Simplified uniform GPU selection in CMake

[gromacs.git] / src / gromacs / fileio / gmxfio.cpp

/*
 * This file is part of the GROMACS molecular simulation package.
 *
 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
 * Copyright (c) 2001-2004, The GROMACS development team.
 * Copyright (c) 2013,2014,2015,2016,2017 by the GROMACS development team.
 * Copyright (c) 2018,2019,2020, by the GROMACS development team, led by
 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
 * and including many others, as listed in the AUTHORS file in the
 * top-level source directory and at http://www.gromacs.org.
 *
 * GROMACS is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public License
 * as published by the Free Software Foundation; either version 2.1
 * of the License, or (at your option) any later version.
 *
 * GROMACS is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with GROMACS; if not, see
 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
 *
 * If you want to redistribute modifications to GROMACS, please
 * consider that scientific software is very special. Version
 * control is crucial - bugs must be traceable. We will be happy to
 * consider code for inclusion in the official distribution, but
 * derived work must not be called official GROMACS. Details are found
 * in the README & COPYING files - if they are missing, get the
 * official version at http://www.gromacs.org.
 *
 * To help us fund GROMACS development, we humbly ask that you cite
 * the research papers on the package. Check out http://www.gromacs.org.
 */
#include "gmxpre.h"

#include "gmxfio.h"

#include "config.h"

#include <cerrno>
#include <cstdio>
#include <cstring>

#include <vector>

#if HAVE_IO_H
#    include <io.h>
#endif
#ifdef HAVE_UNISTD_H
#    include <unistd.h>
#endif

#include "thread_mpi/threads.h"

#include "gromacs/fileio/filetypes.h"
#include "gromacs/fileio/md5.h"
#include "gromacs/utility/fatalerror.h"
#include "gromacs/utility/futil.h"
#include "gromacs/utility/mutex.h"
#include "gromacs/utility/smalloc.h"

#include "gmxfio_impl.h"

/* This is the new improved and thread safe version of gmxfio. */


/* the list of open files is a linked list, with a dummy element at its head;
       it is initialized when the first file is opened. */
static t_fileio* open_files = nullptr;


/* this mutex locks the open_files structure so that no two threads can
   modify it.

   For now, we use this as a coarse grained lock on all file
   insertion/deletion operations because it makes avoiding deadlocks
   easier, and adds almost no overhead: the only overhead is during
   opening and closing of files, or during global operations like
   iterating along all open files. All these cases should be rare
   during the simulation. */
static gmx::Mutex open_file_mutex;

using Lock = gmx::lock_guard<gmx::Mutex>;

/******************************************************************
 *
 * Internal functions:
 *
 ******************************************************************/

static int gmx_fio_int_flush(t_fileio* fio)
{
    int rc = 0;

    if (fio->fp)
    {
        rc = fflush(fio->fp);
    }

    return rc;
}

/* lock the mutex associated with this fio. This needs to be done for every
   type of access to the fio's elements. */
void gmx_fio_lock(t_fileio* fio)
{
    tMPI_Lock_lock(&(fio->mtx));
}
/* unlock the mutex associated with this fio.  */
void gmx_fio_unlock(t_fileio* fio)
{
    tMPI_Lock_unlock(&(fio->mtx));
}

/* make a dummy head element, assuming we locked everything. */
static void gmx_fio_make_dummy()
{
    if (!open_files)
    {
        open_files       = new t_fileio{};
        open_files->fp   = nullptr;
        open_files->fn   = nullptr;
        open_files->next = open_files;
        open_files->prev = open_files;
        tMPI_Lock_init(&(open_files->mtx));
    }
}


/***********************************************************************
 *
 * FILE LIST OPERATIONS
 *
 ***********************************************************************/


/* insert a new t_fileio into the list */
static void gmx_fio_insert(t_fileio* fio)
{
    t_fileio* prev;
    Lock      openFilesLock(open_file_mutex);
    gmx_fio_make_dummy();

    /* and lock the fio we got and the list's head **/
    gmx_fio_lock(fio);
    gmx_fio_lock(open_files);
    prev = open_files->prev;
    /* lock the element after the current one */
    if (prev != open_files)
    {
        gmx_fio_lock(prev);
    }

    /* now do the actual insertion: */
    fio->next        = open_files;
    open_files->prev = fio;
    prev->next       = fio;
    fio->prev        = prev;

    /* now unlock all our locks */
    if (prev != open_files)
    {
        gmx_fio_unlock(prev);
    }
    gmx_fio_unlock(open_files);
    gmx_fio_unlock(fio);
}

