greg/stasis-aries-wal
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
stasis-aries-wal/stasis/page.h
Sears Russell 7660467f8e Refactor page implementation to use multiple include trick (so the backing representation can be changed). 
Also, fix apparent bug in stasis_page_fixed_set_record_type().

Otherwise, this change should not make any semantic changes.
2011-11-09 21:37:38 +00:00

674 lines
25 KiB
C
Raw Blame History

/*---
This software is copyrighted by the Regents of the University of
California, and other parties. The following terms apply to all files
associated with the software unless explicitly disclaimed in
individual files.
The authors hereby grant permission to use, copy, modify, distribute,
and license this software and its documentation for any purpose,
provided that existing copyright notices are retained in all copies
and that this notice is included verbatim in any distributions. No
written agreement, license, or royalty fee is required for any of the
authorized uses. Modifications to this software may be copyrighted by
their authors and need not follow the licensing terms described here,
provided that the new terms are clearly indicated on the first page of
each file where they apply.
IN NO EVENT SHALL THE AUTHORS OR DISTRIBUTORS BE LIABLE TO ANY PARTY
FOR DIRECT, INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES
ARISING OUT OF THE USE OF THIS SOFTWARE, ITS DOCUMENTATION, OR ANY
DERIVATIVES THEREOF, EVEN IF THE AUTHORS HAVE BEEN ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
THE AUTHORS AND DISTRIBUTORS SPECIFICALLY DISCLAIM ANY WARRANTIES,
INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND
NON-INFRINGEMENT. THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, AND
THE AUTHORS AND DISTRIBUTORS HAVE NO OBLIGATION TO PROVIDE
MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR MODIFICATIONS.
GOVERNMENT USE: If you are acquiring this software on behalf of the
U.S. government, the Government shall have only "Restricted Rights" in
the software and related documentation as defined in the Federal
Acquisition Regulations (FARs) in Clause 52.227.19 (c) (2). If you are
acquiring the software on behalf of the Department of Defense, the
software shall be classified as "Commercial Computer Software" and the
Government shall have only "Restricted Rights" as defined in Clause
252.227-7013 (c) (1) of DFARs. Notwithstanding the foregoing, the
authors grant the U.S. Government and others acting in its behalf
permission to use and distribute the software in accordance with the
terms specified in this license.
---*/
#ifndef __PAGE_H__
#define __PAGE_H__
#include <stasis/common.h>
#include <stasis/constants.h>
#include <stasis/util/latches.h>
#include <stasis/dirtyPageTable.h>
#include <assert.h>
BEGIN_C_DECLS
/**
The page type contains in-memory information about pages. This
information is used by Stasis to track the page while it is in
memory, and is never written to disk.
@todo Page_s shouldn't hardcode doubly linked lists.
@todo The Page struct should be tuned for better memory utilization.
@todo Remove next and prev from Page_s
*/
struct Page_s {
pageid_t id;
/**
* The type of this page. Set when the page is loaded from disk. If the page contains
* a header, this will be set automatically. Otherwise, it must be passed in by the code
* that pinned the page.
*/
int pageType;
/**
* The LSN of the page (or an estimate). Set when page is loaded from disk.
* The on-page LSN (if any) is set at page writeback.
*/
lsn_t LSN;
byte *memAddr;
byte dirty;
byte needsFlush;
/** The next item in the replacement policy's queue */
struct Page_s *next;
/** The previous item in the replacement policy's queue. */
struct Page_s *prev;
/** How many times has the page been pinned? */
int pinCount;
/** Is the page pending I/O? (Used by some buffer managers) */
int pending;
/** Is the page in the cache at all? */
int inCache;
/** Used for page-level latching.
Each page has an associated read/write lock. This lock only
protects the internal layout of the page, and the members of the
page struct. Here is how rwlatch is held in various circumstances:
Record allocation: Write lock
Record read: Read lock
Read LSN Read lock
Record write *READ LOCK*
Write LSN Write lock
Write page to disk No lock
Read page from disk No lock
Any circumstance where one these locks are held during an I/O
operation is a bug.
For the 'no lock' cases, @see loadlatch
*/
rwl * rwlatch;
/**
Since the bufferManager re-uses page structs, this lock is used
to ensure that the page is in one of two consistent states,
depending on whether a read lock or a write lock is being held.
