fis-gtm/sr_unix/dsk_read.c

221 lines
10 KiB
C
Raw Normal View History

/****************************************************************
* *
* Copyright 2001, 2012 Fidelity Information Services, Inc *
* *
* This source code contains the intellectual property *
* of its copyright holder(s), and is made available *
* under a license. If you do not know the terms of *
* the license, please stop and do not read further. *
* *
****************************************************************/
#include "mdef.h"
#include <sys/types.h>
#include "gtm_unistd.h"
#include "gtm_string.h"
#include <signal.h>
#include <errno.h>
#ifdef DEBUG
#include "gtm_stdio.h"
#endif
#include "gdsroot.h"
#include "gdsblk.h"
#include "gtm_facility.h"
#include "fileinfo.h"
#include "gdsbt.h"
#include "gdsfhead.h"
#include "filestruct.h"
#include "iosp.h"
#include "error.h"
#include "gtmio.h"
#include "gds_blk_upgrade.h"
#include "gdsbml.h"
#ifdef GTM_CRYPT
#include "gtmcrypt.h"
#endif
#include "gdsdbver.h"
#include "min_max.h"
#include "gtmimagename.h"
#include "memcoherency.h"
GBLREF gd_region *gv_cur_region;
GBLREF sgmnt_addrs *cs_addrs;
GBLREF sgmnt_data_ptr_t cs_data;
GBLREF volatile int4 fast_lock_count;
GBLREF boolean_t dse_running;
error_def(ERR_DYNUPGRDFAIL);
int4 dsk_read (block_id blk, sm_uc_ptr_t buff, enum db_ver *ondsk_blkver, boolean_t blk_free)
{
unix_db_info *udi;
int4 size, save_errno;
enum db_ver tmp_ondskblkver;
sm_uc_ptr_t save_buff = NULL, enc_save_buff;
boolean_t fully_upgraded, buff_is_modified_after_lseekread;
int bsiz;
DEBUG_ONLY(
blk_hdr_ptr_t blk_hdr_val;
static int in_dsk_read;
)
/* It is possible that the block that we read in from disk is a V4 format block. The database block scanning routines
* (gvcst_*search*.c) that might be concurrently running currently assume all global buffers (particularly the block
* headers) are V5 format. They are not robust enough to handle a V4 format block. Therefore we do not want to
* risk reading a potential V4 format block directly into the cache and then upgrading it. Instead we read it into
* a private buffer, upgrade it there and then copy it over to the cache in V5 format. This is the static variable
* read_reformat_buffer. We could have as well used the global variable "reformat_buffer" for this purpose. But
* that would then prevent dsk_reads and concurrent dsk_writes from proceeding. We dont want that loss of asynchronocity.
* Hence we keep them separate. Note that while "reformat_buffer" is used by a lot of routines, "read_reformat_buffer"
* is used only by this routine and hence is a static instead of a GBLDEF.
*/
static sm_uc_ptr_t read_reformat_buffer;
static int read_reformat_buffer_len;
GTMCRYPT_ONLY(
int req_dec_blk_size;
int crypt_status;
boolean_t is_encrypted;
)
/* Note: Even in snapshots, only INTEG requires dsk_read to read FREE blocks. The assert below should be modified
* if we later introduce a scheme where we can figure out as to who started the snapshots and assert accordingly
*/
assert(!blk_free || SNAPSHOTS_IN_PROG(cs_addrs)); /* Only SNAPSHOTS require dsk_read to read a FREE block from the disk */
assert(0 == in_dsk_read); /* dsk_read should never be nested. the read_reformat_buffer logic below relies on this */
DEBUG_ONLY(in_dsk_read++;)
udi = (unix_db_info *)(gv_cur_region->dyn.addr->file_cntl->file_info);
assert(cs_addrs->hdr == cs_data);
size = cs_data->blk_size;
assert (cs_data->acc_meth == dba_bg);
/* Since cs_data->fully_upgraded is referenced more than once in this module (once explicitly and once in
* GDS_BLK_UPGRADE_IF_NEEDED macro used below), take a copy of it and use that so all usages see the same value.
* Not doing this, for example, can cause us to see the database as fully upgraded in the first check causing us
* not to allocate save_buff (a temporary buffer to hold a V4 format block) at all but later in the macro
* we might see the database as NOT fully upgraded so we might choose to call the function gds_blk_upgrade which
* does expect a temporary buffer to have been pre-allocated. It is ok if the value of cs_data->fully_upgraded
* changes after we took a copy of it since we have a buffer locked for this particular block (at least in BG)
* so no concurrent process could be changing the format of this block. For MM there might be an issue.
