| 1 | |
| 2 | |
| 3 | drbd_actlog.c
|
| 4 |
|
| 5 | This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
|
| 6 |
|
| 7 | Copyright (C) 2003-2008, LINBIT Information Technologies GmbH.
|
| 8 | Copyright (C) 2003-2008, Philipp Reisner <[email protected]>.
|
| 9 | Copyright (C) 2003-2008, Lars Ellenberg <[email protected]>.
|
| 10 |
|
| 11 |
|
| 12 |
|
| 13 |
|
| 14 | #include <linux/slab.h>
|
| 15 | #include <linux/crc32c.h>
|
| 16 | #include <linux/drbd.h>
|
| 17 | #include <linux/drbd_limits.h>
|
| 18 | #include "drbd_int.h"
|
| 19 |
|
| 20 |
|
| 21 | enum al_transaction_types {
|
| 22 | AL_TR_UPDATE = 0,
|
| 23 | AL_TR_INITIALIZED = 0xffff
|
| 24 | };
|
| 25 |
|
| 26 | struct __packed al_transaction_on_disk {
|
| 27 |
|
| 28 | __be32 magic;
|
| 29 |
|
| 30 | |
| 31 | * in the on disk ring buffer
|
| 32 | __be32 tr_number;
|
| 33 |
|
| 34 |
|
| 35 | __be32 crc32c;
|
| 36 |
|
| 37 | |
| 38 | * purge-all, set-all-idle, set-all-active, ... to-be-defined
|
| 39 | * see also enum al_transaction_types
|
| 40 | __be16 transaction_type;
|
| 41 |
|
| 42 | |
| 43 | * so 16bit will be enough for the slot number.
|
| 44 |
|
| 45 |
|
| 46 | __be16 n_updates;
|
| 47 |
|
| 48 | |
| 49 | * Having this in each transaction should make reconfiguration
|
| 50 | * of that parameter easier.
|
| 51 | __be16 context_size;
|
| 52 |
|
| 53 |
|
| 54 | __be16 context_start_slot_nr;
|
| 55 |
|
| 56 | |
| 57 | * sectors-written since device creation, and other data generation tag
|
| 58 | * supporting usage
|
| 59 | __be32 __reserved[4];
|
| 60 |
|
| 61 |
|
| 62 |
|
| 63 | |
| 64 | * in one transaction, then use the remaining byte in the 4k block for
|
| 65 | * context information. "Flexible" number of updates per transaction
|
| 66 | * does not help, as we have to account for the case when all update
|
| 67 | * slots are used anyways, so it would only complicate code without
|
| 68 | * additional benefit.
|
| 69 |
|
| 70 | __be16 update_slot_nr[AL_UPDATES_PER_TRANSACTION];
|
| 71 |
|
| 72 | |
| 73 | * allows to cover device sizes of up to 2**54 Byte (16 PiB)
|
| 74 | __be32 update_extent_nr[AL_UPDATES_PER_TRANSACTION];
|
| 75 |
|
| 76 |
|
| 77 |
|
| 78 |
|
| 79 | __be32 context[AL_CONTEXT_PER_TRANSACTION];
|
| 80 | };
|
| 81 |
|
| 82 | void *drbd_md_get_buffer(struct drbd_device *device, const char *intent)
|
| 83 | {
|
| 84 | int r;
|
| 85 |
|
| 86 | wait_event(device->misc_wait,
|
| 87 | (r = atomic_cmpxchg(&device->md_io.in_use, 0, 1)) == 0 ||
|
| 88 | device->state.disk <= D_FAILED);
|
| 89 |
|
| 90 | if (r)
|
| 91 | return NULL;
|
| 92 |
|
| 93 | device->md_io.current_use = intent;
|
| 94 | device->md_io.start_jif = jiffies;
|
| 95 | device->md_io.submit_jif = device->md_io.start_jif - 1;
|
| 96 | return page_address(device->md_io.page);
|
| 97 | }
|
| 98 |
|
| 99 | void drbd_md_put_buffer(struct drbd_device *device)
|
| 100 | {
|
| 101 | if (atomic_dec_and_test(&device->md_io.in_use))
|
| 102 | wake_up(&device->misc_wait);
|
| 103 | }
|
| 104 |
|
| 105 | void wait_until_done_or_force_detached(struct drbd_device *device, struct drbd_backing_dev *bdev,
|
| 106 | unsigned int *done)
|
| 107 | {
|
| 108 | long dt;
|
| 109 |
|
| 110 | rcu_read_lock();
|
| 111 | dt = rcu_dereference(bdev->disk_conf)->disk_timeout;
|
| 112 | rcu_read_unlock();
|
| 113 | dt = dt * HZ / 10;
|
| 114 | if (dt == 0)
|
| 115 | dt = MAX_SCHEDULE_TIMEOUT;
|
| 116 |
|
| 117 | dt = wait_event_timeout(device->misc_wait,
|
| 118 | *done || test_bit(FORCE_DETACH, &device->flags), dt);
|
| 119 | if (dt == 0) {
|
| 120 | drbd_err(device, "meta-data IO operation timed out\n");
|
| 121 | drbd_chk_io_error(device, 1, DRBD_FORCE_DETACH);
|
| 122 | }
|
| 123 | }
|
| 124 |
|
| 125 | static int _drbd_md_sync_page_io(struct drbd_device *device,
|
| 126 | struct drbd_backing_dev *bdev,
|
| 127 | sector_t sector, enum req_op op)
|
| 128 | {
|
| 129 | struct bio *bio;
|
| 130 |
|
| 131 | const int size = 4096;
|
| 132 | int err;
|
| 133 | blk_opf_t op_flags = 0;
|
| 134 |
|
| 135 | device->md_io.done = 0;
|
| 136 | device->md_io.error = -ENODEV;
|
| 137 |
|
| 138 | if ((op == REQ_OP_WRITE) && !test_bit(MD_NO_FUA, &device->flags))
|
| 139 | op_flags |= REQ_FUA | REQ_PREFLUSH;
|
| 140 | op_flags |= REQ_SYNC;
|
| 141 |
|
| 142 | bio = bio_alloc_bioset(bdev->md_bdev, 1, op | op_flags, GFP_NOIO,
|
| 143 | &drbd_md_io_bio_set);
|
| 144 | bio->bi_iter.bi_sector = sector;
|
| 145 | err = -EIO;
|
| 146 | if (bio_add_page(bio, device->md_io.page, size, 0) != size)
|
| 147 | goto out;
|
| 148 | bio->bi_private = device;
|
| 149 | bio->bi_end_io = drbd_md_endio;
|
| 150 |
|
| 151 | if (op != REQ_OP_WRITE && device->state.disk == D_DISKLESS && device->ldev == NULL)
|
| 152 |
|
| 153 | ;
|
| 154 | else if (!get_ldev_if_state(device, D_ATTACHING)) {
|
| 155 |
|
| 156 | drbd_err(device, "ASSERT FAILED: get_ldev_if_state() == 1 in _drbd_md_sync_page_io()\n");
|
| 157 | err = -ENODEV;
|
| 158 | goto out;
|
| 159 | }
|
| 160 |
|
| 161 | bio_get(bio);
|
| 162 | atomic_inc(&device->md_io.in_use);
|
| 163 | device->md_io.submit_jif = jiffies;
|
| 164 | if (drbd_insert_fault(device, (op == REQ_OP_WRITE) ? DRBD_FAULT_MD_WR : DRBD_FAULT_MD_RD))
|
| 165 | bio_io_error(bio);
|
| 166 | else
|
| 167 | submit_bio(bio);
|
| 168 | wait_until_done_or_force_detached(device, bdev, &device->md_io.done);
|
| 169 | if (!bio->bi_status)
|
| 170 | err = device->md_io.error;
|
| 171 |
|
| 172 | out:
|
| 173 | bio_put(bio);
|
| 174 | return err;
|
| 175 | }
|
| 176 |
|
| 177 | int drbd_md_sync_page_io(struct drbd_device *device, struct drbd_backing_dev *bdev,
|
| 178 | sector_t sector, enum req_op op)
|
