREF_TYPE_FREEMAP:
2034 case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2035 case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2036 /* not applicable to recovery scan */
2044 * Defer operation if depth limit reached or if we are crossing a
2047 if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH || pfs_boundary) {
2048 struct hammer2_recovery_elm *elm;
2050 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2051 elm->chain = parent;
2052 elm->sync_tid = sync_tid;
2053 hammer2_chain_ref(parent);
2054 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2055 /* unlocked by caller */
2062 * Recursive scan of the last flushed transaction only. We are
2063 * doing this without pmp assignments so don't leave the chains
2064 * hanging around after we are done with them.
2067 chain = hammer2_chain_scan(parent, NULL, &cache_index,
2068 HAMMER2_LOOKUP_NODATA);
2070 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2071 if (chain->bref.mirror_tid >= sync_tid) {
2073 error = hammer2_recovery_scan(trans, hmp, chain,
2077 cumulative_error = error;
2079 chain = hammer2_chain_scan(parent, chain, &cache_index,
2080 HAMMER2_LOOKUP_NODATA);
2083 return cumulative_error;
2087 * Sync the entire filesystem; this is called from the filesystem syncer
2088 * process periodically and whenever a user calls sync(1) on the hammer
2091 * Currently is actually called from the syncer! \o/
2093 * This task will have to snapshot the state of the dirty inode chain.
2094 * From that, it will have to make sure all of the inodes on the dirty
2095 * chain have IO initiated. We make sure that io is initiated for the root
2098 * If waitfor is set, we wait for media to acknowledge the new rootblock.
2100 * THINKS: side A vs side B, to have sync not stall all I/O?
2103 hammer2_vfs_sync(struct mount *mp, int waitfor)
2105 struct hammer2_sync_info info;
2106 hammer2_inode_t *iroot;
2107 hammer2_chain_t *chain;
2108 hammer2_chain_t *parent;
2109 hammer2_pfsmount_t *pmp;
2110 hammer2_mount_t *hmp;
2121 KKASSERT(iroot->pmp == pmp);
2124 * We can't acquire locks on existing vnodes while in a transaction
2125 * without risking a deadlock. This assumes that vfsync() can be
2126 * called without the vnode locked (which it can in DragonFly).
2127 * Otherwise we'd have to implement a multi-pass or flag the lock
2128 * failures and retry.
2130 * The reclamation code interlocks with the sync list's token
2131 * (by removing the vnode from the scan list) before unlocking
2132 * the inode, giving us time to ref the inode.
2134 /*flags = VMSC_GETVP;*/
2136 if (waitfor & MNT_LAZY)
2137 flags |= VMSC_ONEPASS;
2140 * Start our flush transaction. This does not return until all
2141 * concurrent transactions have completed and will prevent any
2142 * new transactions from running concurrently, except for the
2143 * buffer cache transactions.
2145 * For efficiency do an async pass before making sure with a
2146 * synchronous pass on all related buffer cache buffers. It
2147 * should theoretically not be possible for any new file buffers
2148 * to be instantiated during this sequence.
2150 hammer2_trans_init(&info.trans, pmp, HAMMER2_TRANS_ISFLUSH |
2151 HAMMER2_TRANS_PREFLUSH);
2152 hammer2_run_unlinkq(&info.trans, pmp);
2155 info.waitfor = MNT_NOWAIT;
2156 vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2157 info.waitfor = MNT_WAIT;
2158 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2161 * Clear PREFLUSH. This prevents (or asserts on) any new logical
2162 * buffer cache flushes which occur during the flush. Device buffers
2167 if (info.error == 0 && (waitfor & MNT_WAIT)) {
2168 info.waitfor = waitfor;
2169 vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2173 hammer2_bioq_sync(info.trans.pmp);
2174 atomic_clear_int(&info.trans.flags, HAMMER2_TRANS_PREFLUSH);
2179 * Flush all storage elements making up the cluster
2181 * We must also flush any deleted siblings because the super-root
2182 * flush won't do it for us. They all must be staged or the
2183 * super-root flush will not be able to update its block table
2186 * XXX currently done serially instead of concurrently
2188 for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2189 chain = iroot->cluster.array[i];
2191 hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
2192 hammer2_flush(&info.trans, &chain);
2193 hammer2_chain_unlock(chain);
2196 hammer2_chain_t *nchain;
2197 chain = TAILQ_FIRST(&chain->core->ownerq);
2198 hammer2_chain_ref(chain);
2200 hammer2_chain_lock(chain,
2201 HAMMER2_RESOLVE_ALWAYS);
2202 hammer2_flush(&info.trans, &chain);
2203 hammer2_chain_unlock(chain);
2204 nchain = TAILQ_NEXT(chain, core_entry);
2206 hammer2_chain_ref(nchain);
2207 hammer2_chain_drop(chain);
2213 hammer2_trans_done(&info.trans);
2217 * Flush all volume roots to synchronize PFS flushes with the
2218 * storage media. Use a super-root transaction for each one.
