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/device.h>
37 #include <linux/cdev.h>
38 #include <linux/idr.h>
40 #include <lustre_log.h>
41 #include <uapi/linux/lustre/lustre_ioctl.h>
43 #include "mdc_internal.h"
47 * -- Changelog delivery through character device --
51 * Mutex to protect chlg_registered_devices below
53 static DEFINE_MUTEX(chlg_registered_dev_lock);
56 * Global linked list of all registered devices (one per MDT).
58 static LIST_HEAD(chlg_registered_devices);
61 struct chlg_registered_dev {
62 /* Device name of the form "changelog-{MDTNAME}" */
64 /* changelog char device */
66 struct device ced_device;
67 /* OBDs referencing this device (multiple mount point) */
68 struct list_head ced_obds;
69 /* Reference counter for proper deregistration */
71 /* Link within the global chlg_registered_devices */
72 struct list_head ced_link;
75 struct chlg_reader_state {
76 /* Shortcut to the corresponding OBD device */
77 struct obd_device *crs_obd;
78 /* the corresponding chlg_registered_dev */
79 struct chlg_registered_dev *crs_ced;
80 /* Producer thread (if any) */
81 struct task_struct *crs_prod_task;
82 /* An error occurred that prevents from reading further */
84 /* EOF, no more records available */
86 /* Desired start position */
87 __u64 crs_start_offset;
88 /* Wait queue for the catalog processing thread */
89 wait_queue_head_t crs_waitq_prod;
90 /* Wait queue for the record copy threads */
91 wait_queue_head_t crs_waitq_cons;
92 /* Mutex protecting crs_rec_count and crs_rec_queue */
93 struct mutex crs_lock;
94 /* Number of item in the list */
96 /* List of prefetched enqueued_record::enq_linkage_items */
97 struct list_head crs_rec_queue;
98 unsigned int crs_last_catidx;
99 unsigned int crs_last_idx;
103 struct chlg_rec_entry {
104 /* Link within the chlg_reader_state::crs_rec_queue list */
105 struct list_head enq_linkage;
106 /* Data (enq_record) field length */
108 /* Copy of a changelog record (see struct llog_changelog_rec) */
109 struct changelog_rec enq_record[];
113 /* Number of records to prefetch locally. */
114 CDEV_CHLG_MAX_PREFETCH = 1024,
117 DEFINE_IDR(mdc_changelog_minor_idr);
118 static DEFINE_SPINLOCK(chlg_minor_lock);
120 static int chlg_minor_alloc(int *pminor)
122 void *minor_allocated = (void *)-1;
125 idr_preload(GFP_KERNEL);
126 spin_lock(&chlg_minor_lock);
127 minor = idr_alloc(&mdc_changelog_minor_idr, minor_allocated, 0,
128 MDC_CHANGELOG_DEV_COUNT, GFP_NOWAIT);
129 spin_unlock(&chlg_minor_lock);
139 static void chlg_minor_free(int minor)
141 spin_lock(&chlg_minor_lock);
142 idr_remove(&mdc_changelog_minor_idr, minor);
143 spin_unlock(&chlg_minor_lock);
146 static void chlg_device_release(struct device *dev)
148 struct chlg_registered_dev *entry = dev_get_drvdata(dev);
150 chlg_minor_free(MINOR(entry->ced_cdev.dev));
155 * Deregister a changelog character device whose refcount has reached zero.