/* remove a t_fileio into the list. We assume the fio is locked, and we leave
   it locked.
   NOTE: We also assume that the open_file_mutex has been locked */
static void gmx_fio_remove(t_fileio* fio)
{
    /* lock prev, because we're changing it */
    gmx_fio_lock(fio->prev);

    /* now set the prev's pointer */
    fio->prev->next = fio->next;
    gmx_fio_unlock(fio->prev);

    /* with the next ptr, we can simply lock while the original was locked */
    gmx_fio_lock(fio->next);
    fio->next->prev = fio->prev;
    gmx_fio_unlock(fio->next);

    /* and make sure we point nowhere in particular */
    fio->next = fio->prev = fio;
}


/* get the first open file, or NULL if there is none.
   Returns a locked fio. Assumes open_files_mutex is locked. */
static t_fileio* gmx_fio_get_first()
{
    t_fileio* ret;

    gmx_fio_make_dummy();

    gmx_fio_lock(open_files);
    ret = open_files->next;


    /* check whether there were any to begin with */
    if (ret == open_files)
    {
        /* after this, the open_file pointer should never change */
        ret = nullptr;
    }
    else
    {
        gmx_fio_lock(open_files->next);
    }
    gmx_fio_unlock(open_files);


    return ret;
}

/* get the next open file, or NULL if there is none.
   Unlocks the previous fio and locks the next one.
   Assumes open_file_mutex is locked. */
static t_fileio* gmx_fio_get_next(t_fileio* fio)
{
    t_fileio* ret;

    ret = fio->next;
    /* check if that was the last one */
    if (fio->next == open_files)
    {
        ret = nullptr;
    }
    else
    {
        gmx_fio_lock(ret);
    }
    gmx_fio_unlock(fio);

    return ret;
}

/* Stop looping through the open_files. Assumes open_file_mutex is locked. */
static void gmx_fio_stop_getting_next(t_fileio* fio)
{
    gmx_fio_unlock(fio);
}


/*****************************************************************
 *
 *                     EXPORTED SECTION
 *
 *****************************************************************/
t_fileio* gmx_fio_open(const char* fn, const char* mode)
{
    t_fileio* fio = nullptr;
    char      newmode[5];
    gmx_bool  bRead, bReadWrite;

    /* sanitize the mode string */
    if (std::strncmp(mode, "r+", 2) == 0)
    {
        std::strcpy(newmode, "r+");
    }
    else if (mode[0] == 'r')
    {
        std::strcpy(newmode, "r");
    }
    else if (strncmp(mode, "w+", 2) == 0)
    {
        std::strcpy(newmode, "w+");
    }
    else if (mode[0] == 'w')
    {
        std::strcpy(newmode, "w");
    }
    else if (strncmp(mode, "a+", 2) == 0)
    {
        std::strcpy(newmode, "a+");
    }
    else if (mode[0] == 'a')
    {
        std::strcpy(newmode, "a");
    }
    else
    {
        gmx_fatal(FARGS, "DEATH HORROR in gmx_fio_open, mode is '%s'", mode);
    }

    /* Check if it should be opened as a binary file */
    if (!ftp_is_text(fn2ftp(fn)))
    {
        strcat(newmode, "b");
    }

    fio = new t_fileio{};
    tMPI_Lock_init(&(fio->mtx));
    bRead      = (newmode[0] == 'r' && newmode[1] != '+');
    bReadWrite = (newmode[1] == '+');
    fio->fp    = nullptr;
    fio->xdr   = nullptr;
    if (fn)
    {
        if (fn2ftp(fn) == efTNG)
        {
            gmx_incons("gmx_fio_open may not be used to open TNG files");
        }
        fio->iFTP = fn2ftp(fn);
        fio->fn   = gmx_strdup(fn);

        fio->fp = gmx_ffopen(fn, newmode);
        /* If this file type is in the list of XDR files, open it like that */
        if (ftp_is_xdr(fio->iFTP))
        {
            /* determine the XDR direction */
            if (newmode[0] == 'w' || newmode[0] == 'a')
            {
                fio->xdrmode = XDR_ENCODE;
            }
            else
            {
                fio->xdrmode = XDR_DECODE;
            }
            snew(fio->xdr, 1);
            xdrstdio_create(fio->xdr, fio->fp, fio->xdrmode);
        }

        /* for appending seek to end of file to make sure ftell gives correct position
         * important for checkpointing */
        if (newmode[0] == 'a')
        {
            gmx_fseek(fio->fp, 0, SEEK_END);
        }
    }
    else
    {
        gmx_fatal(FARGS, "Cannot open file with NULL filename string");
    }

    fio->bRead      = bRead;
    fio->bReadWrite = bReadWrite;
    fio->bDouble    = (sizeof(real) == sizeof(double));