2220 * The flush code will detect super-root -> pfs-root chain
2221 * transitions using the last pfs-root flush.
2223 for (i = 0; iroot && i < iroot->cluster.nchains; ++i) {
2224 chain = iroot->cluster.array[i];
2231 * We only have to flush each hmp once
2233 for (j = i - 1; j >= 0; --j) {
2234 if (iroot->cluster.array[j] &&
2235 iroot->cluster.array[j]->hmp == hmp)
2240 hammer2_trans_spmp(&info.trans, hmp->spmp);
2243 * Force an update of the XID from the PFS root to the
2244 * topology root. We couldn't do this from the PFS
2245 * transaction because a SPMP transaction is needed.
2246 * This does not modify blocks, instead what it does is
2247 * allow the flush code to find the transition point and
2248 * then update on the way back up.
2250 parent = TAILQ_LAST(&chain->above->ownerq, h2_core_list);
2251 KKASSERT(chain->pmp != parent->pmp);
2252 hammer2_chain_setsubmod(&info.trans, parent);
2255 * Media mounts have two 'roots', vchain for the topology
2256 * and fchain for the free block table. Flush both.
2258 * Note that the topology and free block table are handled
2259 * independently, so the free block table can wind up being
2260 * ahead of the topology. We depend on the bulk free scan
2261 * code to deal with any loose ends.
2263 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2264 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2265 if ((hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) ||
2266 hmp->fchain.update_xhi > hmp->fchain.update_xlo) {
2268 * This will also modify vchain as a side effect,
2269 * mark vchain as modified now.
2271 hammer2_voldata_modify(hmp);
2272 chain = &hmp->fchain;
2273 hammer2_flush(&info.trans, &chain);
2274 KKASSERT(chain == &hmp->fchain);
2276 hammer2_chain_unlock(&hmp->fchain);
2277 hammer2_chain_unlock(&hmp->vchain);
2279 hammer2_chain_lock(&hmp->vchain, HAMMER2_RESOLVE_ALWAYS);
2280 if ((hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) ||
2281 hmp->vchain.update_xhi > hmp->vchain.update_xlo) {
2282 chain = &hmp->vchain;
2283 hammer2_flush(&info.trans, &chain);
2284 KKASSERT(chain == &hmp->vchain);
2289 hammer2_chain_unlock(&hmp->vchain);
2292 hammer2_chain_lock(&hmp->fchain, HAMMER2_RESOLVE_ALWAYS);
2293 if ((hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) ||
2294 hmp->fchain.update_xhi > hmp->fchain.update_xlo ||
2296 /* this will also modify vchain as a side effect */
2297 chain = &hmp->fchain;
2298 hammer2_flush(&info.trans, &chain);
2299 KKASSERT(chain == &hmp->fchain);
2301 hammer2_chain_unlock(&hmp->fchain);
2307 * We can't safely flush the volume header until we have
2308 * flushed any device buffers which have built up.
2310 * XXX this isn't being incremental
2312 vn_lock(hmp->devvp, LK_EXCLUSIVE | LK_RETRY);
2313 error = VOP_FSYNC(hmp->devvp, MNT_WAIT, 0);
2314 vn_unlock(hmp->devvp);
2317 * The flush code sets CHAIN_VOLUMESYNC to indicate that the
2318 * volume header needs synchronization via hmp->volsync.
