1.. SPDX-License-Identifier: GPL-2.0
   4XFS Delayed Logging Design
   7Introduction to Re-logging in XFS
  10XFS logging is a combination of logical and physical logging. Some objects,
  11such as inodes and dquots, are logged in logical format where the details
  12logged are made up of the changes to in-core structures rather than on-disk
  13structures. Other objects - typically buffers - have their physical changes
  14logged. The reason for these differences is to reduce the amount of log space
  15required for objects that are frequently logged. Some parts of inodes are more
  16frequently logged than others, and inodes are typically more frequently logged
  17than any other object (except maybe the superblock buffer) so keeping the
  18amount of metadata logged low is of prime importance.
  20The reason that this is such a concern is that XFS allows multiple separate
  21modifications to a single object to be carried in the log at any given time.
  22This allows the log to avoid needing to flush each change to disk before
  23recording a new change to the object. XFS does this via a method called
  24"re-logging". Conceptually, this is quite simple - all it requires is that any
  25new change to the object is recorded with a *new copy* of all the existing
  26changes in the new transaction that is written to the log.
  28That is, if we have a sequence of changes A through to F, and the object was
  29written to disk after change D, we would see in the log the following series
  30of transactions, their contents and the log sequence number (LSN) of the
  33        Transaction             Contents        LSN
  34           A                       A               X
  35           B                      A+B             X+n
  36           C                     A+B+C           X+n+m
  37           D                    A+B+C+D         X+n+m+o
  38            <object written to disk>
  39           E                       E               Y (> X+n+m+o)
  40           F                      E+F             Y+p
  42In other words, each time an object is relogged, the new transaction contains
  43the aggregation of all the previous changes currently held only in the log.
  45This relogging technique also allows objects to be moved forward in the log so
  46that an object being relogged does not prevent the tail of the log from ever
  47moving forward.  This can be seen in the table above by the changing
  48(increasing) LSN of each subsequent transaction - the LSN is effectively a
  49direct encoding of the location in the log of the transaction.
  51This relogging is also used to implement long-running, multiple-commit
  52transactions.  These transaction are known as rolling transactions, and require
  53a special log reservation known as a permanent transaction reservation. A
  54typical example of a rolling transaction is the removal of extents from an
  55inode which can only be done at a rate of two extents per transaction because
  56of reservation size limitations. Hence a rolling extent removal transaction
  57keeps relogging the inode and btree buffers as they get modified in each
  58removal operation. This keeps them moving forward in the log as the operation
  59progresses, ensuring that current operation never gets blocked by itself if the
  60log wraps around.
  62Hence it can be seen that the relogging operation is fundamental to the correct
  63working of the XFS journalling subsystem. From the above description, most
  64people should be able to see why the XFS metadata operations writes so much to
  65the log - repeated operations to the same objects write the same changes to
  66the log over and over again. Worse is the fact that objects tend to get
  67dirtier as they get relogged, so each subsequent transaction is writing more
  68metadata into the log.
  70Another feature of the XFS transaction subsystem is that most transactions are
  71asynchronous. That is, they don't commit to disk until either a log buffer is
  72filled (a log buffer can hold multiple transactions) or a synchronous operation
  73forces the log buffers holding the transactions to disk. This means that XFS is
  74doing aggregation of transactions in memory - batching them, if you like - to
  75minimise the impact of the log IO on transaction throughput.
  77The limitation on asynchronous transaction throughput is the number and size of
  78log buffers made available by the log manager. By default there are 8 log
  79buffers available and the size of each is 32kB - the size can be increased up
  80to 256kB by use of a mount option.
  82Effectively, this gives us the maximum bound of outstanding metadata changes
  83that can be made to the filesystem at any point in time - if all the log
  84buffers are full and under IO, then no more transactions can be committed until
  85the current batch completes. It is now common for a single current CPU core to
  86be to able to issue enough transactions to keep the log buffers full and under
  87IO permanently. Hence the XFS journalling subsystem can be considered to be IO
  90Delayed Logging: Concepts
  93The key thing to note about the asynchronous logging combined with the
  94relogging technique XFS uses is that we can be relogging changed objects
  95multiple times before they are committed to disk in the log buffers. If we
  96return to the previous relogging example, it is entirely possible that
  97transactions A through D are committed to disk in the same log buffer.