| 179 | {
|
| 180 | int err;
|
| 181 | D_ASSERT(device, atomic_read(&device->md_io.in_use) == 1);
|
| 182 |
|
| 183 | BUG_ON(!bdev->md_bdev);
|
| 184 |
|
| 185 | dynamic_drbd_dbg(device, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
|
| 186 | current->comm, current->pid, __func__,
|
| 187 | (unsigned long long)sector, (op == REQ_OP_WRITE) ? "WRITE" : "READ",
|
| 188 | (void*)_RET_IP_ );
|
| 189 |
|
| 190 | if (sector < drbd_md_first_sector(bdev) ||
|
| 191 | sector + 7 > drbd_md_last_sector(bdev))
|
| 192 | drbd_alert(device, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
|
| 193 | current->comm, current->pid, __func__,
|
| 194 | (unsigned long long)sector,
|
| 195 | (op == REQ_OP_WRITE) ? "WRITE" : "READ");
|
| 196 |
|
| 197 | err = _drbd_md_sync_page_io(device, bdev, sector, op);
|
| 198 | if (err) {
|
| 199 | drbd_err(device, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
|
| 200 | (unsigned long long)sector,
|
| 201 | (op == REQ_OP_WRITE) ? "WRITE" : "READ", err);
|
| 202 | }
|
| 203 | return err;
|
| 204 | }
|
| 205 |
|
| 206 | static struct bm_extent *find_active_resync_extent(struct drbd_device *device, unsigned int enr)
|
| 207 | {
|
| 208 | struct lc_element *tmp;
|
| 209 | tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
|
| 210 | if (unlikely(tmp != NULL)) {
|
| 211 | struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce);
|
| 212 | if (test_bit(BME_NO_WRITES, &bm_ext->flags))
|
| 213 | return bm_ext;
|
| 214 | }
|
| 215 | return NULL;
|
| 216 | }
|
| 217 |
|
| 218 | static struct lc_element *_al_get(struct drbd_device *device, unsigned int enr, bool nonblock)
|
| 219 | {
|
| 220 | struct lc_element *al_ext;
|
| 221 | struct bm_extent *bm_ext;
|
| 222 | int wake;
|
| 223 |
|
| 224 | spin_lock_irq(&device->al_lock);
|
| 225 | bm_ext = find_active_resync_extent(device, enr);
|
| 226 | if (bm_ext) {
|
| 227 | wake = !test_and_set_bit(BME_PRIORITY, &bm_ext->flags);
|
| 228 | spin_unlock_irq(&device->al_lock);
|
| 229 | if (wake)
|
| 230 | wake_up(&device->al_wait);
|
| 231 | return NULL;
|
| 232 | }
|
| 233 | if (nonblock)
|
| 234 | al_ext = lc_try_get(device->act_log, enr);
|
| 235 | else
|
| 236 | al_ext = lc_get(device->act_log, enr);
|
| 237 | spin_unlock_irq(&device->al_lock);
|
| 238 | return al_ext;
|
| 239 | }
|
| 240 |
|
| 241 | bool drbd_al_begin_io_fastpath(struct drbd_device *device, struct drbd_interval *i)
|
| 242 | {
|
| 243 | |
| 244 | * we may need to activate two extents in one go
|
| 245 | unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
|
| 246 | unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
|
| 247 |
|
| 248 | D_ASSERT(device, first <= last);
|
| 249 | D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
|
| 250 |
|
| 251 |
|
| 252 | if (first != last)
|
| 253 | return false;
|
| 254 |
|
| 255 | return _al_get(device, first, true);
|
| 256 | }
|
| 257 |
|
| 258 | bool drbd_al_begin_io_prepare(struct drbd_device *device, struct drbd_interval *i)
|
| 259 | {
|
| 260 | |
| 261 | * we may need to activate two extents in one go
|
| 262 | unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
|
| 263 | unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
|
| 264 | unsigned enr;
|
| 265 | bool need_transaction = false;
|
| 266 |
|
| 267 | D_ASSERT(device, first <= last);
|
| 268 | D_ASSERT(device, atomic_read(&device->local_cnt) > 0);
|
| 269 |
|
| 270 | for (enr = first; enr <= last; enr++) {
|
| 271 | struct lc_element *al_ext;
|
| 272 | wait_event(device->al_wait,
|
| 273 | (al_ext = _al_get(device, enr, false)) != NULL);
|
| 274 | if (al_ext->lc_number != enr)
|
| 275 | need_transaction = true;
|
| 276 | }
|
| 277 | return need_transaction;
|
| 278 | }
|
| 279 |
|
| 280 | #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
|
| 281 | |
| 282 | * are still coupled, or assume too much about their relation.
|
| 283 | * Code below will not work if this is violated.
|
| 284 | * Will be cleaned up with some followup patch.
|
| 285 |
|
| 286 | # error FIXME
|
| 287 | #endif
|
| 288 |
|
| 289 | static unsigned int al_extent_to_bm_page(unsigned int al_enr)
|
| 290 | {
|
| 291 | return al_enr >>
|
| 292 |
|
| 293 | ((PAGE_SHIFT + 3) -
|
| 294 |
|
| 295 | (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT));
|
| 296 | }
|
| 297 |
|
| 298 | static sector_t al_tr_number_to_on_disk_sector(struct drbd_device *device)
|
| 299 | {
|
| 300 | const unsigned int stripes = device->ldev->md.al_stripes;
|
| 301 | const unsigned int stripe_size_4kB = device->ldev->md.al_stripe_size_4k;
|
| 302 |
|
| 303 |
|
| 304 | unsigned int t = device->al_tr_number % (device->ldev->md.al_size_4k);
|
| 305 |
|
| 306 |
|
| 307 | t = ((t % stripes) * stripe_size_4kB) + t/stripes;
|
| 308 |
|
| 309 |
|
| 310 | t *= 8;
|
| 311 |
|
| 312 |
|
| 313 | return device->ldev->md.md_offset + device->ldev->md.al_offset + t;
|
| 314 | }
|
| 315 |
|
| 316 | static int __al_write_transaction(struct drbd_device *device, struct al_transaction_on_disk *buffer)
|
| 317 | {
|
| 318 | struct lc_element *e;
|
| 319 | sector_t sector;
|
| 320 | int i, mx;
|
| 321 | unsigned extent_nr;
|
| 322 | unsigned crc = 0;
|
| 323 | int err = 0;
|
| 324 |
|
| 325 | memset(buffer, 0, sizeof(*buffer));
|
| 326 | buffer->magic = cpu_to_be32(DRBD_AL_MAGIC);
|
| 327 | buffer->tr_number = cpu_to_be32(device->al_tr_number);
|
| 328 |
|
| 329 | i = 0;
|
| 330 |
|
| 331 | drbd_bm_reset_al_hints(device);
|
| 332 |
|
| 333 | |
| 334 | * once we set the LC_LOCKED -- from drbd_al_begin_io(),
|
| 335 | * lc_try_lock_for_transaction() --, someone may still
|
| 336 | * be in the process of changing it.