2320 * XXX synchronize the flag & data with only this flush XXX
2323 (hmp->vchain.flags & HAMMER2_CHAIN_VOLUMESYNC)) {
2327 * Synchronize the disk before flushing the volume
2331 bp->b_bio1.bio_offset = 0;
2334 bp->b_cmd = BUF_CMD_FLUSH;
2335 bp->b_bio1.bio_done = biodone_sync;
2336 bp->b_bio1.bio_flags |= BIO_SYNC;
2337 vn_strategy(hmp->devvp, &bp->b_bio1);
2338 biowait(&bp->b_bio1, "h2vol");
2342 * Then we can safely flush the version of the
2343 * volume header synchronized by the flush code.
2345 i = hmp->volhdrno + 1;
2346 if (i >= HAMMER2_NUM_VOLHDRS)
2348 if (i * HAMMER2_ZONE_BYTES64 + HAMMER2_SEGSIZE >
2349 hmp->volsync.volu_size) {
2352 kprintf("sync volhdr %d %jd\n",
2353 i, (intmax_t)hmp->volsync.volu_size);
2354 bp = getblk(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2355 HAMMER2_PBUFSIZE, 0, 0);
2356 atomic_clear_int(&hmp->vchain.flags,
2357 HAMMER2_CHAIN_VOLUMESYNC);
2358 bcopy(&hmp->volsync, bp->b_data, HAMMER2_PBUFSIZE);
2363 total_error = error;
2366 hammer2_trans_done(&info.trans);
2369 hammer2_trans_done(&info.trans);
2371 return (total_error);
2378 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2380 struct hammer2_sync_info *info = data;
2381 hammer2_inode_t *ip;
2390 if (vp->v_type == VNON || vp->v_type == VBAD) {
2394 if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2395 RB_EMPTY(&vp->v_rbdirty_tree)) {
2401 * VOP_FSYNC will start a new transaction so replicate some code
2402 * here to do it inline (see hammer2_vop_fsync()).
2404 * WARNING: The vfsync interacts with the buffer cache and might
2405 * block, we can't hold the inode lock at that time.
2406 * However, we MUST ref ip before blocking to ensure that
2407 * it isn't ripped out from under us (since we do not
2408 * hold a lock on the vnode).
2410 hammer2_inode_ref(ip);
2411 atomic_clear_int(&ip->flags, HAMMER2_INODE_MODIFIED);
2413 vfsync(vp, MNT_NOWAIT, 1, NULL, NULL);
2415 hammer2_inode_drop(ip);
2419 info->error = error;
2426 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2433 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2434 struct fid *fhp, struct vnode **vpp)
2441 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2442 int *exflagsp, struct ucred **credanonp)
2448 * Support code for hammer2_mount(). Read, verify, and install the volume
2449 * header into the HMP
2451 * XXX read four volhdrs and use the one with the highest TID whos CRC
2456 * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2457 * nonexistant locations.
2459 * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2463 hammer2_install_volume_header(hammer2_mount_t *hmp)
2465 hammer2_volume_data_t *vd;
2467 hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2479 * There are up to 4 copies of the volume header (syncs iterate
2480 * between them so there is no single master). We don't trust the
2481 * volu_size field so we don't know precisely how large the filesystem
2482 * is, so depend on the OS to return an error if we go beyond the
2483 * block device's EOF.
2485 for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2486 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2487 HAMMER2_VOLUME_BYTES, &bp);
2494 vd = (struct hammer2_volume_data *) bp->b_data;
2495 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2496 (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2502 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2503 /* XXX: Reversed-endianness filesystem */
2504 kprintf("hammer2: reverse-endian filesystem detected");
2510 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2511 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2512 HAMMER2_VOLUME_ICRC0_SIZE);
2513 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2514 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2515 HAMMER2_VOLUME_ICRC1_SIZE);
2516 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2517 kprintf("hammer2 volume header crc "
2518 "mismatch copy #%d %08x/%08x\n",
2525 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2534 hmp->volsync = hmp->voldata;
2536 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2537 kprintf("hammer2: using volume header #%d\n",
2542 kprintf("hammer2: no valid volume headers found!\n");
2548 * Reconnect using the passed file pointer. The caller must ref the
2552 hammer2_cluster_reconnect(hammer2_mount_t *hmp, struct file *fp)
2555 const char *name = "disk-volume";
2558 * Closes old comm descriptor, kills threads, cleans up
2559 * states, then installs the new descriptor and creates
2562 kdmsg_iocom_reconnect(&hmp->iocom, fp, "hammer2");
2565 * Setup LNK_CONN fields for autoinitiated state machine. We
2566 * will use SPANs to advertise multiple PFSs so only pass the
2567 * fsid and HAMMER2_PFSTYPE_SUPROOT for the AUTOCONN.
