void set_enospc_flags(gd_addr *addr_ptr, char enospc_enable_list[], boolean_t ok_to_interrupt)
{
gd_region *r_local, *r_top;
int i;
sgmnt_addrs *csa;
const char *syslog_msg;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
for (r_local = addr_ptr->regions, r_top = r_local + addr_ptr->n_regions, i = 0;
r_local < r_top; r_local++, i++)
{
if (!r_local->open || r_local->was_open)
continue;
if ((dba_bg != r_local->dyn.addr->acc_meth) && (dba_mm != r_local->dyn.addr->acc_meth))
continue;
csa = &FILE_INFO(r_local)->s_addrs;
if (ANTICIPATORY_FREEZE_ENABLED(csa))
{
switch(enospc_enable_list[i])
{
case NONE:
syslog_msg = "Turning off fake ENOSPC for both database and journal file.";
csa->nl->fake_db_enospc = FALSE;
csa->nl->fake_jnl_enospc = FALSE;
break;
case DB_ON:
syslog_msg = "Turning on fake ENOSPC only for database file.";
csa->nl->fake_db_enospc = TRUE;
csa->nl->fake_jnl_enospc = FALSE;
break;
case JNL_ON:
syslog_msg = "Turning on fake ENOSPC only for journal file.";
csa->nl->fake_db_enospc = FALSE;
csa->nl->fake_jnl_enospc = TRUE;
break;
case DB_AND_JNL_ON:
syslog_msg = "Turning on fake ENOSPC for both database and journal file.";
csa->nl->fake_db_enospc = TRUE;
csa->nl->fake_jnl_enospc = TRUE;
break;
default:
assert(FALSE);
}
if (ok_to_interrupt)
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_TEXT, 2, DB_LEN_STR(r_local), ERR_TEXT, 2,
LEN_AND_STR(syslog_msg));
}
}
}
uint4 jnl_file_extend(jnl_private_control *jpc, uint4 total_jnl_rec_size)
{
file_control *fc;
boolean_t need_extend;
jnl_buffer_ptr_t jb;
jnl_create_info jnl_info;
jnl_file_header *header;
unsigned char hdr_buff[REAL_JNL_HDR_LEN + MAX_IO_BLOCK_SIZE];
uint4 new_alq;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
char prev_jnl_fn[JNL_NAME_SIZE];
uint4 jnl_status = 0, status;
int new_blocks, warn_blocks, result;
gtm_uint64_t avail_blocks;
uint4 aligned_tot_jrec_size, count;
uint4 jnl_fs_block_size, read_write_size;
DCL_THREADGBL_ACCESS;
switch(jpc->region->dyn.addr->acc_meth)
{
case dba_mm:
case dba_bg:
csa = &FILE_INFO(jpc->region)->s_addrs;
break;
default:
GTMASSERT;
}
csd = csa->hdr;
assert(csa == cs_addrs && csd == cs_data);
assert(csa->now_crit || (csd->clustered && (CCST_CLOSED == csa->nl->ccp_state)));
assert(&FILE_INFO(jpc->region)->s_addrs == csa);
assert(csa->jnl_state == csd->jnl_state);
assertpro(JNL_ENABLED(csa) && (NOJNL != jpc->channel) && (!JNL_FILE_SWITCHED(jpc)));
/* crit and messing with the journal file - how could it have vanished? */
if (!csd->jnl_deq || (csd->jnl_alq + csd->jnl_deq > csd->autoswitchlimit))
{
assert(DIVIDE_ROUND_UP(total_jnl_rec_size, DISK_BLOCK_SIZE) <= csd->jnl_alq);
assert(csd->jnl_alq == csd->autoswitchlimit);
new_blocks = csd->jnl_alq;
} else
/* May cause extension of csd->jnl_deq * n blocks where n > 0 */
new_blocks = ROUND_UP(DIVIDE_ROUND_UP(total_jnl_rec_size, DISK_BLOCK_SIZE), csd->jnl_deq);
jpc->status = SS_NORMAL;
jb = jpc->jnl_buff;
assert(0 <= new_blocks);
DEBUG_ONLY(count = 0);
for (need_extend = (0 != new_blocks); need_extend; )
{
DEBUG_ONLY(count++);
/* usually we will do the loop just once where we do the file extension.
* rarely we might need to do an autoswitch instead after which again rarely
* we might need to do an extension on the new journal to fit in the transaction's journal requirements.