|
| 337 | spin_lock_irq(&device->al_lock);
|
| 338 | list_for_each_entry(e, &device->act_log->to_be_changed, list) {
|
| 339 | if (i == AL_UPDATES_PER_TRANSACTION) {
|
| 340 | i++;
|
| 341 | break;
|
| 342 | }
|
| 343 | buffer->update_slot_nr[i] = cpu_to_be16(e->lc_index);
|
| 344 | buffer->update_extent_nr[i] = cpu_to_be32(e->lc_new_number);
|
| 345 | if (e->lc_number != LC_FREE)
|
| 346 | drbd_bm_mark_for_writeout(device,
|
| 347 | al_extent_to_bm_page(e->lc_number));
|
| 348 | i++;
|
| 349 | }
|
| 350 | spin_unlock_irq(&device->al_lock);
|
| 351 | BUG_ON(i > AL_UPDATES_PER_TRANSACTION);
|
| 352 |
|
| 353 | buffer->n_updates = cpu_to_be16(i);
|
| 354 | for ( ; i < AL_UPDATES_PER_TRANSACTION; i++) {
|
| 355 | buffer->update_slot_nr[i] = cpu_to_be16(-1);
|
| 356 | buffer->update_extent_nr[i] = cpu_to_be32(LC_FREE);
|
| 357 | }
|
| 358 |
|
| 359 | buffer->context_size = cpu_to_be16(device->act_log->nr_elements);
|
| 360 | buffer->context_start_slot_nr = cpu_to_be16(device->al_tr_cycle);
|
| 361 |
|
| 362 | mx = min_t(int, AL_CONTEXT_PER_TRANSACTION,
|
| 363 | device->act_log->nr_elements - device->al_tr_cycle);
|
| 364 | for (i = 0; i < mx; i++) {
|
| 365 | unsigned idx = device->al_tr_cycle + i;
|
| 366 | extent_nr = lc_element_by_index(device->act_log, idx)->lc_number;
|
| 367 | buffer->context[i] = cpu_to_be32(extent_nr);
|
| 368 | }
|
| 369 | for (; i < AL_CONTEXT_PER_TRANSACTION; i++)
|
| 370 | buffer->context[i] = cpu_to_be32(LC_FREE);
|
| 371 |
|
| 372 | device->al_tr_cycle += AL_CONTEXT_PER_TRANSACTION;
|
| 373 | if (device->al_tr_cycle >= device->act_log->nr_elements)
|
| 374 | device->al_tr_cycle = 0;
|
| 375 |
|
| 376 | sector = al_tr_number_to_on_disk_sector(device);
|
| 377 |
|
| 378 | crc = crc32c(0, buffer, 4096);
|
| 379 | buffer->crc32c = cpu_to_be32(crc);
|
| 380 |
|
| 381 | if (drbd_bm_write_hinted(device))
|
| 382 | err = -EIO;
|
| 383 | else {
|
| 384 | bool write_al_updates;
|
| 385 | rcu_read_lock();
|
| 386 | write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
|
| 387 | rcu_read_unlock();
|
| 388 | if (write_al_updates) {
|
| 389 | if (drbd_md_sync_page_io(device, device->ldev, sector, REQ_OP_WRITE)) {
|
| 390 | err = -EIO;
|
| 391 | drbd_chk_io_error(device, 1, DRBD_META_IO_ERROR);
|
| 392 | } else {
|
| 393 | device->al_tr_number++;
|
| 394 | device->al_writ_cnt++;
|
| 395 | }
|
| 396 | }
|
| 397 | }
|
| 398 |
|
| 399 | return err;
|
| 400 | }
|
| 401 |
|
| 402 | static int al_write_transaction(struct drbd_device *device)
|
| 403 | {
|
| 404 | struct al_transaction_on_disk *buffer;
|
| 405 | int err;
|
| 406 |
|
| 407 | if (!get_ldev(device)) {
|
| 408 | drbd_err(device, "disk is %s, cannot start al transaction\n",
|
| 409 | drbd_disk_str(device->state.disk));
|
| 410 | return -EIO;
|
| 411 | }
|
| 412 |
|
| 413 |
|
| 414 | if (device->state.disk < D_INCONSISTENT) {
|
| 415 | drbd_err(device,
|
| 416 | "disk is %s, cannot write al transaction\n",
|
| 417 | drbd_disk_str(device->state.disk));
|
| 418 | put_ldev(device);
|
| 419 | return -EIO;
|
| 420 | }
|
| 421 |
|
| 422 |
|
| 423 | buffer = drbd_md_get_buffer(device, __func__);
|
| 424 | if (!buffer) {
|
| 425 | drbd_err(device, "disk failed while waiting for md_io buffer\n");
|
| 426 | put_ldev(device);
|
| 427 | return -ENODEV;
|
| 428 | }
|
| 429 |
|
| 430 | err = __al_write_transaction(device, buffer);
|
| 431 |
|
| 432 | drbd_md_put_buffer(device);
|
| 433 | put_ldev(device);
|
| 434 |
|
| 435 | return err;
|
| 436 | }
|
| 437 |
|
| 438 |
|
| 439 | void drbd_al_begin_io_commit(struct drbd_device *device)
|
| 440 | {
|
| 441 | bool locked = false;
|
| 442 |
|
| 443 | |
| 444 | * This uses test_and_set_bit, memory barrier is implicit.
|
| 445 |
|
| 446 | wait_event(device->al_wait,
|
| 447 | device->act_log->pending_changes == 0 ||
|
| 448 | (locked = lc_try_lock_for_transaction(device->act_log)));
|
| 449 |
|
| 450 | if (locked) {
|
| 451 | |
| 452 | * while we have been waiting for the lock.