If a read lock is held, then the page is managed by the rwlatch
also defined in this struct. Therefore, it cannot be read from
or written to disk. Furthermore, since we do not impose an
order on operations, the holder of a readlock may not use the
LSN field to determine whether a particular operation has
completed on the page.
The write lock is used to block all writers (other than the one
holding the page), and to ensure that all updates with LSN less
than or equal to the page's LSN have been applied. Therefore,
threads that write the page to disk must hold this lock. Since
it precludes access by all other threads, a write lock also
allows the holder to evict the current page, and replace it.
Examples:
Write page to disk Write lock
Read page from disk Write lock
Allocate a new record Read lock
Write to a record Read lock
Read from a record Read lock
@see rwlatch, getPage(), pageRalloc(), read_page()
*/
rwl * loadlatch;
/**
Page type implementations may store whatever they'd like in this
pointer. It persists from each call to stasis_page_loaded() to the
subsequent call to stasis_page_flushed().
*/
void * impl;
};
/*@}*/
/**
@defgroup PAGE_HEADER Default page header
Most Stasis pages contain an LSN and a page type. These are used
by recovery to determine whether or not to perform redo. At
run time, the page type is used to decide which page implementation
should manipulate the page.
STRUCTURE OF A GENERIC PAGE
<pre>
+----------------------------------------------------------------------+
| |
| USABLE SPACE |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| +-----------+-----+
| | page type | LSN |
+----------------------------------------------------+-----------+-----+
</pre>
*/
/**
* initializes all the global variables needed by the functions
* dealing with pages.
*
* @todo documentation group for page init and deinit?
*/
void stasis_page_init(stasis_dirty_page_table_t * dirtyPages);
/**
* releases all resources held by the page sub-system.
*/
void stasis_page_deinit();
/**
@defgroup PAGE_RECORD_INTERFACE Record interface
@ingroup PAGE_FORMATS
Page formats define the layout of data on pages. Currently, all
pages contain a header with an LSN and a page type in it. This
information is used by recovery and the buffer manager to invoke
callbacks at appropriate times. (LSN-free pages are currently not
supported.)
Stasis' record-oriented page interface uses the page type to determine
which page implementation should be used to access or modify
records. This API's functions begin with "stasis_record". Two
commonly used examples are stasis_read_record() and
stasis_write_record().
This interface is not re-entrant. Rather, page implementations
assume that their callers will latch pages using
readLock(p->rwlatch) and writeLock(p-rwlatch) before attempting to
access the page. A second latch, p->loadlatch, should not be
acquired by page manipulation code. Instead, it is used by the
buffer manager to protect against races during page eviction.
@par Registering new page type implementations
Page implementations are registered with Stasis by passing a
page_impl struct into stasis_page_impl_register(). page_impl.page_type
should contain an integer that is unique across all page types,
while the rest of the fields contain function pointers to the page
type's implementation.
*/
/*@{*/
static const size_t USABLE_SIZE_OF_PAGE = (PAGE_SIZE - sizeof(lsn_t) - sizeof(int));
/**
Stasis records carry type information with them. The type either
encodes the physical size of opaque data (the common case), or
indicates that the record should be handled specially (as the
fixed-length header of a data structure, for example).
This function maps from record type to record size.
@param type If positive, a record type is equal to the record size,
in bytes. If negative, the type has special significance. Its
length is looked up in the SLOT_TYPE_LENGTHS array.
@return The length of the record in bytes.
*/
static inline size_t
stasis_record_type_to_size(ssize_t type) {
if(type >= 0) {
return type;
} else {
assert(SLOT_TYPE_LENGTHS[0 - type] >= 0);
return SLOT_TYPE_LENGTHS[0 - type];
}
}
/**
*
* Write a record. This call will be dispatched to the proper page implementation.
*
* @param xid transaction id
*
* @param page a pointer to an in-memory copy of the page as it
* currently exists. This copy will be updated by writeRecord.
*
* @param rid recordid where you want to write
* @param dat the new value of the record.
*
*/
void stasis_record_write(int xid, Page * page, recordid rid, const byte *dat);
/**
* Read a record. This call will be dispatched to the proper page implementation.
*
* @param xid transaction ID
* @param page a pointer to the pinned page that contains the record.
