Skip to content

M
mariadb
Commit 4f56acb6 authored by Sergey Petrunya's avatar Sergey Petrunya
Browse files
Options

Better comments, move Lifo_buffer to separate file.

parent 3066c377
Show whitespace changes
Inline Side-by-side
Showing with 77 additions and 464 deletions
sql/CMakeLists.txt
View file @ 4f56acb6
......@@ -63,6 +63,7 @@ SET (SQL_SOURCE
sql_cache.cc sql_class.cc sql_client.cc sql_crypt.cc sql_crypt.h
sql_cursor.cc sql_db.cc sql_delete.cc sql_derived.cc sql_do.cc
sql_error.cc sql_handler.cc sql_help.cc sql_insert.cc
sql_lifo_buffer.h
sql_join_cache.cc sql_lex.cc sql_list.cc sql_load.cc sql_manager.cc
sql_map.cc sql_parse.cc sql_partition.cc sql_plugin.cc
sql_prepare.cc sql_rename.cc
......
sql/Makefile.am
View file @ 4f56acb6
......@@ -66,6 +66,7 @@ noinst_HEADERS = item.h item_func.h item_sum.h item_cmpfunc.h \
log.h log_slow.h sql_show.h rpl_rli.h rpl_mi.h \
sql_select.h structs.h table.h sql_udf.h hash_filo.h \
lex.h lex_symbol.h sql_acl.h sql_crypt.h \
sql_lifo_buffer.h \
sql_repl.h slave.h rpl_filter.h rpl_injector.h \
log_event.h rpl_record.h \
log_event_old.h rpl_record_old.h \
......
sql/handler.h
View file @ 4f56acb6
......@@ -1326,6 +1326,10 @@ void get_sweep_read_cost(TABLE *table, ha_rows nrows, bool interrupted,
The MRR user has materialized range keys somewhere in the user's buffer.
This can be used for optimization of the procedure that sorts these keys
since in this case key values don't have to be copied into the MRR buffer.
In other words, it is guaranteed that after RANGE_SEQ_IF::next() call the
pointer in range->start_key.key will point to a key value that will remain
there until the end of the MRR scan.
*/
#define HA_MRR_MATERIALIZED_KEYS 256
......
sql/multi_range_read.cc
View file @ 4f56acb6
......@@ -332,7 +332,7 @@ int DsMrr_impl::dsmrr_init(handler *h_arg, RANGE_SEQ_IF *seq_funcs,
is_mrr_assoc= !test(mode & HA_MRR_NO_ASSOCIATION);
/*
Figure out what steps we'll need to do
Determine whether we'll need to do key sorting and/or rnd_pos() scan
*/
do_sort_keys= FALSE;
if ((mode & HA_MRR_SINGLE_POINT) &&
......@@ -362,8 +362,9 @@ int DsMrr_impl::dsmrr_init(handler *h_arg, RANGE_SEQ_IF *seq_funcs,
status_var_increment(table->in_use->status_var.ha_multi_range_read_init_count);
/*
At start, alloc all of the buffer for rowids. Key sorting code will grab a
piece if necessary.
At start, alloc all of the buffer for rowids. When/if key sorting code
figures how much buffer space it needs, it will call setup_buffer_sizes()
to re-distribute the buffer space.
*/
full_buf= buf->buffer;
full_buf_end= buf->buffer_end;
......@@ -530,22 +531,19 @@ static int rowid_cmp_reverse(void *h, uchar *a, uchar *b)
/**
DS-MRR: Fill and sort the rowid buffer
{This is an internal function of DiskSweep MRR implementation}
Scan the MRR ranges and collect ROWIDs (or {ROWID, range_id} pairs) into
buffer. When the buffer is full or scan is completed, sort the buffer by
rowid and return.
When this function returns, either rowid buffer is not empty, or the source
of lookup keys (i.e. ranges) is exhaused.
dsmrr_eof is set to indicate whether we've exhausted the list of ranges we're
scanning. This function never returns HA_ERR_END_OF_FILE.
post-condition:
rowid buffer is not empty, or key source is exhausted.