2569 * We are not initiating a LNK_SPAN so we do not have to set-up
2570 * iocom.auto_lnk_span.
2572 bzero(&hmp->iocom.auto_lnk_conn.pfs_clid,
2573 sizeof(hmp->iocom.auto_lnk_conn.pfs_clid));
2574 hmp->iocom.auto_lnk_conn.pfs_fsid = hmp->voldata.fsid;
2575 hmp->iocom.auto_lnk_conn.pfs_type = HAMMER2_PFSTYPE_SUPROOT;
2576 hmp->iocom.auto_lnk_conn.proto_version = DMSG_SPAN_PROTO_1;
2578 hmp->iocom.auto_lnk_conn.peer_type = hmp->voldata.peer_type;
2580 hmp->iocom.auto_lnk_conn.peer_type = DMSG_PEER_HAMMER2;
2583 * Filter adjustment. Clients do not need visibility into other
2584 * clients (otherwise millions of clients would present a serious
2585 * problem). The fs_label also serves to restrict the namespace.
2587 hmp->iocom.auto_lnk_conn.peer_mask = 1LLU << DMSG_PEER_HAMMER2;
2588 hmp->iocom.auto_lnk_conn.pfs_mask = (uint64_t)-1;
2591 switch (ipdata->pfs_type) {
2592 case DMSG_PFSTYPE_CLIENT:
2593 hmp->iocom.auto_lnk_conn.peer_mask &=
2594 ~(1LLU << DMSG_PFSTYPE_CLIENT);
2601 name_len = strlen(name);
2602 if (name_len >= sizeof(hmp->iocom.auto_lnk_conn.fs_label))
2603 name_len = sizeof(hmp->iocom.auto_lnk_conn.fs_label) - 1;
2604 bcopy(name, hmp->iocom.auto_lnk_conn.fs_label, name_len);
2605 hmp->iocom.auto_lnk_conn.fs_label[name_len] = 0;
2607 kdmsg_iocom_autoinitiate(&hmp->iocom, hammer2_autodmsg);
2611 hammer2_rcvdmsg(kdmsg_msg_t *msg)
2613 kprintf("RCVMSG %08x\n", msg->tcmd);
2616 case DMSG_DBG_SHELL:
2619 * Execute shell command (not supported atm)
2621 kdmsg_msg_result(msg, DMSG_ERR_NOSUPP);
2623 case DMSG_DBG_SHELL | DMSGF_REPLY:
2627 if (msg->aux_data) {
2628 msg->aux_data[msg->aux_size - 1] = 0;
2629 kprintf("HAMMER2 DBG: %s\n", msg->aux_data);
2634 * Unsupported message received. We only need to
2635 * reply if it's a transaction in order to close our end.
2636 * Ignore any one-way messages or any further messages
2637 * associated with the transaction.
2639 * NOTE: This case also includes DMSG_LNK_ERROR messages
2640 * which might be one-way, replying to those would
2641 * cause an infinite ping-pong.
2643 if (msg->any.head.cmd & DMSGF_CREATE)
2644 kdmsg_msg_reply(msg, DMSG_ERR_NOSUPP);
2651 * This function is called after KDMSG has automatically handled processing
2652 * of a LNK layer message (typically CONN, SPAN, or CIRC).
2654 * We tag off the LNK_CONN to trigger our LNK_VOLCONF messages which
2655 * advertises all available hammer2 super-root volumes.
2658 hammer2_autodmsg(kdmsg_msg_t *msg)
2660 hammer2_mount_t *hmp = msg->iocom->handle;
2663 kprintf("RCAMSG %08x\n", msg->tcmd);
2666 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_REPLY:
2667 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_DELETE | DMSGF_REPLY:
2668 if (msg->any.head.cmd & DMSGF_CREATE) {
2669 kprintf("HAMMER2: VOLDATA DUMP\n");
2672 * Dump the configuration stored in the volume header.
2673 * This will typically be import/export access rights,
2674 * master encryption keys (encrypted), etc.
