sanlock-devel August 2024

sanlock-devel@lists.fedorahosted.org

1 participants
10 discussions

[sanlock] 07/07: sanlock: change log level for resource names
by pagure＠pagure.io 22 Aug '24

22 Aug '24

This is an automated email from the git hooks/post-receive script. teigland pushed a commit to branch master in repository sanlock. commit 88b7726e71da241c801bb23a7bd3833ee0f91a2e Author: David Teigland <teigland(a)redhat.com> AuthorDate: Thu Aug 22 15:41:58 2024 -0500 sanlock: change log level for resource names Stop writing every resource name and location to sanlock.log by default. Also fix names_log_priority (set in sanlock.conf) which controls the log level for the resource names, and was being ignored. --- src/main.c | 2 +- src/resource.c | 3 ++- src/sanlock.8 | 4 ++-- src/sanlock.conf | 2 +- 4 files changed, 6 insertions(+), 5 deletions(-) diff --git a/src/main.c b/src/main.c index 82872fa..6f63bf8 100644 --- a/src/main.c +++ b/src/main.c @@ -4222,7 +4222,7 @@ int main(int argc, char *argv[]) com.kill_grace_seconds = DEFAULT_GRACE_SEC; com.high_priority = DEFAULT_HIGH_PRIORITY; com.mlock_level = DEFAULT_MLOCK_LEVEL; - com.names_log_priority = LOG_WARNING; + com.names_log_priority = LOG_INFO; com.max_worker_threads = DEFAULT_MAX_WORKER_THREADS; com.io_timeout = DEFAULT_IO_TIMEOUT; com.write_init_io_timeout = DEFAULT_WRITE_INIT_IO_TIMEOUT; diff --git a/src/resource.c b/src/resource.c index 388e737..6cbfbf0 100644 --- a/src/resource.c +++ b/src/resource.c @@ -1974,7 +1974,8 @@ int acquire_token(struct task *task, struct token *token, uint32_t cmd_flags, if (new_id) { /* save a record of what this id is for later debugging */ - log_warnt(token, "resource %.48s:%.48s:%.256s:%llu", + log_level(token->space_id, token->res_id, NULL, com.names_log_priority, + "resource %.48s:%.48s:%.256s:%llu", token->r.lockspace_name, token->r.name, token->r.disks[0].path, diff --git a/src/sanlock.8 b/src/sanlock.8 index f91933b..23774a1 100644 --- a/src/sanlock.8 +++ b/src/sanlock.8 @@ -1306,10 +1306,10 @@ syslog_priority = 3 See -S .IP \[bu] 2 -names_log_priority = 4 +names_log_priority = 6 .br Log resource names at this priority level (uses syslog priority numbers). -If this is greater than or equal to logfile_priority, each requested resource +If this number less than or equal to logfile_priority, each requested resource name and location is recorded in sanlock.log. .IP \[bu] 2 diff --git a/src/sanlock.conf b/src/sanlock.conf index 6fae7dd..325c37d 100644 --- a/src/sanlock.conf +++ b/src/sanlock.conf @@ -20,7 +20,7 @@ # syslog_priority = 3 # command line: -S 3 # -# names_log_priority = 4 +# names_log_priority = 6 # command line: n/a # # use_watchdog = 1 -- To stop receiving notification emails like this one, please contact the administrator of this repository.

1 0

[sanlock] 06/07: sanlock: include worker thread info in status
by pagure＠pagure.io 22 Aug '24

22 Aug '24

This is an automated email from the git hooks/post-receive script. teigland pushed a commit to branch master in repository sanlock. commit 025fbb828a7fda53d3db905c96e5c8299b286738 Author: David Teigland <teigland(a)redhat.com> AuthorDate: Wed Aug 21 16:38:58 2024 -0500 sanlock: include worker thread info in status 'sanlock status -D' will show worker thread info. --- src/client_cmd.c | 14 +++++ src/cmd.c | 34 +++++++++++- src/main.c | 137 ++++++++++++++++++++++++++++++++++++++++++++++++- src/sanlock_internal.h | 14 +++++ src/sanlock_sock.h | 1 + 5 files changed, 198 insertions(+), 2 deletions(-) diff --git a/src/client_cmd.c b/src/client_cmd.c index e06d0f7..43a3da8 100644 --- a/src/client_cmd.c +++ b/src/client_cmd.c @@ -180,6 +180,17 @@ static void status_host(struct sanlk_state *st, char *str, int debug) print_debug(str, st->str_len); } +static void status_worker(struct sanlk_state *st, char *str, int debug) +{ + if (!debug) + return; + + printf("worker %u\n", st->data32); + + if (st->str_len) + print_debug(str, st->str_len); +} + static void print_st(struct sanlk_state *st, char *str, char *bin, int debug) { switch (st->type) { @@ -195,6 +206,9 @@ static void print_st(struct sanlk_state *st, char *str, char *bin, int debug) case SANLK_STATE_RESOURCE: status_resource(st, str, bin, debug); break; + case SANLK_STATE_WORKER: + status_worker(st, str, debug); + break; } } diff --git a/src/cmd.c b/src/cmd.c index 01ebb9d..e91da9d 100644 --- a/src/cmd.c +++ b/src/cmd.c @@ -2302,6 +2302,7 @@ static int print_state_daemon(char *str) "max_sectors_kb_align=%d " "max_sectors_kb_num=%d " "max_worker_threads=%d " + "num_worker_threads=%d " "use_hugepages=%s " "write_init_io_timeout=%u " "use_aio=%d " @@ -2334,6 +2335,7 @@ static int print_state_daemon(char *str) com.max_sectors_kb_align, com.max_sectors_kb_num, com.max_worker_threads, + daemon_status_num_workers, com.use_hugepages ? "all" : "none", com.write_init_io_timeout, main_task.use_aio, @@ -2531,6 +2533,30 @@ static void send_state_daemon(int fd) send_all(fd, str, str_len, MSG_NOSIGNAL); } +int print_state_worker(int i, char *str); + +static void send_state_worker(int fd, int i) +{ + struct sanlk_state st; + char str[SANLK_STATE_MAXSTR]; + int str_len; + + memset(&st, 0, sizeof(st)); + st.type = SANLK_STATE_WORKER; + st.data32 = i; + + str_len = print_state_worker(i, str); + + if (!str_len) + return; + + st.str_len = str_len; + + send_all(fd, &st, sizeof(st), MSG_NOSIGNAL); + if (str_len) + send_all(fd, str, str_len, MSG_NOSIGNAL); +} + static void send_state_client(int fd, struct client *cl, int ci) { struct sanlk_state st; @@ -2660,7 +2686,7 @@ static void cmd_status(int ci, int fd, struct sm_header *h_recv, int client_maxi struct sm_header h; struct client *cl; struct space *sp; - int ci_iter; + int ci_iter, i; log_cmd(cmd, "cmd_status %d,%d", ci, fd); @@ -2677,6 +2703,12 @@ static void cmd_status(int ci, int fd, struct sm_header *h_recv, int client_maxi if (h_recv->data == SANLK_STATE_DAEMON) return; + for (i = 0; i < com.max_worker_threads; i++) + send_state_worker(fd, i); + + if (h_recv->data == SANLK_STATE_WORKER) + return; + for (ci_iter = 0; ci_iter <= client_maxi; ci_iter++) { cl = &client[ci_iter]; if (!cl->used) diff --git a/src/main.c b/src/main.c index e8bf7c4..82872fa 100644 --- a/src/main.c +++ b/src/main.c @@ -69,6 +69,7 @@ struct thread_pool { pthread_mutex_t mutex; pthread_cond_t cond; pthread_cond_t quit_wait; + struct worker_info *info; }; /* priorities are LOG_* from syslog.h */ @@ -941,14 +942,43 @@ static int main_loop(void) return 0; } +int print_state_worker(int i, char *str); + +int print_state_worker(int i, char *str) +{ + if (!pool.info[i].work_count) + return 0; + + memset(str, 0, SANLK_STATE_MAXSTR); + + snprintf(str, SANLK_STATE_MAXSTR-1, + "cmd_last=%s " + "busy=%u " + "work_count=%u " + "io_count=%u " + "to_count=%u " + "paxos_hugebuf_created=%u " + "paxos_hugebuf_used=%u", + cmd_num_to_str(pool.info[i].cmd_last), + pool.info[i].busy, + pool.info[i].work_count, + pool.info[i].io_count, + pool.info[i].to_count, + pool.info[i].paxos_hugebuf_created, + pool.info[i].paxos_hugebuf_used); + + return strlen(str) + 1; +} + static void *thread_pool_worker(void *data) { struct task task; struct cmd_args *ca; + long workerid = (long)data; memset(&task, 0, sizeof(struct task)); setup_task_aio(&task, main_task.use_aio, WORKER_AIO_CB_SIZE); - snprintf(task.name, NAME_ID_SIZE, "worker%ld", (long)data); + snprintf(task.name, NAME_ID_SIZE, "worker%ld", workerid); pthread_mutex_lock(&pool.mutex); @@ -962,12 +992,20 @@ static void *thread_pool_worker(void *data) while (!list_empty(&pool.work_data)) { ca = list_first_entry(&pool.work_data, struct cmd_args, list); list_del(&ca->list); + pool.info[workerid].busy = 1; + pool.info[workerid].cmd_last = ca->header.cmd; pthread_mutex_unlock(&pool.mutex); call_cmd_thread(&task, ca); free(ca); pthread_mutex_lock(&pool.mutex); + pool.info[workerid].busy = 0; + pool.info[workerid].work_count++; + pool.info[workerid].io_count = task.io_count; + pool.info[workerid].to_count = task.to_count; + pool.info[workerid].paxos_hugebuf_created = task.paxos_hugebuf_created; + pool.info[workerid].paxos_hugebuf_used = task.paxos_hugebuf_used; } if (pool.quit) @@ -975,6 +1013,7 @@ static void *thread_pool_worker(void *data) } pool.num_workers--; + daemon_status_num_workers--; /* for status reporting */ if (!pool.num_workers) pthread_cond_signal(&pool.quit_wait); pthread_mutex_unlock(&pool.mutex); @@ -1007,6 +1046,7 @@ static int thread_pool_add_work(struct cmd_args *ca) return rv; } pool.num_workers++; + daemon_status_num_workers++; /* for status reporting */ } pthread_cond_signal(&pool.cond); @@ -1036,6 +1076,12 @@ static int thread_pool_create(int min_workers, int max_workers) pthread_cond_init(&pool.cond, NULL); pthread_cond_init(&pool.quit_wait, NULL); pool.max_workers = max_workers; + pool.info = malloc(max_workers * sizeof(struct worker_info)); + + if (!pool.info) + return -ENOMEM; + + memset(pool.info, 0, max_workers * sizeof(struct worker_info)); for (i = 0; i < min_workers; i++) { rv = pthread_create(&th, NULL, thread_pool_worker, @@ -1046,6 +1092,7 @@ static int thread_pool_create(int min_workers, int max_workers) break; } pool.num_workers++; + daemon_status_num_workers++; /* for status reporting */ } if (rv < 0) @@ -2731,6 +2778,94 @@ static int read_command_line(int argc, char *argv[]) return 0; } +const char *cmd_num_to_str(int cmd) +{ + switch (cmd) { + case SM_CMD_INQ_LOCKSPACE: + return "inq_lockspace"; + case SM_CMD_READ_RESOURCE_OWNERS: + return "read_resource_owners"; + case SM_CMD_GET_LOCKSPACES: + return "get_lockspaces"; + case SM_CMD_GET_HOSTS: + return "get_hosts"; + case SM_CMD_REGISTER: + return "register"; + case SM_CMD_ADD_LOCKSPACE: + return "add_lockspace"; + case SM_CMD_REM_LOCKSPACE: + return "rem_lockspace"; + case SM_CMD_SHUTDOWN: + return "shutdown"; + case SM_CMD_STATUS: + return "status"; + case SM_CMD_ACQUIRE: + return "acquire"; + case SM_CMD_RELEASE: + return "release"; + case SM_CMD_INQUIRE: + return "inquire"; + case SM_CMD_RESTRICT: + return "restrict"; + case SM_CMD_REQUEST: + return "request"; + case SM_CMD_ALIGN: + return "align"; + case SM_CMD_EXAMINE_LOCKSPACE: + return "examine_lockspace"; + case SM_CMD_EXAMINE_RESOURCE: + return "examine_resource"; + case SM_CMD_HOST_STATUS: + return "host_status"; + case SM_CMD_KILLPATH: + return "killpath"; + case SM_CMD_WRITE_LOCKSPACE: + return "write_lockspace"; + case SM_CMD_WRITE_RESOURCE: + return "write_resource"; + case SM_CMD_READ_LOCKSPACE: + return "read_lockspace"; + case SM_CMD_READ_RESOURCE: + return "read_resource"; + case SM_CMD_SET_LVB: + return "set_lvb"; + case SM_CMD_GET_LVB: + return "get_lvb"; + case SM_CMD_CONVERT: + return "convert"; + case SM_CMD_VERSION: + return "version"; + case SM_CMD_SHUTDOWN_WAIT: + return "shutdown_wait"; + case SM_CMD_REG_EVENT: + return "reg_event"; + case SM_CMD_END_EVENT: + return "end_event"; + case SM_CMD_SET_EVENT: + return "set_event"; + case SM_CMD_SET_CONFIG: + return "set_config"; + case SM_CMD_RENEWAL: + return "renewal"; + case SM_CMD_FORMAT_RINDEX: + return "format_rindex"; + case SM_CMD_UPDATE_RINDEX: + return "update_rindex"; + case SM_CMD_LOOKUP_RINDEX: + return "lookup_rindex"; + case SM_CMD_CREATE_RESOURCE: + return "create_resource"; + case SM_CMD_DELETE_RESOURCE: + return "delete_resource"; + case SM_CMD_REBUILD_RINDEX: + return "rebuild_rindex"; + case SM_CMD_LOG_DUMP: + return "log_dump"; + default: + return "unknown"; + } +} + uint32_t cmd_str_to_num(const char *str) { if (!strcmp(str, "inq_lockspace")) diff --git a/src/sanlock_internal.h b/src/sanlock_internal.h index 22b7698..ebb0a9c 100644 --- a/src/sanlock_internal.h +++ b/src/sanlock_internal.h @@ -291,6 +291,17 @@ struct task { EXTERN struct task main_task; +struct worker_info { + unsigned int busy; + unsigned int cmd_last; + unsigned int work_count; + unsigned int io_count; + unsigned int to_count; + unsigned int paxos_hugebuf_used; + unsigned int paxos_hugebuf_created; + unsigned int paxos_hugebuf_size; +}; + /* TODO: change used, suspend, need_free, pid_dead to flags */ #define CL_KILLPATH_PID 0x00000001 /* include pid as killpath arg */ @@ -418,6 +429,7 @@ struct command_line { EXTERN struct command_line com; uint32_t cmd_str_to_num(const char *str); +const char *cmd_num_to_str(int cmd); uint64_t cmd_num_to_debug_flag(uint32_t cmd); int is_cmd_debug(uint32_t cmd); void set_cmd_debug(uint32_t cmd); @@ -474,6 +486,8 @@ EXTERN char our_host_name_global[SANLK_NAME_LEN+1]; EXTERN char our_product_uuid[SANLK_NAME_LEN+1]; EXTERN char our_product_uuid_compact[SANLK_NAME_LEN+1]; /* dash chars omitted from uuid */ +EXTERN int daemon_status_num_workers; /* replicate thread pool counter for status reporting */ + EXTERN int kill_count_max; EXTERN int is_helper; EXTERN int helper_ci; diff --git a/src/sanlock_sock.h b/src/sanlock_sock.h index 924468d..e252f83 100644 --- a/src/sanlock_sock.h +++ b/src/sanlock_sock.h @@ -83,6 +83,7 @@ struct sm_header { #define SANLK_STATE_RESOURCE 4 #define SANLK_STATE_HOST 5 #define SANLK_STATE_RENEWAL 6 +#define SANLK_STATE_WORKER 7 struct sanlk_state { uint32_t type; /* SANLK_STATE_ */ -- To stop receiving notification emails like this one, please contact the administrator of this repository.

1 0

[sanlock] 05/07: sanlock: reuse io buffer with hugepages
by pagure＠pagure.io 22 Aug '24

22 Aug '24

This is an automated email from the git hooks/post-receive script. teigland pushed a commit to branch master in repository sanlock. commit 66ed5ea259da54939ae8c02533a4238c6e3e053d Author: David Teigland <teigland(a)redhat.com> AuthorDate: Mon Aug 19 16:22:39 2024 -0500 sanlock: reuse io buffer with hugepages Avoid repeating malloc/madvise/free every time an io buffer is needed to read the full paxos lease area. --- src/paxos_lease.c | 168 ++++++++++++++++++++++++++++++++++++------------- src/sanlock_internal.h | 6 +- src/task.c | 3 + 3 files changed, 133 insertions(+), 44 deletions(-) diff --git a/src/paxos_lease.c b/src/paxos_lease.c index e62a93f..8f7d816 100644 --- a/src/paxos_lease.c +++ b/src/paxos_lease.c @@ -505,7 +505,8 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, int read_len; int iobuf_len; int read_mem_align; - int is_huge = 0; + int use_huge = 0; + int use_task_hugebuf = 0; int phase2 = 0; int d, q, rv = 0; int q_max = -1; @@ -517,34 +518,59 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, iobuf_len = token->align_size; /* can increase (from 1MB) for hugepage */ read_mem_align = getpagesize(); /* increases to 2MB for hugepage */ - /* TODO: for hugepages, the task should keep a huge iobuf around to reuse - rather than alloc+madvise+free each time; after a while of heavy use, - madvise may be less likely to successfully turn the iobuf into a hugepage. */ - - if (com.use_hugepages && ((iobuf_len == ONE_MB_IN_BYTES) || !(iobuf_len % TWO_MB_IN_BYTES))) { + /* Determine if hugepages/hubebuf is applicable. */ + if (com.use_hugepages && (num_disks == 1) && + ((iobuf_len == ONE_MB_IN_BYTES) || !(iobuf_len % TWO_MB_IN_BYTES))) { read_mem_align = TWO_MB_IN_BYTES; if (iobuf_len < TWO_MB_IN_BYTES) iobuf_len = TWO_MB_IN_BYTES; - is_huge = 1; + use_huge = 1; + } + + /* Reuse a previously allocated buffer. */ + if (use_huge && task->paxos_hugebuf && (task->paxos_hugebuf_size >= iobuf_len)) { + use_task_hugebuf = 1; + iobuf_array[0] = task->paxos_hugebuf; + task->paxos_hugebuf_used++; + goto mem_done; } + /* Allocate a new buffer. */ for (d = 0; d < num_disks; d++) { p_iobuf[d] = &iobuf_array[d]; rv = posix_memalign((void *)p_iobuf[d], read_mem_align, iobuf_len); if (rv) return rv; + } + + /* Make the new buffer use hugepages. */ + if (use_huge) { + char *bp; + iobuf = iobuf_array[0]; + madvise(iobuf, iobuf_len, MADV_HUGEPAGE); + /* allocate the pages */ + for (bp = iobuf; bp < iobuf+iobuf_len; bp += TWO_MB_IN_BYTES) + memset(bp, 0, 1); - if (is_huge) { - char *bp; - iobuf = iobuf_array[d]; - madvise(iobuf, iobuf_len, MADV_HUGEPAGE); - /* allocate the pages */ - for (bp = iobuf; bp < iobuf+iobuf_len; bp += TWO_MB_IN_BYTES) - memset(bp, 0, 1); + /* + * Save the new buffer to reuse if the task doesn't + * have one (Note: the task may already have paxos_hugebuf + * with a smaller size than we need here, in which case we + * ignore that, allocate a new one, and free it at the end.) + * TODO: add config min_hugebuf_size to avoid that case? + */ + if (!task->paxos_hugebuf) { + task->paxos_hugebuf = iobuf; + task->paxos_hugebuf_size = iobuf_len; + task->paxos_hugebuf_created++; + use_task_hugebuf = 1; + log_taskd(task, "created hugebuf %p %u", iobuf, task->paxos_hugebuf_created); } } + mem_done: + /* * phase 1 * @@ -570,7 +596,7 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, num_writes = 0; for (d = 0; d < num_disks; d++) { - /* acquire io: write 1 */ + /* acquire io: write 1 (sector) */ rv = write_dblock(task, token, &token->disks[d], token->host_id, &dblock); if (rv < 0) continue; @@ -599,10 +625,17 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, memset(iobuf, 0, read_len); - /* acquire io: read 2 */ + /* acquire io: read 2 (lease area) */ rv = read_iobuf(disk->fd, disk->offset, iobuf, read_len, task, token->io_timeout, NULL); - if (rv == SANLK_AIO_TIMEOUT) + if (rv == SANLK_AIO_TIMEOUT) { iobuf_array[d] = NULL; + if (use_task_hugebuf) { + /* hugebuf will be freed when aio completes */ + task->paxos_hugebuf = NULL; + task->paxos_hugebuf_size = 0; + use_task_hugebuf = 0; + } + } if (rv < 0) continue; num_reads++; @@ -788,7 +821,7 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, num_writes = 0; for (d = 0; d < num_disks; d++) { - /* acquire io: write 2 */ + /* acquire io: write 2 (sector) */ rv = write_dblock(task, token, &token->disks[d], token->host_id, &dblock); if (rv < 0) continue; @@ -817,10 +850,17 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, memset(iobuf, 0, read_len); - /* acquire io: read 3 */ + /* acquire io: read 3 (lease area) */ rv = read_iobuf(disk->fd, disk->offset, iobuf, read_len, task, token->io_timeout, NULL); - if (rv == SANLK_AIO_TIMEOUT) + if (rv == SANLK_AIO_TIMEOUT) { iobuf_array[d] = NULL; + if (use_task_hugebuf) { + /* hugebuf will be freed when aio completes */ + task->paxos_hugebuf = NULL; + task->paxos_hugebuf_size = 0; + use_task_hugebuf = 0; + } + } if (rv < 0) continue; num_reads++; @@ -939,13 +979,19 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, memcpy(dblock_out, &dblock, sizeof(struct paxos_dblock)); error = SANLK_OK; out: - for (d = 0; d < num_disks; d++) { - /* don't free iobufs that have timed out */ - if (!iobuf_array[d]) - continue; - - /* TODO: the task should keep huge iobuf and reuse it. */ - free(iobuf_array[d]); + if (use_task_hugebuf) { + if (rv != SANLK_AIO_TIMEOUT) { + /* Keep buffer to reuse. */ + memset(task->paxos_hugebuf, 0, task->paxos_hugebuf_size); + } + } else { + for (d = 0; d < num_disks; d++) { + /* don't free iobufs that have timed out */ + if (!iobuf_array[d]) + continue; + free(iobuf_array[d]); + iobuf_array[d] = NULL; + } } if (phase2 && (error < 0) && @@ -1415,7 +1461,8 @@ static int _lease_read_one(struct task *task, int read_len; int iobuf_len; int read_mem_align; - int is_huge = 0; + int use_huge = 0; + int use_task_hugebuf = 0; int q, tmp_q = -1, rv; /* FIXME: what case is this? */ @@ -1428,27 +1475,48 @@ static int _lease_read_one(struct task *task, iobuf_len = token->align_size; /* can increase (from 1MB) for hugepage */ read_mem_align = getpagesize(); /* increases to 2MB for hugepage */ + /* Determine if hugepages/hubebuf is applicable. */ if (com.use_hugepages && ((iobuf_len == ONE_MB_IN_BYTES) || !(iobuf_len % TWO_MB_IN_BYTES))) { read_mem_align = TWO_MB_IN_BYTES; if (iobuf_len < TWO_MB_IN_BYTES) iobuf_len = TWO_MB_IN_BYTES; - is_huge = 1; + use_huge = 1; } - p_iobuf = &iobuf; + /* Reuse a previously allocated buffer. */ + if (use_huge && task->paxos_hugebuf && (task->paxos_hugebuf_size >= iobuf_len)) { + use_task_hugebuf = 1; + iobuf = task->paxos_hugebuf; + task->paxos_hugebuf_used++; + goto mem_done; + } + /* Allocate a new buffer. */ + p_iobuf = &iobuf; rv = posix_memalign((void *)p_iobuf, read_mem_align, iobuf_len); if (rv) return rv; - if (is_huge) { + /* Make the new buffer use hugepages. */ + if (use_huge) { char *bp; madvise(iobuf, iobuf_len, MADV_HUGEPAGE); /* allocate the pages */ for (bp = iobuf; bp < iobuf+iobuf_len; bp += TWO_MB_IN_BYTES) memset(bp, 0, 1); + + /* Save the new buffer to reuse if the task doesn't have one. */ + if (!task->paxos_hugebuf) { + task->paxos_hugebuf = iobuf; + task->paxos_hugebuf_size = iobuf_len; + task->paxos_hugebuf_created++; + use_task_hugebuf = 1; + log_taskd(task, "created hugebuf %p %u", iobuf, task->paxos_hugebuf_created); + } } + mem_done: + rv = read_iobuf(disk->fd, disk->offset, iobuf, read_len, task, token->io_timeout, NULL); if (rv < 0) goto out; @@ -1520,8 +1588,20 @@ static int _lease_read_one(struct task *task, bk_debug); out: - if (rv != SANLK_AIO_TIMEOUT) - free(iobuf); + if (use_task_hugebuf) { + if (rv != SANLK_AIO_TIMEOUT) { + /* Keep buffer to reuse. */ + memset(task->paxos_hugebuf, 0, task->paxos_hugebuf_size); + } else { + /* hugebuf will be freed when aio completes */ + task->paxos_hugebuf = NULL; + task->paxos_hugebuf_size = 0; + } + } else { + if (rv != SANLK_AIO_TIMEOUT) + free(iobuf); + /* else iobuf will be freed when aio completes */ + } return rv; } @@ -1717,19 +1797,19 @@ static int write_new_leader(struct task *task, */ /* - * i/o required to acquire a free lease - * (1 disk in token, 512 byte sectors, default num_hosts of 2000) + * acquire io: disk ios needed to acquire a free and uncontended lease * * paxos_lease_acquire() - * paxos_lease_read() 1 read 1 MB (entire lease area) - * run_ballot() - * write_dblock() 1 write 512 bytes (1 dblock sector) - * read_iobuf() 1 read 1 MB (round up num_hosts + 2 sectors) - * write_dblock() 1 write 512 bytes (1 dblock sector) - * read_iobuf() 1 read 1 MB (round up num_hosts + 2 sectors) - * write_new_leader() 1 write 512 bytes (1 leader sector) + * R1 paxos_lease_read() 1 read 1 MB (lease area size) + * run_ballot() + * W1 write_dblock() 1 write 512 bytes (1 dblock sector) + * R2 read_iobuf() 1 read 1 MB (lease area size) + * W2 write_dblock() 1 write 512 bytes (1 dblock sector) + * R3 read_iobuf() 1 read 1 MB (lease area size) + * W3 write_new_leader() 1 write 512 bytes (1 leader sector) * * 6 i/os = 3 1MB reads, 3 512 byte writes + * (assuming 512 sector size and 1MB area size) */ int paxos_lease_acquire(struct task *task, @@ -1778,7 +1858,7 @@ int paxos_lease_acquire(struct task *task, memset(&tmp_leader, 0, sizeof(tmp_leader)); copy_cur_leader = 0; - /* acquire io: read 1 */ + /* acquire io: read 1 (lease area) */ error = paxos_lease_read(task, token, flags, &cur_leader, &max_mbal, "paxos_acquire", 1); if (error < 0) goto out; @@ -2216,6 +2296,7 @@ int paxos_lease_acquire(struct task *task, next_lver = larger_lver + 1; } + /* acquire io: write 1 (sector), read 2 (lease area), write 2 (sector), read 3 (lease area) */ error = run_ballot(task, token, flags, cur_leader.num_hosts, next_lver, our_mbal, &dblock, &other_max_mbal, &larger_lver); if ((error == SANLK_DBLOCK_MBAL) || (error == SANLK_DBLOCK_LVER)) { @@ -2296,6 +2377,7 @@ int paxos_lease_acquire(struct task *task, new_leader.checksum = 0; /* set after leader_record_out */ + /* acquire io: write 3 (sector) */ error = write_new_leader(task, token, &new_leader, "paxos_acquire"); if (error < 0) { /* See comment in run_ballot about this flag. */ diff --git a/src/sanlock_internal.h b/src/sanlock_internal.h index 7725eb6..22b7698 100644 --- a/src/sanlock_internal.h +++ b/src/sanlock_internal.h @@ -276,10 +276,14 @@ struct task { unsigned int io_count; /* stats */ unsigned int to_count; /* stats */ + unsigned int paxos_hugebuf_used; /* stats */ + unsigned int paxos_hugebuf_created; /* stats */ int use_aio; int cb_size; - char *iobuf; + int paxos_hugebuf_size; + char *paxos_hugebuf; /* paxos_lease_acquire reads of lease area */ + char *iobuf; /* delta_lease_renew reads of lease area */ io_context_t aio_ctx; struct aicb *read_iobuf_timeout_aicb; struct aicb *callbacks; diff --git a/src/task.c b/src/task.c index ad2e761..8b73097 100644 --- a/src/task.c +++ b/src/task.c @@ -145,6 +145,9 @@ void close_task_aio(struct task *task) if (task->iobuf) free(task->iobuf); + if (task->paxos_hugebuf) + free(task->paxos_hugebuf); + skip_aio: if (task->callbacks) free(task->callbacks); -- To stop receiving notification emails like this one, please contact the administrator of this repository.

1 0

[sanlock] 04/07: sanlock: use hugepages to avoid splitting read io
by pagure＠pagure.io 22 Aug '24

22 Aug '24

This is an automated email from the git hooks/post-receive script. teigland pushed a commit to branch master in repository sanlock. commit bcd94c521c40676a31175998f5697f2a8d76087d Author: David Teigland <teigland(a)redhat.com> AuthorDate: Mon Aug 12 16:38:33 2024 -0500 sanlock: use hugepages to avoid splitting read io Use 2MB transparent hugepages via madvise(MADV_HUGEPAGE) to reduce or eliminate splitting read io's for lease areas. This requires allocating 2MB of memory for 1MB lease areas, leading to some minimal wasted allocation. --- src/cmd.c | 2 + src/delta_lease.c | 50 +++++++++++++++++----- src/diskio.c | 14 +++--- src/lockspace.c | 3 -- src/main.c | 8 ++++ src/paxos_lease.c | 114 ++++++++++++++++++++++++++++++++++++------------- src/sanlock.8 | 11 +++++ src/sanlock.conf | 3 ++ src/sanlock_internal.h | 5 +++ 9 files changed, 160 insertions(+), 50 deletions(-) diff --git a/src/cmd.c b/src/cmd.c index 561e8d5..01ebb9d 100644 --- a/src/cmd.c +++ b/src/cmd.c @@ -2302,6 +2302,7 @@ static int print_state_daemon(char *str) "max_sectors_kb_align=%d " "max_sectors_kb_num=%d " "max_worker_threads=%d " + "use_hugepages=%s " "write_init_io_timeout=%u " "use_aio=%d " "io_timeout=%d " @@ -2333,6 +2334,7 @@ static int print_state_daemon(char *str) com.max_sectors_kb_align, com.max_sectors_kb_num, com.max_worker_threads, + com.use_hugepages ? "all" : "none", com.write_init_io_timeout, main_task.use_aio, com.io_timeout, diff --git a/src/delta_lease.c b/src/delta_lease.c index 55a5bba..fb4a640 100644 --- a/src/delta_lease.c +++ b/src/delta_lease.c @@ -21,6 +21,7 @@ #include <syslog.h> #include <sys/types.h> #include <sys/time.h> +#include <sys/mman.h> #include "sanlock_internal.h" #include "sanlock.h" @@ -575,7 +576,11 @@ int delta_lease_renew(struct task *task, uint32_t checksum; uint32_t reap_timeout_msec; uint64_t host_id, id_offset, new_ts, now; - int rv, iobuf_len, sector_size; + int sector_size = sp->sector_size; + int align_size = sp->align_size; /* 1,2,4,8 MB depending on sector_size */ + int iobuf_len = sp->align_size; /* may increase (from 1MB) for hugepage */ + int read_len = sp->align_size; /* not increased (from 1MB) for hugepage */ + int rv; if (!leader_last) { log_erros(sp, "delta_renew no leader_last"); @@ -589,15 +594,11 @@ int delta_lease_renew(struct task *task, host_id = leader_last->owner_id; - iobuf_len = sp->align_size; - - sector_size = sp->sector_size; - /* offset of our leader_record */ id_offset = (host_id - 1) * sector_size; - if (id_offset > iobuf_len) { - log_erros(sp, "delta_renew bad offset %llu iobuf_len %d", - (unsigned long long)id_offset, iobuf_len); + if (id_offset > align_size) { + log_erros(sp, "delta_renew bad offset %llu align_size %d", + (unsigned long long)id_offset, align_size); return -EINVAL; } @@ -632,7 +633,7 @@ int delta_lease_renew(struct task *task, clock_gettime(CLOCK_MONOTONIC_RAW, &begin); rv = read_iobuf_reap(disk->fd, disk->offset, - task->iobuf, iobuf_len, task, reap_timeout_msec); + task->iobuf, read_len, task, reap_timeout_msec); log_space(sp, "delta_renew reap %d", rv); @@ -676,23 +677,50 @@ int delta_lease_renew(struct task *task, */ if (!task->iobuf) { + int read_mem_align = getpagesize(); + int is_huge = 0; + /* this will happen the first time renew is called, and after a timed out renewal read fails to be reaped (see task->iobuf = NULL above) */ p_iobuf = &task->iobuf; - rv = posix_memalign((void *)p_iobuf, getpagesize(), iobuf_len); + /* + * Minimum hugepage is 2MB, so for 1MB lease area we increase + * the iobuf size and alignment to 2MB, wasting a MB of space + * to more likely avoid splitting the i/o (read_len remains + * 1MB even though iobuf_len is extended to 2MB.) + * iobuf_len begins being set to align_size (1MB, 2MB, 4MB, 8MB). + */ + if (com.use_hugepages && ((iobuf_len == ONE_MB_IN_BYTES) || !(iobuf_len % TWO_MB_IN_BYTES))) { + read_mem_align = TWO_MB_IN_BYTES; + if (iobuf_len < TWO_MB_IN_BYTES) + iobuf_len = TWO_MB_IN_BYTES; + is_huge = 1; + } + + rv = posix_memalign((void *)p_iobuf, read_mem_align, iobuf_len); if (rv) { log_erros(sp, "dela_renew memalign rv %d", rv); rv = -ENOMEM; } + + if (is_huge) { + char *iobuf = task->iobuf; + char *bp; + + madvise(iobuf, iobuf_len, MADV_HUGEPAGE); + /* allocate the pages */ + for (bp = iobuf; bp < iobuf+iobuf_len; bp += TWO_MB_IN_BYTES) + memset(bp, 0, 1); + } } if (log_renewal_level != -1) log_level(sp->space_id, 0, NULL, log_renewal_level, "delta_renew begin read"); - rv = read_iobuf(disk->fd, disk->offset, task->iobuf, iobuf_len, task, sp->io_timeout, rd_ms); + rv = read_iobuf(disk->fd, disk->offset, task->iobuf, read_len, task, sp->io_timeout, rd_ms); if (rv) { /* the next time delta_lease_renew() is called, prev_result will be this rv. If this rv is SANLK_AIO_TIMEOUT, we'll diff --git a/src/diskio.c b/src/diskio.c index 697bde7..1d84a6c 100644 --- a/src/diskio.c +++ b/src/diskio.c @@ -914,13 +914,13 @@ int write_sectors(const struct sync_disk *disk, int sector_size, uint64_t sector /* read aligned io buffer */ -int read_iobuf(int fd, uint64_t offset, char *iobuf, int iobuf_len, +int read_iobuf(int fd, uint64_t offset, char *iobuf, int len, struct task *task, int ioto, int *rd_ms) { if (task && task->use_aio) - return do_read_aio_linux(fd, offset, iobuf, iobuf_len, task, ioto, rd_ms); + return do_read_aio_linux(fd, offset, iobuf, len, task, ioto, rd_ms); else - return do_read(fd, offset, iobuf, iobuf_len, task, rd_ms); + return do_read(fd, offset, iobuf, len, task, rd_ms); } /* read sector_count sectors starting with sector_nr, where sector_nr @@ -977,7 +977,7 @@ int read_sectors(const struct sync_disk *disk, int sector_size, uint64_t sector_ The aicb used in a task's last timed out read_iobuf is task->read_iobuf_timeout_aicb . */ -int read_iobuf_reap(int fd, uint64_t offset, char *iobuf, int iobuf_len, +int read_iobuf_reap(int fd, uint64_t offset, char *iobuf, int read_len, struct task *task, uint32_t ioto_msec) { struct timespec ts; @@ -995,7 +995,7 @@ int read_iobuf_reap(int fd, uint64_t offset, char *iobuf, int iobuf_len, return -EINVAL; if (iocb->u.c.buf != iobuf) return -EINVAL; - if (iocb->u.c.nbytes != iobuf_len) + if (iocb->u.c.nbytes != read_len) return -EINVAL; if (iocb->u.c.offset != offset) return -EINVAL; @@ -1044,9 +1044,9 @@ int read_iobuf_reap(int fd, uint64_t offset, char *iobuf, int iobuf_len, rv = event.res; goto out; } - if (event.res != iobuf_len) { + if (event.res != read_len) { log_taskw(task, "aio collect %s %p:%p:%p result %ld:%ld match len %d r", - op_str, ev_aicb, ev_iocb, ev_aicb->buf, event.res, event.res2, iobuf_len); + op_str, ev_aicb, ev_iocb, ev_aicb->buf, event.res, event.res2, read_len); rv = -EMSGSIZE; goto out; } diff --git a/src/lockspace.c b/src/lockspace.c index 4d4240d..a2bee3f 100644 --- a/src/lockspace.c +++ b/src/lockspace.c @@ -752,9 +752,6 @@ static void save_renewal_history(struct space *sp, int delta_result, } } -#define ONE_MB_IN_BYTES 1048576 -#define ONE_MB_IN_KB 1024 - static void set_lockspace_max_sectors_kb(struct space *sp, int sector_size, int align_size) { struct stat st; diff --git a/src/main.c b/src/main.c index f9ebe96..e8bf7c4 100644 --- a/src/main.c +++ b/src/main.c @@ -3069,6 +3069,13 @@ static void read_config_file(void) val = DEFAULT_MIN_WORKER_THREADS; com.max_worker_threads = val; + } else if (!strcmp(str, "use_hugepages")) { + memset(str, 0, sizeof(str)); + get_val_str(line, str); + if (!strcmp(str, "none")) + com.use_hugepages = 0; + else if (!strcmp(str, "all")) + com.use_hugepages = 1; } } @@ -4096,6 +4103,7 @@ int main(int argc, char *argv[]) com.max_sectors_kb_ignore = DEFAULT_MAX_SECTORS_KB_IGNORE; com.max_sectors_kb_align = DEFAULT_MAX_SECTORS_KB_ALIGN; com.max_sectors_kb_num = DEFAULT_MAX_SECTORS_KB_NUM; + com.use_hugepages = 1; com.debug_cmds = ~0LL; com.debug_hosts = 1; diff --git a/src/paxos_lease.c b/src/paxos_lease.c index 3f11b30..e62a93f 100644 --- a/src/paxos_lease.c +++ b/src/paxos_lease.c @@ -20,6 +20,7 @@ #include <syslog.h> #include <sys/types.h> #include <sys/time.h> +#include <sys/mman.h> #include "sanlock_internal.h" #include "diskio.h" @@ -80,6 +81,7 @@ static void bk_debug_append(char *bk_debug, char *bk_str) } } +/* static uint32_t roundup_power_of_two(uint32_t val) { val--; @@ -91,6 +93,7 @@ static uint32_t roundup_power_of_two(uint32_t val) val++; return val; } +*/ uint32_t leader_checksum(struct leader_record *lr) { @@ -492,32 +495,54 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, struct paxos_dblock *bk_end; struct paxos_dblock *bk; struct sync_disk *disk; - char *iobuf[SANLK_MAX_DISKS]; + char *iobuf_array[SANLK_MAX_DISKS]; char **p_iobuf[SANLK_MAX_DISKS]; + char *iobuf; uint32_t checksum; int num_disks = token->r.num_disks; int num_writes, num_reads; - int sector_size = token->sector_size; - int sector_count; + int sector_size; + int read_len; int iobuf_len; + int read_mem_align; + int is_huge = 0; int phase2 = 0; int d, q, rv = 0; int q_max = -1; int error; - sector_count = roundup_power_of_two(num_hosts + 2); + sector_size = token->sector_size; + /* token->align_size: 1,2,4,8 MB depending on sector_size */ + read_len = token->align_size; /* not increased (from 1MB) for hugepage */ + iobuf_len = token->align_size; /* can increase (from 1MB) for hugepage */ + read_mem_align = getpagesize(); /* increases to 2MB for hugepage */ - iobuf_len = sector_count * sector_size; + /* TODO: for hugepages, the task should keep a huge iobuf around to reuse + rather than alloc+madvise+free each time; after a while of heavy use, + madvise may be less likely to successfully turn the iobuf into a hugepage. */ - if (!iobuf_len) - return -EINVAL; + if (com.use_hugepages && ((iobuf_len == ONE_MB_IN_BYTES) || !(iobuf_len % TWO_MB_IN_BYTES))) { + read_mem_align = TWO_MB_IN_BYTES; + if (iobuf_len < TWO_MB_IN_BYTES) + iobuf_len = TWO_MB_IN_BYTES; + is_huge = 1; + } for (d = 0; d < num_disks; d++) { - p_iobuf[d] = &iobuf[d]; + p_iobuf[d] = &iobuf_array[d]; - rv = posix_memalign((void *)p_iobuf[d], getpagesize(), iobuf_len); + rv = posix_memalign((void *)p_iobuf[d], read_mem_align, iobuf_len); if (rv) return rv; + + if (is_huge) { + char *bp; + iobuf = iobuf_array[d]; + madvise(iobuf, iobuf_len, MADV_HUGEPAGE); + /* allocate the pages */ + for (bp = iobuf; bp < iobuf+iobuf_len; bp += TWO_MB_IN_BYTES) + memset(bp, 0, 1); + } } /* @@ -567,20 +592,23 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, for (d = 0; d < num_disks; d++) { disk = &token->disks[d]; - if (!iobuf[d]) + if (!iobuf_array[d]) continue; - memset(iobuf[d], 0, iobuf_len); + + iobuf = iobuf_array[d]; + + memset(iobuf, 0, read_len); /* acquire io: read 2 */ - rv = read_iobuf(disk->fd, disk->offset, iobuf[d], iobuf_len, task, token->io_timeout, NULL); + rv = read_iobuf(disk->fd, disk->offset, iobuf, read_len, task, token->io_timeout, NULL); if (rv == SANLK_AIO_TIMEOUT) - iobuf[d] = NULL; + iobuf_array[d] = NULL; if (rv < 0) continue; num_reads++; for (q = 0; q < num_hosts; q++) { - bk_end = (struct paxos_dblock *)(iobuf[d] + ((2 + q)*sector_size)); + bk_end = (struct paxos_dblock *)(iobuf + ((2 + q)*sector_size)); checksum = dblock_checksum(bk_end); @@ -782,20 +810,23 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, for (d = 0; d < num_disks; d++) { disk = &token->disks[d]; - if (!iobuf[d]) + if (!iobuf_array[d]) continue; - memset(iobuf[d], 0, iobuf_len); + + iobuf = iobuf_array[d]; + + memset(iobuf, 0, read_len); /* acquire io: read 3 */ - rv = read_iobuf(disk->fd, disk->offset, iobuf[d], iobuf_len, task, token->io_timeout, NULL); + rv = read_iobuf(disk->fd, disk->offset, iobuf, read_len, task, token->io_timeout, NULL); if (rv == SANLK_AIO_TIMEOUT) - iobuf[d] = NULL; + iobuf_array[d] = NULL; if (rv < 0) continue; num_reads++; for (q = 0; q < num_hosts; q++) { - bk_end = (struct paxos_dblock *)(iobuf[d] + ((2 + q)*sector_size)); + bk_end = (struct paxos_dblock *)(iobuf + ((2 + q)*sector_size)); checksum = dblock_checksum(bk_end); @@ -910,9 +941,11 @@ static int run_ballot(struct task *task, struct token *token, uint32_t flags, out: for (d = 0; d < num_disks; d++) { /* don't free iobufs that have timed out */ - if (!iobuf[d]) + if (!iobuf_array[d]) continue; - free(iobuf[d]); + + /* TODO: the task should keep huge iobuf and reuse it. */ + free(iobuf_array[d]); } if (phase2 && (error < 0) && @@ -1375,25 +1408,48 @@ static int _lease_read_one(struct task *task, struct paxos_dblock bk; char *iobuf, **p_iobuf; uint32_t host_id = token->host_id; - uint32_t sector_size = token->sector_size; uint32_t checksum; struct paxos_dblock *bk_end; uint64_t tmp_mbal = 0; - int q, tmp_q = -1, rv, iobuf_len; + int sector_size; + int read_len; + int iobuf_len; + int read_mem_align; + int is_huge = 0; + int q, tmp_q = -1, rv; - iobuf_len = token->align_size; - if (iobuf_len < 0) - return iobuf_len; + /* FIXME: what case is this? */ + if (token->align_size < 0) + return token->align_size; + + sector_size = token->sector_size; + /* token->align_size: 1,2,4,8 MB depending on sector_size */ + read_len = token->align_size; /* not increased (from 1MB) for hugepage */ + iobuf_len = token->align_size; /* can increase (from 1MB) for hugepage */ + read_mem_align = getpagesize(); /* increases to 2MB for hugepage */ + + if (com.use_hugepages && ((iobuf_len == ONE_MB_IN_BYTES) || !(iobuf_len % TWO_MB_IN_BYTES))) { + read_mem_align = TWO_MB_IN_BYTES; + if (iobuf_len < TWO_MB_IN_BYTES) + iobuf_len = TWO_MB_IN_BYTES; + is_huge = 1; + } p_iobuf = &iobuf; - rv = posix_memalign((void *)p_iobuf, getpagesize(), iobuf_len); + rv = posix_memalign((void *)p_iobuf, read_mem_align, iobuf_len); if (rv) return rv; - memset(iobuf, 0, iobuf_len); + if (is_huge) { + char *bp; + madvise(iobuf, iobuf_len, MADV_HUGEPAGE); + /* allocate the pages */ + for (bp = iobuf; bp < iobuf+iobuf_len; bp += TWO_MB_IN_BYTES) + memset(bp, 0, 1); + } - rv = read_iobuf(disk->fd, disk->offset, iobuf, iobuf_len, task, token->io_timeout, NULL); + rv = read_iobuf(disk->fd, disk->offset, iobuf, read_len, task, token->io_timeout, NULL); if (rv < 0) goto out; diff --git a/src/sanlock.8 b/src/sanlock.8 index 75abcf5..f91933b 100644 --- a/src/sanlock.8 +++ b/src/sanlock.8 @@ -1277,6 +1277,9 @@ paxos algorithm to acquire the paxos_lease, and set a new owner. /etc/sanlock/sanlock.conf +The current settings in use by the sanlock daemon can be seen +in the output of 'sanlock status -D'. + .IP \[bu] 2 quiet_fail = 1 .br @@ -1445,6 +1448,14 @@ Not doing so will invalidate the lease protection provided by sanlock. The io_timeout should usually be tuned along with this value, e.g. watchdog_fire_timeout = 30 with io_timeout = 5. +.IP \[bu] 2 +use_hugepages = <str> +.br +Set to "all" to use transparent hugepages (2MB via MADV_HUGEPAGE.) +This should minimize, or prevent, splitting read io's on lease areas. +2MB is allocated for 1MB lease areas, causing some extra memory usage. +Set to "none" to disable. + .SH SEE ALSO .BR wdmd (8) diff --git a/src/sanlock.conf b/src/sanlock.conf index d55c280..6fae7dd 100644 --- a/src/sanlock.conf +++ b/src/sanlock.conf @@ -78,3 +78,6 @@ # # debug_hosts = 1 # command line: n/a +# +# use_hugepages = all +# command line: n/a diff --git a/src/sanlock_internal.h b/src/sanlock_internal.h index cb41538..7725eb6 100644 --- a/src/sanlock_internal.h +++ b/src/sanlock_internal.h @@ -35,6 +35,10 @@ #include <libaio.h> +#define ONE_MB_IN_BYTES 1048576 +#define TWO_MB_IN_BYTES 2097152 +#define ONE_MB_IN_KB 1024 + /* default max number of hosts supported */ #define DEFAULT_MAX_HOSTS 2000 @@ -377,6 +381,7 @@ struct command_line { int clear_arg; int sector_size; int align_size; + int use_hugepages; char *uname; /* -U */ int uid; /* -U */ char *gname; /* -G */ -- To stop receiving notification emails like this one, please contact the administrator of this repository.

1 0

[sanlock] 03/07: sanlock: don't reduce max_sectors_kb
by pagure＠pagure.io 22 Aug '24

22 Aug '24

This is an automated email from the git hooks/post-receive script. teigland pushed a commit to branch master in repository sanlock. commit 52d095a1c824bf3a591d431afee7a1aba32d5a31 Author: David Teigland <teigland(a)redhat.com> AuthorDate: Wed Aug 14 16:07:00 2024 -0500 sanlock: don't reduce max_sectors_kb If the current value is larger than the desired value, then do nothing, leaving the current value. --- src/diskio.c | 6 +++--- src/lockspace.c | 18 +++++++++++++++--- src/sanlock.8 | 1 + 3 files changed, 19 insertions(+), 6 deletions(-) diff --git a/src/diskio.c b/src/diskio.c index 488427e..697bde7 100644 --- a/src/diskio.c +++ b/src/diskio.c @@ -139,16 +139,16 @@ static int write_sysfs_size(const char *disk_path, const char *name, unsigned in int set_max_sectors_kb(struct sync_disk *disk, uint32_t set_kb) { - unsigned int max_kb = 0; + unsigned int cur_kb = 0; int rv; - rv = read_sysfs_size(disk->path, "max_sectors_kb", &max_kb); + rv = read_sysfs_size(disk->path, "max_sectors_kb", &cur_kb); if (rv < 0) { log_debug("set_max_sectors_kb read error %d %s", rv, disk->path); return rv; } - if (max_kb == set_kb) + if (cur_kb >= set_kb) return 0; rv = write_sysfs_size(disk->path, "max_sectors_kb", set_kb); diff --git a/src/lockspace.c b/src/lockspace.c index 4cd8277..4d4240d 100644 --- a/src/lockspace.c +++ b/src/lockspace.c @@ -54,7 +54,7 @@ static int helper_set_max_sectors_kb(struct sync_disk *disk, uint32_t set_kb) struct helper_msg hm; struct stat st; int ma, mi; - unsigned int max_kb = 0; + unsigned int cur_kb = 0; int rv; rv = stat(disk->path, &st); @@ -68,13 +68,13 @@ static int helper_set_max_sectors_kb(struct sync_disk *disk, uint32_t set_kb) snprintf(sysfs_path, sizeof(sysfs_path), "/sys/dev/block/%d:%d/queue/max_sectors_kb", ma, mi); sysfs_path[SANLK_HELPER_PATH_LEN-1] = '\0'; - rv = read_sysfs_uint(sysfs_path, &max_kb); + rv = read_sysfs_uint(sysfs_path, &cur_kb); if (rv < 0) { /* max_sectors_kb setting may not exist */ return -1; } - if (max_kb == set_kb) + if (cur_kb >= set_kb) return 0; if (helper_kill_fd == -1) @@ -760,6 +760,7 @@ static void set_lockspace_max_sectors_kb(struct space *sp, int sector_size, int struct stat st; int align_size_kb = align_size / 1024; /* align_size is in bytes */ unsigned int hw_kb = 0; + unsigned int cur_kb = 0; unsigned int set_kb = 0; int rv; @@ -787,6 +788,17 @@ static void set_lockspace_max_sectors_kb(struct space *sp, int sector_size, int return; } + rv = read_sysfs_size(sp->host_id_disk.path, "max_sectors_kb", &cur_kb); + if (rv < 0 || !cur_kb) { + log_space(sp, "set_lockspace_max_sectors_kb max_sectors_kb unknown %d %u", rv, cur_kb); + return; + } + + if (cur_kb >= set_kb) { + log_space(sp, "set_lockspace_max_sectors_kb keep current %u for set %u", cur_kb, set_kb); + return; + } + if (hw_kb < set_kb) { /* * If the hardware won't support requested size, try setting 1MB. diff --git a/src/sanlock.8 b/src/sanlock.8 index d94e6e5..75abcf5 100644 --- a/src/sanlock.8 +++ b/src/sanlock.8 @@ -1385,6 +1385,7 @@ changing max_sectors_kb for the lockspace disk when starting a lockspace. Set to "align" (no quotes) to set max_sectors_kb for the lockspace disk to the align size of the lockspace. Set to a number to set a specific number of KB for all lockspace disks. +A larger existing max_sectors_kb value will not be reduced by this setting. .IP \[bu] 2 debug_clients = 0 -- To stop receiving notification emails like this one, please contact the administrator of this repository.

1 0

[sanlock] 02/07: sanlock: fix is dead log message for unused host id
by pagure＠pagure.io 22 Aug '24

22 Aug '24

This is an automated email from the git hooks/post-receive script. teigland pushed a commit to branch master in repository sanlock. commit 87038c493a9e8fda2805c855d179674ea2f43208 Author: David Teigland <teigland(a)redhat.com> AuthorDate: Tue Aug 20 11:32:38 2024 -0500 sanlock: fix is dead log message for unused host id Remove log message "host ID GEN 0 is dead" for a host id lease that was released and not currently used. From "sanlock: fix time calculations for dead host" --- src/lockspace.c | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/src/lockspace.c b/src/lockspace.c index ef3336e..4cd8277 100644 --- a/src/lockspace.c +++ b/src/lockspace.c @@ -485,6 +485,12 @@ void check_other_leases(struct space *sp, char *buf) if (sp->host_id == hs->owner_id) continue; + /* + * Other host lease is released, not failing or dead. + */ + if (!hs->timestamp) + continue; + /* * Don't bother logging info about other host leases if * we are quitting/failing ourself. -- To stop receiving notification emails like this one, please contact the administrator of this repository.

1 0

[sanlock] 01/07: sanlock: repeat helper_msg string termination
by pagure＠pagure.io 22 Aug '24

22 Aug '24

This is an automated email from the git hooks/post-receive script. teigland pushed a commit to branch master in repository sanlock. commit 9eae80474fd86193af03148e58010186f9e39fc7 Author: David Teigland <teigland(a)redhat.com> AuthorDate: Thu Aug 15 10:47:12 2024 -0500 sanlock: repeat helper_msg string termination --- src/helper.c | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/src/helper.c b/src/helper.c index 84b5263..987834d 100644 --- a/src/helper.c +++ b/src/helper.c @@ -205,6 +205,10 @@ int run_helper(int in_fd, int out_fd, int log_stderr) if (rv) continue; + /* terminated at sender, but confirm for checker */ + hm.path[SANLK_HELPER_PATH_LEN-1] = '\0'; + hm.args[SANLK_HELPER_ARGS_LEN-1] = '\0'; + if (hm.type == HELPER_MSG_RUNPATH) { pid = fork(); if (!pid) { -- To stop receiving notification emails like this one, please contact the administrator of this repository.

1 0

[sanlock] branch master updated: sanlock: fix invalid strcpy in direct dump
by pagure＠pagure.io 15 Aug '24

15 Aug '24

This is an automated email from the git hooks/post-receive script. teigland pushed a commit to branch master in repository sanlock. The following commit(s) were added to refs/heads/master by this push: new 93a24a4 sanlock: fix invalid strcpy in direct dump 93a24a4 is described below commit 93a24a45f617acee76f1d19c1c3e40fc3be06677 Author: David Teigland <teigland(a)redhat.com> AuthorDate: Wed Aug 14 11:51:17 2024 -0500 sanlock: fix invalid strcpy in direct dump If a lockspace or resource name is the full 48 characters in length, then the ondisk name field does not contain a terminating null character and strcpy fails. --- src/direct.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/src/direct.c b/src/direct.c index f70dc19..3bc7659 100644 --- a/src/direct.c +++ b/src/direct.c @@ -730,8 +730,8 @@ int direct_dump(struct task *task, char *dump_path, int force_mode) if (!lr->owner_id && !lr->owner_generation) continue; - strcpy(sname, lr->space_name); - strcpy(rname, lr->resource_name); + memcpy(sname, lr->space_name, NAME_ID_SIZE); + memcpy(rname, lr->resource_name, NAME_ID_SIZE); printf("%08llu %36s %48s %010llu %04llu %04llu", (unsigned long long)(start_offset + ((sector_nr + i) * sector_size)), @@ -754,8 +754,8 @@ int direct_dump(struct task *task, char *dump_path, int force_mode) leader_record_in(lr_end, &lr_in); lr = &lr_in; - strcpy(sname, lr->space_name); - strcpy(rname, lr->resource_name); + memcpy(sname, lr->space_name, NAME_ID_SIZE); + memcpy(rname, lr->resource_name, NAME_ID_SIZE); printf("%08llu %36s %48s %010llu %04llu %04llu %llu", (unsigned long long)(start_offset + (sector_nr * sector_size)), @@ -806,7 +806,7 @@ int direct_dump(struct task *task, char *dump_path, int force_mode) rindex_header_in(rh_end, &rh_in); rh = &rh_in; - strcpy(sname, rh->lockspace_name); + memcpy(sname, rh->lockspace_name, NAME_ID_SIZE); printf("%08llu %36s rindex_header 0x%x %d %u %llu\n", (unsigned long long)(start_offset + (sector_nr * sector_size)), -- To stop receiving notification emails like this one, please contact the administrator of this repository.

1 0

[sanlock] branch master updated: release 3.9.4
by pagure＠pagure.io 15 Aug '24

15 Aug '24

This is an automated email from the git hooks/post-receive script. teigland pushed a commit to branch master in repository sanlock. The following commit(s) were added to refs/heads/master by this push: new e077f1a release 3.9.4 e077f1a is described below commit e077f1a99a07c32ef460db89ed395b441f5e8b03 Author: David Teigland <teigland(a)redhat.com> AuthorDate: Fri Aug 9 10:03:28 2024 -0500 release 3.9.4 --- VERSION | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/VERSION b/VERSION index 820476a..e0d61b5 100644 --- a/VERSION +++ b/VERSION @@ -1 +1 @@ -3.9.3 +3.9.4 -- To stop receiving notification emails like this one, please contact the administrator of this repository.

1 0

[sanlock] 01/01: sanlock: fix invalid strcpy in direct dump
by pagure＠pagure.io 14 Aug '24

14 Aug '24

This is an automated email from the git hooks/post-receive script. teigland pushed a commit to branch main in repository sanlock. commit 5b34da78d33c2fca7dc3d2c49ae6d395d90c48c1 Author: David Teigland <teigland(a)redhat.com> AuthorDate: Wed Aug 14 11:51:17 2024 -0500 sanlock: fix invalid strcpy in direct dump If a lockspace or resource name is the full 48 characters in length, then the ondisk name field does not contain a terminating null character and strcpy fails. --- src/direct.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/src/direct.c b/src/direct.c index f70dc19..3bc7659 100644 --- a/src/direct.c +++ b/src/direct.c @@ -730,8 +730,8 @@ int direct_dump(struct task *task, char *dump_path, int force_mode) if (!lr->owner_id && !lr->owner_generation) continue; - strcpy(sname, lr->space_name); - strcpy(rname, lr->resource_name); + memcpy(sname, lr->space_name, NAME_ID_SIZE); + memcpy(rname, lr->resource_name, NAME_ID_SIZE); printf("%08llu %36s %48s %010llu %04llu %04llu", (unsigned long long)(start_offset + ((sector_nr + i) * sector_size)), @@ -754,8 +754,8 @@ int direct_dump(struct task *task, char *dump_path, int force_mode) leader_record_in(lr_end, &lr_in); lr = &lr_in; - strcpy(sname, lr->space_name); - strcpy(rname, lr->resource_name); + memcpy(sname, lr->space_name, NAME_ID_SIZE); + memcpy(rname, lr->resource_name, NAME_ID_SIZE); printf("%08llu %36s %48s %010llu %04llu %04llu %llu", (unsigned long long)(start_offset + (sector_nr * sector_size)), @@ -806,7 +806,7 @@ int direct_dump(struct task *task, char *dump_path, int force_mode) rindex_header_in(rh_end, &rh_in); rh = &rh_in; - strcpy(sname, rh->lockspace_name); + memcpy(sname, rh->lockspace_name, NAME_ID_SIZE); printf("%08llu %36s rindex_header 0x%x %d %u %llu\n", (unsigned long long)(start_offset + (sector_nr * sector_size)), -- To stop receiving notification emails like this one, please contact the administrator of this repository.

1 0

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

sanlock-devel August 2024