@retval 0 OK, the next portion of rowids is in the buffer,
properly ordered
@retval other Error
*/
int DsMrr_impl::dsmrr_fill_rowid_buffer()
......@@ -556,15 +554,13 @@ int DsMrr_impl::dsmrr_fill_rowid_buffer()
DBUG_ENTER("DsMrr_impl::dsmrr_fill_rowid_buffer");
DBUG_ASSERT(rowid_buffer.is_empty());
rowid_buffer.reset_for_writing();
rowid_buffer.reset();
rowid_buffer.setup_writing(&h2->ref, h2->ref_length,
is_mrr_assoc? (uchar**)&range_info_ptr: NULL, sizeof(void*));
is_mrr_assoc? (uchar**)&range_info_ptr: NULL,
sizeof(void*));
last_identical_rowid= NULL;
//if (do_sort_keys && key_buffer.is_reverse())
// key_buffer.flip();
while (rowid_buffer.can_write())
{
if (do_sort_keys)
......@@ -652,18 +648,21 @@ equals:
return 0;
}
int DsMrr_impl::key_tuple_cmp_reverse(void* arg, uchar* key1, uchar* key2)
{
return -key_tuple_cmp(arg, key1, key2);
}
/*
Setup key/rowid buffer sizes based on sample_key
DESCRIPTION
/**
Setup key/rowid buffer sizes based on sample_key and its length.
This function must be called when all buffer space is empty.
@param
sample_key A lookup key to use as a sample. It is assumed that
all other keys will have the same length/etc.
@note
This function must be called when all buffers are empty
*/
void DsMrr_impl::setup_buffer_sizes(key_range *sample_key)
......@@ -737,22 +736,19 @@ void DsMrr_impl::setup_buffer_sizes(key_range *sample_key)
}
/*
/**
DS-MRR/CPK: Fill the buffer with (lookup_tuple, range_id) pairs and sort
SYNOPSIS
DsMrr_impl::dsmrr_fill_key_buffer()
Enumerate the input range (=key) sequence, fill the key buffer with
(lookup_key, range_id) pairs and sort it.
DESCRIPTION
DS-MRR/CPK: Enumerate the input range (=key) sequence, fill the key buffer
(lookup_key, range_id) pairs and sort.
When this function returns, either
- key buffer is non-empty, or
- key buffer is empty and source range sequence is exhausted
@note
dsmrr_eof is set to indicate whether we've exhausted the list of ranges
we're scanning.
post-condition:
- key buffer is non-empty
- key buffer is empty and source range sequence is exhausted
*/
void DsMrr_impl::dsmrr_fill_key_buffer()
......@@ -778,7 +774,7 @@ void DsMrr_impl::dsmrr_fill_key_buffer()
identical_key_it= &backward_key_it;
key_buffer->set_buffer_space(rowid_buffer_end, full_buf_end);
}
key_buffer->reset_for_writing();
key_buffer->reset();
key_buffer->setup_writing(&key_ptr, key_size_in_keybuf,
is_mrr_assoc? (uchar**)&range_info_ptr : NULL,
sizeof(uchar*));
......@@ -825,8 +821,8 @@ void DsMrr_impl::dsmrr_fill_key_buffer()
}
/*
Take unused space from key buffer and give it to rowid buffer.
/**
Take unused space from the key buffer and give it to the rowid buffer
*/
void DsMrr_impl::reallocate_buffer_space()
......@@ -837,30 +833,25 @@ void DsMrr_impl::reallocate_buffer_space()
}
/*
/**
DS-MRR/CPK: multi_range_read_next() function
DESCRIPTION
DsMrr_impl::dsmrr_next_from_index()
range_info OUT identifier of range that the returned record belongs to
DESCRIPTION
@param range_info OUT identifier of range that the returned record belongs to
@note
This function walks over key buffer and does index reads, i.e. it produces
{current_record, range_id} pairs.
The function has the same call contract like multi_range_read_next()'s.
We actually iterate nested sequences:
We actually iterate over nested sequences:
- a disjoint sequence of index ranges
- each range has multiple records
- each record goes into multiple identical ranges.
RETURN
0 OK, next record was successfully read
HA_ERR_END_OF_FILE End of records
Other Some other error
@retval 0 OK, next record was successfully read
@retval HA_ERR_END_OF_FILE End of records
@retval Other Some other error
*/
int DsMrr_impl::dsmrr_next_from_index(char **range_info_arg)
......@@ -1007,7 +998,9 @@ end:
/**
DS-MRR implementation: multi_range_read_next() function
DS-MRR implementation: multi_range_read_next() function.
Calling convention is like multi_range_read_next() has.