*/
fully_upgraded = cs_data->fully_upgraded;
if (!blk_free && !fully_upgraded) /* No V4->V5 translations required if block is FREE */
{
buff_is_modified_after_lseekread = TRUE;
save_buff = buff;
if (size > read_reformat_buffer_len)
{ /* do the same for the reformat_buffer used by dsk_read */
assert(0 == fast_lock_count); /* this is mainline (non-interrupt) code */
++fast_lock_count; /* No interrupts in free/malloc across this change */
if (NULL != read_reformat_buffer)
free(read_reformat_buffer);
read_reformat_buffer = malloc(size);
read_reformat_buffer_len = size;
--fast_lock_count;
}
buff = read_reformat_buffer;
} else
buff_is_modified_after_lseekread = FALSE;
assert(NULL != cs_addrs->nl);
INCR_GVSTATS_COUNTER(cs_addrs, cs_addrs->nl, n_dsk_read, 1);
enc_save_buff = buff;
# ifdef GTM_CRYPT
is_encrypted = cs_data->is_encrypted;
if (is_encrypted)
{
DBG_ENSURE_PTR_IS_VALID_GLOBUFF(cs_addrs, cs_data, buff);
enc_save_buff = GDS_ANY_ENCRYPTGLOBUF(buff, cs_addrs);
DBG_ENSURE_PTR_IS_VALID_ENCTWINGLOBUFF(cs_addrs, cs_data, enc_save_buff);
}
# endif
LSEEKREAD(udi->fd,
(DISK_BLOCK_SIZE * (cs_data->start_vbn - 1) + (off_t)blk * size),
enc_save_buff,
size,
save_errno);
assert((0 == save_errno) GTM_TRUNCATE_ONLY(|| (-1 == save_errno)));
# ifdef GTM_CRYPT
if (is_encrypted && (0 == save_errno))
{
bsiz = (int)((blk_hdr_ptr_t)enc_save_buff)->bsiz;
req_dec_blk_size = MIN(cs_data->blk_size, bsiz) - SIZEOF(blk_hdr);
buff_is_modified_after_lseekread = TRUE;
/* Do not do encryption/decryption if block is FREE */
if (!blk_free && (IS_BLK_ENCRYPTED(((blk_hdr_ptr_t)enc_save_buff)->levl, req_dec_blk_size)))
{
/* The below assert cannot be moved before BLOCK_REQUIRE_ENCRYPTION check done above as tmp_ptr could
* potentially point to a V4 block in which case the assert might fail when a V4 block is casted to
* a V5 block header.
*/
assert((bsiz <= cs_data->blk_size) && (bsiz >= SIZEOF(blk_hdr)));
ASSERT_ENCRYPTION_INITIALIZED;
memcpy(buff, enc_save_buff, SIZEOF(blk_hdr));
GTMCRYPT_DECODE_FAST(cs_addrs->encr_key_handle,
(char *)enc_save_buff + SIZEOF(blk_hdr),
req_dec_blk_size,
(char *)buff + SIZEOF(blk_hdr),
crypt_status);
if (0 != crypt_status)
{
GC_RTS_ERROR(crypt_status, gv_cur_region->dyn.addr->fname);
return crypt_status;
}
} else
memcpy(buff, enc_save_buff, size);
}
# endif
if (!blk_free && (0 == save_errno))
{ /* See if block needs to be converted to current version. Assuming buffer is at least short aligned */
assert(0 == (long)buff % 2);
/* GDSV4 (0) version uses "buff->bver" as a block length so should always be > 0 when M code is running.
* The only exception is if the block has not been initialized (possible if it is BLK_FREE status in the
* bitmap). This is possible due to concurrency issues while traversing down the tree. But if we have
* crit on this region, we should not see these either.
*/
assert(!IS_MCODE_RUNNING || !cs_addrs->now_crit || ((blk_hdr_ptr_t)buff)->bver);
/* Block must be converted to current version (if necessary) for use by internals.
* By definition, all blocks are converted from/to their on-disk version at the IO point.
*/
GDS_BLK_UPGRADE_IF_NEEDED(blk, buff, save_buff, cs_data, &tmp_ondskblkver, save_errno, fully_upgraded);
DEBUG_DYNGRD_ONLY(
if (GDSVCURR != tmp_ondskblkver)
PRINTF("DSK_READ: Block %d being dynamically upgraded on read\n", blk);
)
assert((GDSV5 == tmp_ondskblkver) || (NULL != save_buff)); /* never read a V4 block directly into cache */
if (NULL != ondsk_blkver)
*ondsk_blkver = tmp_ondskblkver;
/* a bitmap block should never be short of space for a dynamic upgrade. assert that. */
assert((NULL == ondsk_blkver) || !IS_BITMAP_BLK(blk) || (ERR_DYNUPGRDFAIL != save_errno));
/* If we didn't run gds_blk_upgrade which would move the block into the cache, we need to do
* it ourselves. Note that buff will be cleared by the GDS_BLK_UPGRADE_IF_NEEDED macro if
* buff and save_buff are different and gds_blk_upgrade was called.
*/
if ((NULL != save_buff) && (NULL != buff)) /* Buffer not moved by upgrade, we must move */
memcpy(save_buff, buff, size);
}
if (buff_is_modified_after_lseekread)
{ /* Normally the disk read (done in LSEEKREAD macro) would do the necessary write memory barrier to make the
* updated shared memory global buffer contents visible to all other processes as long as they see any later
* updates done to shared memory by the reader. But in case of a V4 -> V5 upgrade or reading of an encrypted
* block, the actual disk read would have happened into a different buffer. That would then be used as a
* source for the upgrade or decryption before placing the final contents in the input global buffer.
* We now need a write memory barrier before returning from this function to publish this shared memory
* update to other processes waiting on this read. Note: it is possible in rare cases (e.g. mupip reorg upgrade)
* that the input buffer is NOT a shared memory buffer in which case the write memory barrier is not necessary
* but it is not easily possible to identify that and we want to save if checks on the fast path and so do
* the memory barrier in all cases.
*/
SHM_WRITE_MEMORY_BARRIER;
}
DEBUG_ONLY(
in_dsk_read--;
if (!blk_free && cs_addrs->now_crit && !dse_running && (0 == save_errno))
{ /* Do basic checks on GDS block that was just read. Do it only if holding crit as we could read
* uninitialized blocks otherwise. Also DSE might read bad blocks even inside crit so skip checks.
*/
blk_hdr_val = (NULL != save_buff) ? (blk_hdr_ptr_t)save_buff : (blk_hdr_ptr_t)buff;
GDS_BLK_HDR_CHECK(cs_data, blk_hdr_val, fully_upgraded);
}
)
return save_errno;
}