/*
* shm_destroy - free the struct shmid_kernel
*
* @ns: namespace
* @shp: struct to free
*
* It has to be called with shp and shm_ids.rwsem (writer) locked,
* but returns with shp unlocked and freed.
*/
static void shm_destroy(struct ipc_namespace *ns, struct shmid_kernel *shp)
{
struct file *shm_file;
shm_file = shp->shm_file;
shp->shm_file = NULL;
ns->shm_tot -= (shp->shm_segsz + PAGE_SIZE - 1) >> PAGE_SHIFT;
shm_rmid(ns, shp);
shm_unlock(shp);
if (!is_file_hugepages(shm_file))
shmem_lock(shm_file, 0, shp->mlock_user);
else if (shp->mlock_user)
user_shm_unlock(file_inode(shm_file)->i_size, shp->mlock_user);
fput(shm_file);
ipc_rcu_putref(shp, shm_rcu_free);
}
int mu_rndwn_replpool(replpool_identifier *replpool_id, repl_inst_hdr_ptr_t repl_inst_filehdr, int shm_id, boolean_t *ipc_rmvd)
{
int semval, status, save_errno, nattch;
char *instfilename, pool_type;
sm_uc_ptr_t start_addr;
struct shmid_ds shm_buf;
unix_db_info *udi;
sgmnt_addrs *csa;
boolean_t anticipatory_freeze_available, force_attach;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
assert(INVALID_SHMID != shm_id);
instfilename = replpool_id->instfilename;
pool_type = replpool_id->pool_type;
assert((JNLPOOL_SEGMENT == pool_type) || (RECVPOOL_SEGMENT == pool_type));
anticipatory_freeze_available = ANTICIPATORY_FREEZE_AVAILABLE;
force_attach = (jgbl.onlnrlbk || (!jgbl.mur_rollback && !argumentless_rundown && anticipatory_freeze_available));
if (-1 == shmctl(shm_id, IPC_STAT, &shm_buf))
{
save_errno = errno;
ISSUE_REPLPOOLINST_AND_RETURN(save_errno, shm_id, instfilename, "shmctl()");
}
nattch = shm_buf.shm_nattch;
if ((0 != nattch) && !force_attach)
{
util_out_print("Replpool segment (id = !UL) for replication instance !AD is in use by another process.",
TRUE, shm_id, LEN_AND_STR(instfilename));
return -1;
}
if (-1 == (sm_long_t)(start_addr = (sm_uc_ptr_t) do_shmat(shm_id, 0, 0)))
{
save_errno = errno;
ISSUE_REPLPOOLINST_AND_RETURN(save_errno, shm_id, instfilename, "shmat()");
}
ESTABLISH_RET(mu_rndwn_replpool_ch, -1);
/* assert that the identifiers are at the top of replpool control structure */
assert(0 == offsetof(jnlpool_ctl_struct, jnlpool_id));
assert(0 == offsetof(recvpool_ctl_struct, recvpool_id));
memcpy((void *)replpool_id, (void *)start_addr, SIZEOF(replpool_identifier));
if (memcmp(replpool_id->label, GDS_RPL_LABEL, GDS_LABEL_SZ - 1))
{
if (!memcmp(replpool_id->label, GDS_RPL_LABEL, GDS_LABEL_SZ - 3))
util_out_print(
"Incorrect version for the replpool segment (id = !UL) belonging to replication instance !AD",
TRUE, shm_id, LEN_AND_STR(instfilename));
else
util_out_print("Incorrect replpool format for the segment (id = !UL) belonging to replication instance !AD",
TRUE, shm_id, LEN_AND_STR(instfilename));
DETACH_AND_RETURN(start_addr, shm_id, instfilename);
}
if (memcmp(replpool_id->now_running, gtm_release_name, gtm_release_name_len + 1))
{
util_out_print("Attempt to access with version !AD, while already using !AD for replpool segment (id = !UL)"
" belonging to replication instance !AD.", TRUE, gtm_release_name_len, gtm_release_name,
LEN_AND_STR(replpool_id->now_running), shm_id, LEN_AND_STR(instfilename));
DETACH_AND_RETURN(start_addr, shm_id, instfilename);
}
/* Assert that if we haven't yet attached to the journal pool yet, we have the corresponding global vars set to NULL */
assert((JNLPOOL_SEGMENT != pool_type) || ((NULL == jnlpool.jnlpool_ctl) && (NULL == jnlpool_ctl)));
if (JNLPOOL_SEGMENT == pool_type)
{ /* Initialize variables to simulate a "jnlpool_init". This is required by "repl_inst_flush_jnlpool" called below */
jnlpool_ctl = jnlpool.jnlpool_ctl = (jnlpool_ctl_ptr_t)start_addr;
assert(NULL != jnlpool.jnlpool_dummy_reg);
udi = FILE_INFO(jnlpool.jnlpool_dummy_reg);
csa = &udi->s_addrs;
csa->critical = (mutex_struct_ptr_t)((sm_uc_ptr_t)jnlpool.jnlpool_ctl + JNLPOOL_CTL_SIZE);
csa->nl = (node_local_ptr_t)((sm_uc_ptr_t)csa->critical + CRIT_SPACE + SIZEOF(mutex_spin_parms_struct));
/* secshr_db_clnup uses this relationship */
assert(jnlpool.jnlpool_ctl->filehdr_off);
assert(jnlpool.jnlpool_ctl->srclcl_array_off > jnlpool.jnlpool_ctl->filehdr_off);
assert(jnlpool.jnlpool_ctl->sourcelocal_array_off > jnlpool.jnlpool_ctl->srclcl_array_off);
/* Initialize "jnlpool.repl_inst_filehdr" and related fields as "repl_inst_flush_jnlpool" relies on that */
jnlpool.repl_inst_filehdr = (repl_inst_hdr_ptr_t)((sm_uc_ptr_t)jnlpool.jnlpool_ctl
+ jnlpool.jnlpool_ctl->filehdr_off);
jnlpool.gtmsrc_lcl_array = (gtmsrc_lcl_ptr_t)((sm_uc_ptr_t)jnlpool.jnlpool_ctl
+ jnlpool.jnlpool_ctl->srclcl_array_off);
jnlpool.gtmsource_local_array = (gtmsource_local_ptr_t)((sm_uc_ptr_t)jnlpool.jnlpool_ctl
+ jnlpool.jnlpool_ctl->sourcelocal_array_off);
if (0 == nattch)
{ /* No one attached. So, we can safely flush the journal pool so that the gtmsrc_lcl structures in the
* jnlpool and disk are in sync with each other. More importantly we are about to remove the jnlpool
* so we better get things in sync before that. If anticipatory freeze scheme is in effect, then we
* need to keep the journal pool up and running. So, don't reset the crash field in the instance file
* header (dictated by the second parameter to repl_inst_flush_jnlpool below).
* Note:
* If mu_rndwn_repl_instance created new semaphores (in mu_replpool_remove_sem), we need to flush those
* to the instance file as well. So, override the jnlpool_semid and jnlpool_semid_ctime with the new
* values.
*/
assert((INVALID_SEMID != repl_inst_filehdr->jnlpool_semid)
&& (0 != repl_inst_filehdr->jnlpool_semid_ctime));
jnlpool.repl_inst_filehdr->jnlpool_semid = repl_inst_filehdr->jnlpool_semid;
jnlpool.repl_inst_filehdr->jnlpool_semid_ctime = repl_inst_filehdr->jnlpool_semid_ctime;
repl_inst_flush_jnlpool(FALSE, !anticipatory_freeze_available);
assert(!jnlpool.repl_inst_filehdr->crash || anticipatory_freeze_available);
/* Refresh local copy (repl_inst_filehdr) with the copy that was just flushed (jnlpool.repl_inst_filehdr) */
memcpy(repl_inst_filehdr, jnlpool.repl_inst_filehdr, SIZEOF(repl_inst_hdr));
if (!anticipatory_freeze_available)
{ /* Now that jnlpool has been flushed and there is going to be no journal pool, reset
* "jnlpool.repl_inst_filehdr" as otherwise other routines (e.g. "repl_inst_recvpool_reset") are
* affected by whether this is NULL or not.
*/
jnlpool.jnlpool_ctl = NULL;
jnlpool_ctl = NULL;
jnlpool.gtmsrc_lcl_array = NULL;
jnlpool.gtmsource_local_array = NULL;
jnlpool.jnldata_base = NULL;
jnlpool.repl_inst_filehdr = NULL;
}
} /* else we are ONLINE ROLLBACK. repl_inst_flush_jnlpool will be done later after gvcst_init in mur_open_files */
}
if ((0 == nattch) && (!anticipatory_freeze_available || (RECVPOOL_SEGMENT == pool_type)))
{
if (-1 == shmdt((caddr_t)start_addr))
{
save_errno = errno;
ISSUE_REPLPOOLINST_AND_RETURN(save_errno, shm_id, instfilename, "shmdt()");
}
if (0 != shm_rmid(shm_id))
{
save_errno = errno;
ISSUE_REPLPOOLINST_AND_RETURN(save_errno, shm_id, instfilename, "shm_rmid()");
}
if (JNLPOOL_SEGMENT == pool_type)
{
repl_inst_filehdr->jnlpool_shmid = INVALID_SHMID;
repl_inst_filehdr->jnlpool_shmid_ctime = 0;
assert((NULL == jnlpool.jnlpool_ctl) && (NULL == jnlpool_ctl));
*ipc_rmvd = TRUE;
} else
{
repl_inst_filehdr->recvpool_shmid = INVALID_SHMID;
repl_inst_filehdr->recvpool_shmid_ctime = 0;
*ipc_rmvd = TRUE;
}
} else
{ /* Else we are ONLINE ROLLBACK or anticipatory freeze is in effect and so we want to keep the journal pool available
* for the duration of the rollback. Do not remove and/or reset the fields in the file header
*/
assert((JNLPOOL_SEGMENT != pool_type) || ((NULL != jnlpool.jnlpool_ctl) && (NULL != jnlpool_ctl)));
if (JNLPOOL_SEGMENT == pool_type)
*ipc_rmvd = FALSE;
if (RECVPOOL_SEGMENT == pool_type)
*ipc_rmvd = FALSE;
}
REVERT;
return 0;
}
/* Takes an entry from 'ipcs -m' and checks for its validity to be a GT.M db segment.
* Returns TRUE if the shared memory segment is a valid GT.M db segment
* (based on a check on some fields in the shared memory) else FALSE.
* If the segment belongs to GT.M it returns the database file name by the second argument.
* Sets exit_stat to ERR_MUNOTALLSEC if appropriate.
*/
boolean_t validate_db_shm_entry(shm_parms *parm_buff, char *fname, int *exit_stat)
{
boolean_t remove_shmid;
file_control *fc;
int fname_len, save_errno, status, shmid;
node_local_ptr_t nl_addr;
sm_uc_ptr_t start_addr;
struct stat st_buff;
struct shmid_ds shmstat;
sgmnt_data tsd;
unix_db_info *udi;
char msgbuff[OUT_BUFF_SIZE];
if (NULL == parm_buff)
return FALSE;
/* check for the bare minimum size of the shared memory segment that we expect
* (with no fileheader related information at hand) */
if (MIN_NODE_LOCAL_SPACE + SHMPOOL_SECTION_SIZE > parm_buff->sgmnt_siz)
return FALSE;
if (IPC_PRIVATE != parm_buff->key)
return FALSE;
shmid = parm_buff->shmid;
/* we do not need to lock the shm for reading the rundown information as
* the other rundowns (if any) can also be allowed to share reading the
* same info concurrently.
*/
if (-1 == (sm_long_t)(start_addr = (sm_uc_ptr_t) do_shmat(shmid, 0, SHM_RND)))
return FALSE;
nl_addr = (node_local_ptr_t)start_addr;
memcpy(fname, nl_addr->fname, MAX_FN_LEN + 1);
fname[MAX_FN_LEN] = '\0'; /* make sure the fname is null terminated */
fname_len = STRLEN(fname);
msgbuff[0] = '\0';
if (memcmp(nl_addr->label, GDS_LABEL, GDS_LABEL_SZ - 1))
{
if (!memcmp(nl_addr->label, GDS_LABEL, GDS_LABEL_SZ - 3))
{
util_out_print("Cannot rundown shmid = !UL for database !AD as it has format !AD "
"but this mupip uses format !AD", TRUE, shmid,
fname_len, fname, GDS_LABEL_SZ - 1, nl_addr->label, GDS_LABEL_SZ - 1, GDS_LABEL);
*exit_stat = ERR_MUNOTALLSEC;
}
shmdt((void *)start_addr);
return FALSE;
}
if (-1 == shmctl(shmid, IPC_STAT, &shmstat))
{
save_errno = errno;
assert(FALSE);/* we were able to attach to this shmid before so should be able to get stats on it */
util_out_print("!AD -> Error with shmctl for shmid = !UL",
TRUE, fname_len, fname, shmid);
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) save_errno);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
remove_shmid = FALSE;
/* Check if db filename reported in shared memory still exists. If not, clean this shared memory section
* without even invoking "mu_rndwn_file" as that expects the db file to exist. Same case if shared memory
* points back to a database whose file header does not have this shmid.
*/
if (-1 == Stat(fname, &st_buff))
{
if (ENOENT == errno)
{
SNPRINTF(msgbuff, OUT_BUFF_SIZE, "File %s does not exist", fname);
if (1 < shmstat.shm_nattch)
{
PRINT_AND_SEND_DBRNDWN_FAILURE_MSG(msgbuff, fname, shmid);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
remove_shmid = TRUE;
} else
{ /* Stat errored out e.g. due to file permissions. Log that */
save_errno = errno;
util_out_print("Cannot rundown shmid !UL for database file !AD as stat() on the file"
" returned the following error", TRUE, shmid, fname_len, fname);
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) save_errno);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
} else
{
mu_gv_cur_reg_init();
gv_cur_region->dyn.addr->fname_len = strlen(fname);
STRNCPY_STR(gv_cur_region->dyn.addr->fname, fname, gv_cur_region->dyn.addr->fname_len);
fc = gv_cur_region->dyn.addr->file_cntl;
fc->op = FC_OPEN;
status = dbfilop(fc);
if (SS_NORMAL != status)
{
util_out_print("!AD -> Error with dbfilop for shmid = !UL", TRUE, fname_len, fname, shmid);
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(5) status, 2, DB_LEN_STR(gv_cur_region), errno);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
udi = FILE_INFO(gv_cur_region);
LSEEKREAD(udi->fd, 0, &tsd, SIZEOF(sgmnt_data), status);
if (0 != status)
{
save_errno = errno;
util_out_print("!AD -> Error with LSEEKREAD for shmid = !UL", TRUE, fname_len, fname, shmid);
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) save_errno);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
mu_gv_cur_reg_free();
if (tsd.shmid != shmid)
{
SNPRINTF(msgbuff, OUT_BUFF_SIZE, "Shared memory ID (%d) in the DB file header does not match with the one"
" reported by \"ipcs\" command (%d)", tsd.shmid, shmid);
if (1 < shmstat.shm_nattch)
{
PRINT_AND_SEND_DBRNDWN_FAILURE_MSG(msgbuff, fname, shmid);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
remove_shmid = TRUE;
} else if (tsd.gt_shm_ctime.ctime != shmstat.shm_ctime)
{
SNPRINTF(msgbuff, OUT_BUFF_SIZE, "Shared memory creation time in the DB file header does not match with"
" the one reported by shmctl");
if (1 < shmstat.shm_nattch)
{
PRINT_AND_SEND_DBRNDWN_FAILURE_MSG(msgbuff, fname, shmid);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
remove_shmid = TRUE;
}
}
shmdt((void *)start_addr);
if (remove_shmid)
{
assert('\0' != msgbuff[0]);
if (0 != shm_rmid(shmid))
{
save_errno = errno;
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(8) ERR_DBFILERR, 2, fname_len, fname,
ERR_TEXT, 2, RTS_ERROR_TEXT("Error removing shared memory"));
util_out_print("!AD -> Error removing shared memory for shmid = !UL", TRUE, fname_len, fname, shmid);
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) save_errno);
*exit_stat = ERR_MUNOTALLSEC;
return FALSE;
}
PRINT_AND_SEND_SHMREMOVED_MSG(msgbuff, fname_len, fname, shmid);
*exit_stat = ERR_SHMREMOVED;
} else
*exit_stat = SS_NORMAL;
return TRUE;
}
/* Takes an entry from 'ipcs -am' and checks for its validity to be a GT.M replication segment.
* Returns TRUE if the shared memory segment is a valid GT.M replication segment
* (based on a check on some fields in the shared memory) else FALSE.
* If the segment belongs to GT.M, it returns the replication id of the segment
* by the second argument.
* Sets exit_stat to ERR_MUNOTALLSEC if appropriate.
*/
boolean_t validate_replpool_shm_entry(shm_parms *parm_buff, replpool_id_ptr_t replpool_id, int *exit_stat)
{
boolean_t remove_shmid, jnlpool_segment;
int fd;
repl_inst_hdr repl_instance;
sm_uc_ptr_t start_addr;
int save_errno, status, shmid;
struct shmid_ds shmstat;
char msgbuff[OUT_BUFF_SIZE], *instfilename;
if (NULL == parm_buff)
return FALSE;
/* Check for the bare minimum size of the replic shared segment that we expect */
/* if (parm_buff->sgmnt_siz < (SIZEOF(replpool_identifier) + MIN(MIN_JNLPOOL_SIZE, MIN_RECVPOOL_SIZE))) */
if (parm_buff->sgmnt_siz < MIN(MIN_JNLPOOL_SIZE, MIN_RECVPOOL_SIZE))
return FALSE;
if (IPC_PRIVATE != parm_buff->key)
return FALSE;
shmid = parm_buff->shmid;
/* we do not need to lock the shm for reading the rundown information as
* the other rundowns (if any) can also be allowed to share reading the
* same info concurrently.
*/
if (-1 == (sm_long_t)(start_addr = (sm_uc_ptr_t) do_shmat(shmid, 0, SHM_RND)))
return FALSE;
memcpy((void *)replpool_id, (void *)start_addr, SIZEOF(replpool_identifier));
instfilename = replpool_id->instfilename;
/* Even though we could be looking at a replication pool structure that has been created by an older version
* or newer version of GT.M, the format of the "replpool_identifier" structure is expected to be the same
* across all versions so we can safely dereference the "label" and "instfilename" fields in order to generate
* user-friendly error messages. Asserts for the layout are in "mu_rndwn_repl_instance" (not here) with a
* comment there as to why that location was chosen.
*/
if (memcmp(replpool_id->label, GDS_RPL_LABEL, GDS_LABEL_SZ - 1))
{
if (!memcmp(replpool_id->label, GDS_RPL_LABEL, GDS_LABEL_SZ - 3))
{
util_out_print("Cannot rundown replpool shmid = !UL as it has format !AD "
"created by !AD but this mupip is version and uses format !AD",
TRUE, shmid, GDS_LABEL_SZ - 1, replpool_id->label,
LEN_AND_STR(replpool_id->now_running), gtm_release_name_len, gtm_release_name,
GDS_LABEL_SZ - 1, GDS_RPL_LABEL);
*exit_stat = ERR_MUNOTALLSEC;
}
shmdt((void *)start_addr);
return FALSE;
}
assert(JNLPOOL_SEGMENT == replpool_id->pool_type || RECVPOOL_SEGMENT == replpool_id->pool_type);
if(JNLPOOL_SEGMENT != replpool_id->pool_type && RECVPOOL_SEGMENT != replpool_id->pool_type)
{
shmdt((void *)start_addr);
return FALSE;
}
jnlpool_segment = (JNLPOOL_SEGMENT == replpool_id->pool_type);
if (-1 == shmctl(shmid, IPC_STAT, &shmstat))
{
save_errno = errno;
assert(FALSE);/* we were able to attach to this shmid before so should be able to get stats on it */
util_out_print("!AD -> Error with shmctl for shmid = !UL",
TRUE, LEN_AND_STR(instfilename), shmid);
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) save_errno);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
/* Check if instance filename reported in shared memory still exists. If not, clean this
* shared memory section without even invoking "mu_rndwn_repl_instance" as that expects
* the instance file to exist. Same case if shared memory points back to an instance file
* whose file header does not have this shmid.
*/
OPENFILE(instfilename, O_RDONLY, fd); /* check if we can open it */
msgbuff[0] = '\0';
remove_shmid = FALSE;
if (FD_INVALID == fd)
{
if (ENOENT == errno)
{
SNPRINTF(msgbuff, OUT_BUFF_SIZE, "File %s does not exist", instfilename);
if (1 < shmstat.shm_nattch)
{
PRINT_AND_SEND_REPLPOOL_FAILURE_MSG(msgbuff, replpool_id, shmid);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
remove_shmid = TRUE;
} else
{ /* open() errored out e.g. due to file permissions. Log that */
save_errno = errno;
util_out_print("Cannot rundown replpool shmid !UL for instance file"
" !AD as open() on the file returned the following error",
TRUE, shmid, LEN_AND_STR(instfilename));
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) save_errno);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
} else
{
LSEEKREAD(fd, 0, &repl_instance, SIZEOF(repl_inst_hdr), status);
if (0 != status)
{
save_errno = errno;
util_out_print("!AD -> Error with LSEEKREAD for shmid = !UL", TRUE,
LEN_AND_STR(instfilename), shmid);
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) save_errno);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
if ((jnlpool_segment && (repl_instance.jnlpool_shmid != shmid))
|| (!jnlpool_segment && (repl_instance.recvpool_shmid != shmid)))
{
SNPRINTF(msgbuff, OUT_BUFF_SIZE, "%s SHMID (%d) in the instance file header does not match with the"
" one reported by \"ipcs\" command (%d)", jnlpool_segment ? "Journal Pool" : "Receive Pool",
jnlpool_segment ? repl_instance.jnlpool_shmid : repl_instance.recvpool_shmid, shmid);
if (1 < shmstat.shm_nattch)
{
PRINT_AND_SEND_REPLPOOL_FAILURE_MSG(msgbuff, replpool_id, shmid);
*exit_stat = ERR_MUNOTALLSEC;
shmdt((void *)start_addr);
return FALSE;
}
remove_shmid = TRUE;
}
CLOSEFILE_RESET(fd, status); /* resets "fd" to FD_INVALID */
}
shmdt((void *)start_addr);
if (remove_shmid)
{
assert('\0' != msgbuff[0]);
if (0 != shm_rmid(shmid))
{
save_errno = errno;
util_out_print("!AD -> Error removing shared memory for shmid = !UL",
TRUE, LEN_AND_STR(instfilename), shmid);
gtm_putmsg_csa(CSA_ARG(NULL) VARLSTCNT(1) save_errno);
*exit_stat = ERR_MUNOTALLSEC;
return FALSE;
}
PRINT_AND_SEND_SHMREMOVED_MSG(msgbuff, STRLEN(instfilename), instfilename, shmid);
*exit_stat = ERR_SHMREMOVED;
} else
*exit_stat = SS_NORMAL;
return TRUE;
}
int4 gds_rundown(void)
{
boolean_t canceled_dbsync_timer, canceled_flush_timer, ok_to_write_pfin;
boolean_t have_standalone_access, ipc_deleted, err_caught;
boolean_t is_cur_process_ss_initiator, remove_shm, vermismatch, we_are_last_user, we_are_last_writer, is_mm;
boolean_t unsafe_last_writer;
char time_str[CTIME_BEFORE_NL + 2]; /* for GET_CUR_TIME macro */
gd_region *reg;
int save_errno, status, rc;
int4 semval, ftok_semval, sopcnt, ftok_sopcnt;
short crash_count;
sm_long_t munmap_len;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
node_local_ptr_t cnl;
struct shmid_ds shm_buf;
struct sembuf sop[2], ftok_sop[2];
uint4 jnl_status;
unix_db_info *udi;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
shm_snapshot_t *ss_shm_ptr;
uint4 ss_pid, onln_rlbk_pid, holder_pid;
boolean_t was_crit;
boolean_t safe_mode; /* Do not flush or take down shared memory. */
boolean_t bypassed_ftok = FALSE, bypassed_access = FALSE, may_bypass_ftok, inst_is_frozen,
ftok_counter_halted,
access_counter_halted;
int secshrstat;
intrpt_state_t prev_intrpt_state;
DCL_THREADGBL_ACCESS;
SETUP_THREADGBL_ACCESS;
jnl_status = 0;
reg = gv_cur_region; /* Local copy */
/* early out for cluster regions
* to avoid tripping the assert below.
* Note:
* This early out is consistent with VMS. It has been
* noted that all of the gtcm assignments
* to gv_cur_region should use the TP_CHANGE_REG
* macro. This would also avoid the assert problem
* and should be done eventually.
*/
if (dba_cm == reg->dyn.addr->acc_meth)
return EXIT_NRM;
udi = FILE_INFO(reg);
csa = &udi->s_addrs;
csd = csa->hdr;
assert(csa == cs_addrs && csd == cs_data);
if ((reg->open) && (dba_usr == csd->acc_meth))
{
change_reg();
gvusr_rundown();
return EXIT_NRM;
}
/* If the process has standalone access, it has udi->grabbed_access_sem set to TRUE at this point. Note that down in a local
* variable as the udi->grabbed_access_sem is set to TRUE even for non-standalone access below and hence we can't rely on
* that later to determine if the process had standalone access or not when it entered this function. We need to guarantee
* that none else access database file header when semid/shmid fields are reset. We already have created ftok semaphore in
* db_init or, mu_rndwn_file and did not remove it. So just lock it. We do it in blocking mode.
*/
have_standalone_access = udi->grabbed_access_sem; /* process holds standalone access */
DEFER_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
ESTABLISH_NORET(gds_rundown_ch, err_caught);
if (err_caught)
{
REVERT;
WITH_CH(gds_rundown_ch, gds_rundown_err_cleanup(have_standalone_access), 0);
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
DEBUG_ONLY(ok_to_UNWIND_in_exit_handling = FALSE);
return EXIT_ERR;
}
assert(reg->open); /* if we failed to open, dbinit_ch should have taken care of proper clean up */
assert(!reg->opening); /* see comment above */
assert((dba_bg == csd->acc_meth) || (dba_mm == csd->acc_meth));
is_mm = (dba_bg != csd->acc_meth);
assert(!csa->hold_onto_crit || (csa->now_crit && jgbl.onlnrlbk));
/* If we are online rollback, we should already be holding crit and should release it only at the end of this module. This
* is usually done by noting down csa->now_crit in a local variable (was_crit) and using it whenever we are about to
* grab_crit. But, there are instances (like mupip_set_journal.c) where we grab_crit but invoke gds_rundown without any
* preceeding rel_crit. Such code relies on the fact that gds_rundown does rel_crit unconditionally (to get locks to a known
* state). So, augment csa->now_crit with jgbl.onlnrlbk to track if we can rel_crit unconditionally or not in gds_rundown.
*/
was_crit = (csa->now_crit && jgbl.onlnrlbk);
/* Cancel any pending flush timer for this region by this task */
canceled_flush_timer = FALSE;
canceled_dbsync_timer = FALSE;
CANCEL_DB_TIMERS(reg, csa, canceled_flush_timer, canceled_dbsync_timer);
we_are_last_user = FALSE;
inst_is_frozen = IS_REPL_INST_FROZEN && REPL_ALLOWED(csa->hdr);
if (!csa->persistent_freeze)
region_freeze(reg, FALSE, FALSE, FALSE);
if (!was_crit)
{
rel_crit(reg); /* get locks to known state */
mutex_cleanup(reg);
}
/* The only process that can invoke gds_rundown while holding access control semaphore is RECOVER/ROLLBACK. All the others
* (like MUPIP SET -FILE/MUPIP EXTEND would have invoked db_ipcs_reset() before invoking gds_rundown (from
* mupip_exit_handler). The only exception is when these processes encounter a terminate signal and they reach
* mupip_exit_handler while holding access control semaphore. Assert accordingly.
*/
assert(!have_standalone_access || mupip_jnl_recover || process_exiting);
/* If we have standalone access, then ensure that a concurrent online rollback cannot be running at the same time as it
* needs the access control lock as well. The only expection is we are online rollback and currently running down.
*/
cnl = csa->nl;
onln_rlbk_pid = cnl->onln_rlbk_pid;
assert(!have_standalone_access || mupip_jnl_recover || !onln_rlbk_pid || !is_proc_alive(onln_rlbk_pid, 0));
if (!have_standalone_access)
{
if (-1 == (ftok_semval = semctl(udi->ftok_semid, DB_COUNTER_SEM, GETVAL))) /* Check # of procs counted on FTOK */
{
save_errno = errno;
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get ftok_semval"), CALLFROM, errno);
}
may_bypass_ftok = CAN_BYPASS(ftok_semval, csd, inst_is_frozen); /* Do we need a blocking wait? */
/* We need to guarantee that no one else access database file header when semid/shmid fields are reset.
* We already have created ftok semaphore in db_init or mu_rndwn_file and did not remove it. So just lock it.
*/
if (!ftok_sem_lock(reg, may_bypass_ftok))
{
if (may_bypass_ftok)
{ /* We did a non-blocking wait. It's ok to proceed without locking */
bypassed_ftok = TRUE;
holder_pid = semctl(udi->ftok_semid, DB_CONTROL_SEM, GETPID);
if ((uint4)-1 == holder_pid)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get holder_pid"),
CALLFROM, errno);
if (!IS_GTM_IMAGE) /* MUMPS processes should not flood syslog with bypass messages. */
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_RESRCINTRLCKBYPAS, 10,
LEN_AND_STR(gtmImageNames[image_type].imageName), process_id, LEN_AND_LIT("FTOK"),
REG_LEN_STR(reg), DB_LEN_STR(reg), holder_pid);
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_TEXT, 2,
LEN_AND_LIT("FTOK bypassed at rundown"));
}
} else
{ /* We did a blocking wait but something bad happened. */
FTOK_TRACE(csa, csa->ti->curr_tn, ftok_ops_lock, process_id);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
}
}
sop[0].sem_num = DB_CONTROL_SEM; sop[0].sem_op = 0; /* Wait for 0 */
sop[1].sem_num = DB_CONTROL_SEM; sop[1].sem_op = 1; /* Lock */
sopcnt = 2;
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO | IPC_NOWAIT; /* Don't wait the first time thru */
SEMOP(udi->semid, sop, sopcnt, status, NO_WAIT);
if (0 != status)
{
save_errno = errno;
/* Check # of processes counted on access sem. */
if (-1 == (semval = semctl(udi->semid, DB_COUNTER_SEM, GETVAL)))
{
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get semval"), CALLFROM, errno);
}
bypassed_access = CAN_BYPASS(semval, csd, inst_is_frozen) || onln_rlbk_pid || csd->file_corrupt;
/* Before attempting again in the blocking mode, see if the holding process is an online rollback.
* If so, it is likely we won't get the access control semaphore anytime soon. In that case, we
* are better off skipping rundown and continuing with sanity cleanup and exit.
*/
holder_pid = semctl(udi->semid, DB_CONTROL_SEM, GETPID);
if ((uint4)-1 == holder_pid)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get holder_pid"), CALLFROM, errno);
if (!bypassed_access)
{ /* We couldn't get it in one shot-- see if we already have it */
if (holder_pid == process_id)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(5) MAKE_MSG_INFO(ERR_CRITSEMFAIL), 2, DB_LEN_STR(reg),
ERR_RNDWNSEMFAIL);
REVERT;
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
assert(FALSE);
return EXIT_ERR;
}
if (EAGAIN != save_errno)
{
assert(FALSE);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMOP on access control semaphore"),
CALLFROM, save_errno);
}
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO; /* Try again - blocking this time */
SEMOP(udi->semid, sop, 2, status, FORCED_WAIT);
if (-1 == status) /* We couldn't get it at all.. */
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMOP on access control semaphore"),
CALLFROM, errno);
} else if (!IS_GTM_IMAGE)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_RESRCINTRLCKBYPAS, 10,
LEN_AND_STR(gtmImageNames[image_type].imageName), process_id,
LEN_AND_LIT("access control"), REG_LEN_STR(reg), DB_LEN_STR(reg), holder_pid);
send_msg_csa(CSA_ARG(NULL) VARLSTCNT(4) ERR_TEXT, 2,
LEN_AND_LIT("Access control bypassed at rundown"));
}
udi->grabbed_access_sem = !bypassed_access;
}
} /* else we we hold the access control semaphore and therefore have standalone access. We do not release it now - we
* release it later in mupip_exit_handler.c. Since we already hold the access control semaphore, we don't need the
* ftok semaphore and trying it could cause deadlock
*/
/* Note that in the case of online rollback, "udi->grabbed_access_sem" (and in turn "have_standalone_access") is TRUE.
* But there could be other processes still having the database open so we cannot safely reset the halted fields.
*/
if (have_standalone_access && !jgbl.onlnrlbk)
csd->ftok_counter_halted = csd->access_counter_halted = FALSE;
ftok_counter_halted = csd->ftok_counter_halted;
access_counter_halted = csd->access_counter_halted;
/* If we bypassed any of the semaphores, activate safe mode.
* Also, if the replication instance is frozen and this db has replication turned on (which means
* no flushes of dirty buffers to this db can happen while the instance is frozen) activate safe mode.
*/
ok_to_write_pfin = !(bypassed_access || bypassed_ftok || inst_is_frozen);
safe_mode = !ok_to_write_pfin || ftok_counter_halted || access_counter_halted;
/* At this point we are guaranteed no one else is doing a db_init/rundown as we hold the access control semaphore */
assert(csa->ref_cnt); /* decrement private ref_cnt before shared ref_cnt decrement. */
csa->ref_cnt--; /* Currently journaling logic in gds_rundown() in VMS relies on this order to detect last writer */
assert(!csa->ref_cnt);
--cnl->ref_cnt;
if (memcmp(cnl->now_running, gtm_release_name, gtm_release_name_len + 1))
{ /* VERMISMATCH condition. Possible only if DSE */
assert(dse_running);
vermismatch = TRUE;
} else
vermismatch = FALSE;
if (-1 == shmctl(udi->shmid, IPC_STAT, &shm_buf))
{
save_errno = errno;
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown shmctl"), CALLFROM, save_errno);
} else
we_are_last_user = (1 == shm_buf.shm_nattch) && !vermismatch && !safe_mode;
/* recover => one user except ONLINE ROLLBACK, or standalone with frozen instance */
assert(!have_standalone_access || we_are_last_user || jgbl.onlnrlbk || inst_is_frozen);
if (-1 == (semval = semctl(udi->semid, DB_COUNTER_SEM, GETVAL)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get semval"), CALLFROM, errno);
/* There's one writer left and I am it */
assert(reg->read_only || semval >= 0);
unsafe_last_writer = (DB_COUNTER_SEM_INCR == semval) && (FALSE == reg->read_only) && !vermismatch;
we_are_last_writer = unsafe_last_writer && !safe_mode;
assert(!we_are_last_writer || !safe_mode);
assert(!we_are_last_user || !safe_mode);
/* recover + R/W region => one writer except ONLINE ROLLBACK, or standalone with frozen instance, leading to safe_mode */
assert(!(have_standalone_access && !reg->read_only) || we_are_last_writer || jgbl.onlnrlbk || inst_is_frozen);
GTM_WHITE_BOX_TEST(WBTEST_ANTIFREEZE_JNLCLOSE, we_are_last_writer, 1); /* Assume we are the last writer to invoke wcs_flu */
if (!have_standalone_access && (-1 == (ftok_semval = semctl(udi->ftok_semid, DB_COUNTER_SEM, GETVAL))))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown SEMCTL failed to get ftok_semval"), CALLFROM, errno);
if (NULL != csa->ss_ctx)
ss_destroy_context(csa->ss_ctx);
/* SS_MULTI: If multiple snapshots are supported, then we have to run through each of the snapshots */
assert(1 == MAX_SNAPSHOTS);
ss_shm_ptr = (shm_snapshot_ptr_t)SS_GETSTARTPTR(csa);
ss_pid = ss_shm_ptr->ss_info.ss_pid;
is_cur_process_ss_initiator = (process_id == ss_pid);
if (ss_pid && (is_cur_process_ss_initiator || we_are_last_user))
{
/* Try getting snapshot crit latch. If we don't get latch, we won't hang for eternity and will skip
* doing the orphaned snapshot cleanup. It will be cleaned up eventually either by subsequent MUPIP
* INTEG or by a MUPIP RUNDOWN.
*/
if (ss_get_lock_nowait(reg) && (ss_pid == ss_shm_ptr->ss_info.ss_pid)
&& (is_cur_process_ss_initiator || !is_proc_alive(ss_pid, 0)))
{
ss_release(NULL);
ss_release_lock(reg);
}
}
/* If cnl->donotflush_dbjnl is set, it means mupip recover/rollback was interrupted and therefore we need not flush
* shared memory contents to disk as they might be in an inconsistent state. Moreover, any more flushing will only cause
* future rollback to undo more journal records (PBLKs). In this case, we will go ahead and remove shared memory (without
* flushing the contents) in this routine. A reissue of the recover/rollback command will restore the database to a
* consistent state.
*/
if (!cnl->donotflush_dbjnl && !reg->read_only && !vermismatch)
{ /* If we had an orphaned block and were interrupted, set wc_blocked so we can invoke wcs_recover. Do it ONLY
* if there is NO concurrent online rollback running (as we need crit to set wc_blocked)
*/
if (csa->wbuf_dqd && !is_mm)
{ /* If we had an orphaned block and were interrupted, mupip_exit_handler will invoke secshr_db_clnup which
* will clear this field and so we should never come to gds_rundown with a non-zero wbuf_dqd. The only
* exception is if we are recover/rollback in which case gds_rundown (from mur_close_files) is invoked
* BEFORE secshr_db_clnup in mur_close_files.
* Note: It is NOT possible for online rollback to reach here with wbuf_dqd being non-zero. This is because
* the moment we apply the first PBLK, we stop all interrupts and hence can never be interrupted in
* wcs_wtstart or wcs_get_space. Assert accordingly.
*/
assert(mupip_jnl_recover && !jgbl.onlnrlbk && !safe_mode);
if (!was_crit)
grab_crit(reg);
SET_TRACEABLE_VAR(cnl->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wcb_gds_rundown);
send_msg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_WCBLOCKED, 6, LEN_AND_LIT("wcb_gds_rundown"),
process_id, &csa->ti->curr_tn, DB_LEN_STR(reg));
csa->wbuf_dqd = 0;
wcs_recover(reg);
BG_TRACE_PRO_ANY(csa, lost_block_recovery);
if (!was_crit)
rel_crit(reg);
}
if (JNL_ENABLED(csd) && IS_GTCM_GNP_SERVER_IMAGE)
originator_prc_vec = NULL;
/* If we are the last writing user, then everything must be flushed */
if (we_are_last_writer)
{ /* Time to flush out all of our buffers */
assert(!safe_mode);
if (is_mm)
{
MM_DBFILEXT_REMAP_IF_NEEDED(csa, reg);
cnl->remove_shm = TRUE;
}
if (cnl->wc_blocked && jgbl.onlnrlbk)
{ /* if the last update done by online rollback was not committed in the normal code-path but was
* completed by secshr_db_clnup, wc_blocked will be set to TRUE. But, since online rollback never
* invokes grab_crit (since csa->hold_onto_crit is set to TRUE), wcs_recover is never invoked. This
* could result in the last update never getting flushed to the disk and if online rollback happened
* to be the last writer then the shared memory will be flushed and removed and the last update will
* be lost. So, force wcs_recover if we find ourselves in such a situation. But, wc_blocked is
* possible only if phase1 or phase2 errors are induced using white box test cases
*/
assert(WB_COMMIT_ERR_ENABLED);
wcs_recover(reg);
}
/* Note WCSFLU_SYNC_EPOCH ensures the epoch is synced to the journal and indirectly
* also ensures that the db is fsynced. We don't want to use it in the calls to
* wcs_flu() from t_end() and tp_tend() since we can defer it to out-of-crit there.
* In this case, since we are running down, we don't have any such option.
*/
cnl->remove_shm = wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
/* Since we_are_last_writer, we should be guaranteed that wcs_flu() did not change csd, (in
* case of MM for potential file extension), even if it did a grab_crit(). Therefore, make
* sure that's true.
*/
assert(csd == csa->hdr);
assert(0 == memcmp(csd->label, GDS_LABEL, GDS_LABEL_SZ - 1));
} else if (((canceled_flush_timer && (0 > cnl->wcs_timers)) || canceled_dbsync_timer) && !inst_is_frozen)
{ /* canceled pending db or jnl flush timers - flush database and journal buffers to disk */
if (!was_crit)
grab_crit(reg);
/* we need to sync the epoch as the fact that there is no active pending flush timer implies
* there will be noone else who will flush the dirty buffers and EPOCH to disk in a timely fashion
*/
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
if (!was_crit)
rel_crit(reg);
assert((dba_mm == cs_data->acc_meth) || (csd == cs_data));
csd = cs_data; /* In case this is MM and wcs_flu() remapped an extended database, reset csd */
}
/* Do rundown journal processing after buffer flushes since they require jnl to be open */
if (JNL_ENABLED(csd))
{ /* the following tp_change_reg() is not needed due to the assert csa == cs_addrs at the beginning
* of gds_rundown(), but just to be safe. To be removed by 2002!! --- nars -- 2001/04/25.
*/
tp_change_reg(); /* call this because jnl_ensure_open checks cs_addrs rather than gv_cur_region */
jpc = csa->jnl;
jbp = jpc->jnl_buff;
if (jbp->fsync_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->fsync_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if (jbp->io_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->io_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if ((((NOJNL != jpc->channel) && !JNL_FILE_SWITCHED(jpc))
|| we_are_last_writer && (0 != cnl->jnl_file.u.inode)) && ok_to_write_pfin)
{ /* We need to close the journal file cleanly if we have the latest generation journal file open
* or if we are the last writer and the journal file is open in shared memory (not necessarily
* by ourselves e.g. the only process that opened the journal got shot abnormally)
* Note: we should not infer anything from the shared memory value of cnl->jnl_file.u.inode
* if we are not the last writer as it can be concurrently updated.
*/
if (!was_crit)
grab_crit(reg);
if (JNL_ENABLED(csd))
{
SET_GBL_JREC_TIME; /* jnl_ensure_open/jnl_put_jrt_pini/pfin/jnl_file_close all need it */
/* Before writing to jnlfile, adjust jgbl.gbl_jrec_time if needed to maintain time order
* of jnl records. This needs to be done BEFORE the jnl_ensure_open as that could write
* journal records (if it decides to switch to a new journal file).
*/
ADJUST_GBL_JREC_TIME(jgbl, jbp);
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{ /* If we_are_last_writer, we would have already done a wcs_flu() which would
* have written an epoch record and we are guaranteed no further updates
* since we are the last writer. So, just close the journal.
* If the freeaddr == post_epoch_freeaddr, wcs_flu may have skipped writing
* a pini, so allow for that.
*/
assert(!jbp->before_images || is_mm
|| !we_are_last_writer || (0 != jpc->pini_addr) || jgbl.mur_extract
|| (jpc->jnl_buff->freeaddr == jpc->jnl_buff->post_epoch_freeaddr));
/* If we haven't written a pini, let jnl_file_close write the pini/pfin. */
if (!jgbl.mur_extract && (0 != jpc->pini_addr))
jnl_put_jrt_pfin(csa);
/* If not the last writer and no pending flush timer left, do jnl flush now */
if (!we_are_last_writer && (0 > cnl->wcs_timers))
{
if (SS_NORMAL == (jnl_status = jnl_flush(reg)))
{
assert(jbp->freeaddr == jbp->dskaddr);
jnl_fsync(reg, jbp->dskaddr);
assert(jbp->fsync_dskaddr == jbp->dskaddr);
} else
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_JNLFLUSH, 2,
JNL_LEN_STR(csd), ERR_TEXT, 2,
RTS_ERROR_TEXT("Error with journal flush in gds_rundown"),
jnl_status);
assert(NOJNL == jpc->channel);/* jnl file lost has been triggered */
/* In this routine, all code that follows from here on does not
* assume anything about the journaling characteristics of this
* database so it is safe to continue execution even though
* journaling got closed in the middle.
*/
}
}
jnl_file_close(reg, we_are_last_writer, FALSE);
} else
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd),
DB_LEN_STR(reg));
}
if (!was_crit)
rel_crit(reg);
}
}
if (we_are_last_writer) /* Flush the fileheader last and harden the file to disk */
{
if (!was_crit)
grab_crit(reg); /* To satisfy crit requirement in fileheader_sync() */
memset(csd->machine_name, 0, MAX_MCNAMELEN); /* clear the machine_name field */
if (!have_standalone_access && we_are_last_user)
{ /* mupip_exit_handler will do this after mur_close_file */
csd->semid = INVALID_SEMID;
csd->shmid = INVALID_SHMID;
csd->gt_sem_ctime.ctime = 0;
csd->gt_shm_ctime.ctime = 0;
}
fileheader_sync(reg);
if (!was_crit)
rel_crit(reg);
if (!is_mm)
{
GTM_DB_FSYNC(csa, udi->fd, rc); /* Sync it all */
if (-1 == rc)
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
} else
{ /* Now do final MM file sync before exit */
assert(csa->ti->total_blks == csa->total_blks);
#ifdef _AIX
GTM_DB_FSYNC(csa, udi->fd, rc);
if (-1 == rc)
#else
if (-1 == MSYNC((caddr_t)csa->db_addrs[0], (caddr_t)csa->db_addrs[1]))
#endif
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
}
} else if (unsafe_last_writer && !cnl->lastwriterbypas_msg_issued)
{
send_msg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_LASTWRITERBYPAS, 2, DB_LEN_STR(reg));
cnl->lastwriterbypas_msg_issued = TRUE;
}
} /* end if (!reg->read_only && !cnl->donotflush_dbjnl) */
/* We had canceled all db timers at start of rundown. In case as part of rundown (wcs_flu above), we had started
* any timers, cancel them BEFORE setting reg->open to FALSE (assert in wcs_clean_dbsync relies on this).
*/
CANCEL_DB_TIMERS(reg, csa, canceled_flush_timer, canceled_dbsync_timer);
if (reg->read_only && we_are_last_user && !have_standalone_access && cnl->remove_shm)
{ /* mupip_exit_handler will do this after mur_close_file */
db_ipcs.semid = INVALID_SEMID;
db_ipcs.shmid = INVALID_SHMID;
db_ipcs.gt_sem_ctime = 0;
db_ipcs.gt_shm_ctime = 0;
db_ipcs.fn_len = reg->dyn.addr->fname_len;
memcpy(db_ipcs.fn, reg->dyn.addr->fname, reg->dyn.addr->fname_len);
db_ipcs.fn[reg->dyn.addr->fname_len] = 0;
/* request gtmsecshr to flush. read_only cannot flush itself */
WAIT_FOR_REPL_INST_UNFREEZE_SAFE(csa);
if (!csa->read_only_fs)
{
secshrstat = send_mesg2gtmsecshr(FLUSH_DB_IPCS_INFO, 0, (char *)NULL, 0);
if (0 != secshrstat)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gtmsecshr failed to update database file header"));
}
}
/* Done with file now, close it */
CLOSEFILE_RESET(udi->fd, rc); /* resets "udi->fd" to FD_INVALID */
if (-1 == rc)
{
rts_error_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error during file close"), errno);
}
/* Unmap storage if mm mode but only the part that is not the fileheader (so shows up in dumps) */
# if !defined(_AIX)
if (is_mm && (NULL != csa->db_addrs[0]))
{
assert(csa->db_addrs[1] > csa->db_addrs[0]);
munmap_len = (sm_long_t)(csa->db_addrs[1] - csa->db_addrs[0]);
if (0 < munmap_len)
munmap((caddr_t)(csa->db_addrs[0]), (size_t)(munmap_len));
}
# endif
/* Detach our shared memory while still under lock so reference counts will be correct for the next process to run down
* this region. In the process also get the remove_shm status from node_local before detaching.
* If cnl->donotflush_dbjnl is TRUE, it means we can safely remove shared memory without compromising data
* integrity as a reissue of recover will restore the database to a consistent state.
*/
remove_shm = !vermismatch && (cnl->remove_shm || cnl->donotflush_dbjnl);
/* We are done with online rollback on this region. Indicate to other processes by setting the onln_rlbk_pid to 0.
* Do it before releasing crit (t_end relies on this ordering when accessing cnl->onln_rlbk_pid).
*/
if (jgbl.onlnrlbk)
cnl->onln_rlbk_pid = 0;
rel_crit(reg); /* Since we are about to detach from the shared memory, release crit and reset onln_rlbk_pid */
/* If we had skipped flushing journal and database buffers due to a concurrent online rollback, increment the counter
* indicating that in the shared memory so that online rollback can report the # of such processes when it shuts down.
* The same thing is done for both FTOK and access control semaphores when there are too many MUMPS processes.
*/
if (safe_mode) /* indicates flushing was skipped */
{
if (bypassed_access)
cnl->dbrndwn_access_skip++; /* Access semaphore can be bypassed during online rollback */
if (bypassed_ftok)
cnl->dbrndwn_ftok_skip++;
}
if (jgbl.onlnrlbk)
csa->hold_onto_crit = FALSE;
GTM_WHITE_BOX_TEST(WBTEST_HOLD_SEM_BYPASS, cnl->wbox_test_seq_num, 0);
status = shmdt((caddr_t)cnl);
csa->nl = NULL; /* dereferencing nl after detach is not right, so we set it to NULL so that we can test before dereference*/
/* Note that although csa->nl is NULL, we use CSA_ARG(csa) below (not CSA_ARG(NULL)) to be consistent with similar
* usages before csa->nl became NULL. The "is_anticipatory_freeze_needed" function (which is in turn called by the
* CHECK_IF_FREEZE_ON_ERROR_NEEDED macro) does a check of csa->nl before dereferencing shared memory contents so
* we are safe passing "csa".
*/
if (-1 == status)
send_msg_csa(CSA_ARG(csa) VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg), ERR_TEXT, 2,
LEN_AND_LIT("Error during shmdt"), errno);
REMOVE_CSA_FROM_CSADDRSLIST(csa); /* remove "csa" from list of open regions (cs_addrs_list) */
reg->open = FALSE;
/* If file is still not in good shape, die here and now before we get rid of our storage */
assertpro(0 == csa->wbuf_dqd);
ipc_deleted = FALSE;
/* If we are the very last user, remove shared storage id and the semaphores */
if (we_are_last_user)
{ /* remove shared storage, only if last writer to rundown did a successful wcs_flu() */
assert(!vermismatch);
if (remove_shm)
{
ipc_deleted = TRUE;
if (0 != shm_rmid(udi->shmid))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove shared memory"));
/* Note that we no longer have a new shared memory. Currently only used/usable for standalone rollback. */
udi->new_shm = FALSE;
/* mupip recover/rollback don't release the semaphore here, but do it later in db_ipcs_reset (invoked from
* mur_close_files())
*/
if (!have_standalone_access)
{
if (0 != sem_rmid(udi->semid))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove semaphore"));
udi->new_sem = FALSE; /* Note that we no longer have a new semaphore */
udi->grabbed_access_sem = FALSE;
udi->counter_acc_incremented = FALSE;
}
} else if (is_src_server || is_updproc)
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
send_msg_csa(CSA_ARG(csa) VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
{
assert(!have_standalone_access || jgbl.onlnrlbk || safe_mode);
if (!jgbl.onlnrlbk && !have_standalone_access)
{ /* If we were writing, get rid of our writer access count semaphore */
if (!reg->read_only)
{
if (!access_counter_halted)
{
save_errno = do_semop(udi->semid, DB_COUNTER_SEM, -DB_COUNTER_SEM_INCR, SEM_UNDO);
if (0 != save_errno)
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown access control semaphore decrement"),
CALLFROM, save_errno);
}
udi->counter_acc_incremented = FALSE;
}
assert(safe_mode || !bypassed_access);
/* Now remove the rundown lock */
if (!bypassed_access)
{
if (0 != (save_errno = do_semop(udi->semid, DB_CONTROL_SEM, -1, SEM_UNDO)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(12) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_SYSCALL, 5,
RTS_ERROR_TEXT("gds_rundown access control semaphore release"),
CALLFROM, save_errno);
udi->grabbed_access_sem = FALSE;
}
} /* else access control semaphore will be released in db_ipcs_reset */
}
if (!have_standalone_access)
{
if (bypassed_ftok)
{
if (!ftok_counter_halted)
if (0 != (save_errno = do_semop(udi->ftok_semid, DB_COUNTER_SEM, -DB_COUNTER_SEM_INCR, SEM_UNDO)))
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
} else if (!ftok_sem_release(reg, !ftok_counter_halted, FALSE))
{
FTOK_TRACE(csa, csa->ti->curr_tn, ftok_ops_release, process_id);
rts_error_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
}
udi->grabbed_ftok_sem = FALSE;
udi->counter_ftok_incremented = FALSE;
}
ENABLE_INTERRUPTS(INTRPT_IN_GDS_RUNDOWN, prev_intrpt_state);
if (!ipc_deleted)
{
GET_CUR_TIME(time_str);
if (is_src_server)
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Source server"), REG_LEN_STR(reg));
if (is_updproc)
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Update process"), REG_LEN_STR(reg));
if (mupip_jnl_recover && (!jgbl.onlnrlbk || !we_are_last_user))
{
gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
send_msg_csa(CSA_ARG(csa) VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_str,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
}
}
REVERT;
return EXIT_NRM;
}
void gds_rundown(void)
{
bool is_mm, we_are_last_user, we_are_last_writer;
boolean_t ipc_deleted, remove_shm, cancelled_timer, cancelled_dbsync_timer, vermismatch;
now_t now; /* for GET_CUR_TIME macro */
char *time_ptr, time_str[CTIME_BEFORE_NL + 2]; /* for GET_CUR_TIME macro */
gd_region *reg;
int save_errno, status;
int4 semval, ftok_semval, sopcnt, ftok_sopcnt;
short crash_count;
sm_long_t munmap_len;
sgmnt_addrs *csa;
sgmnt_data_ptr_t csd;
struct shmid_ds shm_buf;
struct sembuf sop[2], ftok_sop[2];
uint4 jnl_status;
unix_db_info *udi;
jnl_private_control *jpc;
jnl_buffer_ptr_t jbp;
error_def(ERR_CRITSEMFAIL);
error_def(ERR_DBCCERR);
error_def(ERR_DBFILERR);
error_def(ERR_DBRNDWNWRN);
error_def(ERR_ERRCALL);
error_def(ERR_GBLOFLOW);
error_def(ERR_GTMASSERT);
error_def(ERR_IPCNOTDEL);
error_def(ERR_JNLFLUSH);
error_def(ERR_RNDWNSEMFAIL);
error_def(ERR_TEXT);
error_def(ERR_WCBLOCKED);
forced_exit = FALSE; /* Okay, we're dying already -- let rel_crit live in peace now.
* If coming through a DAL, not necessarily dying. what to do then? -- nars -- 8/15/2001
*/
grabbed_access_sem = FALSE;
jnl_status = 0;
reg = gv_cur_region; /* Local copy */
/*
* early out for cluster regions
* to avoid tripping the assert below.
* Note:
* This early out is consistent with VMS. It has been
* noted that all of the gtcm assignments
* to gv_cur_region should use the TP_CHANGE_REG
* macro. This would also avoid the assert problem
* and should be done eventually.
*/
if (dba_cm == reg->dyn.addr->acc_meth)
return;
udi = FILE_INFO(reg);
csa = &udi->s_addrs;
csd = csa->hdr;
assert(csa == cs_addrs && csd == cs_data);
if ((reg->open) && (dba_usr == csd->acc_meth))
{
change_reg();
gvusr_rundown();
return;
}
ESTABLISH(gds_rundown_ch);
if (!reg->open) /* Not open, no point to rundown */
{
if (reg->opening) /* Died partway open, kill rest of way */
{
rel_crit(reg);
mutex_cleanup(reg);
/* revist this to handle MM properly SMW 98/12/16
if (NULL != csa->nl)
{
status = shmdt((caddr_t)csa->nl);
if (-1 == status)
send_msg(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error during shmdt"), errno);
}
*/
shmdt((caddr_t)csa->nl);
csa->nl = NULL;
}
REVERT;
return;
}
switch(csd->acc_meth)
{ /* Pass mm and bg through */
case dba_bg:
is_mm = FALSE;
break;
case dba_mm:
is_mm = TRUE;
break;
case dba_usr:
assert(FALSE);
default:
REVERT;
return;
}
/* Cancel any pending flush timer for this region by this task */
CANCEL_DB_TIMERS(reg, cancelled_timer, cancelled_dbsync_timer);
we_are_last_user = FALSE;
if (!csa->persistent_freeze)
region_freeze(reg, FALSE, FALSE, FALSE);
assert(!csa->read_lock);
rel_crit(reg); /* get locks to known state */
mutex_cleanup(reg);
/*
* We need to guarantee that none else access database file header when semid/shmid fields are reset.
* We already have created ftok semaphore in db_init or, mu_rndwn_file and did not remove it.
* So just lock it. We do it in blocking mode.
*/
if (!ftok_sem_lock(reg, FALSE, FALSE))
rts_error(VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
/*
* For mupip_jnl_recover we already have database access control semaphore.
* We do not release it. We release it from mur_close_files.
*/
if (!mupip_jnl_recover)
{
sop[0].sem_num = 0; sop[0].sem_op = 0; /* Wait for 0 */
sop[1].sem_num = 0; sop[1].sem_op = 1; /* Lock */
sopcnt = 2;
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO | IPC_NOWAIT; /* Don't wait the first time thru */
SEMOP(udi->semid, sop, sopcnt, status);
if (-1 == status) /* We couldn't get it in one shot -- see if we already have it */
{
save_errno = errno;
/* see comment about Linux specific difference in behaviour of semctl() with GETPID in gds_rundown_ch() */
if (semctl(udi->semid, 0, GETPID) == process_id)
{
send_msg(VARLSTCNT(5) MAKE_MSG_INFO(ERR_CRITSEMFAIL), 2,
DB_LEN_STR(reg),
ERR_RNDWNSEMFAIL);
REVERT;
return; /* Already in rundown for this region */
}
if (EAGAIN != save_errno)
{
assert(FALSE);
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown first semop/semctl"), save_errno);
}
sop[0].sem_flg = sop[1].sem_flg = SEM_UNDO; /* Try again - blocking this time */
SEMOP(udi->semid, sop, 2, status);
if (-1 == status) /* We couldn't get it at all.. */
rts_error(VARLSTCNT(5) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), errno);
}
}
grabbed_access_sem = TRUE;
/*
* We now have the dbinit/rundown lock, so we are alone in this code for this region
* and nobody else can attach.
* See if we are all alone in accessing this database shared memory.
*/
assert(csa->ref_cnt); /* decrement private ref_cnt before shared ref_cnt decrement. */
csa->ref_cnt--; /* Currently journaling logic in gds_rundown() in VMS relies on this order to detect last writer */
assert(!csa->ref_cnt);
--csa->nl->ref_cnt;
if (memcmp(csa->nl->now_running, gtm_release_name, gtm_release_name_len + 1))
{ /* VERMISMATCH condition. Possible only if DSE */
assert(dse_running);
vermismatch = TRUE;
} else
vermismatch = FALSE;
if (-1 == shmctl(udi->shmid, IPC_STAT, &shm_buf))
{
save_errno = errno;
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown shmctl"), save_errno);
} else
we_are_last_user = (1 == shm_buf.shm_nattch) && !vermismatch;
assert(!mupip_jnl_recover || we_are_last_user); /* recover => one user */
if (-1 == (semval = semctl(udi->semid, 1, GETVAL)))
rts_error(VARLSTCNT(5) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), errno);
we_are_last_writer = (1 == semval) && (FALSE == reg->read_only) && !vermismatch;/* There's one writer left and I am it */
assert(!(mupip_jnl_recover && !reg->read_only) || we_are_last_writer); /* recover + R/W region => one writer */
if (-1 == (ftok_semval = semctl(udi->ftok_semid, 1, GETVAL)))
rts_error(VARLSTCNT(5) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg), errno);
/* If csa->nl->donotflush_dbjnl is set, it means mupip recover/rollback was interrupted and therefore we should
* not flush shared memory contents to disk as they might be in an inconsistent state.
* In this case, we will go ahead and remove shared memory (without flushing the contents) in this routine.
* A reissue of the recover/rollback command will restore the database to a consistent state.
* Otherwise, if we have write access to this region, let us perform a few writing tasks.
*/
if (csa->nl->donotflush_dbjnl)
csa->wbuf_dqd = 0; /* ignore csa->wbuf_dqd status as we do not care about the cache contents */
else if (!reg->read_only && !vermismatch)
{ /* If we had an orphaned block and were interrupted, set wc_blocked so we can invoke wcs_recover */
if (csa->wbuf_dqd)
{
grab_crit(reg);
SET_TRACEABLE_VAR(csd->wc_blocked, TRUE);
BG_TRACE_PRO_ANY(csa, wcb_gds_rundown);
send_msg(VARLSTCNT(8) ERR_WCBLOCKED, 6, LEN_AND_LIT("wcb_gds_rundown"),
process_id, &csa->ti->curr_tn, DB_LEN_STR(reg));
csa->wbuf_dqd = 0;
wcs_recover(reg);
if (is_mm)
{
assert(FALSE);
csd = csa->hdr;
}
BG_TRACE_PRO_ANY(csa, lost_block_recovery);
rel_crit(reg);
}
if (JNL_ENABLED(csd) && (GTCM_GNP_SERVER_IMAGE == image_type))
originator_prc_vec = NULL;
/* If we are the last writing user, then everything must be flushed */
if (we_are_last_writer)
{ /* Time to flush out all of our buffers */
if (is_mm)
{
if (csa->total_blks != csa->ti->total_blks) /* do remap if file had been extended */
{
grab_crit(reg);
wcs_mm_recover(reg);
csd = csa->hdr;
rel_crit(reg);
}
csa->nl->remove_shm = TRUE;
}
/* Note WCSFLU_SYNC_EPOCH ensures the epoch is synced to the journal and indirectly
* also ensures that the db is fsynced. We don't want to use it in the calls to
* wcs_flu() from t_end() and tp_tend() since we can defer it to out-of-crit there.
* In this case, since we are running down, we don't have any such option.
*/
csa->nl->remove_shm = wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
/* Since we_are_last_writer, we should be guaranteed that wcs_flu() did not change csd, (in
* case of MM for potential file extension), even if it did a grab_crit(). Therefore, make
* sure that's true.
*/
assert(csd == csa->hdr);
assert(0 == memcmp(csd->label, GDS_LABEL, GDS_LABEL_SZ - 1));
csd->trans_hist.header_open_tn = csd->trans_hist.curr_tn;
} else if ((cancelled_timer && (0 > csa->nl->wcs_timers)) || cancelled_dbsync_timer)
{ /* cancelled pending db or jnl flush timers - flush database and journal buffers to disk */
grab_crit(reg);
/* we need to sync the epoch as the fact that there is no active pending flush timer implies
* there will be noone else who will flush the dirty buffers and EPOCH to disk in a timely fashion
*/
wcs_flu(WCSFLU_FLUSH_HDR | WCSFLU_WRITE_EPOCH | WCSFLU_SYNC_EPOCH);
rel_crit(reg);
assert((dba_mm == cs_data->acc_meth) || (csd == cs_data));
csd = cs_data; /* In case this is MM and wcs_flu() remapped an extended database, reset csd */
}
/* Do rundown journal processing after buffer flushes since they require jnl to be open */
if (JNL_ENABLED(csd))
{ /* the following tp_change_reg() is not needed due to the assert csa == cs_addrs at the beginning
* of gds_rundown(), but just to be safe. To be removed by 2002!! --- nars -- 2001/04/25.