*/
int DsMrr_impl::dsmrr_next(char **range_info)
......@@ -1237,17 +1230,12 @@ bool key_uses_partial_cols(TABLE *table, uint keyno)
/*
Check if key/flags allow DS-MRR/CPK strategy to be used
SYNOPSIS
DsMrr_impl::check_cpk_scan()
keyno Index that will be used
mrr_flags
@param thd
@param keyno Index that will be used
@param mrr_flags
DESCRIPTION
Check if key/flags allow DS-MRR/CPK strategy to be used.
RETURN
TRUE DS-MRR/CPK should be used
FALSE Otherwise
@retval TRUE DS-MRR/CPK should be used
@retval FALSE Otherwise
*/
bool DsMrr_impl::check_cpk_scan(THD *thd, uint keyno, uint mrr_flags)
......@@ -1414,16 +1402,13 @@ bool DsMrr_impl::get_disk_sweep_mrr_cost(uint keynr, ha_rows rows, uint flags,
/*
Get cost of one sort-and-sweep step
SYNOPSIS
get_sort_and_sweep_cost()
table Table being accessed
nrows Number of rows to be sorted and retrieved
cost OUT The cost
DESCRIPTION
Get cost of these operations:
It consists of two parts:
- sort an array of #nrows ROWIDs using qsort
- read #nrows records from table in a sweep.
@param table Table being accessed
@param nrows Number of rows to be sorted and retrieved
@param cost OUT The cost of scan
*/
static
......
sql/multi_range_read.h
View file @ 4f56acb6
/*
This file contains declarations for Disk-Sweep MultiRangeRead (DS-MRR)
implementation
/**
@defgroup DS-MRR declarations
@{
*/
/**
......@@ -46,387 +46,7 @@
storage and has better performance when reading data in rowid order.
*/
class Forward_lifo_buffer;
class Backward_lifo_buffer;
class Lifo_buffer
{
protected:
/*
Data to be written. write() call will assume that (*write_ptr1) points to
size1 bytes of data to be written.
If write_ptr2 != NULL then the buffer stores pairs, and (*write_ptr2)
points to size2 bytes of data that form the second component.
*/
uchar **write_ptr1;
size_t size1;
uchar **write_ptr2;
size_t size2;
/*
read() will do reading by storing pointer to read data into *read_ptr1 (if
the buffer stores atomic elements), or into {*read_ptr1, *read_ptr2} (if
the buffer stores pairs).
*/
uchar **read_ptr1;
uchar **read_ptr2;
uchar *start; /* points to start of buffer space */
uchar *end; /* points to just beyond the end of buffer space */
public:
enum enum_direction {
BACKWARD=-1, /* buffer is filled/read from bigger to smaller memory addresses */
FORWARD=1 /* buffer is filled/read from smaller to bigger memory addresses */
};
virtual enum_direction type() = 0;
/* Buffer space control functions */
void set_buffer_space(uchar *start_arg, uchar *end_arg)
{
start= start_arg;
end= end_arg;
TRASH(start, end - start);
reset_for_writing();
}
void setup_writing(uchar **data1, size_t len1, uchar **data2, size_t len2)
{
write_ptr1= data1;
size1= len1;
write_ptr2= data2;
size2= len2;
}
void setup_reading(uchar **data1, size_t len1, uchar **data2, size_t len2)
{
read_ptr1= data1;
DBUG_ASSERT(len1 == size1);
read_ptr2= data2;
DBUG_ASSERT(len2 == size2);
}
//virtual void write_bytes(const uchar *data, size_t bytes)=0;
virtual bool read() = 0;
virtual void write() = 0;
bool can_write()
{
return have_space_for(size1 + (write_ptr2 ? size2 : 0));
}
bool is_empty() { return used_size() == 0; }
virtual size_t used_size() = 0;
void sort(qsort2_cmp cmp_func, void *cmp_func_arg)
{
uint elem_size= size1 + (write_ptr2 ? size2 : 0);
uint n_elements= used_size() / elem_size;
my_qsort2(used_area(), n_elements, elem_size, cmp_func, cmp_func_arg);
}
virtual void reset_for_writing() = 0;
virtual uchar *end_of_space() = 0;
bool have_data(size_t bytes)
{
return (used_size() >= bytes);
}
virtual bool have_space_for(size_t bytes) = 0;
//virtual uchar *read_bytes(size_t bytes) = 0;
virtual void remove_unused_space(uchar **unused_start, uchar **unused_end)=0;
virtual uchar *used_area() = 0;
class Iterator
{
public:
virtual void init(Lifo_buffer *buf) = 0;
/*
Read the next value. The calling convention is the same as buf->read()
has.