  99That is, a single log buffer may contain multiple copies of the same object,
 100but only one of those copies needs to be there - the last one "D", as it
 101contains all the changes from the previous changes. In other words, we have one
 102necessary copy in the log buffer, and three stale copies that are simply
 103wasting space. When we are doing repeated operations on the same set of
 104objects, these "stale objects" can be over 90% of the space used in the log
 105buffers. It is clear that reducing the number of stale objects written to the
 106log would greatly reduce the amount of metadata we write to the log, and this
 107is the fundamental goal of delayed logging.
 109From a conceptual point of view, XFS is already doing relogging in memory (where
 110memory == log buffer), only it is doing it extremely inefficiently. It is using
 111logical to physical formatting to do the relogging because there is no
 112infrastructure to keep track of logical changes in memory prior to physically
 113formatting the changes in a transaction to the log buffer. Hence we cannot avoid
 114accumulating stale objects in the log buffers.
 116Delayed logging is the name we've given to keeping and tracking transactional
 117changes to objects in memory outside the log buffer infrastructure. Because of
 118the relogging concept fundamental to the XFS journalling subsystem, this is
 119actually relatively easy to do - all the changes to logged items are already
 120tracked in the current infrastructure. The big problem is how to accumulate
 121them and get them to the log in a consistent, recoverable manner.
 122Describing the problems and how they have been solved is the focus of this
 125One of the key changes that delayed logging makes to the operation of the
 126journalling subsystem is that it disassociates the amount of outstanding
 127metadata changes from the size and number of log buffers available. In other
 128words, instead of there only being a maximum of 2MB of transaction changes not
 129written to the log at any point in time, there may be a much greater amount
 130being accumulated in memory. Hence the potential for loss of metadata on a
 131crash is much greater than for the existing logging mechanism.
 133It should be noted that this does not change the guarantee that log recovery
 134will result in a consistent filesystem. What it does mean is that as far as the
 135recovered filesystem is concerned, there may be many thousands of transactions
 136that simply did not occur as a result of the crash. This makes it even more
 137important that applications that care about their data use fsync() where they
 138need to ensure application level data integrity is maintained.
 140It should be noted that delayed logging is not an innovative new concept that
 141warrants rigorous proofs to determine whether it is correct or not. The method
 142of accumulating changes in memory for some period before writing them to the
 143log is used effectively in many filesystems including ext3 and ext4. Hence
 144no time is spent in this document trying to convince the reader that the
 145concept is sound. Instead it is simply considered a "solved problem" and as
 146such implementing it in XFS is purely an exercise in software engineering.
 148The fundamental requirements for delayed logging in XFS are simple:
 150        1. Reduce the amount of metadata written to the log by at least
 151           an order of magnitude.
 152        2. Supply sufficient statistics to validate Requirement #1.
 153        3. Supply sufficient new tracing infrastructure to be able to debug
 154           problems with the new code.
 155        4. No on-disk format change (metadata or log format).
 156        5. Enable and disable with a mount option.
 157        6. No performance regressions for synchronous transaction workloads.
 159Delayed Logging: Design
 162Storing Changes
 165The problem with accumulating changes at a logical level (i.e. just using the
 166existing log item dirty region tracking) is that when it comes to writing the
 167changes to the log buffers, we need to ensure that the object we are formatting
 168is not changing while we do this. This requires locking the object to prevent
 169concurrent modification. Hence flushing the logical changes to the log would
 170require us to lock every object, format them, and then unlock them again.
 172This introduces lots of scope for deadlocks with transactions that are already
 173running. For example, a transaction has object A locked and modified, but needs
 174the delayed logging tracking lock to commit the transaction. However, the
 175flushing thread has the delayed logging tracking lock already held, and is
 176trying to get the lock on object A to flush it to the log buffer. This appears
 177to be an unsolvable deadlock condition, and it was solving this problem that
 178was the barrier to implementing delayed logging for so long.
 180The solution is relatively simple - it just took a long time to recognise it.
 181Put simply, the current logging code formats the changes to each item into an
 182vector array that points to the changed regions in the item. The log write code
 183simply copies the memory these vectors point to into the log buffer during
 184transaction commit while the item is locked in the transaction. Instead of
 185using the log buffer as the destination of the formatting code, we can use an
 186allocated memory buffer big enough to fit the formatted vector.
 188If we then copy the vector into the memory buffer and rewrite the vector to
 189point to the memory buffer rather than the object itself, we now have a copy of
 190the changes in a format that is compatible with the log buffer writing code.
 191that does not require us to lock the item to access. This formatting and
 192rewriting can all be done while the object is locked during transaction commit,
 193resulting in a vector that is transactionally consistent and can be accessed
 194without needing to lock the owning item.