* therefore we should do this loop a maximum of twice. hence the assert below.
*/
assert(count <= 2);
need_extend = FALSE;
if (SS_NORMAL == (status = disk_block_available(jpc->channel, &avail_blocks, TRUE)))
{
warn_blocks = (csd->jnl_alq + csd->jnl_deq > csd->autoswitchlimit)
? ((csd->jnl_deq > csd->autoswitchlimit) ? csd->jnl_deq : csd->autoswitchlimit)
: new_blocks;
if ((warn_blocks * EXTEND_WARNING_FACTOR) > avail_blocks)
{
if (new_blocks > avail_blocks)
{ /* If we cannot satisfy the request, it is an error, unless the anticipatory freeze
* scheme is in effect in which case, we will assume space is available even if
* it is not and go ahead with writes to the disk. If the writes fail with ENOSPC
* we will freeze the instance and wait for space to become available and keep
* retrying the writes. Therefore, we make the NOSPACEEXT a warning in this case.
*/
SETUP_THREADGBL_ACCESS;
if (!ANTICIPATORY_FREEZE_ENABLED(csa))
{
send_msg(VARLSTCNT(6) ERR_NOSPACEEXT, 4,
JNL_LEN_STR(csd), new_blocks, avail_blocks);
new_blocks = 0;
jpc->status = SS_NORMAL;
break;
} else
send_msg(VARLSTCNT(6) MAKE_MSG_WARNING(ERR_NOSPACEEXT), 4,
JNL_LEN_STR(csd), new_blocks, avail_blocks);
} else
send_msg(VARLSTCNT(5) ERR_DSKSPACEFLOW, 3, JNL_LEN_STR(csd), (avail_blocks - warn_blocks));
}
} else
send_msg(VARLSTCNT(5) ERR_JNLFILEXTERR, 2, JNL_LEN_STR(csd), status);
new_alq = jb->filesize + new_blocks;
/* ensure current journal file size is well within autoswitchlimit --> design constraint */
assert(csd->autoswitchlimit >= jb->filesize);
if (csd->autoswitchlimit < (jb->filesize + (EXTEND_WARNING_FACTOR * new_blocks))) /* close to max */
send_msg(VARLSTCNT(5) ERR_JNLSPACELOW, 3, JNL_LEN_STR(csd), csd->autoswitchlimit - jb->filesize);
if (csd->autoswitchlimit < new_alq)
{ /* Reached max, need to autoswitch */
/* Ensure new journal file can hold the entire current transaction's journal record requirements */
assert(csd->autoswitchlimit >= MAX_REQD_JNL_FILE_SIZE(total_jnl_rec_size));
memset(&jnl_info, 0, SIZEOF(jnl_info));
jnl_info.prev_jnl = &prev_jnl_fn[0];
set_jnl_info(gv_cur_region, &jnl_info);
assert(JNL_ENABLED(csa) && (NOJNL != jpc->channel) && !(JNL_FILE_SWITCHED(jpc)));
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{ /* flush the cache and jnl-buffer-contents to current journal file before
* switching to a new journal. Set a global variable in_jnl_file_autoswitch
* so jnl_write can know not to do the padding check. But because this is a global
* variable, we also need to make sure it is reset in case of errors during the
* autoswitch (or else calls to jnl_write after we are out of the autoswitch logic
* will continue to incorrectly not do the padding check. Hence a condition handler.
*/
assert(!in_jnl_file_autoswitch);
in_jnl_file_autoswitch = TRUE;
/* Also make sure time is not changed. This way if "jnl_write" as part of writing a
* journal record invokes jnl_file_extend, when the autoswitch is done and writing
* of the parent jnl_write resumes, we want it to continue with the same timestamp
* and not have to reset its time (non-trivial task) to reflect any changes since then.
*/
assert(!jgbl.save_dont_reset_gbl_jrec_time);
jgbl.save_dont_reset_gbl_jrec_time = jgbl.dont_reset_gbl_jrec_time;
jgbl.dont_reset_gbl_jrec_time = TRUE;
/* Establish a condition handler so we reset a few global variables that have
* temporarily been modified in case of errors inside wcs_flu/jnl_file_close.
*/
ESTABLISH_RET(jnl_file_autoswitch_ch, EXIT_ERR);
/* It is possible we still have not written a PINI record in this journal file
* (e.g. mupip extend saw the need to do jnl_file_extend inside jnl_write while
* trying to write a PINI record). Write a PINI record in that case before closing
* the journal file that way the EOF record will have a non-zero pini_addr.
