4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2017, Commissariat a l'Energie Atomique et aux Energies
26 * Copyright (c) 2017, Intel Corporation.
28 * Author: Henri Doreau <henri.doreau@cea.fr>
31 #define DEBUG_SUBSYSTEM S_MDC
33 #include <linux/init.h>
34 #include <linux/kthread.h>
35 #include <linux/poll.h>
36 #include <linux/miscdevice.h>
38 #include <lustre_log.h>
40 #include "mdc_internal.h"
44 * -- Changelog delivery through character device --
48 * Mutex to protect chlg_registered_devices below
50 static DEFINE_MUTEX(chlg_registered_dev_lock);
53 * Global linked list of all registered devices (one per MDT).
55 static LIST_HEAD(chlg_registered_devices);
58 struct chlg_registered_dev {
59 /* Device name of the form "changelog-{MDTNAME}" */
61 /* Misc device descriptor */
62 struct miscdevice ced_misc;
63 /* OBDs referencing this device (multiple mount point) */
64 struct list_head ced_obds;
65 /* Reference counter for proper deregistration */
67 /* Link within the global chlg_registered_devices */
68 struct list_head ced_link;
71 struct chlg_reader_state {
72 /* Shortcut to the corresponding OBD device */
73 struct obd_device *crs_obd;
74 /* Producer thread (if any) */
75 struct task_struct *crs_prod_task;
76 /* An error occurred that prevents from reading further */
78 /* EOF, no more records available */
80 /* Desired start position */
81 __u64 crs_start_offset;
82 /* Wait queue for the catalog processing thread */
83 wait_queue_head_t crs_waitq_prod;
84 /* Wait queue for the record copy threads */
85 wait_queue_head_t crs_waitq_cons;
86 /* Mutex protecting crs_rec_count and crs_rec_queue */
87 struct mutex crs_lock;
88 /* Number of item in the list */
90 /* List of prefetched enqueued_record::enq_linkage_items */
91 struct list_head crs_rec_queue;
94 struct chlg_rec_entry {
95 /* Link within the chlg_reader_state::crs_rec_queue list */
96 struct list_head enq_linkage;
97 /* Data (enq_record) field length */
99 /* Copy of a changelog record (see struct llog_changelog_rec) */
100 struct changelog_rec enq_record[];
104 /* Number of records to prefetch locally. */
105 CDEV_CHLG_MAX_PREFETCH = 1024,
109 * ChangeLog catalog processing callback invoked on each record.
110 * If the current record is eligible to userland delivery, push
111 * it into the crs_rec_queue where the consumer code will fetch it.
113 * @param[in] env (unused)
114 * @param[in] llh Client-side handle used to identify the llog
115 * @param[in] hdr Header of the current llog record
116 * @param[in,out] data chlg_reader_state passed from caller
118 * @return 0 or LLOG_PROC_* control code on success, negated error on failure.