 196Hence we avoid the need to lock items when we need to flush outstanding
 197asynchronous transactions to the log. The differences between the existing
 198formatting method and the delayed logging formatting can be seen in the
 199diagram below.
 201Current format log vector::
 203    Object    +---------------------------------------------+
 204    Vector 1      +----+
 205    Vector 2                    +----+
 206    Vector 3                                   +----------+
 208After formatting::
 210    Log Buffer    +-V1-+-V2-+----V3----+
 212Delayed logging vector::
 214    Object    +---------------------------------------------+
 215    Vector 1      +----+
 216    Vector 2                    +----+
 217    Vector 3                                   +----------+
 219After formatting::
 221    Memory Buffer +-V1-+-V2-+----V3----+
 222    Vector 1      +----+
 223    Vector 2           +----+
 224    Vector 3                +----------+
 226The memory buffer and associated vector need to be passed as a single object,
 227but still need to be associated with the parent object so if the object is
 228relogged we can replace the current memory buffer with a new memory buffer that
 229contains the latest changes.
 231The reason for keeping the vector around after we've formatted the memory
 232buffer is to support splitting vectors across log buffer boundaries correctly.
 233If we don't keep the vector around, we do not know where the region boundaries
 234are in the item, so we'd need a new encapsulation method for regions in the log
 235buffer writing (i.e. double encapsulation). This would be an on-disk format
 236change and as such is not desirable.  It also means we'd have to write the log
 237region headers in the formatting stage, which is problematic as there is per
 238region state that needs to be placed into the headers during the log write.
 240Hence we need to keep the vector, but by attaching the memory buffer to it and
 241rewriting the vector addresses to point at the memory buffer we end up with a
 242self-describing object that can be passed to the log buffer write code to be
 243handled in exactly the same manner as the existing log vectors are handled.
 244Hence we avoid needing a new on-disk format to handle items that have been
 245relogged in memory.
 248Tracking Changes
 251Now that we can record transactional changes in memory in a form that allows
 252them to be used without limitations, we need to be able to track and accumulate
 253them so that they can be written to the log at some later point in time.  The
 254log item is the natural place to store this vector and buffer, and also makes sense
 255to be the object that is used to track committed objects as it will always
 256exist once the object has been included in a transaction.
 258The log item is already used to track the log items that have been written to
 259the log but not yet written to disk. Such log items are considered "active"
 260and as such are stored in the Active Item List (AIL) which is a LSN-ordered
 261double linked list. Items are inserted into this list during log buffer IO
 262completion, after which they are unpinned and can be written to disk. An object
 263that is in the AIL can be relogged, which causes the object to be pinned again
 264and then moved forward in the AIL when the log buffer IO completes for that
 267Essentially, this shows that an item that is in the AIL can still be modified
 268and relogged, so any tracking must be separate to the AIL infrastructure. As
 269such, we cannot reuse the AIL list pointers for tracking committed items, nor
 270can we store state in any field that is protected by the AIL lock. Hence the
 271committed item tracking needs it's own locks, lists and state fields in the log
 274Similar to the AIL, tracking of committed items is done through a new list
 275called the Committed Item List (CIL).  The list tracks log items that have been
 276committed and have formatted memory buffers attached to them. It tracks objects
 277in transaction commit order, so when an object is relogged it is removed from
 278it's place in the list and re-inserted at the tail. This is entirely arbitrary
 279and done to make it easy for debugging - the last items in the list are the
 280ones that are most recently modified. Ordering of the CIL is not necessary for
 281transactional integrity (as discussed in the next section) so the ordering is
 282done for convenience/sanity of the developers.
 285Delayed Logging: Checkpoints
 288When we have a log synchronisation event, commonly known as a "log force",
 289all the items in the CIL must be written into the log via the log buffers.
 290We need to write these items in the order that they exist in the CIL, and they
 291need to be written as an atomic transaction. The need for all the objects to be
 292written as an atomic transaction comes from the requirements of relogging and
 293log replay - all the changes in all the objects in a given transaction must
 294either be completely replayed during log recovery, or not replayed at all. If
 295a transaction is not replayed because it is not complete in the log, then
 296no later transactions should be replayed, either.