*/
if (0 == jpc->pini_addr)
jnl_put_jrt_pini(csa);
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SPEEDUP_NOBEFORE);
jnl_file_close(gv_cur_region, TRUE, TRUE);
REVERT;
in_jnl_file_autoswitch = FALSE;
jgbl.dont_reset_gbl_jrec_time = jgbl.save_dont_reset_gbl_jrec_time;
DEBUG_ONLY(jgbl.save_dont_reset_gbl_jrec_time = FALSE);
assert((dba_mm == cs_data->acc_meth) || (csd == cs_data));
csd = cs_data; /* In MM, wcs_flu() can remap an extended DB, so reset csd to be sure */
} else
{
if (SS_NORMAL != jpc->status)
rts_error(VARLSTCNT(7) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region),
jpc->status);
else
rts_error(VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region));
}
assert(!jgbl.forw_phase_recovery || (NULL != jgbl.mur_pini_addr_reset_fnptr));
assert(jgbl.forw_phase_recovery || (NULL == jgbl.mur_pini_addr_reset_fnptr));
if (NULL != jgbl.mur_pini_addr_reset_fnptr)
(*jgbl.mur_pini_addr_reset_fnptr)(csa);
assert(!jnl_info.no_rename);
assert(!jnl_info.no_prev_link);
if (EXIT_NRM == cre_jnl_file(&jnl_info))
{
assert(0 == memcmp(csd->jnl_file_name, jnl_info.jnl, jnl_info.jnl_len));
assert(csd->jnl_file_name[jnl_info.jnl_len] == '\0');
assert(csd->jnl_file_len == jnl_info.jnl_len);
assert(csd->jnl_buffer_size == jnl_info.buffer);
assert(csd->jnl_alq == jnl_info.alloc);
assert(csd->jnl_deq == jnl_info.extend);
assert(csd->jnl_before_image == jnl_info.before_images);
csd->jnl_checksum = jnl_info.checksum;
csd->jnl_eovtn = csd->trans_hist.curr_tn;
send_msg(VARLSTCNT(4) ERR_NEWJNLFILECREAT, 2, JNL_LEN_STR(csd));
fc = gv_cur_region->dyn.addr->file_cntl;
fc->op = FC_WRITE;
fc->op_buff = (sm_uc_ptr_t)csd;
fc->op_len = SGMNT_HDR_LEN;
fc->op_pos = 1;
status = dbfilop(fc);
if (SS_NORMAL != status)
send_msg(VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), status);
assert(JNL_ENABLED(csa));
/* call jnl_ensure_open instead of jnl_file_open to make sure jpc->pini_addr is set to 0 */
jnl_status = jnl_ensure_open(); /* sets jpc->status */
if (0 != jnl_status)
{
if (jpc->status)
rts_error(VARLSTCNT(7) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region),
jpc->status);
else
rts_error(VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(gv_cur_region));
}
assert(jb->filesize == csd->jnl_alq);
if (csd->jnl_alq + csd->jnl_deq <= csd->autoswitchlimit)
{
aligned_tot_jrec_size = ALIGNED_ROUND_UP(MAX_REQD_JNL_FILE_SIZE(total_jnl_rec_size),
csd->jnl_alq, csd->jnl_deq);
if (aligned_tot_jrec_size > csd->jnl_alq)
{ /* need to extend more than initial allocation in the new journal file
* to accommodate the current transaction.