* @param rid the record to be written
* @param dat buffer for data
* @return 0 on success, Stasis error code on failure
*/
int stasis_record_read(int xid, Page * page, recordid rid, byte *dat);
const byte * stasis_record_read_begin(int xid, Page * p, recordid rid);
byte * stasis_record_write_begin(int xid, Page * p, recordid rid);
void stasis_record_read_done(int xid, Page *p, recordid rid, const byte* buf);
void stasis_record_write_done(int xid, Page *p, recordid rid, byte *buf);
int stasis_record_type_read(int xid, Page * p, recordid rid);
void stasis_record_type_write(int xid, Page * p, recordid rid, int type);
int stasis_record_length_read(int xid, Page *p, recordid rid);
recordid stasis_record_first(int xid, Page * p);
recordid stasis_record_next(int xid, Page * p, recordid prev);
recordid stasis_record_last(int xid, Page * p);
recordid stasis_record_alloc_begin(int xid, Page * p, int size);
void stasis_record_alloc_done(int xid, Page * p, recordid rid);
void stasis_record_splice(int xid, Page * p, slotid_t first, slotid_t second);
void stasis_record_free(int xid, Page * p, recordid rid);
int stasis_block_supported(int xid, Page * p);
int stasis_record_freespace(int xid, Page * p);
void stasis_record_compact(Page * p);
void stasis_record_compact_slotids(int xid, Page * p);
void stasis_uninitialized_page_loaded(int xid, Page * p);
void stasis_page_loaded(Page * p, pagetype_t type);
void stasis_page_flushed(Page * p);
void stasis_page_cleanup(Page * p);
/**
@todo XXX stasis_record_dereference should be dispatched via page_impl[]
*/
recordid stasis_record_dereference(int xid, Page *p, recordid rid);
/*@}*/
/**
@param a block returned by stasis_block_first() or stasis_block_next().
is*() methods return zero for false, non-zero for true.
recordFixedLen() returns the (single) length of every record in the
block, or BLOCK_VARIABLE_LENGTH
methods that take int * size as an argument return a record size by
setting the pointer (if it's not null).
@see Abadi, et. al, Integrating Compression and Execution in Column-Oriented Database Systems, VLDB 2006.
*/
typedef struct block_t {
int (*isValid) (struct block_t *b);
int (*isOneValue) (struct block_t *b);
int (*isValueSorted) (struct block_t *b);
int (*isPosContig) (struct block_t *b);
const byte * (*recordFirst) (struct block_t *b);
const byte * (*recordNext) (struct block_t *b);
int (*recordSize) (struct block_t *b);
int (*recordCount) (struct block_t *b);
const byte **(*recordPtrArray) (struct block_t *b);
int * (*recordSizeArray) (struct block_t *b);
// These two are not in paper
int (*recordFixedLen) (struct block_t *b);
/**
This returns a pointer to an array that contains the records in
the page. It only makes sense for pages that many values of the
same length, and that can implement this more efficiently than
repeated calls to recordNext.
@param block the block
@param count will be set to the number of slots in the page
@param stride will be set to the size of each value in the page
@return a pointer to an array of record contents. The array is
ordered according to slot id. The page implementation manages
the memory; hopefully, this points into the buffer manager, and
this function call is O(1). If it would be expensive to return a
packed array of every record in the page, then only some of the
records might be returned. Calling stasis_record_next() on { page->id,
off+ count } will tell the caller if it's received the entire
page's contents.
@todo Should recordArray be forced to return records in slot order?
*/
byte * (*recordPackedArray) (struct block_t *b, int * count, int * stride);
/**
This must be called when you're done with the block (before
releasePage) */
void (*blockRelease)(struct block_t *b);
void * impl;
} block_t;
/**
This function should work with any valid page implementation, but
it might be less efficient than a custom implementation.
This is a convenience function for page implementers. Other code
should call stasis_block_first() instead.
*/
block_t *stasis_block_first_default_impl(int xid, Page *p);
/**
This function should work with any valid page implementation, but
it might be less efficient than a custom implementation.
This is a convenience function for page implementers. Other code
should call stasis_block_next() instead.
*/
block_t * stasis_block_next_default_impl(int xid, Page *p, block_t *prev);
/**
This function should work with any valid page implementation, but
it might be less efficient than a custom implementation.