2676 hammer2_voldata_lock(hmp);
2678 while (copyid < HAMMER2_COPYID_COUNT) {
2679 if (hmp->voldata.copyinfo[copyid].copyid)
2680 hammer2_volconf_update(hmp, copyid);
2683 hammer2_voldata_unlock(hmp);
2685 kprintf("HAMMER2: INITIATE SPANs\n");
2686 hammer2_update_spans(hmp);
2688 if ((msg->any.head.cmd & DMSGF_DELETE) &&
2689 msg->state && (msg->state->txcmd & DMSGF_DELETE) == 0) {
2690 kprintf("HAMMER2: CONN WAS TERMINATED\n");
2699 * Update LNK_SPAN state
2702 hammer2_update_spans(hammer2_mount_t *hmp)
2704 const hammer2_inode_data_t *ipdata;
2705 hammer2_cluster_t *cparent;
2706 hammer2_cluster_t *cluster;
2707 hammer2_pfsmount_t *spmp;
2708 hammer2_key_t key_next;
2714 * Lookup mount point under the media-localized super-root.
2716 * cluster->pmp will incorrectly point to spmp and must be fixed
2720 cparent = hammer2_inode_lock_ex(spmp->iroot);
2721 cluster = hammer2_cluster_lookup(cparent, &key_next,
2726 if (hammer2_cluster_type(cluster) != HAMMER2_BREF_TYPE_INODE)
2728 ipdata = &hammer2_cluster_data(cluster)->ipdata;
2729 kprintf("UPDATE SPANS: %s\n", ipdata->filename);
2731 rmsg = kdmsg_msg_alloc(&hmp->iocom, NULL,
2732 DMSG_LNK_SPAN | DMSGF_CREATE,
2733 hammer2_lnk_span_reply, NULL);
2734 rmsg->any.lnk_span.pfs_clid = ipdata->pfs_clid;
2735 rmsg->any.lnk_span.pfs_fsid = ipdata->pfs_fsid;
2736 rmsg->any.lnk_span.pfs_type = ipdata->pfs_type;
2737 rmsg->any.lnk_span.peer_type = DMSG_PEER_HAMMER2;
2738 rmsg->any.lnk_span.proto_version = DMSG_SPAN_PROTO_1;
2739 name_len = ipdata->name_len;
2740 if (name_len >= sizeof(rmsg->any.lnk_span.fs_label))
2741 name_len = sizeof(rmsg->any.lnk_span.fs_label) - 1;
2742 bcopy(ipdata->filename, rmsg->any.lnk_span.fs_label, name_len);
2744 kdmsg_msg_write(rmsg);
2746 cluster = hammer2_cluster_next(cparent, cluster,
2752 hammer2_inode_unlock_ex(spmp->iroot, cparent);
2757 hammer2_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg)
2759 if ((state->txcmd & DMSGF_DELETE) == 0 &&
2760 (msg->any.head.cmd & DMSGF_DELETE)) {
2761 kdmsg_msg_reply(msg, 0);
2767 * Volume configuration updates are passed onto the userland service
2768 * daemon via the open LNK_CONN transaction.
2771 hammer2_volconf_update(hammer2_mount_t *hmp, int index)
2775 /* XXX interlock against connection state termination */
2776 kprintf("volconf update %p\n", hmp->iocom.conn_state);
2777 if (hmp->iocom.conn_state) {
2778 kprintf("TRANSMIT VOLCONF VIA OPEN CONN TRANSACTION\n");
2779 msg = kdmsg_msg_alloc_state(hmp->iocom.conn_state,
2780 DMSG_LNK_HAMMER2_VOLCONF,
2782 H2_LNK_VOLCONF(msg)->copy = hmp->voldata.copyinfo[index];
2783 H2_LNK_VOLCONF(msg)->mediaid = hmp->voldata.fsid;
2784 H2_LNK_VOLCONF(msg)->index = index;
2785 kdmsg_msg_write(msg);
2790 * This handles hysteresis on regular file flushes. Because the BIOs are
2791 * routed to a thread it is possible for an excessive number to build up
2792 * and cause long front-end stalls long before the runningbuffspace limit
2793 * is hit, so we implement hammer2_flush_pipe to control the
2796 * This is a particular problem when compression is used.