RETURN
FALSE - Ok
TRUE - EOF, reached the end of the buffer
*/
virtual bool read_next()= 0;
virtual ~Iterator() {}
protected:
Lifo_buffer *buf;
virtual uchar *get_next(size_t nbytes)=0;
};
virtual ~Lifo_buffer() {};
friend class Forward_iterator;
friend class Backward_iterator;
};
class Forward_lifo_buffer: public Lifo_buffer
{
uchar *pos;
public:
enum_direction type() { return FORWARD; }
size_t used_size()
{
return pos - start;
}
void reset_for_writing()
{
pos= start;
}
uchar *end_of_space() { return pos; }
bool have_space_for(size_t bytes)
{
return (pos + bytes < end);
}
void write()
{
write_bytes(*write_ptr1, size1);
if (write_ptr2)
write_bytes(*write_ptr2, size2);
}
void write_bytes(const uchar *data, size_t bytes)
{
DBUG_ASSERT(have_space_for(bytes));
memcpy(pos, data, bytes);
pos += bytes;
}
uchar *read_bytes(size_t bytes)
{
DBUG_ASSERT(have_data(bytes));
pos= pos - bytes;
return pos;
}
bool read()
{
if (!have_data(size1 + (read_ptr2 ? size2 : 0)))
return TRUE;
if (read_ptr2)
*read_ptr2= read_bytes(size2);
*read_ptr1= read_bytes(size1);
return FALSE;
}
/*
Stop using/return the unneded space (the one that we have already wrote
to read from).
*/
void remove_unused_space(uchar **unused_start, uchar **unused_end)
{
DBUG_ASSERT(0); /* Don't need this yet */
}
void grow(uchar *unused_start, uchar *unused_end)
{
/*
Passed memory area can be meaningfully used for growing the buffer if:
- it is adjacent to buffer space we're using
- it is on the end towards which we grow.
*/
DBUG_ASSERT(unused_end >= unused_start);
TRASH(unused_start, unused_end - unused_start);
DBUG_ASSERT(end == unused_start);
end= unused_end;
}
/* Return pointer to start of the memory area that is occupied by the data */
uchar *used_area() { return start; }
friend class Forward_iterator;
};
class Forward_iterator : public Lifo_buffer::Iterator
{
uchar *pos;
/* Return pointer to next chunk of nbytes bytes and avance over it */
uchar *get_next(size_t nbytes)
{
if (pos - nbytes < ((Forward_lifo_buffer*)buf)->start)
return NULL;
pos -= nbytes;
return pos;
}
public:
bool read_next()
{
uchar *res;
if (buf->read_ptr2)
{
if ((res= get_next(buf->size2)))
{
*(buf->read_ptr2)= res;
*buf->read_ptr1= get_next(buf->size1);
return FALSE;
}
}
else
{
if ((res= get_next(buf->size1)))
{
*(buf->read_ptr1)= res;
return FALSE;
}
}
return TRUE; /* EOF */
}
void init(Lifo_buffer *buf_arg)
{
DBUG_ASSERT(buf_arg->type() == Lifo_buffer::FORWARD);
buf= buf_arg;
pos= ((Forward_lifo_buffer*)buf)->pos;
}
};
class Backward_lifo_buffer: public Lifo_buffer
{
uchar *pos;
public:
enum_direction type() { return BACKWARD; }
size_t used_size()
{
return end - pos;
}
void reset_for_writing()
{
pos= end;
}
uchar *end_of_space() { return end; }
bool have_space_for(size_t bytes)
{
return (pos - bytes >= start);
}
void write()
{
if (write_ptr2)
write_bytes(*write_ptr2, size2);
write_bytes(*write_ptr1, size1);
}
void write_bytes(const uchar *data, size_t bytes)
{
DBUG_ASSERT(have_space_for(bytes));
pos -= bytes;
memcpy(pos, data, bytes);
}
bool read()
{
if (!have_data(size1 + (read_ptr2 ? size2 : 0)))
return TRUE;
*read_ptr1= read_bytes(size1);
if (read_ptr2)
*read_ptr2= read_bytes(size2);
return FALSE;
}
uchar *read_bytes(size_t bytes)
{
DBUG_ASSERT(have_data(bytes));
uchar *ret= pos;
pos= pos + bytes;
return ret;
}
/*
Stop using/return the unneded space (the one that we have already wrote
to and have read from).