120 static int chlg_read_cat_process_cb(const struct lu_env *env,
121 struct llog_handle *llh,
122 struct llog_rec_hdr *hdr, void *data)
124 struct llog_changelog_rec *rec;
125 struct chlg_reader_state *crs = data;
126 struct chlg_rec_entry *enq;
131 LASSERT(crs != NULL);
132 LASSERT(hdr != NULL);
134 rec = container_of(hdr, struct llog_changelog_rec, cr_hdr);
136 if (rec->cr_hdr.lrh_type != CHANGELOG_REC) {
138 CERROR("%s: not a changelog rec %x/%d in llog "DFID" rc = %d\n",
139 crs->crs_obd->obd_name, rec->cr_hdr.lrh_type,
141 PFID(lu_object_fid(&llh->lgh_obj->do_lu)), rc);
145 /* Skip undesired records */
146 if (rec->cr.cr_index < crs->crs_start_offset)
149 CDEBUG(D_HSM, "%llu %02d%-5s %llu 0x%x t="DFID" p="DFID" %.*s\n",
150 rec->cr.cr_index, rec->cr.cr_type,
151 changelog_type2str(rec->cr.cr_type), rec->cr.cr_time,
152 rec->cr.cr_flags & CLF_FLAGMASK,
153 PFID(&rec->cr.cr_tfid), PFID(&rec->cr.cr_pfid),
154 rec->cr.cr_namelen, changelog_rec_name(&rec->cr));
156 wait_event_interruptible(crs->crs_waitq_prod,
157 crs->crs_rec_count < CDEV_CHLG_MAX_PREFETCH ||
158 kthread_should_stop());
160 if (kthread_should_stop())
161 RETURN(LLOG_PROC_BREAK);
163 len = changelog_rec_size(&rec->cr) + rec->cr.cr_namelen;
164 OBD_ALLOC(enq, sizeof(*enq) + len);
168 INIT_LIST_HEAD(&enq->enq_linkage);
169 enq->enq_length = len;
170 memcpy(enq->enq_record, &rec->cr, len);
172 mutex_lock(&crs->crs_lock);
173 list_add_tail(&enq->enq_linkage, &crs->crs_rec_queue);
174 crs->crs_rec_count++;
175 mutex_unlock(&crs->crs_lock);
177 wake_up_all(&crs->crs_waitq_cons);
183 * Remove record from the list it is attached to and free it.
185 static void enq_record_delete(struct chlg_rec_entry *rec)
187 list_del(&rec->enq_linkage);
188 OBD_FREE(rec, sizeof(*rec) + rec->enq_length);
192 * Record prefetch thread entry point. Opens the changelog catalog and starts
195 * @param[in,out] args chlg_reader_state passed from caller.
196 * @return 0 on success, negated error code on failure.
198 static int chlg_load(void *args)
200 struct chlg_reader_state *crs = args;
201 struct obd_device *obd = crs->crs_obd;
202 struct llog_ctxt *ctx = NULL;
203 struct llog_handle *llh = NULL;
207 ctx = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
209 GOTO(err_out, rc = -ENOENT);
211 rc = llog_open(NULL, ctx, &llh, NULL, CHANGELOG_CATALOG,
214 CERROR("%s: fail to open changelog catalog: rc = %d\n",
219 rc = llog_init_handle(NULL, llh,
222 LLOG_F_EXT_EXTRA_FLAGS |
223 LLOG_F_EXT_X_UIDGID |
227 CERROR("%s: fail to init llog handle: rc = %d\n",
232 rc = llog_cat_process(NULL, llh, chlg_read_cat_process_cb, crs, 0, 0);
234 CERROR("%s: fail to process llog: rc = %d\n", obd->obd_name, rc);
244 wake_up_all(&crs->crs_waitq_cons);
247 llog_cat_close(NULL, llh);
252 wait_event_interruptible(crs->crs_waitq_prod, kthread_should_stop());
258 * Read handler, dequeues records from the chlg_reader_state if any.
259 * No partial records are copied to userland so this function can return less
260 * data than required (short read).
262 * @param[in] file File pointer to the character device.
263 * @param[out] buff Userland buffer where to copy the records.
264 * @param[in] count Userland buffer size.
265 * @param[out] ppos File position, updated with the index number of the next
267 * @return number of copied bytes on success, negated error code on failure.