2799 hammer2_lwinprog_ref(hammer2_pfsmount_t *pmp)
2801 atomic_add_int(&pmp->count_lwinprog, 1);
2805 hammer2_lwinprog_drop(hammer2_pfsmount_t *pmp)
2809 lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2810 if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2811 (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2812 atomic_clear_int(&pmp->count_lwinprog,
2813 HAMMER2_LWINPROG_WAITING);
2814 wakeup(&pmp->count_lwinprog);
2819 hammer2_lwinprog_wait(hammer2_pfsmount_t *pmp)
2824 lwinprog = pmp->count_lwinprog;
2826 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2828 tsleep_interlock(&pmp->count_lwinprog, 0);
2829 atomic_set_int(&pmp->count_lwinprog, HAMMER2_LWINPROG_WAITING);
2830 lwinprog = pmp->count_lwinprog;
2831 if ((lwinprog & HAMMER2_LWINPROG_MASK) < hammer2_flush_pipe)
2833 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2838 * Manage excessive memory resource use for chain and related
2842 hammer2_pfs_memory_wait(hammer2_pfsmount_t *pmp)
2852 * Atomic check condition and wait. Also do an early speedup of
2853 * the syncer to try to avoid hitting the wait.
2856 waiting = pmp->inmem_dirty_chains;
2858 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2860 limit = pmp->mp->mnt_nvnodelistsize / 10;
2861 if (limit < hammer2_limit_dirty_chains)
2862 limit = hammer2_limit_dirty_chains;
2867 if ((int)(ticks - zzticks) > hz) {
2869 kprintf("count %ld %ld\n", count, limit);
2874 * Block if there are too many dirty chains present, wait
2875 * for the flush to clean some out.
2877 if (count > limit) {
2878 tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2879 if (atomic_cmpset_long(&pmp->inmem_dirty_chains,
2881 waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2882 speedup_syncer(pmp->mp);
2883 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2886 continue; /* loop on success or fail */
2890 * Try to start an early flush before we are forced to block.
2892 if (count > limit * 7 / 10)
2893 speedup_syncer(pmp->mp);
2899 hammer2_pfs_memory_inc(hammer2_pfsmount_t *pmp)
2902 atomic_add_long(&pmp->inmem_dirty_chains, 1);
2906 hammer2_pfs_memory_wakeup(hammer2_pfsmount_t *pmp)
2914 waiting = pmp->inmem_dirty_chains;
2916 if (atomic_cmpset_long(&pmp->inmem_dirty_chains,
2919 ~HAMMER2_DIRTYCHAIN_WAITING)) {
2924 if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2925 wakeup(&pmp->inmem_dirty_chains);
2932 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2934 hammer2_chain_t *scan;
2935 hammer2_chain_t *first_parent;
2939 kprintf("%*.*s...\n", tab, tab, "");
2944 first_parent = chain->core ? TAILQ_FIRST(&chain->core->ownerq) : NULL;
2945 kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2947 chain, chain->bref.type,
2948 chain->bref.key, chain->bref.keybits,
2949 chain->bref.mirror_tid);
2951 kprintf("%*.*s [%08x] (%s) mod=%08x del=%08x "
2952 "lo=%08x hi=%08x refs=%d\n",
2955 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2956 chain->data) ? (char *)chain->data->ipdata.filename : "?"),
2963 kprintf("%*.*s core %p [%08x]",
2965 chain->core, (chain->core ? chain->core->flags : 0));
2968 kprintf("\n%*.*s fp=%p np=%p [fpflags %08x fprefs %d",
2971 (first_parent ? TAILQ_NEXT(first_parent, core_entry) :
2973 first_parent->flags,
2974 first_parent->refs);
2975 if (chain->core == NULL || RB_EMPTY(&chain->core->rbtree))
2980 RB_FOREACH(scan, hammer2_chain_tree, &chain->core->rbtree)
2981 hammer2_dump_chain(scan, tab + 4, countp, 'a');
2982 RB_FOREACH(scan, hammer2_chain_tree, &chain->core->dbtree)
2983 hammer2_dump_chain(scan, tab + 4, countp, 'r');
2984 TAILQ_FOREACH(scan, &chain->core->dbq, db_entry)
2985 hammer2_dump_chain(scan, tab + 4, countp, 'd');
2987 if (chain->core && !RB_EMPTY(&chain->core->rbtree)) {
2988 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2989 kprintf("%*.*s}(%s)\n", tab, tab, "",
2990 chain->data->ipdata.filename);
2992 kprintf("%*.*s}\n", tab, tab, "");