This is a convenience function for page implementers. Other code
should call stasis_block_done() instead.
*/
void stasis_block_done_default_impl(int xid, Page *p, block_t *b);
block_t * stasis_block_first(int xid, Page * p);
block_t * stasis_block_next(int xid, Page * p, block_t * prev);
void stasis_block_done(int xid, Page * p, block_t * done);
/**
None of these functions obtain latches. Calling them without
holding rwlatch is an error. (Exception: recordDereference grabs the
latch for you... XXX does it?)
The function pointer should be null if your page implementation does
not support the method in question.
@todo Figure out what to do about readlock vs writelock...
@todo Get rid of .size in recordid?
New allocation sequence (normal operation)
pin
latch
stasis_record_alloc_begin
stasis_record_alloc_done
unlatch
(There is a race here; other transactions can see that the page
contains a new slot, but that the LSN hasn't been updated. This
seems to be benign. stasis_page_lsn_write() refuses to decrement
LSN's... If the lock manager is using LSNs for versioning, it
might get into a situation where the page has changed (the slot was
allocated), but the version wasn't bumped. I can't imagine this
causing trouble, unless the application is using the presence or
absence of an uninitialized slot as some sort of side channel....)
lsn = Tupdate(...)
latch
stasis_page_lsn_write
unlatch
unpin
New allocation sequence (recovery)
pin
latch
stasis_record_alloc_done
stasis_page_lsn_write
unlatch
unpin
*/
typedef struct page_impl {
int page_type;
int has_header;
// ---------- Record access
/**
@param size If non-null, will be set to the size of the record
that has been read.
@return pointer to read region. The pointer will be guaranteed
valid while the page is read latched by this caller, or while
the page is write latched, and no other method has been called on
this page. Return null on error. (XXX current implementations
abort/crash)
*/
const byte* (*recordRead)(int xid, Page *p, recordid rid);
/**
Like recordRead, but marks the page dirty, and provides a
non-const pointer.
@return a pointer to the buffer manager's copy of the record.
*/
byte* (*recordWrite)(int xid, Page *p, recordid rid);
void (*recordReadDone)(int xid, Page *p, recordid rid, const byte *b);
void (*recordWriteDone)(int xid, Page *p, recordid rid, byte *b);
/**
Check to see if a slot is a normal slot, or something else, such
as a blob. This is stored in the size field in the slotted page
structure.
@param p the page of interest
@param slot the slot in p that we're checking.
@return The type of this slot.
*/
int (*recordGetType)(int xid, Page *p, recordid rid);
/**
Change the type of a slot. Doing so must not change the record's
type.
*/
void (*recordSetType)(int xid, Page *p, recordid rid, int type);
/**
@return the length of a record in bytes. This value can be used
to safely access the pointers returned by page_impl.recordRead()
and page_impl.recordWrite()
*/
int (*recordGetLength)(int xid, Page *p, recordid rid);
/**
@param p the page that will be iterated over
@return the first slot on the page (in slot order), or NULLRID if the page is empty.
*/
recordid (*recordFirst)(int xid, Page *p);
/**
This returns the next slot in the page (in slot order).
@param rid If NULLRID, start at the beginning of the page.
@return The next record in the sequence, or NULLRID if done.
*/
recordid (*recordNext)(int xid, Page *p, recordid rid);
/**
This returns the last potentially occupied slot in the page (in slot order).
The slot is guaranteed to be valid if pageCompactSlotIDs was just called,
or if no records have been freed since the page was created or compacted.
@param p the page of interest.
@return a slot id greater than or equal to all valid slots on the page.
*/
recordid (*recordLast)(int xid, Page *p);
// -------- "Exotic" (optional) access methods.
/*
@return non-zero if this page implementation supports the block API
*/
int (*isBlockSupported)(int xid, Page *p);
/** Returns a block from this page. For some page implementations,
this is essentially the same as recordRead. Others can produce
more sophisticated blocks.
*/
block_t* (*blockFirst)(int xid, Page *p);
block_t* (*blockNext)(int xid, Page *p, block_t *prev);
void (*blockDone)(int xid, Page *p, block_t *done);
// -------- Allocation methods.
/**
This returns a lower bound on the amount of freespace in the
page
@param p the page whose freespace will be estimated.