 298To fulfill this requirement, we need to write the entire CIL in a single log
 299transaction. Fortunately, the XFS log code has no fixed limit on the size of a
 300transaction, nor does the log replay code. The only fundamental limit is that
 301the transaction cannot be larger than just under half the size of the log.  The
 302reason for this limit is that to find the head and tail of the log, there must
 303be at least one complete transaction in the log at any given time. If a
 304transaction is larger than half the log, then there is the possibility that a
 305crash during the write of a such a transaction could partially overwrite the
 306only complete previous transaction in the log. This will result in a recovery
 307failure and an inconsistent filesystem and hence we must enforce the maximum
 308size of a checkpoint to be slightly less than a half the log.
 310Apart from this size requirement, a checkpoint transaction looks no different
 311to any other transaction - it contains a transaction header, a series of
 312formatted log items and a commit record at the tail. From a recovery
 313perspective, the checkpoint transaction is also no different - just a lot
 314bigger with a lot more items in it. The worst case effect of this is that we
 315might need to tune the recovery transaction object hash size.
 317Because the checkpoint is just another transaction and all the changes to log
 318items are stored as log vectors, we can use the existing log buffer writing
 319code to write the changes into the log. To do this efficiently, we need to
 320minimise the time we hold the CIL locked while writing the checkpoint
 321transaction. The current log write code enables us to do this easily with the
 322way it separates the writing of the transaction contents (the log vectors) from
 323the transaction commit record, but tracking this requires us to have a
 324per-checkpoint context that travels through the log write process through to
 325checkpoint completion.
 327Hence a checkpoint has a context that tracks the state of the current
 328checkpoint from initiation to checkpoint completion. A new context is initiated
 329at the same time a checkpoint transaction is started. That is, when we remove
 330all the current items from the CIL during a checkpoint operation, we move all
 331those changes into the current checkpoint context. We then initialise a new
 332context and attach that to the CIL for aggregation of new transactions.
 334This allows us to unlock the CIL immediately after transfer of all the
 335committed items and effectively allow new transactions to be issued while we
 336are formatting the checkpoint into the log. It also allows concurrent
 337checkpoints to be written into the log buffers in the case of log force heavy
 338workloads, just like the existing transaction commit code does. This, however,
 339requires that we strictly order the commit records in the log so that
 340checkpoint sequence order is maintained during log replay.
 342To ensure that we can be writing an item into a checkpoint transaction at
 343the same time another transaction modifies the item and inserts the log item
 344into the new CIL, then checkpoint transaction commit code cannot use log items
 345to store the list of log vectors that need to be written into the transaction.
 346Hence log vectors need to be able to be chained together to allow them to be
 347detached from the log items. That is, when the CIL is flushed the memory
 348buffer and log vector attached to each log item needs to be attached to the
 349checkpoint context so that the log item can be released. In diagrammatic form,
 350the CIL would look like this before the flush::
 352        CIL Head
 353           |
 354           V
 355        Log Item <-> log vector 1       -> memory buffer
 356           |                            -> vector array
 357           V
 358        Log Item <-> log vector 2       -> memory buffer
 359           |                            -> vector array
 360           V
 361        ......
 362           |
 363           V
 364        Log Item <-> log vector N-1     -> memory buffer
 365           |                            -> vector array
 366           V
 367        Log Item <-> log vector N       -> memory buffer
 368                                        -> vector array
 370And after the flush the CIL head is empty, and the checkpoint context log
 371vector list would look like::
 373        Checkpoint Context
 374           |
 375           V
 376        log vector 1    -> memory buffer
 377           |            -> vector array
 378           |            -> Log Item
 379           V
 380        log vector 2    -> memory buffer
 381           |            -> vector array
 382           |            -> Log Item
 383           V
 384        ......
 385           |
 386           V
 387        log vector N-1  -> memory buffer
 388           |            -> vector array
 389           |            -> Log Item
 390           V
 391        log vector N    -> memory buffer
 392                        -> vector array
 393                        -> Log Item
 395Once this transfer is done, the CIL can be unlocked and new transactions can
 396start, while the checkpoint flush code works over the log vector chain to
 397commit the checkpoint.
 399Once the checkpoint is written into the log buffers, the checkpoint context is
 400attached to the log buffer that the commit record was written to along with a
 401completion callback. Log IO completion will call that callback, which can then
 402run transaction committed processing for the log items (i.e. insert into AIL
 403and unpin) in the log vector chain and then free the log vector chain and
 404checkpoint context.