157 static void chlg_dev_clear(struct kref *kref)
159 struct chlg_registered_dev *entry;
162 entry = container_of(kref, struct chlg_registered_dev,
165 list_del(&entry->ced_link);
166 cdev_device_del(&entry->ced_cdev, &entry->ced_device);
167 put_device(&entry->ced_device);
171 static inline struct obd_device* chlg_obd_get(struct chlg_registered_dev *dev)
173 struct obd_device *obd;
175 mutex_lock(&chlg_registered_dev_lock);
176 if (list_empty(&dev->ced_obds))
179 obd = list_first_entry(&dev->ced_obds, struct obd_device,
180 u.cli.cl_chg_dev_linkage);
181 class_incref(obd, "changelog", dev);
182 mutex_unlock(&chlg_registered_dev_lock);
186 static inline void chlg_obd_put(struct chlg_registered_dev *dev,
187 struct obd_device *obd)
189 class_decref(obd, "changelog", dev);
193 * ChangeLog catalog processing callback invoked on each record.
194 * If the current record is eligible to userland delivery, push
195 * it into the crs_rec_queue where the consumer code will fetch it.
197 * @param[in] env (unused)
198 * @param[in] llh Client-side handle used to identify the llog
199 * @param[in] hdr Header of the current llog record
200 * @param[in,out] data chlg_reader_state passed from caller
202 * @return 0 or LLOG_PROC_* control code on success, negated error on failure.
204 static int chlg_read_cat_process_cb(const struct lu_env *env,
205 struct llog_handle *llh,
206 struct llog_rec_hdr *hdr, void *data)
208 struct llog_changelog_rec *rec;
209 struct chlg_reader_state *crs = data;
210 struct chlg_rec_entry *enq;
215 LASSERT(crs != NULL);
216 LASSERT(hdr != NULL);
218 rec = container_of(hdr, struct llog_changelog_rec, cr_hdr);
220 crs->crs_last_catidx = llh->lgh_hdr->llh_cat_idx;
221 crs->crs_last_idx = hdr->lrh_index;
223 if (rec->cr_hdr.lrh_type != CHANGELOG_REC) {
225 CERROR("%s: not a changelog rec %x/%d in llog : rc = %d\n",
226 crs->crs_obd->obd_name, rec->cr_hdr.lrh_type,
227 rec->cr.cr_type, rc);
231 /* Check if we can skip the entire llog plain */
232 if (llog_is_plain_skipable(llh->lgh_hdr, hdr, rec->cr.cr_index,
233 crs->crs_start_offset))
234 RETURN(LLOG_SKIP_PLAIN);
236 /* Skip undesired records */
237 if (rec->cr.cr_index < crs->crs_start_offset)
240 CDEBUG(D_HSM, "%llu %02d%-5s %llu 0x%x t="DFID" p="DFID" %.*s\n",
241 rec->cr.cr_index, rec->cr.cr_type,
242 changelog_type2str(rec->cr.cr_type), rec->cr.cr_time,
243 rec->cr.cr_flags & CLF_FLAGMASK,
244 PFID(&rec->cr.cr_tfid), PFID(&rec->cr.cr_pfid),
245 rec->cr.cr_namelen, changelog_rec_name(&rec->cr));
247 wait_event_interruptible(crs->crs_waitq_prod,
248 crs->crs_rec_count < CDEV_CHLG_MAX_PREFETCH ||
249 kthread_should_stop());
251 if (kthread_should_stop())
252 RETURN(LLOG_PROC_BREAK);
254 len = changelog_rec_size(&rec->cr) + rec->cr.cr_namelen;
255 OBD_ALLOC(enq, sizeof(*enq) + len);
259 INIT_LIST_HEAD(&enq->enq_linkage);
260 enq->enq_length = len;
261 memcpy(enq->enq_record, &rec->cr, len);
263 mutex_lock(&crs->crs_lock);
264 list_add_tail(&enq->enq_linkage, &crs->crs_rec_queue);
265 crs->crs_rec_count++;
266 mutex_unlock(&crs->crs_lock);
268 wake_up(&crs->crs_waitq_cons);
274 * Remove record from the list it is attached to and free it.
276 static void enq_record_delete(struct chlg_rec_entry *rec)
278 list_del(&rec->enq_linkage);
279 OBD_FREE(rec, sizeof(*rec) + rec->enq_length);
283 * Record prefetch thread entry point. Opens the changelog catalog and starts
286 * @param[in,out] args chlg_reader_state passed from caller.