@return The number of bytes of free space on the page, or (for
efficiency's sake) an underestimate.
*/
int (*pageFreespace)(int xid, Page * p);
/**
Compact the page in place. This operation should not change the
slot id's allocated to the records on the page. Instead, it
should update the estimate returned by page_impl.freespace().
Depending on the page implementation, this function may have
other side effects.
*/
void(*pageCompact)(Page *p);
/**
* Compact the slotids on the page; the order of valid records will
* be preserved, but holes due to invalid slots will be closed.
*/
void(*pageCompactSlotIDs)(int xid, Page *p);
/**
Generate a new, appropriately sized recordid. This is the first
of two allocation phases, and does not actually modify the page.
The caller of this function must call stasis_record_alloc_done() before
unlatching the page.
@see page_impl.recordPostAlloc()
@see Talloc() for an overview of allocation
@param xid The active transaction.
@param size The size of the new record
@return A new recordid, or NULLRID if there is not enough space
*/
recordid (*recordPreAlloc)(int xid, Page *p, int size);
/** Allocate a new record with the supplied rid. This is the second
of two allocation phases. The supplied record must use an
unoccupied slot, and there must be enough freespace on the page.
@param xid The active transaction
@param p The page that will be allocated from
@param rid A new recordid that is (usually) from stasis_record_alloc_begin()
@see Talloc(), page_impl.recordPreAlloc()
*/
void (*recordPostAlloc)(int xid, Page *p, recordid rid);
/** Reorder records within a page. The second slot will be moved before the
* first slot; all intermediate slots will be shifted up one slotid. The
* intermediate records must always be valid; this can be ensured by calling
* pageCompact(). As long as no records are freed after the page is allocated
* or compacted, all intermediate records will be valid.
*
* @param xid The active transaction
* @param p The page that will be modified. The page must be latched against writes.
* @param first The first slot to be shifted to higher slots in the page.
* @param second The slot which will be moved to first's original location. Second
* must be greater than first.
*/
void (*recordSplice)(int xid, Page *p, slotid_t first, slotid_t second);
/** Free a record. The page implementation doesn't need to worry
about uncommitted deallocations; that is handled by a higher
level.
@param xid The active transaction
@param page The page that contains the record that will be freed.
@param rid the recordid to be freed.
@see Tdealloc() for example usage
*/
void (*recordFree)(int xid, Page *p, recordid rid);
/**
@todo Is recordDereference a page-implementation specific
operation, or should it be implemented once and for all in
page.c?
indirect.c suggests this should be specific to the page type;
blobs suggest this should be specific to record type. Perhaps
two levels of dereferencing are in order... (First: page
specific; second record specific...)
*/
recordid (*recordDereference)(int xid, Page *p, recordid rid);
// -------- Page maintenance
/** This is called when the page is read from disk.
This function should set p->LSN to an appropriate value.
*/
void (*pageLoaded)(Page * p);
/** This is called before the page is written back to disk.
This function should record p->LSN somewhere appropriate
(perhaps via stasis_page_lsn_ptr()), and should prepare the page
to be written back to disk.
*/
void (*pageFlushed)(Page * p);
/** This is called before the page is evicted from memory.
At this point the page has already been written back to disk
(if necessary). Any resources held on behalf of this page
should be released.
*/
void (*pageCleanup)(Page * p);
} page_impl;
/**
Register a new page type with Stasis. The pageType field of impl
will be checked for uniqueness. If the type is not unique, this
function will return non-zero.
(Since error handling isn't really working yet, it aborts if the
page type is not unique.)
*/
int stasis_page_impl_register(page_impl impl);
/**
Get the page_impl for a particular page type. This isn't set
const, so you can write to the function pointers. However, no
attempt has been made to make such things thread safe, so stasis'
worker threads can cause all sorts of undefined trouble if you poke
this.
*/
page_impl* stasis_page_impl_get(int id);
page_impl slottedLsnFreeImpl();
page_impl segmentImpl();
#define stasis_mempage_memaddr(p) (p)
#define stasis_page(x) stasis_mempage_##x
#define PAGE byte
#include "page/page-impl.h"
#undef stasis_mempage_memaddr
#undef stasis_page
#undef PAGE
#include "page/page-impl.h"
END_C_DECLS
#endif
Reference in a new issue View git blame Copy permalink