*/
tp_change_reg(); /* call this because jnl_ensure_open checks cs_addrs rather than gv_cur_region */
jpc = csa->jnl;
jbp = jpc->jnl_buff;
if (jbp->fsync_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->fsync_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if (jbp->io_in_prog_latch.u.parts.latch_pid == process_id)
{
assert(FALSE);
COMPSWAP_UNLOCK(&jbp->io_in_prog_latch, process_id, 0, LOCK_AVAILABLE, 0);
}
if (((NOJNL != jpc->channel) && !JNL_FILE_SWITCHED(jpc))
|| we_are_last_writer && (0 != csa->nl->jnl_file.u.inode))
{ /* We need to close the journal file cleanly if we have the latest generation journal file open
* or if we are the last writer and the journal file is open in shared memory (not necessarily
* by ourselves e.g. the only process that opened the journal got shot abnormally)
* Note: we should not infer anything from the shared memory value of csa->nl->jnl_file.u.inode
* if we are not the last writer as it can be concurrently updated.
*/
grab_crit(reg);
if (JNL_ENABLED(csd))
{
SET_GBL_JREC_TIME; /* jnl_ensure_open/jnl_put_jrt_pini/pfin/jnl_file_close all need it */
/* Before writing to jnlfile, adjust jgbl.gbl_jrec_time if needed to maintain time order
* of jnl records. This needs to be done BEFORE the jnl_ensure_open as that could write
* journal records (if it decides to switch to a new journal file).
*/
ADJUST_GBL_JREC_TIME(jgbl, jbp);
jnl_status = jnl_ensure_open();
if (0 == jnl_status)
{ /* If we_are_last_writer, we would have already done a wcs_flu() which would
* have written an epoch record and we are guaranteed no further updates
* since we are the last writer. So, just close the journal.
* Although we assert pini_addr should be non-zero for last_writer, we
* play it safe in PRO and write a PINI record if not written already.
*/
assert(!jbp->before_images || is_mm
|| !we_are_last_writer || 0 != jpc->pini_addr);
if (we_are_last_writer && 0 == jpc->pini_addr)
jnl_put_jrt_pini(csa);
if (0 != jpc->pini_addr)
jnl_put_jrt_pfin(csa);
/* If not the last writer and no pending flush timer left, do jnl flush now */
if (!we_are_last_writer && (0 > csa->nl->wcs_timers))
{
if (SS_NORMAL == (jnl_status = jnl_flush(reg)))
{
assert(jbp->freeaddr == jbp->dskaddr);
jnl_fsync(reg, jbp->dskaddr);
assert(jbp->fsync_dskaddr == jbp->dskaddr);
} else
{
send_msg(VARLSTCNT(9) ERR_JNLFLUSH, 2, JNL_LEN_STR(csd),
ERR_TEXT, 2,
RTS_ERROR_TEXT("Error with journal flush in gds_rundown"),
jnl_status);
assert(NOJNL == jpc->channel);/* jnl file lost has been triggered */
/* In this routine, all code that follows from here on does not
* assume anything about the journaling characteristics of this
* database so it is safe to continue execution even though
* journaling got closed in the middle.
*/
}
}
jnl_file_close(reg, we_are_last_writer, FALSE);
} else
send_msg(VARLSTCNT(6) jnl_status, 4, JNL_LEN_STR(csd), DB_LEN_STR(reg));
}
rel_crit(reg);
}
}
if (we_are_last_writer) /* Flush the fileheader last and harden the file to disk */
{
grab_crit(reg); /* To satisfy crit requirement in fileheader_sync() */
memset(csd->machine_name, 0, MAX_MCNAMELEN); /* clear the machine_name field */
if (!mupip_jnl_recover && we_are_last_user)
{ /* mupip_jnl_recover will do this after mur_close_file */
csd->semid = INVALID_SEMID;
csd->shmid = INVALID_SHMID;
csd->gt_sem_ctime.ctime = 0;
csd->gt_shm_ctime.ctime = 0;
}
fileheader_sync(reg);
rel_crit(reg);
if (FALSE == is_mm)
{
if (-1 == fsync(udi->fd)) /* Sync it all */
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
} else
{ /* Now do final MM file sync before exit */
#if !defined(TARGETED_MSYNC) && !defined(NO_MSYNC)
if (-1 == fsync(udi->fd)) /* Sync it all */
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file sync at close"), errno);
}
#else
if (-1 == msync((caddr_t)csa->db_addrs[0], (size_t)(csa->db_addrs[1] - csa->db_addrs[0]), MS_SYNC))
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Error during file msync at close"), errno);
}
#endif
}
}
} /* end if (!reg->read_only && !csa->nl->donotflush_dbjnl) */
if (reg->read_only && we_are_last_user && !mupip_jnl_recover)
{ /* mupip_jnl_recover will do this after mur_close_file */
db_ipcs.semid = INVALID_SEMID;
db_ipcs.shmid = INVALID_SHMID;
db_ipcs.gt_sem_ctime = 0;
db_ipcs.gt_shm_ctime = 0;
db_ipcs.fn_len = reg->dyn.addr->fname_len;
memcpy(db_ipcs.fn, reg->dyn.addr->fname, reg->dyn.addr->fname_len);
db_ipcs.fn[reg->dyn.addr->fname_len] = 0;
/* request gtmsecshr to flush. read_only cannot flush itself */
if (0 != send_mesg2gtmsecshr(FLUSH_DB_IPCS_INFO, 0, (char *)NULL, 0))
rts_error(VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gtmsecshr failed to update database file header"));
}
/* Done with file now, close it */
if (-1 == close(udi->fd))
{
rts_error(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, LEN_AND_LIT("Error during file close"), errno);
}
/* Unmap storage if mm mode but only the part that is not the fileheader (so shows up in dumps) */
if (is_mm)
{
munmap_len = (sm_long_t)((csa->db_addrs[1] - csa->db_addrs[0]) - ROUND_UP(SIZEOF_FILE_HDR(csa->hdr),
MSYNC_ADDR_INCS));
if (munmap_len > 0)
{
munmap((caddr_t)(csa->db_addrs[0] + ROUND_UP(SIZEOF_FILE_HDR(csa->hdr), MSYNC_ADDR_INCS)),
(size_t)(munmap_len));
#ifdef DEBUG_DB64
rel_mmseg((caddr_t)csa->db_addrs[0]);
#endif
}
}
/* Detach our shared memory while still under lock so reference counts will be
* correct for the next process to run down this region.
* In the process also get the remove_shm status from node_local before detaching.
* If csa->nl->donotflush_dbjnl is TRUE, it means we can safely remove shared memory without compromising data
* integrity as a reissue of recover will restore the database to a consistent state.
*/
remove_shm = !vermismatch && (csa->nl->remove_shm || csa->nl->donotflush_dbjnl);
status = shmdt((caddr_t)csa->nl);
csa->nl = NULL; /* dereferencing nl after detach is not right, so we set it to NULL so that we can test before dereference*/
if (-1 == status)
send_msg(VARLSTCNT(9) ERR_DBFILERR, 2, DB_LEN_STR(reg), ERR_TEXT, 2, LEN_AND_LIT("Error during shmdt"), errno);
reg->open = FALSE;
/* If file is still not in good shape, die here and now before we get rid of our storage */
if (csa->wbuf_dqd)
GTMASSERT;
ipc_deleted = FALSE;
/* If we are the very last user, remove shared storage id and the semaphores */
if (we_are_last_user)
{ /* remove shared storage, only if last writer to rundown did a successful wcs_flu() */
assert(!vermismatch);
if (remove_shm)
{
ipc_deleted = TRUE;
if (0 != shm_rmid(udi->shmid))
rts_error(VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove shared memory"));
} else if (is_src_server || is_updproc)
{
gtm_putmsg(VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
send_msg(VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
} else
send_msg(VARLSTCNT(6) ERR_DBRNDWNWRN, 4, DB_LEN_STR(reg), process_id, process_id);
/*
* Don't release semaphore in case of mupip recover/rollback; since it has standalone access.
* It will release the semaphore in mur_close_files.