287 * @return 0 on success, negated error code on failure.
289 static int chlg_load(void *args)
291 struct chlg_reader_state *crs = args;
292 struct chlg_registered_dev *ced = crs->crs_ced;
293 struct obd_device *obd = NULL;
294 struct llog_ctxt *ctx = NULL;
295 struct llog_handle *llh = NULL;
299 crs->crs_last_catidx = 0;
300 crs->crs_last_idx = 0;
303 obd = chlg_obd_get(ced);
309 ctx = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
311 GOTO(err_out, rc = -ENOENT);
313 rc = llog_open(NULL, ctx, &llh, NULL, CHANGELOG_CATALOG,
316 CERROR("%s: fail to open changelog catalog: rc = %d\n",
322 rc = llog_init_handle(NULL, llh,
325 LLOG_F_EXT_EXTRA_FLAGS |
326 LLOG_F_EXT_X_UIDGID |
332 CERROR("%s: fail to init llog handle: rc = %d\n",
337 rc = llog_cat_process(NULL, llh, chlg_read_cat_process_cb, crs,
338 crs->crs_last_catidx, crs->crs_last_idx);
340 CERROR("%s: fail to process llog: rc = %d\n", obd->obd_name, rc);
343 if (!kthread_should_stop() && crs->crs_poll) {
344 llog_cat_close(NULL, llh);
346 class_decref(obd, "changelog", crs);
347 schedule_timeout_interruptible(cfs_time_seconds(1));
357 wake_up(&crs->crs_waitq_cons);
360 llog_cat_close(NULL, llh);
366 chlg_obd_put(ced, obd);
367 wait_event_interruptible(crs->crs_waitq_prod, kthread_should_stop());
372 static int chlg_start_thread(struct file *file)
374 struct chlg_reader_state *crs = file->private_data;
375 struct task_struct *task;
378 if (likely(crs->crs_prod_task))
380 if (unlikely(file->f_mode & FMODE_READ) == 0)
383 mutex_lock(&crs->crs_lock);
384 if (crs->crs_prod_task == NULL) {
385 task = kthread_run(chlg_load, crs, "chlg_load_thread");
388 CERROR("%s: cannot start changelog thread: rc = %d\n",
389 crs->crs_ced->ced_name, rc);
392 crs->crs_prod_task = task;
395 mutex_unlock(&crs->crs_lock);
400 * Read handler, dequeues records from the chlg_reader_state if any.
401 * No partial records are copied to userland so this function can return less
402 * data than required (short read).
404 * @param[in] file File pointer to the character device.
405 * @param[out] buff Userland buffer where to copy the records.
406 * @param[in] count Userland buffer size.
407 * @param[out] ppos File position, updated with the index number of the next
409 * @return number of copied bytes on success, negated error code on failure.
411 static ssize_t chlg_read(struct file *file, char __user *buff, size_t count,
414 struct chlg_reader_state *crs = file->private_data;
415 struct chlg_rec_entry *rec;
416 struct chlg_rec_entry *tmp;
417 size_t written_total = 0;
422 if (file->f_flags & O_NONBLOCK && crs->crs_rec_count == 0) {
423 if (crs->crs_err < 0)
424 RETURN(crs->crs_err);
425 else if (crs->crs_eof)
431 rc = chlg_start_thread(file);
435 rc = wait_event_interruptible(crs->crs_waitq_cons,
436 crs->crs_rec_count > 0 || crs->crs_eof || crs->crs_err);
438 mutex_lock(&crs->crs_lock);
439 list_for_each_entry_safe(rec, tmp, &crs->crs_rec_queue, enq_linkage) {
440 if (written_total + rec->enq_length > count)
443 if (copy_to_user(buff, rec->enq_record, rec->enq_length)) {
448 buff += rec->enq_length;
449 written_total += rec->enq_length;
451 crs->crs_rec_count--;
452 list_move_tail(&rec->enq_linkage, &consumed);
454 crs->crs_start_offset = rec->enq_record->cr_index + 1;
456 mutex_unlock(&crs->crs_lock);
458 if (written_total > 0) {
460 wake_up(&crs->crs_waitq_prod);
461 } else if (rc == 0) {
465 list_for_each_entry_safe(rec, tmp, &consumed, enq_linkage)
466 enq_record_delete(rec);
468 *ppos = crs->crs_start_offset;
474 * Jump to a given record index. Helper for chlg_llseek().