*/
new_blocks = aligned_tot_jrec_size - csd->jnl_alq;
assert(new_blocks);
assert(0 == new_blocks % csd->jnl_deq);
need_extend = TRUE;
}
}
} else
{
send_msg(VARLSTCNT(4) ERR_JNLNOCREATE, 2, JNL_LEN_STR(csd));
jpc->status = ERR_JNLNOCREATE;
new_blocks = -1;
}
} else
{
assert(!need_extend); /* ensure we won't go through the for loop again */
/* Virtually extend currently used journal file */
jnl_fs_block_size = jb->fs_block_size;
header = (jnl_file_header *)(ROUND_UP2((uintszofptr_t)hdr_buff, jnl_fs_block_size));
read_write_size = ROUND_UP2(REAL_JNL_HDR_LEN, jnl_fs_block_size);
assert((unsigned char *)header + read_write_size <= ARRAYTOP(hdr_buff));
DO_FILE_READ(jpc->channel, 0, header, read_write_size, jpc->status, jpc->status2);
if (SS_NORMAL != jpc->status)
{
assert(FALSE);
rts_error(VARLSTCNT(5) ERR_JNLRDERR, 2, JNL_LEN_STR(csd), jpc->status);
}
assert((header->virtual_size + new_blocks) == new_alq);
jb->filesize = new_alq; /* Actually this is virtual file size blocks */
header->virtual_size = new_alq;
JNL_DO_FILE_WRITE(csa, csd->jnl_file_name, jpc->channel, 0,
header, read_write_size, jpc->status, jpc->status2);
if (SS_NORMAL != jpc->status)
{
assert(FALSE);
rts_error(VARLSTCNT(5) ERR_JNLWRERR, 2, JNL_LEN_STR(csd), jpc->status);
}
}
if (0 >= new_blocks)
break;
}
if (0 < new_blocks)
{
INCR_GVSTATS_COUNTER(csa, csa->nl, n_jnl_extends, 1);
return EXIT_NRM;
}
jpc->status = ERR_JNLREADEOF;
jnl_file_lost(jpc, ERR_JNLEXTEND);
return EXIT_ERR;
}
OS_PAGE_SIZE_DECLARE
uint4 gdsfilext(uint4 blocks, uint4 filesize, boolean_t trans_in_prog)
{
sm_uc_ptr_t old_base[2], mmap_retaddr;
boolean_t was_crit, is_mm;
char buff[DISK_BLOCK_SIZE];
int result, save_errno, status;
uint4 new_bit_maps, bplmap, map, new_blocks, new_total, max_tot_blks;
uint4 jnl_status, to_wait, to_msg, wait_period;
gtm_uint64_t avail_blocks, mmap_sz;
off_t new_eof;
trans_num curr_tn;
unix_db_info *udi;
inctn_opcode_t save_inctn_opcode;
int4 prev_extend_blks_to_upgrd;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
DCL_THREADGBL_ACCESS;
assert(!IS_DSE_IMAGE);
assert((cs_addrs->nl == NULL) || (process_id != cs_addrs->nl->trunc_pid)); /* mu_truncate shouldn't extend file... */
assert(!process_exiting);
DEBUG_ONLY(old_base[0] = old_base[1] = NULL);
assert(!gv_cur_region->read_only);
udi = FILE_INFO(gv_cur_region);
is_mm = (dba_mm == cs_addrs->hdr->acc_meth);
# if !defined(MM_FILE_EXT_OK)
if (!udi->grabbed_access_sem && is_mm)
return (uint4)(NO_FREE_SPACE); /* should this be changed to show extension not allowed ? */
# endif
/* Both blocks and total blocks are unsigned ints so make sure we aren't asking for huge numbers that will
overflow and end up doing silly things.
*/
assert((blocks <= (MAXTOTALBLKS(cs_data) - cs_data->trans_hist.total_blks)) || WBTEST_ENABLED(WBTEST_FILE_EXTEND_ERROR));
if (!blocks)
return (uint4)(NO_FREE_SPACE); /* should this be changed to show extension not enabled ? */
bplmap = cs_data->bplmap;
/* New total of non-bitmap blocks will be number of current, non-bitmap blocks, plus new blocks desired
* There are (bplmap - 1) non-bitmap blocks per bitmap, so add (bplmap - 2) to number of non-bitmap blocks
* and divide by (bplmap - 1) to get total number of bitmaps for expanded database. (must round up in this
* manner as every non-bitmap block must have an associated bitmap)
* Current number of bitmaps is (total number of current blocks + bplmap - 1) / bplmap.
* Subtract current number of bitmaps from number needed for expanded database to get number of new bitmaps needed.
*/
new_bit_maps = DIVIDE_ROUND_UP(cs_data->trans_hist.total_blks
- DIVIDE_ROUND_UP(cs_data->trans_hist.total_blks, bplmap) + blocks, bplmap - 1)
- DIVIDE_ROUND_UP(cs_data->trans_hist.total_blks, bplmap);
new_blocks = blocks + new_bit_maps;
assert(0 < (int)new_blocks);
if (new_blocks + cs_data->trans_hist.total_blks > MAXTOTALBLKS(cs_data))
{
assert(FALSE);
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(1) ERR_TOTALBLKMAX);
return (uint4)(NO_FREE_SPACE);
}
if (0 != (save_errno = disk_block_available(udi->fd, &avail_blocks, FALSE)))
{
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), save_errno);
rts_error_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_DBFILERR, 2, DB_LEN_STR(gv_cur_region), save_errno);
} else
{
if (!(gtmDebugLevel & GDL_IgnoreAvailSpace))
{ /* Bypass this space check if debug flag above is on. Allows us to create a large sparce DB
* in space it could never fit it if wasn't sparse. Needed for some tests.
