Commit 71a96a05 authored by Bobi Jam's avatar Bobi Jam Committed by Greg Kroah-Hartman

staging/lustre: update comments after cl_lock simplification

Update comments to reflect current cl_lock situations.
Signed-off-by: default avatarBobi Jam <bobijam.xu@intel.com>
Reviewed-on: http://review.whamcloud.com/13137
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-6046Reviewed-by: default avatarJohn L. Hammond <john.hammond@intel.com>
Reviewed-by: default avatarJinshan Xiong <jinshan.xiong@intel.com>
Signed-off-by: default avatarOleg Drokin <green@linuxhacker.ru>
Signed-off-by: default avatarGreg Kroah-Hartman <gregkh@linuxfoundation.org>
parent 06563b56
...@@ -1117,111 +1117,29 @@ static inline struct page *cl_page_vmpage(struct cl_page *page) ...@@ -1117,111 +1117,29 @@ static inline struct page *cl_page_vmpage(struct cl_page *page)
* *
* LIFE CYCLE * LIFE CYCLE
* *
* cl_lock is reference counted. When reference counter drops to 0, lock is * cl_lock is a cacheless data container for the requirements of locks to
* placed in the cache, except when lock is in CLS_FREEING state. CLS_FREEING * complete the IO. cl_lock is created before I/O starts and destroyed when the
* lock is destroyed when last reference is released. Referencing between * I/O is complete.
* top-lock and its sub-locks is described in the lov documentation module. *
* * cl_lock depends on LDLM lock to fulfill lock semantics. LDLM lock is attached
* STATE MACHINE * to cl_lock at OSC layer. LDLM lock is still cacheable.
*
* Also, cl_lock is a state machine. This requires some clarification. One of
* the goals of client IO re-write was to make IO path non-blocking, or at
* least to make it easier to make it non-blocking in the future. Here
* `non-blocking' means that when a system call (read, write, truncate)
* reaches a situation where it has to wait for a communication with the
* server, it should --instead of waiting-- remember its current state and
* switch to some other work. E.g,. instead of waiting for a lock enqueue,
* client should proceed doing IO on the next stripe, etc. Obviously this is
* rather radical redesign, and it is not planned to be fully implemented at
* this time, instead we are putting some infrastructure in place, that would
* make it easier to do asynchronous non-blocking IO easier in the
* future. Specifically, where old locking code goes to sleep (waiting for
* enqueue, for example), new code returns cl_lock_transition::CLO_WAIT. When
* enqueue reply comes, its completion handler signals that lock state-machine
* is ready to transit to the next state. There is some generic code in
* cl_lock.c that sleeps, waiting for these signals. As a result, for users of
* this cl_lock.c code, it looks like locking is done in normal blocking
* fashion, and it the same time it is possible to switch to the non-blocking
* locking (simply by returning cl_lock_transition::CLO_WAIT from cl_lock.c
* functions).
*
* For a description of state machine states and transitions see enum
* cl_lock_state.
*
* There are two ways to restrict a set of states which lock might move to:
*
* - placing a "hold" on a lock guarantees that lock will not be moved
* into cl_lock_state::CLS_FREEING state until hold is released. Hold
* can be only acquired on a lock that is not in
* cl_lock_state::CLS_FREEING. All holds on a lock are counted in
* cl_lock::cll_holds. Hold protects lock from cancellation and
* destruction. Requests to cancel and destroy a lock on hold will be
* recorded, but only honored when last hold on a lock is released;
*
* - placing a "user" on a lock guarantees that lock will not leave
* cl_lock_state::CLS_NEW, cl_lock_state::CLS_QUEUING,
* cl_lock_state::CLS_ENQUEUED and cl_lock_state::CLS_HELD set of
* states, once it enters this set. That is, if a user is added onto a
* lock in a state not from this set, it doesn't immediately enforce
* lock to move to this set, but once lock enters this set it will
* remain there until all users are removed. Lock users are counted in
* cl_lock::cll_users.
*
* User is used to assure that lock is not canceled or destroyed while
* it is being enqueued, or actively used by some IO.
*
* Currently, a user always comes with a hold (cl_lock_invariant()
* checks that a number of holds is not less than a number of users).
*
* CONCURRENCY
*
* This is how lock state-machine operates. struct cl_lock contains a mutex
* cl_lock::cll_guard that protects struct fields.
*
* - mutex is taken, and cl_lock::cll_state is examined.
*
* - for every state there are possible target states where lock can move
* into. They are tried in order. Attempts to move into next state are
* done by _try() functions in cl_lock.c:cl_{enqueue,unlock,wait}_try().
*
* - if the transition can be performed immediately, state is changed,
* and mutex is released.
*
* - if the transition requires blocking, _try() function returns
* cl_lock_transition::CLO_WAIT. Caller unlocks mutex and goes to
* sleep, waiting for possibility of lock state change. It is woken
* up when some event occurs, that makes lock state change possible
* (e.g., the reception of the reply from the server), and repeats
* the loop.
*
* Top-lock and sub-lock has separate mutexes and the latter has to be taken
* first to avoid dead-lock.