 406Discussion Point: I am uncertain as to whether the log item is the most
 407efficient way to track vectors, even though it seems like the natural way to do
 408it. The fact that we walk the log items (in the CIL) just to chain the log
 409vectors and break the link between the log item and the log vector means that
 410we take a cache line hit for the log item list modification, then another for
 411the log vector chaining. If we track by the log vectors, then we only need to
 412break the link between the log item and the log vector, which means we should
 413dirty only the log item cachelines. Normally I wouldn't be concerned about one
 414vs two dirty cachelines except for the fact I've seen upwards of 80,000 log
 415vectors in one checkpoint transaction. I'd guess this is a "measure and
 416compare" situation that can be done after a working and reviewed implementation
 417is in the dev tree....
 419Delayed Logging: Checkpoint Sequencing
 422One of the key aspects of the XFS transaction subsystem is that it tags
 423committed transactions with the log sequence number of the transaction commit.
 424This allows transactions to be issued asynchronously even though there may be
 425future operations that cannot be completed until that transaction is fully
 426committed to the log. In the rare case that a dependent operation occurs (e.g.
 427re-using a freed metadata extent for a data extent), a special, optimised log
 428force can be issued to force the dependent transaction to disk immediately.
 430To do this, transactions need to record the LSN of the commit record of the
 431transaction. This LSN comes directly from the log buffer the transaction is
 432written into. While this works just fine for the existing transaction
 433mechanism, it does not work for delayed logging because transactions are not
 434written directly into the log buffers. Hence some other method of sequencing
 435transactions is required.
 437As discussed in the checkpoint section, delayed logging uses per-checkpoint
 438contexts, and as such it is simple to assign a sequence number to each
 439checkpoint. Because the switching of checkpoint contexts must be done
 440atomically, it is simple to ensure that each new context has a monotonically
 441increasing sequence number assigned to it without the need for an external
 442atomic counter - we can just take the current context sequence number and add
 443one to it for the new context.
 445Then, instead of assigning a log buffer LSN to the transaction commit LSN
 446during the commit, we can assign the current checkpoint sequence. This allows
 447operations that track transactions that have not yet completed know what
 448checkpoint sequence needs to be committed before they can continue. As a
 449result, the code that forces the log to a specific LSN now needs to ensure that
 450the log forces to a specific checkpoint.
 452To ensure that we can do this, we need to track all the checkpoint contexts
 453that are currently committing to the log. When we flush a checkpoint, the
 454context gets added to a "committing" list which can be searched. When a
 455checkpoint commit completes, it is removed from the committing list. Because
 456the checkpoint context records the LSN of the commit record for the checkpoint,
 457we can also wait on the log buffer that contains the commit record, thereby
 458using the existing log force mechanisms to execute synchronous forces.
 460It should be noted that the synchronous forces may need to be extended with
 461mitigation algorithms similar to the current log buffer code to allow
 462aggregation of multiple synchronous transactions if there are already
 463synchronous transactions being flushed. Investigation of the performance of the
 464current design is needed before making any decisions here.
 466The main concern with log forces is to ensure that all the previous checkpoints
 467are also committed to disk before the one we need to wait for. Therefore we
 468need to check that all the prior contexts in the committing list are also
 469complete before waiting on the one we need to complete. We do this
 470synchronisation in the log force code so that we don't need to wait anywhere
 471else for such serialisation - it only matters when we do a log force.
 473The only remaining complexity is that a log force now also has to handle the
 474case where the forcing sequence number is the same as the current context. That
 475is, we need to flush the CIL and potentially wait for it to complete. This is a
 476simple addition to the existing log forcing code to check the sequence numbers
 477and push if required. Indeed, placing the current sequence checkpoint flush in
 478the log force code enables the current mechanism for issuing synchronous
 479transactions to remain untouched (i.e. commit an asynchronous transaction, then
 480force the log at the LSN of that transaction) and so the higher level code
 481behaves the same regardless of whether delayed logging is being used or not.
 483Delayed Logging: Checkpoint Log Space Accounting
 486The big issue for a checkpoint transaction is the log space reservation for the
 487transaction. We don't know how big a checkpoint transaction is going to be
 488ahead of time, nor how many log buffers it will take to write out, nor the
 489number of split log vector regions are going to be used. We can track the
 490amount of log space required as we add items to the commit item list, but we
 491still need to reserve the space in the log for the checkpoint.
 493A typical transaction reserves enough space in the log for the worst case space
 494usage of the transaction. The reservation accounts for log record headers,
 495transaction and region headers, headers for split regions, buffer tail padding,
 496etc. as well as the actual space for all the changed metadata in the
 497transaction. While some of this is fixed overhead, much of it is dependent on
 498the size of the transaction and the number of regions being logged (the number
 499of log vectors in the transaction).