|
| 453 | if (device->act_log->pending_changes) {
|
| 454 | bool write_al_updates;
|
| 455 |
|
| 456 | rcu_read_lock();
|
| 457 | write_al_updates = rcu_dereference(device->ldev->disk_conf)->al_updates;
|
| 458 | rcu_read_unlock();
|
| 459 |
|
| 460 | if (write_al_updates)
|
| 461 | al_write_transaction(device);
|
| 462 | spin_lock_irq(&device->al_lock);
|
| 463 | |
| 464 | if (err)
|
| 465 | we need an "lc_cancel" here;
|
| 466 |
|
| 467 | lc_committed(device->act_log);
|
| 468 | spin_unlock_irq(&device->al_lock);
|
| 469 | }
|
| 470 | lc_unlock(device->act_log);
|
| 471 | wake_up(&device->al_wait);
|
| 472 | }
|
| 473 | }
|
| 474 |
|
| 475 | |
| 476 | * @delegate: delegate activity log I/O to the worker thread
|
| 477 |
|
| 478 | void drbd_al_begin_io(struct drbd_device *device, struct drbd_interval *i)
|
| 479 | {
|
| 480 | if (drbd_al_begin_io_prepare(device, i))
|
| 481 | drbd_al_begin_io_commit(device);
|
| 482 | }
|
| 483 |
|
| 484 | int drbd_al_begin_io_nonblock(struct drbd_device *device, struct drbd_interval *i)
|
| 485 | {
|
| 486 | |
| 487 | * we may need to activate two extents in one go
|
| 488 | unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
|
| 489 | unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
|
| 490 | unsigned enr;
|
| 491 |
|
| 492 | if (i->partially_in_al_next_enr) {
|
| 493 | D_ASSERT(device, first < i->partially_in_al_next_enr);
|
| 494 | D_ASSERT(device, last >= i->partially_in_al_next_enr);
|
| 495 | first = i->partially_in_al_next_enr;
|
| 496 | }
|
| 497 |
|
| 498 | D_ASSERT(device, first <= last);
|
| 499 |
|
| 500 |
|
| 501 | for (enr = first; enr <= last; enr++) {
|
| 502 | struct lc_element *tmp;
|
| 503 | tmp = lc_find(device->resync, enr/AL_EXT_PER_BM_SECT);
|
| 504 | if (unlikely(tmp != NULL)) {
|
| 505 | struct bm_extent *bm_ext = lc_entry(tmp, struct bm_extent, lce);
|
| 506 | if (test_bit(BME_NO_WRITES, &bm_ext->flags)) {
|
| 507 | if (!test_and_set_bit(BME_PRIORITY, &bm_ext->flags))
|
| 508 | return -EBUSY;
|
| 509 | return -EWOULDBLOCK;
|
| 510 | }
|
| 511 | }
|
| 512 | }
|
| 513 |
|
| 514 |
|
| 515 | for (enr = first; enr <= last; enr++) {
|
| 516 | struct lc_element *al_ext;
|
| 517 | al_ext = lc_get_cumulative(device->act_log, enr);
|
| 518 |
|
| 519 | if (!al_ext) {
|
| 520 | |
| 521 | * changes per transaction. Or raced with someone
|
| 522 | * "locking" it against changes.
|
| 523 | * Remember where to continue from.
|
| 524 |
|
| 525 | if (enr > first)
|
| 526 | i->partially_in_al_next_enr = enr;
|
| 527 | return -ENOBUFS;
|
| 528 | }
|
| 529 | }
|
| 530 | return 0;
|
| 531 | }
|
| 532 |
|
| 533 | void drbd_al_complete_io(struct drbd_device *device, struct drbd_interval *i)
|
| 534 | {
|
| 535 | |
| 536 | * we may need to activate two extents in one go
|
| 537 | unsigned first = i->sector >> (AL_EXTENT_SHIFT-9);
|
| 538 | unsigned last = i->size == 0 ? first : (i->sector + (i->size >> 9) - 1) >> (AL_EXTENT_SHIFT-9);
|
| 539 | unsigned enr;
|
| 540 | struct lc_element *extent;
|
| 541 | unsigned long flags;
|
| 542 |
|
| 543 | D_ASSERT(device, first <= last);
|
| 544 | spin_lock_irqsave(&device->al_lock, flags);
|
| 545 |
|
| 546 | for (enr = first; enr <= last; enr++) {
|
| 547 | extent = lc_find(device->act_log, enr);
|
| 548 | |
| 549 | * operation this is harmless, so no need to crash the kernel
|
| 550 | * by the BUG_ON(refcount == 0) in lc_put().
|
| 551 |
|
| 552 | if (!extent || extent->refcnt == 0) {
|
| 553 | drbd_err(device, "al_complete_io() called on inactive extent %u\n", enr);
|
| 554 | continue;
|
| 555 | }
|
| 556 | lc_put(device->act_log, extent);
|
| 557 | }
|
| 558 | spin_unlock_irqrestore(&device->al_lock, flags);
|
| 559 | wake_up(&device->al_wait);
|
| 560 | }
|
| 561 |
|
| 562 | static int _try_lc_del(struct drbd_device *device, struct lc_element *al_ext)
|
| 563 | {
|
| 564 | int rv;
|
| 565 |
|
| 566 | spin_lock_irq(&device->al_lock);
|
| 567 | rv = (al_ext->refcnt == 0);
|
| 568 | if (likely(rv))
|
| 569 | lc_del(device->act_log, al_ext);
|
| 570 | spin_unlock_irq(&device->al_lock);
|
| 571 |
|
| 572 | return rv;
|
| 573 | }
|
| 574 |
|
| 575 | |
| 576 | * drbd_al_shrink() - Removes all active extents form the activity log
|
| 577 | * @device: DRBD device.
|
| 578 | *
|
| 579 | * Removes all active extents form the activity log, waiting until
|
| 580 | * the reference count of each entry dropped to 0 first, of course.
|
| 581 | *
|
| 582 | * You need to lock device->act_log with lc_try_lock() / lc_unlock()
|
| 583 |
|
| 584 | void drbd_al_shrink(struct drbd_device *device)
|
| 585 | {
|
| 586 | struct lc_element *al_ext;
|
| 587 | int i;
|
| 588 |
|
| 589 | D_ASSERT(device, test_bit(__LC_LOCKED, &device->act_log->flags));
|
| 590 |
|
| 591 | for (i = 0; i < device->act_log->nr_elements; i++) {
|
| 592 | al_ext = lc_element_by_index(device->act_log, i);
|
| 593 | if (al_ext->lc_number == LC_FREE)
|
| 594 | continue;
|
| 595 | wait_event(device->al_wait, _try_lc_del(device, al_ext));
|
| 596 | }
|
| 597 |
|
| 598 | wake_up(&device->al_wait);
|
| 599 | }
|
| 600 |
|
| 601 | int drbd_al_initialize(struct drbd_device *device, void *buffer)
|
| 602 | {
|
| 603 | struct al_transaction_on_disk *al = buffer;
|
| 604 | struct drbd_md *md = &device->ldev->md;
|
| 605 | int al_size_4k = md->al_stripes * md->al_stripe_size_4k;
|
| 606 | int i;
|
| 607 |
|
| 608 | __al_write_transaction(device, al);
|
| 609 |
|
| 610 | spin_lock_irq(&device->al_lock);
|
| 611 | lc_committed(device->act_log);
|
| 612 | spin_unlock_irq(&device->al_lock);
|
| 613 |
|
| 614 | |
| 615 | * are written out only to provide the context, and to initialize the
|
| 616 | * on-disk ring buffer.
|
| 617 | for (i = 1; i < al_size_4k; i++) {
|
| 618 | int err = __al_write_transaction(device, al);
|
| 619 | if (err)
|
| 620 | return err;
|
| 621 | }
|
| 622 | return 0;
|
| 623 | }
|
| 624 |
|
| 625 | static const char *drbd_change_sync_fname[] = {
|
| 626 | [RECORD_RS_FAILED] = "drbd_rs_failed_io",
|
| 627 | [SET_IN_SYNC] = "drbd_set_in_sync",
|
| 628 | [SET_OUT_OF_SYNC] = "drbd_set_out_of_sync"
|
| 629 | };
|
| 630 |
|
| 631 | |
| 632 | * resync LRU-cache are 16MB each.
|
| 633 | * The caller of this function has to hold an get_ldev() reference.
|
| 634 | *
|
| 635 | * Adjusts the caching members ->rs_left (success) or ->rs_failed (!success),
|
| 636 | * potentially pulling in (and recounting the corresponding bits)
|
| 637 | * this resync extent into the resync extent lru cache.
