Merge branch 'slurm-24.05' into 24.05.ug-reduce-patch

Author: itkovian

src/api/step_io.c

@@ -39,25 +39,27 @@
 #include <string.h>
 #include <time.h>
 
+#include "src/api/step_io.h"
+#include "src/api/step_launch.h"
+
+#include "src/common/eio.h"
 #include "src/common/fd.h"
 #include "src/common/hostlist.h"
+#include "src/common/io_hdr.h"
 #include "src/common/log.h"
 #include "src/common/macros.h"
+#include "src/common/net.h"
 #include "src/common/pack.h"
 #include "src/common/read_config.h"
 #include "src/common/slurm_protocol_defs.h"
 #include "src/common/slurm_protocol_pack.h"
-#include "src/interfaces/cred.h"
+#include "src/common/write_labelled_message.h"
 #include "src/common/xassert.h"
 #include "src/common/xmalloc.h"
 #include "src/common/xsignal.h"
-#include "src/common/eio.h"
-#include "src/common/io_hdr.h"
-#include "src/common/net.h"
-#include "src/common/write_labelled_message.h"
+#include "src/common/xstring.h"
 
-#include "src/api/step_io.h"
-#include "src/api/step_launch.h"
+#include "src/interfaces/cred.h"
 
 #define STDIO_MAX_FREE_BUF 1024
 
@@ -1071,7 +1073,7 @@ _estimate_nports(int nclients, int cli_per_port)
 }
 
 client_io_t *client_io_handler_create(slurm_step_io_fds_t fds, int num_tasks,
-				      int num_nodes, slurm_cred_t *cred,
+				      int num_nodes, char *io_key,
 				      bool label, uint32_t het_job_offset,
 				      uint32_t het_job_task_offset)
 {
@@ -1090,7 +1092,7 @@ client_io_t *client_io_handler_create(slurm_step_io_fds_t fds, int num_tasks,
 	else
 		cio->taskid_width = 0;
 
-	cio->io_key = slurm_cred_get_signature(cred);
+	cio->io_key = xstrdup(io_key);
 
 	cio->eio = eio_handle_create(slurm_conf.eio_timeout);
 

src/api/step_io.h

@@ -27,17 +27,17 @@
 #ifndef _HAVE_STEP_IO_H
 #define _HAVE_STEP_IO_H
 
-#include <stdint.h>
 #include <pthread.h>
+#include <stdint.h>
 
 #include "slurm/slurm.h"
 
+#include "src/common/bitstring.h"
 #include "src/common/eio.h"
 #include "src/common/list.h"
-#include "src/common/bitstring.h"
 #include "src/common/slurm_step_layout.h"
-struct step_launch_state;
 
+struct step_launch_state;
 
 typedef struct {
 	/* input parameters - set (indirectly) by user */
@@ -96,16 +96,8 @@ typedef struct {
 				       I/O problem.  */
 } client_io_t;
 
-/*
- * IN cred - cred need not be a real job credential, it may be a "fake"
- *	credential generated with slurm_cred_faker().  The credential is
- *	sent to the slurmstepd (via the slurmd) which generates a signature
- *	string from the credential.  The slurmstepd sends the signature back
- *	back to the client when it establishes the IO connection as a sort
- *	of validity check.
- */
 client_io_t *client_io_handler_create(slurm_step_io_fds_t fds, int num_tasks,
-				      int num_nodes, slurm_cred_t *cred,
+				      int num_nodes, char *io_key,
 				      bool label, uint32_t het_job_offset,
 				      uint32_t het_job_task_offset);
 

src/api/step_launch.c

@@ -48,8 +48,8 @@
 #include <pthread.h>
 #include <signal.h>
 #include <stdarg.h>
-#include <stdlib.h>
 #include <stdio.h>
+#include <stdlib.h>
 #include <string.h>
 #include <sys/socket.h>
 #include <sys/stat.h>
@@ -59,6 +59,9 @@
 
 #include "slurm/slurm.h"
 
+#include "src/api/pmi_server.h"
+#include "src/api/step_launch.h"
+
 #include "src/common/cpu_frequency.h"
 #include "src/common/eio.h"
 #include "src/common/fd.h"
@@ -68,9 +71,6 @@
 #include "src/common/macros.h"
 #include "src/common/net.h"
 #include "src/common/read_config.h"
-#include "src/interfaces/auth.h"
-#include "src/interfaces/cred.h"
-#include "src/interfaces/mpi.h"
 #include "src/common/slurm_protocol_api.h"
 #include "src/common/slurm_protocol_defs.h"
 #include "src/common/slurm_time.h"
@@ -80,8 +80,9 @@
 #include "src/common/xmalloc.h"
 #include "src/common/xstring.h"
 
-#include "src/api/step_launch.h"
-#include "src/api/pmi_server.h"
+#include "src/interfaces/auth.h"
+#include "src/interfaces/cred.h"
+#include "src/interfaces/mpi.h"
 
 #include "src/srun/step_ctx.h"
 
@@ -201,6 +202,7 @@ extern int slurm_step_launch(slurm_step_ctx_t *ctx,
 	int rc = SLURM_SUCCESS;
 	bool preserve_env = params->preserve_env;
 	uint32_t mpi_plugin_id;
+	char *io_key = NULL;
 
 	debug("Entering %s", __func__);
 	memset(&launch, 0, sizeof(launch));
@@ -211,6 +213,8 @@ extern int slurm_step_launch(slurm_step_ctx_t *ctx,
 		return SLURM_ERROR;
 	}
 
+	io_key = slurm_cred_get_signature(ctx->step_resp->cred);
+
 	/* Initialize the callback pointers */
 	if (callbacks != NULL) {
 		/* copy the user specified callback pointers */
@@ -360,8 +364,7 @@ extern int slurm_step_launch(slurm_step_ctx_t *ctx,
 	ctx->launch_state->io =
 		client_io_handler_create(params->local_fds,
 					 ctx->step_req->num_tasks,
-					 launch.nnodes,
-					 ctx->step_resp->cred,
+					 launch.nnodes, io_key,
 					 params->labelio,
 					 params->het_job_offset,
 					 params->het_job_task_offset);
@@ -401,6 +404,7 @@ extern int slurm_step_launch(slurm_step_ctx_t *ctx,
 	xfree(launch.io_port);
 
 fail1:
+	xfree(io_key);
 	xfree(launch.complete_nodelist);
 	xfree(launch.cwd);
 	xfree(launch.stepmgr);
@@ -430,6 +434,7 @@ extern int slurm_step_launch_add(slurm_step_ctx_t *ctx,
 	uint16_t resp_port = 0;
 	bool preserve_env = params->preserve_env;
 	uint32_t mpi_plugin_id;
+	char *io_key = NULL;
 
 	debug("Entering %s", __func__);
 
@@ -439,6 +444,8 @@ extern int slurm_step_launch_add(slurm_step_ctx_t *ctx,
 		return SLURM_ERROR;
 	}
 
+	io_key = slurm_cred_get_signature(ctx->step_resp->cred);
+
 	mpi_plugin_id = mpi_g_client_init((char **)&params->mpi_plugin_name);
 	if (!mpi_plugin_id) {
 		slurm_seterrno(SLURM_MPI_PLUGIN_NAME_INVALID);
@@ -544,8 +551,7 @@ extern int slurm_step_launch_add(slurm_step_ctx_t *ctx,
 	ctx->launch_state->io =
 		client_io_handler_create(params->local_fds,
 					 ctx->step_req->num_tasks,
-					 launch.nnodes,
-					 ctx->step_resp->cred,
+					 launch.nnodes, io_key,
 					 params->labelio,
 					 params->het_job_offset,
 					 params->het_job_task_offset);
@@ -586,6 +592,7 @@ extern int slurm_step_launch_add(slurm_step_ctx_t *ctx,
 
 fail1:
 	/* clean up */
+	xfree(io_key);
 	xfree(launch.resp_port);
 	xfree(launch.io_port);
 

src/common/pack.c

@@ -214,10 +214,17 @@ void grow_buf(buf_t *buffer, uint32_t size)
 
 extern int try_grow_buf(buf_t *buffer, uint32_t size)
 {
-	uint64_t new_size = ((uint64_t) size) + buffer->size;
+	uint64_t new_size = buffer->size + BUF_SIZE;
 
 	xassert(buffer->magic == BUF_MAGIC);
 
+	/*
+	 * Force increase to always be at least BUF_SIZE to reduce number of
+	 * successive xrealloc()s that get called while packing larger RPCs
+	 */
+	if (size >= BUF_SIZE)
+		new_size += size;
+
 	if (buffer->mmaped || buffer->shadow)
 		return EINVAL;
 	if (new_size > MAX_BUF_SIZE) {

src/plugins/data_parser/v0.0.40/parsing.c

@@ -1268,7 +1268,12 @@ static int _dump_pointer(const parser_t *const parser, void *src, data_t *dst,
 	const parser_t *pt = find_parser_by_type(parser->pointer_type);
 	void **ptr = src;
 
-	if (!*ptr && !is_complex_mode(args)) {
+	if (!*ptr) {
+		if (is_complex_mode(args)) {
+			xassert(data_get_type(dst) == DATA_TYPE_NULL);
+			return SLURM_SUCCESS;
+		}
+
 		/* Fully resolve pointer on NULL to use correct model */
 		while (pt->pointer_type)
 			pt = find_parser_by_type(pt->pointer_type);

src/sattach/sattach.c

@@ -68,15 +68,13 @@ static void _mpir_cleanup(void);
 static void _mpir_dump_proctable(void);
 static void _pty_restore(void);
 static void print_layout_info(slurm_step_layout_t *layout);
-static slurm_cred_t *_generate_fake_cred(slurm_step_id_t stepid,
-					uid_t uid, char *nodelist,
-					uint32_t node_cnt);
+static char *_generate_io_key(void);
 static uint32_t _nodeid_from_layout(slurm_step_layout_t *layout,
 				    uint32_t taskid);
 static void _set_exit_code(void);
 static int _attach_to_tasks(slurm_step_id_t stepid,
 			    slurm_step_layout_t *layout,
-			    slurm_cred_t *fake_cred,
+			    char *io_key,
 			    uint16_t num_resp_ports,
 			    uint16_t *resp_ports,
 			    int num_io_ports,
@@ -118,11 +116,10 @@ int sattach(int argc, char **argv)
 {
 	log_options_t logopt = LOG_OPTS_STDERR_ONLY;
 	slurm_step_layout_t *layout;
-	slurm_cred_t *fake_cred;
 	message_thread_state_t *mts;
 	uint32_t jobid, stepid;
 	client_io_t *io;
-	char *hosts;
+	char *io_key = NULL;
 
 	slurm_init(NULL);
 	log_init(xbasename(argv[0]), logopt, 0, NULL);
@@ -169,16 +166,11 @@ int sattach(int argc, char **argv)
 			_nodeid_from_layout(layout, opt.fds.input.taskid);
 	}
 
-	if (layout->front_end)
-		hosts = layout->front_end;
-	else
-		hosts = layout->node_list;
-	fake_cred = _generate_fake_cred(opt.selected_step->step_id,
-					opt.uid, hosts, layout->node_cnt);
+	io_key = _generate_io_key();
 	mts = _msg_thr_create(layout->node_cnt, layout->task_cnt);
 
 	io = client_io_handler_create(opt.fds, layout->task_cnt,
-				      layout->node_cnt, fake_cred,
+				      layout->node_cnt, io_key,
 				      opt.labelio, NO_VAL, NO_VAL);
 	client_io_handler_start(io);
 
@@ -197,7 +189,7 @@ int sattach(int argc, char **argv)
 		xsignal_block(pty_sigarray);
 	}
 
-	_attach_to_tasks(opt.selected_step->step_id, layout, fake_cred,
+	_attach_to_tasks(opt.selected_step->step_id, layout, io_key,
 			 mts->num_resp_port, mts->resp_port,
 			 io->num_listen, io->listenport,
 			 mts->tasks_started);
@@ -213,6 +205,7 @@ int sattach(int argc, char **argv)
 	client_io_handler_finish(io);
 	client_io_handler_destroy(io);
 	_mpir_cleanup();
+	xfree(io_key);
 
 	return global_rc;
 }
@@ -278,44 +271,22 @@ static void print_layout_info(slurm_step_layout_t *layout)
 	hostlist_destroy(nl);
 }
 
-
-/* return a faked job credential */
-static slurm_cred_t *_generate_fake_cred(slurm_step_id_t stepid,
-					uid_t uid, char *nodelist,
-					uint32_t node_cnt)
+/*
+ * The io_key requires a modest amount of entropy to prevent someone guessing
+ * it, then racing to initiate a connection to the sattach command.
+ * By (ab)using the auth token generation mechanisms, the key should be
+ * sufficiently random for our purposes. (An attacker would need to request
+ * an auth key be generated at the same time by the same uid on the same host.)
+ */
+static char *_generate_io_key(void)
 {
-	slurm_cred_t *cred;
-	slurm_cred_arg_t *arg = xmalloc(sizeof(*arg));
-
-	arg->step_id.job_id = stepid.job_id;
-	arg->step_id.step_id = stepid.step_id;
-	arg->step_id.step_het_comp = stepid.step_het_comp;
-	arg->uid = uid;
-
-	arg->job_hostlist = nodelist;
-	arg->job_nhosts = node_cnt;
-
-	arg->step_hostlist = nodelist;
-
-	arg->job_core_bitmap = bit_alloc(node_cnt);
-	bit_set_all(arg->job_core_bitmap);
-	arg->step_core_bitmap = bit_alloc(node_cnt);
-	bit_set_all(arg->step_core_bitmap);
-
-	arg->cores_per_socket = xmalloc(sizeof(uint16_t));
-	arg->cores_per_socket[0] = 1;
-	arg->sockets_per_node = xmalloc(sizeof(uint16_t));
-	arg->sockets_per_node[0] = 1;
-	arg->sock_core_rep_count = xmalloc(sizeof(uint32_t));
-	arg->sock_core_rep_count[0] = node_cnt;
-
-	cred = slurm_cred_faker(arg);
-
-	/* Don't free, this memory will be free'd later */
-	arg->job_hostlist = NULL;
-	arg->step_hostlist = NULL;
-	slurm_cred_free_args(arg);
-	return cred;
+	char *key = auth_g_create(AUTH_DEFAULT_INDEX, slurm_conf.authinfo,
+				  0, NULL, 0);
+
+	if (!key)
+		fatal("failed to generate a suitable io_key");
+
+	return key;
 }
 
 void _handle_response_msg(slurm_msg_type_t msg_type, void *msg,
@@ -383,7 +354,7 @@ void _handle_response_msg_list(List other_nodes_resp, bitstr_t *tasks_started)
  */
 static int _attach_to_tasks(slurm_step_id_t stepid,
 			    slurm_step_layout_t *layout,
-			    slurm_cred_t *fake_cred,
+			    char *io_key,
 			    uint16_t num_resp_ports,
 			    uint16_t *resp_ports,
 			    int num_io_ports,
@@ -402,7 +373,7 @@ static int _attach_to_tasks(slurm_step_id_t stepid,
 	reattach_msg.num_resp_port = num_resp_ports;
 	reattach_msg.resp_port = resp_ports; /* array of response ports */
 	reattach_msg.num_io_port = num_io_ports;
-	reattach_msg.io_key = slurm_cred_get_signature(fake_cred);
+	reattach_msg.io_key = xstrdup(io_key);
 	reattach_msg.io_port = io_ports;
 
 	slurm_msg_set_r_uid(&msg, SLURM_AUTH_UID_ANY);