*/
avail_blocks = avail_blocks / (cs_data->blk_size / DISK_BLOCK_SIZE);
if ((blocks * EXTEND_WARNING_FACTOR) > avail_blocks)
{
if (blocks > (uint4)avail_blocks)
{
SETUP_THREADGBL_ACCESS;
if (!ANTICIPATORY_FREEZE_ENABLED(cs_addrs))
return (uint4)(NO_FREE_SPACE);
else
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(6) MAKE_MSG_WARNING(ERR_NOSPACEEXT), 4,
DB_LEN_STR(gv_cur_region), new_blocks, (uint4)avail_blocks);
} else
send_msg_csa(CSA_ARG(cs_addrs) VARLSTCNT(5) ERR_DSKSPACEFLOW, 3, DB_LEN_STR(gv_cur_region),
(uint4)(avail_blocks - ((new_blocks <= avail_blocks) ? new_blocks : 0)));
}
}
}
/* From here on, we need to use GDSFILEXT_CLNUP before returning to the caller */
was_crit = cs_addrs->now_crit;
assert(!cs_addrs->hold_onto_crit || was_crit);
/* If we are coming from mupip_extend (which gets crit itself) we better have waited for any unfreezes to occur.
* If we are coming from online rollback (when that feature is available), we will come in holding crit and in
* the final retry. In that case too, we expect to have waited for unfreezes to occur in the caller itself.
* Therefore if we are coming in holding crit from MUPIP, we expect the db to be unfrozen so no need to wait for
* freeze.
* If we are coming from GT.M and final retry (in which case we come in holding crit) we expect to have waited
* for any unfreezes (by invoking tp_crit_all_regions) to occur (TP or non-TP) before coming into this
* function. However, there is one exception. In the final retry, if tp_crit_all_regions notices that
* at least one of the participating regions did ONLY READs, it will not wait for any freeze on THAT region
* to complete before grabbing crit. Later, in the final retry, if THAT region did an update which caused
* op_tcommit to invoke bm_getfree->gdsfilext, then we would have come here with a frozen region on which
* we hold crit.
*/
assert(!was_crit || !cs_data->freeze || (dollar_tlevel && (CDB_STAGNATE <= t_tries)));
/*
* If we are in the final retry and already hold crit, it is possible that csa->nl->wc_blocked is also set to TRUE
* (by a concurrent process in phase2 which encountered an error in the midst of commit and secshr_db_clnup
* finished the job for it). In this case we do NOT want to invoke wcs_recover as that will update the "bt"
* transaction numbers without correspondingly updating the history transaction numbers (effectively causing
* a cdb_sc_blkmod type of restart). Therefore do NOT call grab_crit (which unconditionally invokes wcs_recover)
* if we already hold crit.
*/
if (!was_crit)
{
for ( ; ; )
{
grab_crit(gv_cur_region);
if (!cs_data->freeze && !IS_REPL_INST_FROZEN)
break;
rel_crit(gv_cur_region);
while (cs_data->freeze || IS_REPL_INST_FROZEN)
hiber_start(1000);
}
} else if ((cs_data->freeze || IS_REPL_INST_FROZEN) && dollar_tlevel)
{ /* We don't want to continue with file extension as explained above. Hence return with an error code which
* op_tcommit will recognize (as a cdb_sc_needcrit type of restart) and restart accordingly.
*/
assert(CDB_STAGNATE <= t_tries);
GDSFILEXT_CLNUP;
return (uint4)(FINAL_RETRY_FREEZE_PROG);
}
assert(cs_addrs->ti->total_blks == cs_data->trans_hist.total_blks);
if (cs_data->trans_hist.total_blks != filesize)
{ /* Somebody else has already extended it, since we are in crit, this is trust-worthy. However, in case of MM,
* we still need to remap the database
*/
assert((cs_data->trans_hist.total_blks > filesize) GTM_TRUNCATE_ONLY( || !is_mm));
/* For BG, someone else could have truncated or extended - we have no idea */
GDSFILEXT_CLNUP;
return (SS_NORMAL);
}