*
* To see an example of interaction of all these issues, take a look at the
* lov_cl.c:lov_lock_enqueue() function. It is called as a part of
* cl_enqueue_try(), and tries to advance top-lock to ENQUEUED state, by
* advancing state-machines of its sub-locks (lov_lock_enqueue_one()). Note
* also, that it uses trylock to grab sub-lock mutex to avoid dead-lock. It
* also has to handle CEF_ASYNC enqueue, when sub-locks enqueues have to be
* done in parallel, rather than one after another (this is used for glimpse
* locks, that cannot dead-lock).
* *
* INTERFACE AND USAGE * INTERFACE AND USAGE
* *
* struct cl_lock_operations provide a number of call-backs that are invoked * Two major methods are supported for cl_lock: clo_enqueue and clo_cancel. A
* when events of interest occurs. Layers can intercept and handle glimpse, * cl_lock is enqueued by cl_lock_request(), which will call clo_enqueue()
* blocking, cancel ASTs and a reception of the reply from the server. * methods for each layer to enqueue the lock. At the LOV layer, if a cl_lock
* consists of multiple sub cl_locks, each sub locks will be enqueued
* correspondingly. At OSC layer, the lock enqueue request will tend to reuse
* cached LDLM lock; otherwise a new LDLM lock will have to be requested from
* OST side.
* *
* One important difference with the old client locking model is that new * cl_lock_cancel() must be called to release a cl_lock after use. clo_cancel()
* client has a representation for the top-lock, whereas in the old code only * method will be called for each layer to release the resource held by this
* sub-locks existed as real data structures and file-level locks are * lock. At OSC layer, the reference count of LDLM lock, which is held at
* represented by "request sets" that are created and destroyed on each and * clo_enqueue time, is released.
* every lock creation.
* *
* Top-locks are cached, and can be found in the cache by the system calls. It * LDLM lock can only be canceled if there is no cl_lock using it.
* is possible that top-lock is in cache, but some of its sub-locks were
* canceled and destroyed. In that case top-lock has to be enqueued again
* before it can be used.
* *
* Overall process of the locking during IO operation is as following: * Overall process of the locking during IO operation is as following:
* *
...@@ -1234,7 +1152,7 @@ static inline struct page *cl_page_vmpage(struct cl_page *page) ...@@ -1234,7 +1152,7 @@ static inline struct page *cl_page_vmpage(struct cl_page *page)
* *
* - when all locks are acquired, IO is performed; * - when all locks are acquired, IO is performed;
* *
* - locks are released into cache. * - locks are released after IO is complete.
* *
* Striping introduces major additional complexity into locking. The * Striping introduces major additional complexity into locking. The
* fundamental problem is that it is generally unsafe to actively use (hold) * fundamental problem is that it is generally unsafe to actively use (hold)
...@@ -1256,16 +1174,6 @@ static inline struct page *cl_page_vmpage(struct cl_page *page) ...@@ -1256,16 +1174,6 @@ static inline struct page *cl_page_vmpage(struct cl_page *page)
* buf is a part of memory mapped Lustre file, a lock or locks protecting buf * buf is a part of memory mapped Lustre file, a lock or locks protecting buf
* has to be held together with the usual lock on [offset, offset + count]. * has to be held together with the usual lock on [offset, offset + count].
* *
* As multi-stripe locks have to be allowed, it makes sense to cache them, so
* that, for example, a sequence of O_APPEND writes can proceed quickly
* without going down to the individual stripes to do lock matching. On the
* other hand, multi-stripe locks shouldn't be used by normal read/write
* calls. To achieve this, every layer can implement ->clo_fits_into() method,
* that is called by lock matching code (cl_lock_lookup()), and that can be
* used to selectively disable matching of certain locks for certain IOs. For
* example, lov layer implements lov_lock_fits_into() that allow multi-stripe
* locks to be matched only for truncates and O_APPEND writes.
*
* Interaction with DLM * Interaction with DLM
* *
* In the expected setup, cl_lock is ultimately backed up by a collection of * In the expected setup, cl_lock is ultimately backed up by a collection of
......
...@@ -73,19 +73,6 @@ ...@@ -73,19 +73,6 @@
* - top-page keeps a reference to its sub-page, and destroys it when it * - top-page keeps a reference to its sub-page, and destroys it when it
* is destroyed. * is destroyed.
* *
* - sub-lock keep a reference to its top-locks. Top-lock keeps a
* reference (and a hold, see cl_lock_hold()) on its sub-locks when it
* actively using them (that is, in cl_lock_state::CLS_QUEUING,
* cl_lock_state::CLS_ENQUEUED, cl_lock_state::CLS_HELD states). When
* moving into cl_lock_state::CLS_CACHED state, top-lock releases a
* hold. From this moment top-lock has only a 'weak' reference to its
* sub-locks. This reference is protected by top-lock
* cl_lock::cll_guard, and will be automatically cleared by the sub-lock
* when the latter is destroyed. When a sub-lock is canceled, a
* reference to it is removed from the top-lock array, and top-lock is
* moved into CLS_NEW state. It is guaranteed that all sub-locks exist
* while their top-lock is in CLS_HELD or CLS_CACHED states.
*
* - IO's are not reference counted. * - IO's are not reference counted.
* *
* To implement a connection between top and sub entities, lov layer is split * To implement a connection between top and sub entities, lov layer is split
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment