2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
48 #include "types/enums.h"
49 #include "types/hw_info.h"
50 #include "types/commrec.h"
51 #include "gmx_fatal.h"
52 #include "gmx_fatal_collective.h"
54 #include "gpu_utils.h"
56 #include "gmx_detect_hardware.h"
58 #include "md_logging.h"
59 #include "gromacs/utility/gmxomp.h"
61 #include "thread_mpi/threads.h"
63 #ifdef GMX_NATIVE_WINDOWS
68 const gmx_bool bGPUBinary = TRUE;
70 const gmx_bool bGPUBinary = FALSE;
73 static const char * invalid_gpuid_hint =
74 "A delimiter-free sequence of valid numeric IDs of available GPUs is expected.";
76 /* The globally shared hwinfo structure. */
77 static gmx_hw_info_t *hwinfo_g;
78 /* A reference counter for the hwinfo structure */
79 static int n_hwinfo = 0;
80 /* A lock to protect the hwinfo structure */
81 static tMPI_Thread_mutex_t hw_info_lock = TMPI_THREAD_MUTEX_INITIALIZER;
85 static void limit_num_gpus_used(gmx_gpu_opt_t *gpu_opt, int count);
86 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
87 const gmx_gpu_opt_t *gpu_opt);
89 static void sprint_gpus(char *sbuf, const gmx_gpu_info_t *gpu_info)
94 ndev = gpu_info->ncuda_dev;
97 for (i = 0; i < ndev; i++)
99 get_gpu_device_info_string(stmp, gpu_info, i);
109 static void print_gpu_detection_stats(FILE *fplog,
110 const gmx_gpu_info_t *gpu_info,
113 char onhost[266], stmp[STRLEN];
116 if (!gpu_info->bDetectGPUs)
118 /* We skipped the detection, so don't print detection stats */
122 ngpu = gpu_info->ncuda_dev;
124 #if defined GMX_MPI && !defined GMX_THREAD_MPI
125 /* We only print the detection on one, of possibly multiple, nodes */
126 strncpy(onhost, " on host ", 10);
127 gmx_gethostname(onhost+9, 256);
129 /* We detect all relevant GPUs */
130 strncpy(onhost, "", 1);
135 sprint_gpus(stmp, gpu_info);
136 md_print_warn(cr, fplog, "%d GPU%s detected%s:\n%s\n",
137 ngpu, (ngpu > 1) ? "s" : "", onhost, stmp);
141 md_print_warn(cr, fplog, "No GPUs detected%s\n", onhost);
145 static void print_gpu_use_stats(FILE *fplog,
146 const gmx_gpu_info_t *gpu_info,
147 const gmx_gpu_opt_t *gpu_opt,
150 char sbuf[STRLEN], stmp[STRLEN];
151 int i, ngpu_comp, ngpu_use;
153 ngpu_comp = gpu_info->ncuda_dev_compatible;
154 ngpu_use = gpu_opt->ncuda_dev_use;
156 /* Issue a note if GPUs are available but not used */
157 if (ngpu_comp > 0 && ngpu_use < 1)
160 "%d compatible GPU%s detected in the system, but none will be used.\n"
161 "Consider trying GPU acceleration with the Verlet scheme!",
162 ngpu_comp, (ngpu_comp > 1) ? "s" : "");
168 ngpu_use_uniq = gmx_count_gpu_dev_unique(gpu_info, gpu_opt);
170 sprintf(sbuf, "%d GPU%s %sselected for this run.\n"
171 "Mapping of GPU%s to the %d PP rank%s in this node: ",
172 ngpu_use_uniq, (ngpu_use_uniq > 1) ? "s" : "",
173 gpu_opt->bUserSet ? "user-" : "auto-",
174 (ngpu_use > 1) ? "s" : "",
175 cr->nrank_pp_intranode,
176 (cr->nrank_pp_intranode > 1) ? "s" : "");
178 for (i = 0; i < ngpu_use; i++)
180 sprintf(stmp, "#%d", get_gpu_device_id(gpu_info, gpu_opt, i));
181 if (i < ngpu_use - 1)
188 md_print_info(cr, fplog, "%s\n\n", sbuf);
191 /* Parse a "plain" GPU ID string which contains a sequence of digits corresponding
192 * to GPU IDs; the order will indicate the process/tMPI thread - GPU assignment. */
193 static void parse_gpu_id_plain_string(const char *idstr, int *nid, int **idlist)
197 *nid = strlen(idstr);
201 for (i = 0; i < *nid; i++)
203 if (idstr[i] < '0' || idstr[i] > '9')
205 gmx_fatal(FARGS, "Invalid character in GPU ID string: '%c'\n%s\n",
206 idstr[i], invalid_gpuid_hint);
208 (*idlist)[i] = idstr[i] - '0';
212 static void parse_gpu_id_csv_string(const char gmx_unused *idstr, int gmx_unused *nid, int gmx_unused *idlist)
214 /* XXX implement cvs format to support more than 10 different GPUs in a box. */
215 gmx_incons("Not implemented yet");
218 void gmx_check_hw_runconf_consistency(FILE *fplog,
219 const gmx_hw_info_t *hwinfo,
221 const gmx_hw_opt_t *hw_opt,
225 char sbuf[STRLEN], th_or_proc[STRLEN], th_or_proc_plural[STRLEN], pernode[STRLEN];
226 gmx_bool btMPI, bMPI, bMaxMpiThreadsSet, bNthreadsAuto, bEmulateGPU;
231 /* Below we only do consistency checks for PP and GPUs,
232 * this is irrelevant for PME only nodes, so in that case we return
235 if (!(cr->duty & DUTY_PP))
240 btMPI = bMPI = FALSE;
241 bNthreadsAuto = FALSE;
242 #if defined(GMX_THREAD_MPI)
244 bNthreadsAuto = (hw_opt->nthreads_tmpi < 1);
245 #elif defined(GMX_LIB_MPI)
249 /* GPU emulation detection is done later, but we need here as well
250 * -- uncool, but there's no elegant workaround */
251 bEmulateGPU = (getenv("GMX_EMULATE_GPU") != NULL);
252 bMaxMpiThreadsSet = (getenv("GMX_MAX_MPI_THREADS") != NULL);
254 /* check the acceleration mdrun is compiled with against hardware
256 /* TODO: Here we assume homogeneous hardware which is not necessarily
257 the case! Might not hurt to add an extra check over MPI. */
258 gmx_cpuid_acceleration_check(hwinfo->cpuid_info, fplog, SIMMASTER(cr));
260 /* NOTE: this print is only for and on one physical node */
261 print_gpu_detection_stats(fplog, &hwinfo->gpu_info, cr);
263 if (hwinfo->gpu_info.ncuda_dev_compatible > 0)
265 /* NOTE: this print is only for and on one physical node */
266 print_gpu_use_stats(fplog, &hwinfo->gpu_info, &hw_opt->gpu_opt, cr);
269 /* Need to ensure that we have enough GPUs:
270 * - need one GPU per PP node
271 * - no GPU oversubscription with tMPI
273 /* number of PP processes per node */
274 npppn = cr->nrank_pp_intranode;
277 th_or_proc_plural[0] = '\0';
280 sprintf(th_or_proc, "thread-MPI thread");
283 sprintf(th_or_proc_plural, "s");
288 sprintf(th_or_proc, "MPI process");
291 sprintf(th_or_proc_plural, "es");
293 sprintf(pernode, " per node");
297 /* neither MPI nor tMPI */
298 sprintf(th_or_proc, "process");
301 if (bUseGPU && hwinfo->gpu_info.ncuda_dev_compatible > 0 &&
304 int ngpu_comp, ngpu_use;
305 char gpu_comp_plural[2], gpu_use_plural[2];
307 ngpu_comp = hwinfo->gpu_info.ncuda_dev_compatible;
308 ngpu_use = hw_opt->gpu_opt.ncuda_dev_use;
310 sprintf(gpu_comp_plural, "%s", (ngpu_comp > 1) ? "s" : "");
311 sprintf(gpu_use_plural, "%s", (ngpu_use > 1) ? "s" : "");
313 /* number of tMPI threads auto-adjusted */
314 if (btMPI && bNthreadsAuto)
316 if (hw_opt->gpu_opt.bUserSet && npppn < ngpu_use)
318 /* The user manually provided more GPUs than threads we
319 could automatically start. */
321 "%d GPU%s provided, but only %d PP thread-MPI thread%s coud be started.\n"
322 "%s requires one PP tread-MPI thread per GPU; use fewer GPUs%s.",
323 ngpu_use, gpu_use_plural,
324 npppn, th_or_proc_plural,
325 ShortProgram(), bMaxMpiThreadsSet ? "\nor allow more threads to be used" : "");
328 if (!hw_opt->gpu_opt.bUserSet && npppn < ngpu_comp)
330 /* There are more GPUs than tMPI threads; we have
331 limited the number GPUs used. */
332 md_print_warn(cr, fplog,
333 "NOTE: %d GPU%s were detected, but only %d PP thread-MPI thread%s can be started.\n"
334 " %s can use one GPU per PP tread-MPI thread, so only %d GPU%s will be used.%s\n",
335 ngpu_comp, gpu_comp_plural,
336 npppn, th_or_proc_plural,
337 ShortProgram(), npppn,
338 npppn > 1 ? "s" : "",
339 bMaxMpiThreadsSet ? "\n Also, you can allow more threads to be used by increasing GMX_MAX_MPI_THREADS" : "");
343 if (hw_opt->gpu_opt.bUserSet)
345 if (ngpu_use != npppn)
348 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
349 "%s was started with %d PP %s%s%s, but you provided %d GPU%s.",
350 th_or_proc, btMPI ? "s" : "es", pernode,
351 ShortProgram(), npppn, th_or_proc,
352 th_or_proc_plural, pernode,
353 ngpu_use, gpu_use_plural);
358 if (ngpu_comp > npppn)
360 md_print_warn(cr, fplog,
361 "NOTE: potentially sub-optimal launch configuration, %s started with less\n"
362 " PP %s%s%s than GPU%s available.\n"
363 " Each PP %s can use only one GPU, %d GPU%s%s will be used.\n",
364 ShortProgram(), th_or_proc,
365 th_or_proc_plural, pernode, gpu_comp_plural,
366 th_or_proc, npppn, gpu_use_plural, pernode);
369 if (ngpu_use != npppn)
371 /* Avoid duplicate error messages.
372 * Unfortunately we can only do this at the physical node
373 * level, since the hardware setup and MPI process count
374 * might differ between physical nodes.
376 if (cr->rank_pp_intranode == 0)
379 "Incorrect launch configuration: mismatching number of PP %s%s and GPUs%s.\n"
380 "%s was started with %d PP %s%s%s, but only %d GPU%s were detected.",
381 th_or_proc, btMPI ? "s" : "es", pernode,
382 ShortProgram(), npppn, th_or_proc,
383 th_or_proc_plural, pernode,
384 ngpu_use, gpu_use_plural);
392 same_count = gmx_count_gpu_dev_shared(&hw_opt->gpu_opt);
396 md_print_info(cr, fplog,
397 "NOTE: You assigned %s to multiple %s%s.\n",
398 same_count > 1 ? "GPUs" : "a GPU", th_or_proc, btMPI ? "s" : "es");
406 /* Avoid other ranks to continue after
408 MPI_Barrier(cr->mpi_comm_mygroup);
414 /* Return 0 if none of the GPU (per node) are shared among PP ranks.
416 * Sharing GPUs among multiple PP ranks is possible when the user passes
417 * GPU IDs. Here we check for sharing and return a non-zero value when
418 * this is detected. Note that the return value represents the number of
419 * PP rank pairs that share a device.
421 int gmx_count_gpu_dev_shared(const gmx_gpu_opt_t *gpu_opt)
424 int ngpu = gpu_opt->ncuda_dev_use;
426 if (gpu_opt->bUserSet)
430 for (i = 0; i < ngpu - 1; i++)
432 for (j = i + 1; j < ngpu; j++)
434 same_count += (gpu_opt->cuda_dev_use[i] ==
435 gpu_opt->cuda_dev_use[j]);
443 /* Count and return the number of unique GPUs (per node) selected.
445 * As sharing GPUs among multiple PP ranks is possible when the user passes
446 * GPU IDs, the number of GPUs user (per node) can be different from the
447 * number of GPU IDs selected.
449 static int gmx_count_gpu_dev_unique(const gmx_gpu_info_t *gpu_info,
450 const gmx_gpu_opt_t *gpu_opt)
452 int i, uniq_count, ngpu;
458 ngpu = gpu_info->ncuda_dev;
461 snew(uniq_ids, ngpu);
463 /* Each element in uniq_ids will be set to 0 or 1. The n-th element set
464 * to 1 indicates that the respective GPU was selected to be used. */
465 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
467 uniq_ids[get_gpu_device_id(gpu_info, gpu_opt, i)] = 1;
469 /* Count the devices used. */
470 for (i = 0; i < ngpu; i++)
472 uniq_count += uniq_ids[i];
481 /* Return the number of hardware threads supported by the current CPU.
482 * We assume that this is equal with the number of CPUs reported to be
483 * online by the OS at the time of the call.
485 static int get_nthreads_hw_avail(FILE gmx_unused *fplog, const t_commrec gmx_unused *cr)
489 #if ((defined(WIN32) || defined( _WIN32 ) || defined(WIN64) || defined( _WIN64 )) && !(defined (__CYGWIN__) || defined (__CYGWIN32__)))
492 GetSystemInfo( &sysinfo );
493 ret = sysinfo.dwNumberOfProcessors;
494 #elif defined HAVE_SYSCONF
495 /* We are probably on Unix.
496 * Now check if we have the argument to use before executing the call
498 #if defined(_SC_NPROCESSORS_ONLN)
499 ret = sysconf(_SC_NPROCESSORS_ONLN);
500 #elif defined(_SC_NPROC_ONLN)
501 ret = sysconf(_SC_NPROC_ONLN);
502 #elif defined(_SC_NPROCESSORS_CONF)
503 ret = sysconf(_SC_NPROCESSORS_CONF);
504 #elif defined(_SC_NPROC_CONF)
505 ret = sysconf(_SC_NPROC_CONF);
507 #warning "No valid sysconf argument value found. Executables will not be able to determine the number of hardware threads: mdrun will use 1 thread by default!"
508 #endif /* End of check for sysconf argument values */
511 /* Neither windows nor Unix. No fscking idea how many CPUs we have! */
517 fprintf(debug, "Detected %d processors, will use this as the number "
518 "of supported hardware threads.\n", ret);
522 if (ret != gmx_omp_get_num_procs())
524 md_print_warn(cr, fplog,
525 "Number of CPUs detected (%d) does not match the number reported by OpenMP (%d).\n"
526 "Consider setting the launch configuration manually!",
527 ret, gmx_omp_get_num_procs());
534 static void gmx_detect_gpus(FILE *fplog, const t_commrec *cr)
538 MPI_Comm physicalnode_comm;
542 /* Under certain circumstances MPI ranks on the same physical node
543 * can not simultaneously access the same GPU(s). Therefore we run
544 * the detection only on one MPI rank per node and broadcast the info.
545 * Note that with thread-MPI only a single thread runs this code.
547 * TODO: We should also do CPU hardware detection only once on each
548 * physical node and broadcast it, instead of do it on every MPI rank.
551 /* A split of MPI_COMM_WORLD over physical nodes is only required here,
552 * so we create and destroy it locally.
554 MPI_Comm_rank(MPI_COMM_WORLD, &rank_world);
555 MPI_Comm_split(MPI_COMM_WORLD, gmx_physicalnode_id_hash(),
556 rank_world, &physicalnode_comm);
557 MPI_Comm_rank(physicalnode_comm, &rank_local);
559 /* Here there should be only one process, check this */
560 assert(cr->nnodes == 1 && cr->sim_nodeid == 0);
567 char detection_error[STRLEN] = "", sbuf[STRLEN];
569 if (detect_cuda_gpus(&hwinfo_g->gpu_info, detection_error) != 0)
571 if (detection_error != NULL && detection_error[0] != '\0')
573 sprintf(sbuf, ":\n %s\n", detection_error);
579 md_print_warn(cr, fplog,
580 "NOTE: Error occurred during GPU detection%s"
581 " Can not use GPU acceleration, will fall back to CPU kernels.\n",
587 /* Broadcast the GPU info to the other ranks within this node */
588 MPI_Bcast(&hwinfo_g->gpu_info.ncuda_dev, 1, MPI_INT, 0, physicalnode_comm);
590 if (hwinfo_g->gpu_info.ncuda_dev > 0)
594 cuda_dev_size = hwinfo_g->gpu_info.ncuda_dev*sizeof_cuda_dev_info();
598 hwinfo_g->gpu_info.cuda_dev =
599 (cuda_dev_info_ptr_t)malloc(cuda_dev_size);
601 MPI_Bcast(hwinfo_g->gpu_info.cuda_dev, cuda_dev_size, MPI_BYTE,
602 0, physicalnode_comm);
603 MPI_Bcast(&hwinfo_g->gpu_info.ncuda_dev_compatible, 1, MPI_INT,
604 0, physicalnode_comm);
607 MPI_Comm_free(&physicalnode_comm);
611 gmx_hw_info_t *gmx_detect_hardware(FILE *fplog, const t_commrec *cr,
612 gmx_bool bDetectGPUs)
617 /* make sure no one else is doing the same thing */
618 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
621 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
624 /* only initialize the hwinfo structure if it is not already initalized */
629 /* detect CPUID info; no fuss, we don't detect system-wide
630 * -- sloppy, but that's it for now */
631 if (gmx_cpuid_init(&hwinfo_g->cpuid_info) != 0)
633 gmx_fatal_collective(FARGS, cr, NULL, "CPUID detection failed!");
636 /* detect number of hardware threads */
637 hwinfo_g->nthreads_hw_avail = get_nthreads_hw_avail(fplog, cr);
640 hwinfo_g->gpu_info.ncuda_dev = 0;
641 hwinfo_g->gpu_info.cuda_dev = NULL;
642 hwinfo_g->gpu_info.ncuda_dev_compatible = 0;
644 /* Run the detection if the binary was compiled with GPU support
645 * and we requested detection.
647 hwinfo_g->gpu_info.bDetectGPUs =
648 (bGPUBinary && bDetectGPUs &&
649 getenv("GMX_DISABLE_GPU_DETECTION") == NULL);
650 if (hwinfo_g->gpu_info.bDetectGPUs)
652 gmx_detect_gpus(fplog, cr);
655 /* increase the reference counter */
658 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
661 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));
667 void gmx_parse_gpu_ids(gmx_gpu_opt_t *gpu_opt)
671 if (gpu_opt->gpu_id != NULL && !bGPUBinary)
673 gmx_fatal(FARGS, "GPU ID string set, but %s was compiled without GPU support!", ShortProgram());
676 env = getenv("GMX_GPU_ID");
677 if (env != NULL && gpu_opt->gpu_id != NULL)
679 gmx_fatal(FARGS, "GMX_GPU_ID and -gpu_id can not be used at the same time");
683 env = gpu_opt->gpu_id;
686 /* parse GPU IDs if the user passed any */
689 parse_gpu_id_plain_string(env,
690 &gpu_opt->ncuda_dev_use,
691 &gpu_opt->cuda_dev_use);
693 if (gpu_opt->ncuda_dev_use == 0)
695 gmx_fatal(FARGS, "Empty GPU ID string encountered.\n%s\n",
699 gpu_opt->bUserSet = TRUE;
703 void gmx_select_gpu_ids(FILE *fplog, const t_commrec *cr,
704 const gmx_gpu_info_t *gpu_info,
705 gmx_bool bForceUseGPU,
706 gmx_gpu_opt_t *gpu_opt)
710 char sbuf[STRLEN], stmp[STRLEN];
712 /* Bail if binary is not compiled with GPU acceleration, but this is either
713 * explicitly (-nb gpu) or implicitly (gpu ID passed) requested. */
714 if (bForceUseGPU && !bGPUBinary)
716 gmx_fatal(FARGS, "GPU acceleration requested, but %s was compiled without GPU support!", ShortProgram());
719 if (gpu_opt->bUserSet)
721 /* Check the GPU IDs passed by the user.
722 * (GPU IDs have been parsed by gmx_parse_gpu_ids before)
727 snew(checkres, gpu_opt->ncuda_dev_use);
729 res = check_selected_cuda_gpus(checkres, gpu_info, gpu_opt);
733 print_gpu_detection_stats(fplog, gpu_info, cr);
735 sprintf(sbuf, "Some of the requested GPUs do not exist, behave strangely, or are not compatible:\n");
736 for (i = 0; i < gpu_opt->ncuda_dev_use; i++)
738 if (checkres[i] != egpuCompatible)
740 sprintf(stmp, " GPU #%d: %s\n",
741 gpu_opt->cuda_dev_use[i],
742 gpu_detect_res_str[checkres[i]]);
746 gmx_fatal(FARGS, "%s", sbuf);
753 pick_compatible_gpus(&hwinfo_g->gpu_info, gpu_opt);
755 if (gpu_opt->ncuda_dev_use > cr->nrank_pp_intranode)
757 /* We picked more GPUs than we can use: limit the number.
758 * We print detailed messages about this later in
759 * gmx_check_hw_runconf_consistency.
761 limit_num_gpus_used(gpu_opt, cr->nrank_pp_intranode);
764 gpu_opt->bUserSet = FALSE;
767 /* If the user asked for a GPU, check whether we have a GPU */
768 if (bForceUseGPU && gpu_info->ncuda_dev_compatible == 0)
770 gmx_fatal(FARGS, "GPU acceleration requested, but no compatible GPUs were detected.");
774 static void limit_num_gpus_used(gmx_gpu_opt_t *gpu_opt, int count)
780 ndev_use = gpu_opt->ncuda_dev_use;
782 if (count > ndev_use)
784 /* won't increase the # of GPUs */
791 sprintf(sbuf, "Limiting the number of GPUs to <1 doesn't make sense (detected %d, %d requested)!",
796 /* TODO: improve this implementation: either sort GPUs or remove the weakest here */
797 gpu_opt->ncuda_dev_use = count;
800 void gmx_hardware_info_free(gmx_hw_info_t *hwinfo)
804 ret = tMPI_Thread_mutex_lock(&hw_info_lock);
807 gmx_fatal(FARGS, "Error locking hwinfo mutex: %s", strerror(errno));
810 /* decrease the reference counter */
814 if (hwinfo != hwinfo_g)
816 gmx_incons("hwinfo < hwinfo_g");
821 gmx_incons("n_hwinfo < 0");
826 gmx_cpuid_done(hwinfo_g->cpuid_info);
827 free_gpu_info(&hwinfo_g->gpu_info);
831 ret = tMPI_Thread_mutex_unlock(&hw_info_lock);
834 gmx_fatal(FARGS, "Error unlocking hwinfo mutex: %s", strerror(errno));