 501An example of the differences would be logging directory changes versus logging
 502inode changes. If you modify lots of inode cores (e.g. ``chmod -R g+w *``), then
 503there are lots of transactions that only contain an inode core and an inode log
 504format structure. That is, two vectors totaling roughly 150 bytes. If we modify
 50510,000 inodes, we have about 1.5MB of metadata to write in 20,000 vectors. Each
 506vector is 12 bytes, so the total to be logged is approximately 1.75MB. In
 507comparison, if we are logging full directory buffers, they are typically 4KB
 508each, so we in 1.5MB of directory buffers we'd have roughly 400 buffers and a
 509buffer format structure for each buffer - roughly 800 vectors or 1.51MB total
 510space.  From this, it should be obvious that a static log space reservation is
 511not particularly flexible and is difficult to select the "optimal value" for
 512all workloads.
 514Further, if we are going to use a static reservation, which bit of the entire
 515reservation does it cover? We account for space used by the transaction
 516reservation by tracking the space currently used by the object in the CIL and
 517then calculating the increase or decrease in space used as the object is
 518relogged. This allows for a checkpoint reservation to only have to account for
 519log buffer metadata used such as log header records.
 521However, even using a static reservation for just the log metadata is
 522problematic. Typically log record headers use at least 16KB of log space per
 5231MB of log space consumed (512 bytes per 32k) and the reservation needs to be
 524large enough to handle arbitrary sized checkpoint transactions. This
 525reservation needs to be made before the checkpoint is started, and we need to
 526be able to reserve the space without sleeping.  For a 8MB checkpoint, we need a
 527reservation of around 150KB, which is a non-trivial amount of space.
 529A static reservation needs to manipulate the log grant counters - we can take a
 530permanent reservation on the space, but we still need to make sure we refresh
 531the write reservation (the actual space available to the transaction) after
 532every checkpoint transaction completion. Unfortunately, if this space is not
 533available when required, then the regrant code will sleep waiting for it.
 535The problem with this is that it can lead to deadlocks as we may need to commit
 536checkpoints to be able to free up log space (refer back to the description of
 537rolling transactions for an example of this).  Hence we *must* always have
 538space available in the log if we are to use static reservations, and that is
 539very difficult and complex to arrange. It is possible to do, but there is a
 540simpler way.
 542The simpler way of doing this is tracking the entire log space used by the
 543items in the CIL and using this to dynamically calculate the amount of log
 544space required by the log metadata. If this log metadata space changes as a
 545result of a transaction commit inserting a new memory buffer into the CIL, then
 546the difference in space required is removed from the transaction that causes
 547the change. Transactions at this level will *always* have enough space
 548available in their reservation for this as they have already reserved the
 549maximal amount of log metadata space they require, and such a delta reservation
 550will always be less than or equal to the maximal amount in the reservation.
 552Hence we can grow the checkpoint transaction reservation dynamically as items
 553are added to the CIL and avoid the need for reserving and regranting log space
 554up front. This avoids deadlocks and removes a blocking point from the
 555checkpoint flush code.
 557As mentioned early, transactions can't grow to more than half the size of the
 558log. Hence as part of the reservation growing, we need to also check the size
 559of the reservation against the maximum allowed transaction size. If we reach
 560the maximum threshold, we need to push the CIL to the log. This is effectively
 561a "background flush" and is done on demand. This is identical to
 562a CIL push triggered by a log force, only that there is no waiting for the
 563checkpoint commit to complete. This background push is checked and executed by
 564transaction commit code.
 566If the transaction subsystem goes idle while we still have items in the CIL,
 567they will be flushed by the periodic log force issued by the xfssyncd. This log
 568force will push the CIL to disk, and if the transaction subsystem stays idle,
 569allow the idle log to be covered (effectively marked clean) in exactly the same
 570manner that is done for the existing logging method. A discussion point is
 571whether this log force needs to be done more frequently than the current rate
 572which is once every 30s.
 575Delayed Logging: Log Item Pinning
 578Currently log items are pinned during transaction commit while the items are
 579still locked. This happens just after the items are formatted, though it could
 580be done any time before the items are unlocked. The result of this mechanism is
 581that items get pinned once for every transaction that is committed to the log
 582buffers. Hence items that are relogged in the log buffers will have a pin count
 583for every outstanding transaction they were dirtied in. When each of these
 584transactions is completed, they will unpin the item once. As a result, the item
 585only becomes unpinned when all the transactions complete and there are no
 586pending transactions. Thus the pinning and unpinning of a log item is symmetric
 587as there is a 1:1 relationship with transaction commit and log item completion.
 589For delayed logging, however, we have an asymmetric transaction commit to
 590completion relationship. Every time an object is relogged in the CIL it goes
 591through the commit process without a corresponding completion being registered.
 592That is, we now have a many-to-one relationship between transaction commit and
 593log item completion. The result of this is that pinning and unpinning of the
 594log items becomes unbalanced if we retain the "pin on transaction commit, unpin
 595on transaction completion" model.
 597To keep pin/unpin symmetry, the algorithm needs to change to a "pin on
 598insertion into the CIL, unpin on checkpoint completion". In other words, the
 599pinning and unpinning becomes symmetric around a checkpoint context. We have to
 600pin the object the first time it is inserted into the CIL - if it is already in
 601the CIL during a transaction commit, then we do not pin it again. Because there
 602can be multiple outstanding checkpoint contexts, we can still see elevated pin
 603counts, but as each checkpoint completes the pin count will retain the correct
 604value according to it's context.
 606Just to make matters more slightly more complex, this checkpoint level context
 607for the pin count means that the pinning of an item must take place under the
 608CIL commit/flush lock. If we pin the object outside this lock, we cannot
 609guarantee which context the pin count is associated with. This is because of
 610the fact pinning the item is dependent on whether the item is present in the
 611current CIL or not. If we don't pin the CIL first before we check and pin the
 612object, we have a race with CIL being flushed between the check and the pin
 613(or not pinning, as the case may be). Hence we must hold the CIL flush/commit
 614lock to guarantee that we pin the items correctly.
 616Delayed Logging: Concurrent Scalability
 619A fundamental requirement for the CIL is that accesses through transaction
 620commits must scale to many concurrent commits. The current transaction commit
 621code does not break down even when there are transactions coming from 2048
 622processors at once. The current transaction code does not go any faster than if
 623there was only one CPU using it, but it does not slow down either.
 625As a result, the delayed logging transaction commit code needs to be designed
 626for concurrency from the ground up. It is obvious that there are serialisation
 627points in the design - the three important ones are:
 629        1. Locking out new transaction commits while flushing the CIL
 630        2. Adding items to the CIL and updating item space accounting
 631        3. Checkpoint commit ordering
 633Looking at the transaction commit and CIL flushing interactions, it is clear
 634that we have a many-to-one interaction here. That is, the only restriction on
 635the number of concurrent transactions that can be trying to commit at once is
 636the amount of space available in the log for their reservations. The practical
 637limit here is in the order of several hundred concurrent transactions for a
 638128MB log, which means that it is generally one per CPU in a machine.
 640The amount of time a transaction commit needs to hold out a flush is a
 641relatively long period of time - the pinning of log items needs to be done
 642while we are holding out a CIL flush, so at the moment that means it is held
 643across the formatting of the objects into memory buffers (i.e. while memcpy()s
 644are in progress). Ultimately a two pass algorithm where the formatting is done
 645separately to the pinning of objects could be used to reduce the hold time of
 646the transaction commit side.
 648Because of the number of potential transaction commit side holders, the lock
 649really needs to be a sleeping lock - if the CIL flush takes the lock, we do not
 650want every other CPU in the machine spinning on the CIL lock. Given that
 651flushing the CIL could involve walking a list of tens of thousands of log
 652items, it will get held for a significant time and so spin contention is a
 653significant concern. Preventing lots of CPUs spinning doing nothing is the
 654main reason for choosing a sleeping lock even though nothing in either the
 655transaction commit or CIL flush side sleeps with the lock held.
 657It should also be noted that CIL flushing is also a relatively rare operation
 658compared to transaction commit for asynchronous transaction workloads - only
 659time will tell if using a read-write semaphore for exclusion will limit
 660transaction commit concurrency due to cache line bouncing of the lock on the
 661read side.
 663The second serialisation point is on the transaction commit side where items
 664are inserted into the CIL. Because transactions can enter this code
 665concurrently, the CIL needs to be protected separately from the above
 666commit/flush exclusion. It also needs to be an exclusive lock but it is only
 667held for a very short time and so a spin lock is appropriate here. It is
 668possible that this lock will become a contention point, but given the short
 669hold time once per transaction I think that contention is unlikely.