476 * @param[in,out] crs Internal reader state.
477 * @param[in] offset Desired offset (index record).
478 * @return 0 on success, negated error code on failure.
480 static int chlg_set_start_offset(struct chlg_reader_state *crs, __u64 offset)
482 struct chlg_rec_entry *rec;
483 struct chlg_rec_entry *tmp;
485 mutex_lock(&crs->crs_lock);
486 if (offset < crs->crs_start_offset) {
487 mutex_unlock(&crs->crs_lock);
491 crs->crs_start_offset = offset;
492 list_for_each_entry_safe(rec, tmp, &crs->crs_rec_queue, enq_linkage) {
493 struct changelog_rec *cr = rec->enq_record;
495 if (cr->cr_index >= crs->crs_start_offset)
498 crs->crs_rec_count--;
499 enq_record_delete(rec);
502 mutex_unlock(&crs->crs_lock);
503 wake_up(&crs->crs_waitq_prod);
508 * Move read pointer to a certain record index, encoded as an offset.
510 * @param[in,out] file File pointer to the changelog character device
511 * @param[in] off Offset to skip, actually a record index, not byte count
512 * @param[in] whence Relative/Absolute interpretation of the offset
513 * @return the resulting position on success or negated error code on failure.
515 static loff_t chlg_llseek(struct file *file, loff_t off, int whence)
517 struct chlg_reader_state *crs = file->private_data;
526 pos = file->f_pos + off;
533 /* We cannot go backward */
534 if (pos < file->f_pos)
537 rc = chlg_set_start_offset(crs, pos);
546 * Clear record range for a given changelog reader.
548 * @param[in] crs Current internal state.
549 * @param[in] reader Changelog reader ID (cl1, cl2...)
550 * @param[in] record Record index up which to clear
551 * @return 0 on success, negated error code on failure.
553 static int chlg_clear(struct chlg_reader_state *crs, __u32 reader, __u64 record)
555 struct obd_device *obd = NULL;
556 struct changelog_setinfo cs = {
562 obd = chlg_obd_get(crs->crs_ced);
566 rc = obd_set_info_async(NULL, obd->obd_self_export,
567 strlen(KEY_CHANGELOG_CLEAR),
568 KEY_CHANGELOG_CLEAR, sizeof(cs), &cs, NULL);
570 chlg_obd_put(crs->crs_ced, obd);
574 /** Maximum changelog control command size */
575 #define CHLG_CONTROL_CMD_MAX 64
578 * Handle writes() into the changelog character device. Write() can be used
579 * to request special control operations.
581 * @param[in] file File pointer to the changelog character device
582 * @param[in] buff User supplied data (written data)
583 * @param[in] count Number of written bytes
584 * @param[in] off (unused)
585 * @return number of written bytes on success, negated error code on failure.