269 static ssize_t chlg_read(struct file *file, char __user *buff, size_t count,
272 struct chlg_reader_state *crs = file->private_data;
273 struct chlg_rec_entry *rec;
274 struct chlg_rec_entry *tmp;
275 size_t written_total = 0;
280 if (file->f_flags & O_NONBLOCK && crs->crs_rec_count == 0) {
281 if (crs->crs_err < 0)
282 RETURN(crs->crs_err);
283 else if (crs->crs_eof)
289 rc = wait_event_interruptible(crs->crs_waitq_cons,
290 crs->crs_rec_count > 0 || crs->crs_eof || crs->crs_err);
292 mutex_lock(&crs->crs_lock);
293 list_for_each_entry_safe(rec, tmp, &crs->crs_rec_queue, enq_linkage) {
294 if (written_total + rec->enq_length > count)
297 if (copy_to_user(buff, rec->enq_record, rec->enq_length)) {
302 buff += rec->enq_length;
303 written_total += rec->enq_length;
305 crs->crs_rec_count--;
306 list_move_tail(&rec->enq_linkage, &consumed);
308 crs->crs_start_offset = rec->enq_record->cr_index + 1;
310 mutex_unlock(&crs->crs_lock);
312 if (written_total > 0) {
314 wake_up_all(&crs->crs_waitq_prod);
315 } else if (rc == 0) {
319 list_for_each_entry_safe(rec, tmp, &consumed, enq_linkage)
320 enq_record_delete(rec);
322 *ppos = crs->crs_start_offset;
328 * Jump to a given record index. Helper for chlg_llseek().
330 * @param[in,out] crs Internal reader state.
331 * @param[in] offset Desired offset (index record).
332 * @return 0 on success, negated error code on failure.
334 static int chlg_set_start_offset(struct chlg_reader_state *crs, __u64 offset)
336 struct chlg_rec_entry *rec;
337 struct chlg_rec_entry *tmp;
339 mutex_lock(&crs->crs_lock);
340 if (offset < crs->crs_start_offset) {
341 mutex_unlock(&crs->crs_lock);
345 crs->crs_start_offset = offset;
346 list_for_each_entry_safe(rec, tmp, &crs->crs_rec_queue, enq_linkage) {
347 struct changelog_rec *cr = rec->enq_record;
349 if (cr->cr_index >= crs->crs_start_offset)
352 crs->crs_rec_count--;
353 enq_record_delete(rec);
356 mutex_unlock(&crs->crs_lock);
357 wake_up_all(&crs->crs_waitq_prod);
362 * Move read pointer to a certain record index, encoded as an offset.
364 * @param[in,out] file File pointer to the changelog character device
365 * @param[in] off Offset to skip, actually a record index, not byte count
366 * @param[in] whence Relative/Absolute interpretation of the offset
367 * @return the resulting position on success or negated error code on failure.
369 static loff_t chlg_llseek(struct file *file, loff_t off, int whence)
371 struct chlg_reader_state *crs = file->private_data;
380 pos = file->f_pos + off;
387 /* We cannot go backward */
388 if (pos < file->f_pos)
391 rc = chlg_set_start_offset(crs, pos);
400 * Clear record range for a given changelog reader.
402 * @param[in] crs Current internal state.
403 * @param[in] reader Changelog reader ID (cl1, cl2...)
404 * @param[in] record Record index up which to clear
405 * @return 0 on success, negated error code on failure.
407 static int chlg_clear(struct chlg_reader_state *crs, __u32 reader, __u64 record)
409 struct obd_device *obd = crs->crs_obd;
410 struct changelog_setinfo cs = {
415 return obd_set_info_async(NULL, obd->obd_self_export,
416 strlen(KEY_CHANGELOG_CLEAR),
417 KEY_CHANGELOG_CLEAR, sizeof(cs), &cs, NULL);
420 /** Maximum changelog control command size */
421 #define CHLG_CONTROL_CMD_MAX 64
424 * Handle writes() into the changelog character device. Write() can be used
425 * to request special control operations.
427 * @param[in] file File pointer to the changelog character device
428 * @param[in] buff User supplied data (written data)
429 * @param[in] count Number of written bytes
430 * @param[in] off (unused)
431 * @return number of written bytes on success, negated error code on failure.