*/
void remove_unused_space(uchar **unused_start, uchar **unused_end)
{
*unused_start= start;
*unused_end= pos;
start= pos;
}
void grow(uchar *unused_start, uchar *unused_end)
{
/*
Passed memory area can be meaningfully used for growing the buffer if:
- it is adjacent to buffer space we're using
- it is on the end towards which we grow.
*/
/*
DBUG_ASSERT(unused_end >= unused_start);
TRASH(unused_start, unused_end - unused_start);
DBUG_ASSERT(start == unused_end);
start= unused_start;
*/
DBUG_ASSERT(0); //Not used
}
/* Return pointer to start of the memory area that is occupied by the data */
uchar *used_area() { return pos; }
friend class Backward_iterator;
};
class Backward_iterator : public Lifo_buffer::Iterator
{
uchar *pos;
/* Return pointer to next chunk of nbytes bytes and advance over it */
uchar *get_next(size_t nbytes)
{
if (pos + nbytes > ((Backward_lifo_buffer*)buf)->end)
return NULL;
uchar *res= pos;
pos += nbytes;
return res;
}
public:
bool read_next()
{
/*
Always read the first component first (if the buffer is backwards, we
have written the second component first).
*/
uchar *res;
if ((res= get_next(buf->size1)))
{
*(buf->read_ptr1)= res;
if (buf->read_ptr2)
*buf->read_ptr2= get_next(buf->size2);
return FALSE;
}
return TRUE; /* EOF */
}
void init(Lifo_buffer *buf_arg)
{
DBUG_ASSERT(buf_arg->type() == Lifo_buffer::BACKWARD);
buf= buf_arg;
pos= ((Backward_lifo_buffer*)buf)->pos;
}
};
/*
An in-memory buffer used by DS-MRR implementation.
- The buffer contains fixed-size elements. The elements are either atomic
byte sequences or pairs.
- The buffer resides in memory provided by the user. It is possible to
= dynamically (ie. between write operations) add ajacent memory space to
the buffer
= dynamically remove unused space from the buffer.
- Buffer can be set to be either "forward" or "backward".
The intent of the last two properties is to allow to have two buffers on
adjacent memory space, one is being read from (and so its space shrinks)
while the other is being written to (and so it needs more and more space).
Illustration of forward buffer operation:
+-- next read will read from here
|
| +-- next write will write to here
v v
*--------------*===============*----------------*
| ^ | ^ | |
| | read_pos | write_pos |
start | | end
| |
usused space user data
For reverse buffer, start/end have the same meaning, but reading and
writing is done from end to start.
*/
#include "sql_lifo_buffer.h"
/*
DS-MRR implementation for one table. Create/use one object of this class for
......@@ -440,7 +60,7 @@ public:
- Key-Ordered Retrieval
- Rowid-Ordered Retrieval
DsMrr_impl will use one of the above strategies, or combination of them,
DsMrr_impl will use one of the above strategies, or a combination of them,
according to the following diagram:
(mrr function calls)
......@@ -470,7 +90,7 @@ public:
(table records and range_ids)
The choice of strategy depends on MRR scan properties, table properties
(whether we're scanning clustered primary key), and @@optimizer_flag
(whether we're scanning clustered primary key), and @@optimizer_switch
settings.
Key-Ordered Retrieval
......@@ -541,7 +161,7 @@ private:
/*
Secondary handler object. (created when needed, we need it when we need
to run both index scan and rnd_pos() at the same time)
to run both index scan and rnd_pos() scan at the same time)
*/
handler *h2;
......@@ -568,14 +188,13 @@ private:
uchar *full_buf_end;
/*
When using both rowid and key buffers: the bound between key and rowid
When using both rowid and key buffers: the boundary between key and rowid
parts of the buffer. This is the "original" value, actual memory ranges
used by key and rowid parts may be different because of dynamic space
reallocation between them.
*/
uchar *rowid_buffer_end;
/** Index scaning and key buffer-related members **/
/* TRUE <=> We can get at most one index tuple for a lookup key */
......@@ -689,4 +308,7 @@ private:
static uint key_buf_seq_next(range_seq_t rseq, KEY_MULTI_RANGE *range);
};
/**
@} (end of group DS-MRR declarations)
*/
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
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7