587 static ssize_t chlg_write(struct file *file, const char __user *buff,
588 size_t count, loff_t *off)
590 struct chlg_reader_state *crs = file->private_data;
597 if (count > CHLG_CONTROL_CMD_MAX)
600 OBD_ALLOC(kbuf, CHLG_CONTROL_CMD_MAX);
604 if (copy_from_user(kbuf, buff, count))
605 GOTO(out_kbuf, rc = -EFAULT);
607 kbuf[CHLG_CONTROL_CMD_MAX - 1] = '\0';
609 if (sscanf(kbuf, "clear:cl%u:%llu", &reader, &record) == 2)
610 rc = chlg_clear(crs, reader, record);
616 OBD_FREE(kbuf, CHLG_CONTROL_CMD_MAX);
617 return rc < 0 ? rc : count;
621 * Open handler, initialize internal CRS state and spawn prefetch thread if
623 * @param[in] inode Inode struct for the open character device.
624 * @param[in] file Corresponding file pointer.
625 * @return 0 on success, negated error code on failure.
627 static int chlg_open(struct inode *inode, struct file *file)
629 struct chlg_reader_state *crs;
630 struct chlg_registered_dev *dev;
633 dev = container_of(inode->i_cdev, struct chlg_registered_dev, ced_cdev);
639 kref_get(&dev->ced_refs);
641 crs->crs_err = false;
642 crs->crs_eof = false;
644 mutex_init(&crs->crs_lock);
645 INIT_LIST_HEAD(&crs->crs_rec_queue);
646 init_waitqueue_head(&crs->crs_waitq_prod);
647 init_waitqueue_head(&crs->crs_waitq_cons);
648 crs->crs_prod_task = NULL;
650 file->private_data = crs;
655 * Close handler, release resources.
657 * @param[in] inode Inode struct for the open character device.
658 * @param[in] file Corresponding file pointer.
659 * @return 0 on success, negated error code on failure.
661 static int chlg_release(struct inode *inode, struct file *file)
663 struct chlg_reader_state *crs = file->private_data;
664 struct chlg_rec_entry *rec;
665 struct chlg_rec_entry *tmp;
668 if (crs->crs_prod_task)
669 rc = kthread_stop(crs->crs_prod_task);
671 list_for_each_entry_safe(rec, tmp, &crs->crs_rec_queue, enq_linkage)
672 enq_record_delete(rec);
674 kref_put(&crs->crs_ced->ced_refs, chlg_dev_clear);
681 * Poll handler, indicates whether the device is readable (new records) and
684 * @param[in] file Device file pointer.
685 * @param[in] wait (opaque)
686 * @return combination of the poll status flags.
688 static unsigned int chlg_poll(struct file *file, poll_table *wait)
690 struct chlg_reader_state *crs = file->private_data;
691 unsigned int mask = 0;
694 rc = chlg_start_thread(file);
698 mutex_lock(&crs->crs_lock);
699 poll_wait(file, &crs->crs_waitq_cons, wait);
700 if (crs->crs_rec_count > 0)
701 mask |= POLLIN | POLLRDNORM;
706 mutex_unlock(&crs->crs_lock);
710 static long chlg_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
714 struct chlg_reader_state *crs = file->private_data;
716 case OBD_IOC_CHLG_POLL:
717 crs->crs_poll = !!arg;
727 static const struct file_operations chlg_fops = {
728 .owner = THIS_MODULE,
729 .llseek = chlg_llseek,
733 .release = chlg_release,
735 .unlocked_ioctl = chlg_ioctl,
739 * This uses obd_name of the form: "testfs-MDT0000-mdc-ffff88006501600"
740 * and returns a name of the form: "changelog-testfs-MDT0000".
742 static void get_target_name(char *name, size_t name_len, struct obd_device *obd)
746 snprintf(name, name_len, "%s", obd->obd_name);
748 /* Find the 2nd '-' from the end and truncate on it */
749 for (i = 0; i < 2; i++) {
750 char *p = strrchr(name, '-');
759 * Find a changelog character device by name.
760 * All devices registered during MDC setup are listed in a global list with
761 * their names attached.