 671The final serialisation point is the checkpoint commit record ordering code
 672that is run as part of the checkpoint commit and log force sequencing. The code
 673path that triggers a CIL flush (i.e. whatever triggers the log force) will enter
 674an ordering loop after writing all the log vectors into the log buffers but
 675before writing the commit record. This loop walks the list of committing
 676checkpoints and needs to block waiting for checkpoints to complete their commit
 677record write. As a result it needs a lock and a wait variable. Log force
 678sequencing also requires the same lock, list walk, and blocking mechanism to
 679ensure completion of checkpoints.
 681These two sequencing operations can use the mechanism even though the
 682events they are waiting for are different. The checkpoint commit record
 683sequencing needs to wait until checkpoint contexts contain a commit LSN
 684(obtained through completion of a commit record write) while log force
 685sequencing needs to wait until previous checkpoint contexts are removed from
 686the committing list (i.e. they've completed). A simple wait variable and
 687broadcast wakeups (thundering herds) has been used to implement these two
 688serialisation queues. They use the same lock as the CIL, too. If we see too
 689much contention on the CIL lock, or too many context switches as a result of
 690the broadcast wakeups these operations can be put under a new spinlock and
 691given separate wait lists to reduce lock contention and the number of processes
 692woken by the wrong event.
 695Lifecycle Changes
 698The existing log item life cycle is as follows::
 700        1. Transaction allocate
 701        2. Transaction reserve
 702        3. Lock item
 703        4. Join item to transaction
 704                If not already attached,
 705                        Allocate log item
 706                        Attach log item to owner item
 707                Attach log item to transaction
 708        5. Modify item
 709                Record modifications in log item
 710        6. Transaction commit
 711                Pin item in memory
 712                Format item into log buffer
 713                Write commit LSN into transaction
 714                Unlock item
 715                Attach transaction to log buffer
 717        <log buffer IO dispatched>
 718        <log buffer IO completes>
 720        7. Transaction completion
 721                Mark log item committed
 722                Insert log item into AIL
 723                        Write commit LSN into log item
 724                Unpin log item
 725        8. AIL traversal
 726                Lock item
 727                Mark log item clean
 728                Flush item to disk
 730        <item IO completion>
 732        9. Log item removed from AIL
 733                Moves log tail
 734                Item unlocked
 736Essentially, steps 1-6 operate independently from step 7, which is also
 737independent of steps 8-9. An item can be locked in steps 1-6 or steps 8-9
 738at the same time step 7 is occurring, but only steps 1-6 or 8-9 can occur
 739at the same time. If the log item is in the AIL or between steps 6 and 7
 740and steps 1-6 are re-entered, then the item is relogged. Only when steps 8-9
 741are entered and completed is the object considered clean.
 743With delayed logging, there are new steps inserted into the life cycle::
 745        1. Transaction allocate
 746        2. Transaction reserve
 747        3. Lock item
 748        4. Join item to transaction
 749                If not already attached,
 750                        Allocate log item
 751                        Attach log item to owner item
 752                Attach log item to transaction
 753        5. Modify item
 754                Record modifications in log item
 755        6. Transaction commit
 756                Pin item in memory if not pinned in CIL
 757                Format item into log vector + buffer
 758                Attach log vector and buffer to log item
 759                Insert log item into CIL
 760                Write CIL context sequence into transaction
 761                Unlock item
 763        <next log force>
 765        7. CIL push
 766                lock CIL flush
 767                Chain log vectors and buffers together
 768                Remove items from CIL
 769                unlock CIL flush
 770                write log vectors into log
 771                sequence commit records
 772                attach checkpoint context to log buffer
 774        <log buffer IO dispatched>
 775        <log buffer IO completes>
 777        8. Checkpoint completion
 778                Mark log item committed
 779                Insert item into AIL
 780                        Write commit LSN into log item
 781                Unpin log item
 782        9. AIL traversal
 783                Lock item
 784                Mark log item clean
 785                Flush item to disk
 786        <item IO completion>
 787        10. Log item removed from AIL
 788                Moves log tail
 789                Item unlocked
 791From this, it can be seen that the only life cycle differences between the two
 792logging methods are in the middle of the life cycle - they still have the same
 793beginning and end and execution constraints. The only differences are in the
 794committing of the log items to the log itself and the completion processing.
 795Hence delayed logging should not introduce any constraints on log item
 796behaviour, allocation or freeing that don't already exist.
 798As a result of this zero-impact "insertion" of delayed logging infrastructure
 799and the design of the internal structures to avoid on disk format changes, we
 800can basically switch between delayed logging and the existing mechanism with a
 801mount option. Fundamentally, there is no reason why the log manager would not
 802be able to swap methods automatically and transparently depending on load
 803characteristics, but this should not be necessary if delayed logging works as