|
| 638 | *
|
| 639 | * Returns whether all bits have been cleared for this resync extent,
|
| 640 | * precisely: (rs_left <= rs_failed)
|
| 641 | *
|
| 642 | * TODO will be obsoleted once we have a caching lru of the on disk bitmap
|
| 643 |
|
| 644 | static bool update_rs_extent(struct drbd_device *device,
|
| 645 | unsigned int enr, int count,
|
| 646 | enum update_sync_bits_mode mode)
|
| 647 | {
|
| 648 | struct lc_element *e;
|
| 649 |
|
| 650 | D_ASSERT(device, atomic_read(&device->local_cnt));
|
| 651 |
|
| 652 | |
| 653 | * we don't need it cached (lc_find).
|
| 654 | * But if it is present in the cache,
|
| 655 | * we should update the cached bit count.
|
| 656 | * Otherwise, that extent should be in the resync extent lru cache
|
| 657 | * already -- or we want to pull it in if necessary -- (lc_get),
|
| 658 | * then update and check rs_left and rs_failed.
|
| 659 | if (mode == SET_OUT_OF_SYNC)
|
| 660 | e = lc_find(device->resync, enr);
|
| 661 | else
|
| 662 | e = lc_get(device->resync, enr);
|
| 663 | if (e) {
|
| 664 | struct bm_extent *ext = lc_entry(e, struct bm_extent, lce);
|
| 665 | if (ext->lce.lc_number == enr) {
|
| 666 | if (mode == SET_IN_SYNC)
|
| 667 | ext->rs_left -= count;
|
| 668 | else if (mode == SET_OUT_OF_SYNC)
|
| 669 | ext->rs_left += count;
|
| 670 | else
|
| 671 | ext->rs_failed += count;
|
| 672 | if (ext->rs_left < ext->rs_failed) {
|
| 673 | drbd_warn(device, "BAD! enr=%u rs_left=%d "
|
| 674 | "rs_failed=%d count=%d cstate=%s\n",
|
| 675 | ext->lce.lc_number, ext->rs_left,
|
| 676 | ext->rs_failed, count,
|
| 677 | drbd_conn_str(device->state.conn));
|
| 678 |
|
| 679 | |
| 680 | * than have been set when we originally counted
|
| 681 | * the set bits to cache that value in ext->rs_left.
|
| 682 | * Whatever the reason (disconnect during resync,
|
| 683 | * delayed local completion of an application write),
|
| 684 | * try to fix it up by recounting here.
|
| 685 | ext->rs_left = drbd_bm_e_weight(device, enr);
|
| 686 | }
|
| 687 | } else {
|
| 688 | |
| 689 | * since drbd_rs_begin_io() pulled it already in.
|
| 690 | *
|
| 691 | * But maybe an application write finished, and we set
|
| 692 | * something outside the resync lru_cache in sync.
|
| 693 |
|
| 694 | int rs_left = drbd_bm_e_weight(device, enr);
|
| 695 | if (ext->flags != 0) {
|
| 696 | drbd_warn(device, "changing resync lce: %d[%u;%02lx]"
|
| 697 | " -> %d[%u;00]\n",
|
| 698 | ext->lce.lc_number, ext->rs_left,
|
| 699 | ext->flags, enr, rs_left);
|
| 700 | ext->flags = 0;
|
| 701 | }
|
| 702 | if (ext->rs_failed) {
|
| 703 | drbd_warn(device, "Kicking resync_lru element enr=%u "
|
| 704 | "out with rs_failed=%d\n",
|
| 705 | ext->lce.lc_number, ext->rs_failed);
|
| 706 | }
|
| 707 | ext->rs_left = rs_left;
|
| 708 | ext->rs_failed = (mode == RECORD_RS_FAILED) ? count : 0;
|
| 709 | |
| 710 | * we can commit any change right away.
|
| 711 | lc_committed(device->resync);
|
| 712 | }
|
| 713 | if (mode != SET_OUT_OF_SYNC)
|
| 714 | lc_put(device->resync, &ext->lce);
|
| 715 |
|
| 716 |
|
| 717 | if (ext->rs_left <= ext->rs_failed) {
|
| 718 | ext->rs_failed = 0;
|
| 719 | return true;
|
| 720 | }
|
| 721 | } else if (mode != SET_OUT_OF_SYNC) {
|
| 722 |
|
| 723 | drbd_err(device, "lc_get() failed! locked=%d/%d flags=%lu\n",
|
| 724 | device->resync_locked,
|
| 725 | device->resync->nr_elements,
|
| 726 | device->resync->flags);
|
| 727 | }
|
| 728 | return false;
|
| 729 | }
|
| 730 |
|
| 731 | void drbd_advance_rs_marks(struct drbd_peer_device *peer_device, unsigned long still_to_go)
|
| 732 | {
|
| 733 | struct drbd_device *device = peer_device->device;
|
| 734 | unsigned long now = jiffies;
|
| 735 | unsigned long last = device->rs_mark_time[device->rs_last_mark];
|
| 736 | int next = (device->rs_last_mark + 1) % DRBD_SYNC_MARKS;
|
| 737 | if (time_after_eq(now, last + DRBD_SYNC_MARK_STEP)) {
|
| 738 | if (device->rs_mark_left[device->rs_last_mark] != still_to_go &&
|
| 739 | device->state.conn != C_PAUSED_SYNC_T &&
|
| 740 | device->state.conn != C_PAUSED_SYNC_S) {
|
| 741 | device->rs_mark_time[next] = now;
|
| 742 | device->rs_mark_left[next] = still_to_go;
|
| 743 | device->rs_last_mark = next;
|
| 744 | }
|
| 745 | }
|
| 746 | }
|
| 747 |
|
| 748 |
|
| 749 | static bool lazy_bitmap_update_due(struct drbd_device *device)
|
| 750 | {
|
| 751 | return time_after(jiffies, device->rs_last_bcast + 2*HZ);
|
| 752 | }
|
| 753 |
|
| 754 | static void maybe_schedule_on_disk_bitmap_update(struct drbd_device *device, bool rs_done)
|
| 755 | {
|
| 756 | if (rs_done) {
|
| 757 | struct drbd_connection *connection = first_peer_device(device)->connection;
|
| 758 | if (connection->agreed_pro_version <= 95 ||
|
| 759 | is_sync_target_state(device->state.conn))
|
| 760 | set_bit(RS_DONE, &device->flags);
|
| 761 |
|
| 762 |
|
| 763 | |
| 764 | * to avoid uuids-rotated-too-fast causing full resync
|
| 765 | * in next handshake, in case the replication link breaks
|
| 766 | * at the most unfortunate time...
|
| 767 | } else if (!lazy_bitmap_update_due(device))
|
| 768 | return;
|
| 769 |
|
| 770 | drbd_device_post_work(device, RS_PROGRESS);
|
| 771 | }
|
| 772 |
|
| 773 | static int update_sync_bits(struct drbd_device *device,
|
| 774 | unsigned long sbnr, unsigned long ebnr,
|
| 775 | enum update_sync_bits_mode mode)
|
| 776 | {
|
| 777 | |
| 778 | * We keep a count of set bits per resync-extent in the ->rs_left
|
| 779 | * caching member, so we need to loop and work within the resync extent
|
| 780 | * alignment. Typically this loop will execute exactly once.