433 static ssize_t chlg_write(struct file *file, const char __user *buff,
434 size_t count, loff_t *off)
436 struct chlg_reader_state *crs = file->private_data;
443 if (count > CHLG_CONTROL_CMD_MAX)
446 OBD_ALLOC(kbuf, CHLG_CONTROL_CMD_MAX);
450 if (copy_from_user(kbuf, buff, count))
451 GOTO(out_kbuf, rc = -EFAULT);
453 kbuf[CHLG_CONTROL_CMD_MAX - 1] = '\0';
455 if (sscanf(kbuf, "clear:cl%u:%llu", &reader, &record) == 2)
456 rc = chlg_clear(crs, reader, record);
462 OBD_FREE(kbuf, CHLG_CONTROL_CMD_MAX);
463 return rc < 0 ? rc : count;
467 * Find the OBD device associated to a changelog character device.
468 * @param[in] cdev character device instance descriptor
469 * @return corresponding OBD device or NULL if none was found.
471 static struct obd_device *chlg_obd_get(dev_t cdev)
473 int minor = MINOR(cdev);
474 struct obd_device *obd = NULL;
475 struct chlg_registered_dev *curr;
477 mutex_lock(&chlg_registered_dev_lock);
478 list_for_each_entry(curr, &chlg_registered_devices, ced_link) {
479 if (curr->ced_misc.minor == minor) {
480 /* take the first available OBD device attached */
481 obd = list_first_entry(&curr->ced_obds,
483 u.cli.cl_chg_dev_linkage);
487 mutex_unlock(&chlg_registered_dev_lock);
492 * Open handler, initialize internal CRS state and spawn prefetch thread if
494 * @param[in] inode Inode struct for the open character device.
495 * @param[in] file Corresponding file pointer.
496 * @return 0 on success, negated error code on failure.
498 static int chlg_open(struct inode *inode, struct file *file)
500 struct chlg_reader_state *crs;
501 struct obd_device *obd = chlg_obd_get(inode->i_rdev);
502 struct task_struct *task;
514 crs->crs_err = false;
515 crs->crs_eof = false;
517 mutex_init(&crs->crs_lock);
518 INIT_LIST_HEAD(&crs->crs_rec_queue);
519 init_waitqueue_head(&crs->crs_waitq_prod);
520 init_waitqueue_head(&crs->crs_waitq_cons);
522 if (file->f_mode & FMODE_READ) {
523 task = kthread_run(chlg_load, crs, "chlg_load_thread");
526 CERROR("%s: cannot start changelog thread: rc = %d\n",
530 crs->crs_prod_task = task;
533 file->private_data = crs;
542 * Close handler, release resources.
544 * @param[in] inode Inode struct for the open character device.
545 * @param[in] file Corresponding file pointer.
546 * @return 0 on success, negated error code on failure.
548 static int chlg_release(struct inode *inode, struct file *file)
550 struct chlg_reader_state *crs = file->private_data;
551 struct chlg_rec_entry *rec;
552 struct chlg_rec_entry *tmp;
555 if (crs->crs_prod_task)
556 rc = kthread_stop(crs->crs_prod_task);
558 list_for_each_entry_safe(rec, tmp, &crs->crs_rec_queue, enq_linkage)
559 enq_record_delete(rec);
567 * Poll handler, indicates whether the device is readable (new records) and
570 * @param[in] file Device file pointer.
571 * @param[in] wait (opaque)
572 * @return combination of the poll status flags.
574 static unsigned int chlg_poll(struct file *file, poll_table *wait)
576 struct chlg_reader_state *crs = file->private_data;
577 unsigned int mask = 0;
579 mutex_lock(&crs->crs_lock);
580 poll_wait(file, &crs->crs_waitq_cons, wait);
581 if (crs->crs_rec_count > 0)
582 mask |= POLLIN | POLLRDNORM;
587 mutex_unlock(&crs->crs_lock);
591 static const struct file_operations chlg_fops = {
592 .owner = THIS_MODULE,
593 .llseek = chlg_llseek,
597 .release = chlg_release,
602 * This uses obd_name of the form: "testfs-MDT0000-mdc-ffff88006501600"
603 * and returns a name of the form: "changelog-testfs-MDT0000".