*/
if (!mupip_jnl_recover)
{
if (0 != sem_rmid(udi->semid))
rts_error(VARLSTCNT(8) ERR_DBFILERR, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("Unable to remove semaphore"));
grabbed_access_sem = FALSE;
}
} else
{
assert(!mupip_jnl_recover);
/* If we were writing, get rid of our writer access count semaphore */
if (!reg->read_only)
if (0 != (save_errno = do_semop(udi->semid, 1, -1, SEM_UNDO)))
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown write semaphore release"), save_errno);
/* Now remove the rundown lock */
if (0 != (save_errno = do_semop(udi->semid, 0, -1, SEM_UNDO)))
rts_error(VARLSTCNT(9) ERR_CRITSEMFAIL, 2, DB_LEN_STR(reg),
ERR_TEXT, 2, RTS_ERROR_TEXT("gds_rundown rundown semaphore release"), save_errno);
grabbed_access_sem = FALSE;
}
if (!ftok_sem_release(reg, !mupip_jnl_recover, FALSE))
rts_error(VARLSTCNT(4) ERR_DBFILERR, 2, DB_LEN_STR(reg));
if (!ipc_deleted)
{
GET_CUR_TIME;
if (is_src_server)
gtm_putmsg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Source server"), REG_LEN_STR(reg));
if (is_updproc)
gtm_putmsg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Update process"), REG_LEN_STR(reg));
if (mupip_jnl_recover)
{
gtm_putmsg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
send_msg(VARLSTCNT(8) ERR_IPCNOTDEL, 6, CTIME_BEFORE_NL, time_ptr,
LEN_AND_LIT("Mupip journal process"), REG_LEN_STR(reg));
}
}
REVERT;
}
int gtmsource_ipc_cleanup(boolean_t auto_shutdown, int *exit_status)
{
boolean_t i_am_the_last_user, attempt_ipc_cleanup;
int status, detach_status, remove_status, expected_nattach;
unix_db_info *udi;
struct shmid_ds shm_buf;
/* Attempt cleaning up the IPCs */
attempt_ipc_cleanup = TRUE;
/* Wait for the Source Server to detach and takeover the semaphore */
if (!auto_shutdown && 0 > grab_sem(SOURCE, SRC_SERV_COUNT_SEM))
{
repl_log(stderr, FALSE, TRUE, "Error taking control of source server count semaphore : %s. Shutdown not complete\n",
REPL_SEM_ERROR);
*exit_status = ABNORMAL_SHUTDOWN;
attempt_ipc_cleanup = FALSE;
}
udi = (unix_db_info *)FILE_INFO(jnlpool.jnlpool_dummy_reg);
if (!auto_shutdown || !gtmsource_srv_count)
expected_nattach = 1; /* Self, or parent */
else
expected_nattach = 0; /* Source server already detached */
i_am_the_last_user = (((status = shmctl(udi->shmid, IPC_STAT, &shm_buf)) == 0) && (shm_buf.shm_nattch == expected_nattach));
if (!i_am_the_last_user)
{
if (status < 0)
repl_log(stderr, FALSE, TRUE, "Error in jnlpool shmctl : %s\n", STRERROR(ERRNO));
else
repl_log(stderr, FALSE, TRUE, "Not deleting jnlpool ipcs. %d processes still attached to jnlpool\n",
shm_buf.shm_nattch - expected_nattach);
attempt_ipc_cleanup = FALSE;
*exit_status = ABNORMAL_SHUTDOWN;
}
if (attempt_ipc_cleanup)
{
if (INVALID_SHMID != udi->shmid && (auto_shutdown || (detach_status = SHMDT(jnlpool.jnlpool_ctl)) == 0)
&& (remove_status = shm_rmid(udi->shmid)) == 0)
{
jnlpool.jnlpool_ctl = jnlpool_ctl = NULL;
pool_init = FALSE;
repl_log(stdout, FALSE, FALSE, "Journal pool shared memory removed\n");
if (0 == (status = remove_sem_set(SOURCE)))
repl_log(stdout, FALSE, TRUE, "Journal pool semaphore removed\n");
else
{
repl_log(stderr, FALSE, TRUE, "Error removing jnlpool semaphore : %s\n", STRERROR(status));
*exit_status = ABNORMAL_SHUTDOWN;
}
} else if (INVALID_SHMID != udi->shmid)
{
if (!auto_shutdown && detach_status < 0)
repl_log(stderr, FALSE, FALSE,
"Error detaching from jnlpool shared memory : %s. Shared memory not removed\n",
STRERROR(ERRNO));
else if (remove_status != 0)
{
if (!auto_shutdown)
{
jnlpool.jnlpool_ctl = NULL; /* Detached successfully */
jnlpool_ctl = NULL;
pool_init = FALSE;
}
repl_log(stderr, FALSE, TRUE, "Error removing jnlpool shared memory : %s\n", STRERROR(ERRNO));
}
*exit_status = ABNORMAL_SHUTDOWN;
}
}
return attempt_ipc_cleanup;
}
int gtmrecv_ipc_cleanup(boolean_t auto_shutdown, int *exit_status)
{
boolean_t i_am_the_last_user, attempt_ipc_cleanup;
int status, detach_status, remove_status, expected_nattach;
struct shmid_ds shm_buf;
/* Attempt cleaning up the IPCs */
attempt_ipc_cleanup = TRUE;
/*
* Wait for the Receiver Server and Update Process to detach and
* takeover the semaphores. Note that the Receiver Server has already
* waited for the Update Process to detach. It is done here as a
* precaution against Receiver Server crashes.
*/
if (!auto_shutdown)
status = grab_sem(RECV, RECV_SERV_COUNT_SEM);
else
status = 0;
if (0 == status && 0 > (status = grab_sem(RECV, UPD_PROC_COUNT_SEM)))
rel_sem(RECV, RECV_SERV_COUNT_SEM);
if (status < 0)
{
repl_log(stderr, FALSE, TRUE,
"Error taking control of Receiver Server/Update Process count semaphore : %s. Shutdown not complete\n",
REPL_SEM_ERROR);
*exit_status = ABNORMAL_SHUTDOWN;
attempt_ipc_cleanup = FALSE;
}
/* Now we have locked out all users from the receive pool */
if (!auto_shutdown || !gtmrecv_srv_count)
expected_nattach = 1; /* Self, or parent */
else
expected_nattach = 0; /* Receiver server already detached */
i_am_the_last_user = (((status = shmctl(recvpool_shmid, IPC_STAT, &shm_buf)) == 0)
&& (shm_buf.shm_nattch == expected_nattach));
if (!i_am_the_last_user)
{
if (status < 0)
repl_log(stderr, FALSE, TRUE, "Error in jnlpool shmctl : %s\n", STRERROR(ERRNO));
else
repl_log(stderr, FALSE, TRUE,
"Not deleting receive pool ipcs. %d processes still attached to receive pool\n",
shm_buf.shm_nattch - expected_nattach);
attempt_ipc_cleanup = FALSE;
*exit_status = ABNORMAL_SHUTDOWN;
}
if (attempt_ipc_cleanup)
{
if (INVALID_SHMID != recvpool_shmid && (auto_shutdown || (detach_status = SHMDT(recvpool.recvpool_ctl)) == 0)
&& (remove_status = shm_rmid(recvpool_shmid)) == 0)
{
recvpool.recvpool_ctl = NULL;
repl_log(stdout, FALSE, FALSE, "Receive pool shared memory removed\n");
if (0 == (status = remove_sem_set(RECV)))
repl_log(stdout, FALSE, TRUE, "Receive pool semaphore removed\n");
else
{
repl_log(stderr, FALSE, TRUE, "Error removing receive pool semaphore : %s\n", STRERROR(status));
*exit_status = ABNORMAL_SHUTDOWN;
}
} else if (INVALID_SHMID != recvpool_shmid)
{
if (!auto_shutdown && detach_status < 0)
repl_log(stderr, FALSE, FALSE,
"Error detaching from receive pool shared memory : %s. Shared memory not removed\n",
STRERROR(ERRNO));
else if (remove_status != 0)
{
if (!auto_shutdown)
recvpool.recvpool_ctl = NULL; /* Detached successfully */
repl_log(stderr, FALSE, TRUE, "Error removing receive pool shared memory : %s\n", STRERROR(ERRNO));
}
*exit_status = ABNORMAL_SHUTDOWN;
}
}
return attempt_ipc_cleanup;
}
void db_init_err_cleanup(boolean_t retry_dbinit)
{
unix_db_info *udi;
gd_segment *seg;
sgmnt_addrs *csa;
int rc, lcl_new_dbinit_ipc;
boolean_t ftok_counter_halted, access_counter_halted;
/* Here, we can not rely on the validity of csa->hdr because this function can be triggered anywhere in db_init().Because
* we don't have access to file header, we can not know if counters are disabled so we go by our best guess, not disabled,
* during cleanup.
*/
assert(NULL != db_init_region);
seg = db_init_region->dyn.addr;
udi = NULL;
if (NULL != seg->file_cntl)
udi = FILE_INFO(db_init_region);
if (NULL != udi)
{
if (FD_INVALID != udi->fd && !retry_dbinit)
CLOSEFILE_RESET(udi->fd, rc); /* resets "udi->fd" to FD_INVALID */
assert(FD_INVALID == udi->fd || retry_dbinit);
csa = &udi->s_addrs;
# ifdef _AIX
if ((NULL != csa->hdr) && (dba_mm == db_init_region->dyn.addr->acc_meth))
{
assert((NULL != csa->db_addrs[1]) && (csa->db_addrs[1] > csa->db_addrs[0]));
munmap((caddr_t)csa->db_addrs[0], (size_t)(csa->db_addrs[1] - csa->db_addrs[0]));
}
# endif
if (NULL != csa->jnl)
{
free(csa->jnl);
csa->jnl = NULL;
}
/* If shared memory is not available or if this is a VERMISMATCH error situation (where we do not know the exact
* position of csa->nl->ftok_counter_halted or if it even exists in the other version), we have to be pessimistic
* and assume the counters are halted. This avoids prematurely removing the semaphores.
*/
if ((NULL != csa->nl) && ((int)ERR_VERMISMATCH != SIGNAL))
{
ftok_counter_halted = csa->nl->ftok_counter_halted;
access_counter_halted = csa->nl->access_counter_halted;
shmdt((caddr_t)csa->nl);
csa->nl = (node_local_ptr_t)NULL;
} else
{
ftok_counter_halted = TRUE;
access_counter_halted = TRUE;
}
if (udi->shm_created && (INVALID_SHMID != udi->shmid))
{
shm_rmid(udi->shmid);
udi->shmid = INVALID_SHMID;
udi->shm_created = FALSE;
}
if (udi->sem_created && (INVALID_SEMID != udi->semid))
{
sem_rmid(udi->semid);
udi->semid = INVALID_SEMID;
udi->sem_created = FALSE;
udi->grabbed_access_sem = FALSE;
udi->counter_acc_incremented = FALSE;
}
if (udi->counter_acc_incremented && !access_counter_halted)
{
assert((INVALID_SEMID != udi->semid) && !db_init_region->read_only);
/* decrement the read-write sem */
do_semop(udi->semid, DB_COUNTER_SEM, -DB_COUNTER_SEM_INCR, SEM_UNDO | IPC_NOWAIT);
udi->counter_acc_incremented = FALSE;
}
if (udi->grabbed_access_sem)
{
do_semop(udi->semid, DB_CONTROL_SEM, -1, SEM_UNDO | IPC_NOWAIT); /* release the startup-shutdown sem */
udi->grabbed_access_sem = FALSE;
}
if (udi->grabbed_ftok_sem)
ftok_sem_release(db_init_region, udi->counter_ftok_incremented && !ftok_counter_halted, TRUE);
else if (udi->counter_ftok_incremented && !ftok_counter_halted)
do_semop(udi->ftok_semid, DB_COUNTER_SEM, -DB_COUNTER_SEM_INCR, SEM_UNDO | IPC_NOWAIT);
udi->counter_ftok_incremented = FALSE;
udi->grabbed_ftok_sem = FALSE;
if (!IS_GTCM_GNP_SERVER_IMAGE && !retry_dbinit) /* gtcm_gnp_server reuses file_cntl */
{
free(seg->file_cntl->file_info);
free(seg->file_cntl);
seg->file_cntl = NULL;
}
}
/* Enable interrupts in case we are here with intrpt_ok_state == INTRPT_IN_GVCST_INIT due to an rts error.
* Normally we would have the new state stored in "prev_intrpt_state" but that is not possible here because
* the corresponding DEFER_INTERRUPTS happened in gvcst_init.c (a different function) so we have an assert
* there that the previous state was INTRPT_OK_TO_INTERRUPT and use that instead of prev_intrpt_state here.
*/
if (!retry_dbinit)
ENABLE_INTERRUPTS(INTRPT_IN_GVCST_INIT, INTRPT_OK_TO_INTERRUPT);
}