|
| 781 |
|
| 782 | unsigned long flags;
|
| 783 | unsigned long count = 0;
|
| 784 | unsigned int cleared = 0;
|
| 785 | while (sbnr <= ebnr) {
|
| 786 | |
| 787 | * the resync extent of the current start bit number,
|
| 788 | * but cap at provided end bit number
|
| 789 | unsigned long tbnr = min(ebnr, sbnr | BM_BLOCKS_PER_BM_EXT_MASK);
|
| 790 | unsigned long c;
|
| 791 |
|
| 792 | if (mode == RECORD_RS_FAILED)
|
| 793 | |
| 794 | * supposedly still set. Recount, maybe some
|
| 795 | * of the bits have been successfully cleared
|
| 796 | * by application IO meanwhile.
|
| 797 |
|
| 798 | c = drbd_bm_count_bits(device, sbnr, tbnr);
|
| 799 | else if (mode == SET_IN_SYNC)
|
| 800 | c = drbd_bm_clear_bits(device, sbnr, tbnr);
|
| 801 | else
|
| 802 | c = drbd_bm_set_bits(device, sbnr, tbnr);
|
| 803 |
|
| 804 | if (c) {
|
| 805 | spin_lock_irqsave(&device->al_lock, flags);
|
| 806 | cleared += update_rs_extent(device, BM_BIT_TO_EXT(sbnr), c, mode);
|
| 807 | spin_unlock_irqrestore(&device->al_lock, flags);
|
| 808 | count += c;
|
| 809 | }
|
| 810 | sbnr = tbnr + 1;
|
| 811 | }
|
| 812 | if (count) {
|
| 813 | if (mode == SET_IN_SYNC) {
|
| 814 | unsigned long still_to_go = drbd_bm_total_weight(device);
|
| 815 | bool rs_is_done = (still_to_go <= device->rs_failed);
|
| 816 | drbd_advance_rs_marks(first_peer_device(device), still_to_go);
|
| 817 | if (cleared || rs_is_done)
|
| 818 | maybe_schedule_on_disk_bitmap_update(device, rs_is_done);
|
| 819 | } else if (mode == RECORD_RS_FAILED)
|
| 820 | device->rs_failed += count;
|
| 821 | wake_up(&device->al_wait);
|
| 822 | }
|
| 823 | return count;
|
| 824 | }
|
| 825 |
|
| 826 | static bool plausible_request_size(int size)
|
| 827 | {
|
| 828 | return size > 0
|
| 829 | && size <= DRBD_MAX_BATCH_BIO_SIZE
|
| 830 | && IS_ALIGNED(size, 512);
|
| 831 | }
|
| 832 |
|
| 833 | |
| 834 | * size byte of data starting from sector. Only clear bits of the affected
|
| 835 | * one or more _aligned_ BM_BLOCK_SIZE blocks.
|
| 836 | *
|
| 837 | * called by worker on C_SYNC_TARGET and receiver on SyncSource.
|
| 838 | *
|
| 839 |
|
| 840 | int __drbd_change_sync(struct drbd_peer_device *peer_device, sector_t sector, int size,
|
| 841 | enum update_sync_bits_mode mode)
|
| 842 | {
|
| 843 |
|
| 844 | struct drbd_device *device = peer_device->device;
|
| 845 | unsigned long sbnr, ebnr, lbnr;
|
| 846 | unsigned long count = 0;
|
| 847 | sector_t esector, nr_sectors;
|
| 848 |
|
| 849 |
|
| 850 | if ((mode == SET_OUT_OF_SYNC) && size == 0)
|
| 851 | return 0;
|
| 852 |
|
| 853 | if (!plausible_request_size(size)) {
|
| 854 | drbd_err(device, "%s: sector=%llus size=%d nonsense!\n",
|
| 855 | drbd_change_sync_fname[mode],
|
| 856 | (unsigned long long)sector, size);
|
| 857 | return 0;
|
| 858 | }
|
| 859 |
|
| 860 | if (!get_ldev(device))
|
| 861 | return 0;
|
| 862 |
|
| 863 | nr_sectors = get_capacity(device->vdisk);
|
| 864 | esector = sector + (size >> 9) - 1;
|
| 865 |
|
| 866 | if (!expect(device, sector < nr_sectors))
|
| 867 | goto out;
|
| 868 | if (!expect(device, esector < nr_sectors))
|
| 869 | esector = nr_sectors - 1;
|
| 870 |
|
| 871 | lbnr = BM_SECT_TO_BIT(nr_sectors-1);
|
| 872 |
|
| 873 | if (mode == SET_IN_SYNC) {
|
| 874 | |
| 875 | * we only clear full, aligned, BM_BLOCK_SIZE blocks.
|
| 876 | if (unlikely(esector < BM_SECT_PER_BIT-1))
|
| 877 | goto out;
|
| 878 | if (unlikely(esector == (nr_sectors-1)))
|
| 879 | ebnr = lbnr;
|
| 880 | else
|
| 881 | ebnr = BM_SECT_TO_BIT(esector - (BM_SECT_PER_BIT-1));
|
| 882 | sbnr = BM_SECT_TO_BIT(sector + BM_SECT_PER_BIT-1);
|
| 883 | } else {
|
| 884 | |
| 885 | * Should not round anything here.
|
| 886 | sbnr = BM_SECT_TO_BIT(sector);
|
| 887 | ebnr = BM_SECT_TO_BIT(esector);
|
| 888 | }
|
| 889 |
|
| 890 | count = update_sync_bits(device, sbnr, ebnr, mode);
|
| 891 | out:
|
| 892 | put_ldev(device);
|
| 893 | return count;
|
| 894 | }
|
| 895 |
|
| 896 | static
|
| 897 | struct bm_extent *_bme_get(struct drbd_device *device, unsigned int enr)
|
| 898 | {
|
| 899 | struct lc_element *e;
|
| 900 | struct bm_extent *bm_ext;
|
| 901 | int wakeup = 0;
|
| 902 | unsigned long rs_flags;
|
| 903 |
|
| 904 | spin_lock_irq(&device->al_lock);
|
| 905 | if (device->resync_locked > device->resync->nr_elements/2) {
|
| 906 | spin_unlock_irq(&device->al_lock);
|
| 907 | return NULL;
|
| 908 | }
|
| 909 | e = lc_get(device->resync, enr);
|
| 910 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
|
| 911 | if (bm_ext) {
|
| 912 | if (bm_ext->lce.lc_number != enr) {
|
| 913 | bm_ext->rs_left = drbd_bm_e_weight(device, enr);
|
| 914 | bm_ext->rs_failed = 0;
|
| 915 | lc_committed(device->resync);
|
| 916 | wakeup = 1;
|
| 917 | }
|
| 918 | if (bm_ext->lce.refcnt == 1)
|
| 919 | device->resync_locked++;
|
| 920 | set_bit(BME_NO_WRITES, &bm_ext->flags);
|
| 921 | }
|
| 922 | rs_flags = device->resync->flags;
|
| 923 | spin_unlock_irq(&device->al_lock);
|
| 924 | if (wakeup)
|
| 925 | wake_up(&device->al_wait);
|
| 926 |
|
| 927 | if (!bm_ext) {
|
| 928 | if (rs_flags & LC_STARVING)
|
| 929 | drbd_warn(device, "Have to wait for element"
|
| 930 | " (resync LRU too small?)\n");
|
| 931 | BUG_ON(rs_flags & LC_LOCKED);
|
| 932 | }
|
| 933 |
|
| 934 | return bm_ext;
|
| 935 | }
|
| 936 |
|
| 937 | static int _is_in_al(struct drbd_device *device, unsigned int enr)
|
| 938 | {
|
| 939 | int rv;
|
| 940 |
|
| 941 | spin_lock_irq(&device->al_lock);
|
| 942 | rv = lc_is_used(device->act_log, enr);
|
| 943 | spin_unlock_irq(&device->al_lock);
|
| 944 |
|
| 945 | return rv;
|
| 946 | }
|
| 947 |
|
| 948 | |
| 949 | * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
|
| 950 | * @device: DRBD device.
|
| 951 | * @sector: The sector number.
|
| 952 | *
|
| 953 | * This functions sleeps on al_wait.
|
| 954 | *
|
| 955 | * Returns: %0 on success, -EINTR if interrupted.
|
| 956 |
|
| 957 | int drbd_rs_begin_io(struct drbd_device *device, sector_t sector)
|
| 958 | {
|
| 959 | unsigned int enr = BM_SECT_TO_EXT(sector);
|
| 960 | struct bm_extent *bm_ext;
|
| 961 | int i, sig;
|
| 962 | bool sa;
|
| 963 |
|
| 964 | retry:
|
| 965 | sig = wait_event_interruptible(device->al_wait,
|
| 966 | (bm_ext = _bme_get(device, enr)));
|
| 967 | if (sig)
|
| 968 | return -EINTR;
|
| 969 |
|
| 970 | if (test_bit(BME_LOCKED, &bm_ext->flags))
|
| 971 | return 0;
|
| 972 |
|
| 973 |
|
| 974 | sa = drbd_rs_c_min_rate_throttle(device);
|
| 975 |
|
| 976 | for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
|
| 977 | sig = wait_event_interruptible(device->al_wait,
|
| 978 | !_is_in_al(device, enr * AL_EXT_PER_BM_SECT + i) ||
|
| 979 | (sa && test_bit(BME_PRIORITY, &bm_ext->flags)));
|
| 980 |
|
| 981 | if (sig || (sa && test_bit(BME_PRIORITY, &bm_ext->flags))) {
|
| 982 | spin_lock_irq(&device->al_lock);
|
| 983 | if (lc_put(device->resync, &bm_ext->lce) == 0) {
|
| 984 | bm_ext->flags = 0;
|
| 985 | device->resync_locked--;
|
| 986 | wake_up(&device->al_wait);
|
| 987 | }
|
| 988 | spin_unlock_irq(&device->al_lock);
|
| 989 | if (sig)
|
| 990 | return -EINTR;
|
| 991 | if (schedule_timeout_interruptible(HZ/10))
|
| 992 | return -EINTR;
|
| 993 | goto retry;
|
| 994 | }
|
| 995 | }
|
| 996 | set_bit(BME_LOCKED, &bm_ext->flags);
|
| 997 | return 0;
|
| 998 | }
|
| 999 |
|
| 1000 | |
| 1001 | * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
|
| 1002 | * @peer_device: DRBD device.
|
| 1003 | * @sector: The sector number.
|
| 1004 | *
|
| 1005 | * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
|
| 1006 | * tries to set it to BME_LOCKED.
|
| 1007 | *
|
| 1008 | * Returns: %0 upon success, and -EAGAIN
|
| 1009 | * if there is still application IO going on in this area.
|
| 1010 |
|
| 1011 | int drbd_try_rs_begin_io(struct drbd_peer_device *peer_device, sector_t sector)
|
| 1012 | {
|
| 1013 | struct drbd_device *device = peer_device->device;
|
| 1014 | unsigned int enr = BM_SECT_TO_EXT(sector);
|
| 1015 | const unsigned int al_enr = enr*AL_EXT_PER_BM_SECT;
|
| 1016 | struct lc_element *e;
|
| 1017 | struct bm_extent *bm_ext;
|
| 1018 | int i;
|
| 1019 | bool throttle = drbd_rs_should_slow_down(peer_device, sector, true);
|
| 1020 |
|
| 1021 | |
| 1022 | * not yet BME_LOCKED) extent needs to be kicked out explicitly if we
|
| 1023 | * need to throttle. There is at most one such half-locked extent,
|
| 1024 | * which is remembered in resync_wenr.
|
| 1025 |
|
| 1026 | if (throttle && device->resync_wenr != enr)
|
| 1027 | return -EAGAIN;
|
| 1028 |
|
| 1029 | spin_lock_irq(&device->al_lock);
|
| 1030 | if (device->resync_wenr != LC_FREE && device->resync_wenr != enr) {
|
| 1031 | |
| 1032 | * stall the syncer undefined if we give up the ref count
|
| 1033 | * when we try again and requeue.
|
| 1034 | *
|
| 1035 | * if we don't give up the refcount, but the next time
|
| 1036 | * we are scheduled this extent has been "synced" by new
|
| 1037 | * application writes, we'd miss the lc_put on the
|
| 1038 | * extent we keep the refcount on.
|
| 1039 | * so we remembered which extent we had to try again, and
|
| 1040 | * if the next requested one is something else, we do
|
| 1041 | * the lc_put here...
|
| 1042 | * we also have to wake_up
|
| 1043 |
|
| 1044 | e = lc_find(device->resync, device->resync_wenr);
|
| 1045 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
|
| 1046 | if (bm_ext) {
|
| 1047 | D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
|
| 1048 | D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
|
| 1049 | clear_bit(BME_NO_WRITES, &bm_ext->flags);
|
| 1050 | device->resync_wenr = LC_FREE;
|
| 1051 | if (lc_put(device->resync, &bm_ext->lce) == 0) {
|
| 1052 | bm_ext->flags = 0;
|
| 1053 | device->resync_locked--;
|
| 1054 | }
|
| 1055 | wake_up(&device->al_wait);
|
| 1056 | } else {
|
| 1057 | drbd_alert(device, "LOGIC BUG\n");
|
| 1058 | }
|
| 1059 | }
|
| 1060 |
|
| 1061 | e = lc_try_get(device->resync, enr);
|
| 1062 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
|
| 1063 | if (bm_ext) {
|
| 1064 | if (test_bit(BME_LOCKED, &bm_ext->flags))
|
| 1065 | goto proceed;
|
| 1066 | if (!test_and_set_bit(BME_NO_WRITES, &bm_ext->flags)) {
|
| 1067 | device->resync_locked++;
|
| 1068 | } else {
|
| 1069 | |
| 1070 | * but then could not set BME_LOCKED,
|
| 1071 | * so we tried again.
|
| 1072 | * drop the extra reference.
|
| 1073 | bm_ext->lce.refcnt--;
|
| 1074 | D_ASSERT(device, bm_ext->lce.refcnt > 0);
|
| 1075 | }
|
| 1076 | goto check_al;
|
| 1077 | } else {
|
| 1078 |
|
| 1079 | if (device->resync_locked > device->resync->nr_elements-3)
|
| 1080 | goto try_again;
|
| 1081 |
|
| 1082 | e = lc_get(device->resync, enr);
|
| 1083 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
|
| 1084 | if (!bm_ext) {
|
| 1085 | const unsigned long rs_flags = device->resync->flags;
|
| 1086 | if (rs_flags & LC_STARVING)
|
| 1087 | drbd_warn(device, "Have to wait for element"
|
| 1088 | " (resync LRU too small?)\n");
|
| 1089 | BUG_ON(rs_flags & LC_LOCKED);
|
| 1090 | goto try_again;
|
| 1091 | }
|
| 1092 | if (bm_ext->lce.lc_number != enr) {
|
| 1093 | bm_ext->rs_left = drbd_bm_e_weight(device, enr);
|
| 1094 | bm_ext->rs_failed = 0;
|
| 1095 | lc_committed(device->resync);
|
| 1096 | wake_up(&device->al_wait);
|
| 1097 | D_ASSERT(device, test_bit(BME_LOCKED, &bm_ext->flags) == 0);
|
| 1098 | }
|
| 1099 | set_bit(BME_NO_WRITES, &bm_ext->flags);
|
| 1100 | D_ASSERT(device, bm_ext->lce.refcnt == 1);
|
| 1101 | device->resync_locked++;
|
| 1102 | goto check_al;
|
| 1103 | }
|
| 1104 | check_al:
|
| 1105 | for (i = 0; i < AL_EXT_PER_BM_SECT; i++) {
|
| 1106 | if (lc_is_used(device->act_log, al_enr+i))
|
| 1107 | goto try_again;
|
| 1108 | }
|
| 1109 | set_bit(BME_LOCKED, &bm_ext->flags);
|
| 1110 | proceed:
|
| 1111 | device->resync_wenr = LC_FREE;
|
| 1112 | spin_unlock_irq(&device->al_lock);
|
| 1113 | return 0;
|
| 1114 |
|
| 1115 | try_again:
|
| 1116 | if (bm_ext) {
|
| 1117 | if (throttle) {
|
| 1118 | D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
|
| 1119 | D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
|
| 1120 | clear_bit(BME_NO_WRITES, &bm_ext->flags);
|
| 1121 | device->resync_wenr = LC_FREE;
|
| 1122 | if (lc_put(device->resync, &bm_ext->lce) == 0) {
|
| 1123 | bm_ext->flags = 0;
|
| 1124 | device->resync_locked--;
|
| 1125 | }
|
| 1126 | wake_up(&device->al_wait);
|
| 1127 | } else
|
| 1128 | device->resync_wenr = enr;
|
| 1129 | }
|
| 1130 | spin_unlock_irq(&device->al_lock);
|
| 1131 | return -EAGAIN;
|
| 1132 | }
|
| 1133 |
|
| 1134 | void drbd_rs_complete_io(struct drbd_device *device, sector_t sector)
|
| 1135 | {
|
| 1136 | unsigned int enr = BM_SECT_TO_EXT(sector);
|
| 1137 | struct lc_element *e;
|
| 1138 | struct bm_extent *bm_ext;
|
| 1139 | unsigned long flags;
|
| 1140 |
|
| 1141 | spin_lock_irqsave(&device->al_lock, flags);
|
| 1142 | e = lc_find(device->resync, enr);
|
| 1143 | bm_ext = e ? lc_entry(e, struct bm_extent, lce) : NULL;
|
| 1144 | if (!bm_ext) {
|
| 1145 | spin_unlock_irqrestore(&device->al_lock, flags);
|
| 1146 | if (drbd_ratelimit())
|
| 1147 | drbd_err(device, "drbd_rs_complete_io() called, but extent not found\n");
|
| 1148 | return;
|
| 1149 | }
|
| 1150 |
|
| 1151 | if (bm_ext->lce.refcnt == 0) {
|
| 1152 | spin_unlock_irqrestore(&device->al_lock, flags);
|
| 1153 | drbd_err(device, "drbd_rs_complete_io(,%llu [=%u]) called, "
|
| 1154 | "but refcnt is 0!?\n",
|
| 1155 | (unsigned long long)sector, enr);
|
| 1156 | return;
|
| 1157 | }
|
| 1158 |
|
| 1159 | if (lc_put(device->resync, &bm_ext->lce) == 0) {
|
| 1160 | bm_ext->flags = 0;
|
| 1161 | device->resync_locked--;
|
| 1162 | wake_up(&device->al_wait);
|
| 1163 | }
|
| 1164 |
|
| 1165 | spin_unlock_irqrestore(&device->al_lock, flags);
|
| 1166 | }
|
| 1167 |
|
| 1168 | |
| 1169 | * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
|
| 1170 | * @device: DRBD device.
|
| 1171 |
|
| 1172 | void drbd_rs_cancel_all(struct drbd_device *device)
|
| 1173 | {
|
| 1174 | spin_lock_irq(&device->al_lock);
|
| 1175 |
|
| 1176 | if (get_ldev_if_state(device, D_FAILED)) {
|
| 1177 | lc_reset(device->resync);
|
| 1178 | put_ldev(device);
|
| 1179 | }
|
| 1180 | device->resync_locked = 0;
|
| 1181 | device->resync_wenr = LC_FREE;
|
| 1182 | spin_unlock_irq(&device->al_lock);
|
| 1183 | wake_up(&device->al_wait);
|
| 1184 | }
|
| 1185 |
|
| 1186 | |
| 1187 | * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
|
| 1188 | * @device: DRBD device.
|
| 1189 | *
|
| 1190 | * Returns: %0 upon success, -EAGAIN if at least one reference count was
|
| 1191 | * not zero.
|
| 1192 |
|
| 1193 | int drbd_rs_del_all(struct drbd_device *device)
|
| 1194 | {
|
| 1195 | struct lc_element *e;
|
| 1196 | struct bm_extent *bm_ext;
|
| 1197 | int i;
|
| 1198 |
|
| 1199 | spin_lock_irq(&device->al_lock);
|
| 1200 |
|
| 1201 | if (get_ldev_if_state(device, D_FAILED)) {
|
| 1202 |
|
| 1203 | for (i = 0; i < device->resync->nr_elements; i++) {
|
| 1204 | e = lc_element_by_index(device->resync, i);
|
| 1205 | bm_ext = lc_entry(e, struct bm_extent, lce);
|
| 1206 | if (bm_ext->lce.lc_number == LC_FREE)
|
| 1207 | continue;
|
| 1208 | if (bm_ext->lce.lc_number == device->resync_wenr) {
|
| 1209 | drbd_info(device, "dropping %u in drbd_rs_del_all, apparently"
|
| 1210 | " got 'synced' by application io\n",
|
| 1211 | device->resync_wenr);
|
| 1212 | D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
|
| 1213 | D_ASSERT(device, test_bit(BME_NO_WRITES, &bm_ext->flags));
|
| 1214 | clear_bit(BME_NO_WRITES, &bm_ext->flags);
|
| 1215 | device->resync_wenr = LC_FREE;
|
| 1216 | lc_put(device->resync, &bm_ext->lce);
|
| 1217 | }
|
| 1218 | if (bm_ext->lce.refcnt != 0) {
|
| 1219 | drbd_info(device, "Retrying drbd_rs_del_all() later. "
|
| 1220 | "refcnt=%d\n", bm_ext->lce.refcnt);
|
| 1221 | put_ldev(device);
|
| 1222 | spin_unlock_irq(&device->al_lock);
|
| 1223 | return -EAGAIN;
|
| 1224 | }
|
| 1225 | D_ASSERT(device, !test_bit(BME_LOCKED, &bm_ext->flags));
|
| 1226 | D_ASSERT(device, !test_bit(BME_NO_WRITES, &bm_ext->flags));
|
| 1227 | lc_del(device->resync, &bm_ext->lce);
|
| 1228 | }
|
| 1229 | D_ASSERT(device, device->resync->used == 0);
|
| 1230 | put_ldev(device);
|
| 1231 | }
|
| 1232 | spin_unlock_irq(&device->al_lock);
|
| 1233 | wake_up(&device->al_wait);
|
| 1234 |
|
| 1235 | return 0;
|
| 1236 | }
|