763 static struct chlg_registered_dev *
764 chlg_registered_dev_find_by_name(const char *name)
766 struct chlg_registered_dev *dit;
768 LASSERT(mutex_is_locked(&chlg_registered_dev_lock));
769 list_for_each_entry(dit, &chlg_registered_devices, ced_link)
770 if (strcmp(name, dit->ced_name) == 0)
776 * Find chlg_registered_dev structure for a given OBD device.
777 * This is bad O(n^2) but for each filesystem:
778 * - N is # of MDTs times # of mount points
779 * - this only runs at shutdown
781 static struct chlg_registered_dev *
782 chlg_registered_dev_find_by_obd(const struct obd_device *obd)
784 struct chlg_registered_dev *dit;
785 struct obd_device *oit;
787 LASSERT(mutex_is_locked(&chlg_registered_dev_lock));
788 list_for_each_entry(dit, &chlg_registered_devices, ced_link)
789 list_for_each_entry(oit, &dit->ced_obds,
790 u.cli.cl_chg_dev_linkage)
797 * Changelog character device initialization.
798 * Register a misc character device with a dynamic minor number, under a name
799 * of the form: 'changelog-fsname-MDTxxxx'. Reference this OBD device with it.
801 * @param[in] obd This MDC obd_device.
802 * @return 0 on success, negated error code on failure.
804 int mdc_changelog_cdev_init(struct obd_device *obd)
806 struct chlg_registered_dev *exist;
807 struct chlg_registered_dev *entry;
811 OBD_ALLOC_PTR(entry);
815 get_target_name(entry->ced_name, sizeof(entry->ced_name), obd);
817 kref_init(&entry->ced_refs);
818 INIT_LIST_HEAD(&entry->ced_obds);
819 INIT_LIST_HEAD(&entry->ced_link);
821 mutex_lock(&chlg_registered_dev_lock);
822 exist = chlg_registered_dev_find_by_name(entry->ced_name);
824 kref_get(&exist->ced_refs);
825 list_add_tail(&obd->u.cli.cl_chg_dev_linkage, &exist->ced_obds);
826 GOTO(out_unlock, rc = 0);
829 list_add_tail(&obd->u.cli.cl_chg_dev_linkage, &entry->ced_obds);
830 list_add_tail(&entry->ced_link, &chlg_registered_devices);
832 rc = chlg_minor_alloc(&minor);
834 GOTO(out_listrm, rc);
836 device_initialize(&entry->ced_device);
837 entry->ced_device.devt = MKDEV(MAJOR(mdc_changelog_dev), minor);
838 entry->ced_device.class = mdc_changelog_class;
839 entry->ced_device.release = chlg_device_release;
840 dev_set_drvdata(&entry->ced_device, entry);
841 rc = dev_set_name(&entry->ced_device, "%s-%s", MDC_CHANGELOG_DEV_NAME,
846 /* Register new character device */
847 cdev_init(&entry->ced_cdev, &chlg_fops);
848 entry->ced_cdev.owner = THIS_MODULE;
849 rc = cdev_device_add(&entry->ced_cdev, &entry->ced_device);
851 GOTO(out_device_name, rc);
853 entry = NULL; /* prevent it from being freed below */
854 GOTO(out_unlock, rc = 0);
857 kfree_const(entry->ced_device.kobj.name);
860 chlg_minor_free(minor);
863 list_del_init(&obd->u.cli.cl_chg_dev_linkage);
864 list_del(&entry->ced_link);
867 mutex_unlock(&chlg_registered_dev_lock);
874 * Release OBD, decrease reference count of the corresponding changelog device.
876 void mdc_changelog_cdev_finish(struct obd_device *obd)
878 struct chlg_registered_dev *dev;
881 mutex_lock(&chlg_registered_dev_lock);
882 dev = chlg_registered_dev_find_by_obd(obd);
883 list_del_init(&obd->u.cli.cl_chg_dev_linkage);
884 kref_put(&dev->ced_refs, chlg_dev_clear);
885 mutex_unlock(&chlg_registered_dev_lock);