605 static void get_chlg_name(char *name, size_t name_len, struct obd_device *obd)
609 snprintf(name, name_len, "changelog-%s", obd->obd_name);
611 /* Find the 2nd '-' from the end and truncate on it */
612 for (i = 0; i < 2; i++) {
613 char *p = strrchr(name, '-');
622 * Find a changelog character device by name.
623 * All devices registered during MDC setup are listed in a global list with
624 * their names attached.
626 static struct chlg_registered_dev *
627 chlg_registered_dev_find_by_name(const char *name)
629 struct chlg_registered_dev *dit;
631 list_for_each_entry(dit, &chlg_registered_devices, ced_link)
632 if (strcmp(name, dit->ced_name) == 0)
638 * Find chlg_registered_dev structure for a given OBD device.
639 * This is bad O(n^2) but for each filesystem:
640 * - N is # of MDTs times # of mount points
641 * - this only runs at shutdown
643 static struct chlg_registered_dev *
644 chlg_registered_dev_find_by_obd(const struct obd_device *obd)
646 struct chlg_registered_dev *dit;
647 struct obd_device *oit;
649 list_for_each_entry(dit, &chlg_registered_devices, ced_link)
650 list_for_each_entry(oit, &dit->ced_obds,
651 u.cli.cl_chg_dev_linkage)
658 * Changelog character device initialization.
659 * Register a misc character device with a dynamic minor number, under a name
660 * of the form: 'changelog-fsname-MDTxxxx'. Reference this OBD device with it.
662 * @param[in] obd This MDC obd_device.
663 * @return 0 on success, negated error code on failure.
665 int mdc_changelog_cdev_init(struct obd_device *obd)
667 struct chlg_registered_dev *exist;
668 struct chlg_registered_dev *entry;
672 OBD_ALLOC_PTR(entry);
676 get_chlg_name(entry->ced_name, sizeof(entry->ced_name), obd);
678 entry->ced_misc.minor = MISC_DYNAMIC_MINOR;
679 entry->ced_misc.name = entry->ced_name;
680 entry->ced_misc.fops = &chlg_fops;
682 kref_init(&entry->ced_refs);
683 INIT_LIST_HEAD(&entry->ced_obds);
684 INIT_LIST_HEAD(&entry->ced_link);
686 mutex_lock(&chlg_registered_dev_lock);
687 exist = chlg_registered_dev_find_by_name(entry->ced_name);
689 kref_get(&exist->ced_refs);
690 list_add_tail(&obd->u.cli.cl_chg_dev_linkage, &exist->ced_obds);
691 GOTO(out_unlock, rc = 0);
694 /* Register new character device */
695 rc = misc_register(&entry->ced_misc);
697 GOTO(out_unlock, rc);
699 list_add_tail(&obd->u.cli.cl_chg_dev_linkage, &entry->ced_obds);
700 list_add_tail(&entry->ced_link, &chlg_registered_devices);
702 entry = NULL; /* prevent it from being freed below */
705 mutex_unlock(&chlg_registered_dev_lock);
712 * Deregister a changelog character device whose refcount has reached zero.
714 static void chlg_dev_clear(struct kref *kref)
716 struct chlg_registered_dev *entry = container_of(kref,
717 struct chlg_registered_dev,
721 list_del(&entry->ced_link);
722 misc_deregister(&entry->ced_misc);
728 * Release OBD, decrease reference count of the corresponding changelog device.
730 void mdc_changelog_cdev_finish(struct obd_device *obd)
732 struct chlg_registered_dev *dev = chlg_registered_dev_find_by_obd(obd);
735 mutex_lock(&chlg_registered_dev_lock);
736 list_del_init(&obd->u.cli.cl_chg_dev_linkage);
737 kref_put(&dev->ced_refs, chlg_dev_clear);
738 mutex_unlock(&chlg_registered_dev_lock);