    /* and now insert this file into the list of open files. */
    gmx_fio_insert(fio);
    return fio;
}

static int gmx_fio_close_locked(t_fileio* fio)
{
    int rc = 0;

    if (fio->xdr != nullptr)
    {
        xdr_destroy(fio->xdr);
        sfree(fio->xdr);
    }

    if (fio->fp != nullptr)
    {
        rc = gmx_ffclose(fio->fp); /* fclose returns 0 if happy */
    }

    return rc;
}

int gmx_fio_close(t_fileio* fio)
{
    int rc = 0;

    Lock openFilesLock(open_file_mutex);

    gmx_fio_lock(fio);
    /* first remove it from the list */
    gmx_fio_remove(fio);
    rc = gmx_fio_close_locked(fio);
    gmx_fio_unlock(fio);

    sfree(fio->fn);
    delete fio;

    return rc;
}

/* close only fp but keep FIO entry. */
int gmx_fio_fp_close(t_fileio* fio)
{
    int rc = 0;
    gmx_fio_lock(fio);
    if (fio->xdr == nullptr)
    {
        rc      = gmx_ffclose(fio->fp); /* fclose returns 0 if happy */
        fio->fp = nullptr;
    }
    gmx_fio_unlock(fio);

    return rc;
}

FILE* gmx_fio_fopen(const char* fn, const char* mode)
{
    FILE*     ret;
    t_fileio* fio;

    fio = gmx_fio_open(fn, mode);
    gmx_fio_lock(fio);
    ret = fio->fp;
    gmx_fio_unlock(fio);

    return ret;
}

int gmx_fio_fclose(FILE* fp)
{
    t_fileio* cur;
    int       rc = -1;

    Lock openFilesLock(open_file_mutex);
    cur = gmx_fio_get_first();
    while (cur)
    {
        if (cur->fp == fp)
        {
            rc = gmx_fio_close_locked(cur);
            gmx_fio_remove(cur);
            gmx_fio_stop_getting_next(cur);
            sfree(cur->fn);
            delete cur;
            break;
        }
        cur = gmx_fio_get_next(cur);
    }

    return rc;
}

//! Helper struct for returning the MD5 checksum and the amount of the file that contributed to it.
struct MD5Checksum
{
    //! Checksum md5 digest.
    std::array<unsigned char, 16> checksum;
    //! The length of the file that contributed to the digest.
    gmx_off_t readLength;
};

/*! \brief Internal variant of get_file_md5 that operates on a locked
 * file.
 *
 * \return -1 any time a checksum cannot be computed, otherwise the
 *            length of the data from which the checksum was computed. */
static int gmx_fio_int_get_file_md5(t_fileio* fio, gmx_off_t offset, std::array<unsigned char, 16>* checksum)
{
    /*1MB: large size important to catch almost identical files */
    constexpr size_t maximumChecksumInputSize = 1048576;
    md5_state_t      state;
    gmx_off_t        readLength;
    gmx_off_t        seekOffset;

    seekOffset = offset - maximumChecksumInputSize;
    if (seekOffset < 0)
    {
        seekOffset = 0;
    }
    readLength = offset - seekOffset;

    if (!fio->fp)
    {
        // It's not an error if the file isn't open.
        return -1;
    }
    if (!fio->bReadWrite)
    {
        // It's not an error if the file is open in the wrong mode.
        //
        // TODO It is unclear why this check exists. The bReadWrite
        // flag is true when the file-opening mode included "+" but we
        // only need read and seek to be able to compute the
        // md5sum. Other requirements (e.g. that we can truncate when
        // doing an appending restart) should be expressed in a
        // different way, but it is unclear whether that is part of
        // the logic here.
        return -1;
    }

    if (gmx_fseek(fio->fp, seekOffset, SEEK_SET))
    {
        // It's not an error if file seeking fails. (But it could be
        // an issue when moving a checkpoint from one platform to
        // another, when they differ in their support for seeking, and
        // so can't agree on a checksum for appending).
        gmx_fseek(fio->fp, 0, SEEK_END);
        return -1;
    }

    std::vector<unsigned char> buf(maximumChecksumInputSize);
    // The fread puts the file position back to offset.
    if (static_cast<gmx_off_t>(fread(buf.data(), 1, readLength, fio->fp)) != readLength)
    {
        // Read an unexpected length. This is not a fatal error; the
        // md5sum check to prevent overwriting files is not vital.
        if (ferror(fio->fp))
        {
            fprintf(stderr, "\nTrying to get md5sum: %s: %s\n", fio->fn, strerror(errno));
        }
        else if (!feof(fio->fp))
        {
            fprintf(stderr, "\nTrying to get md5sum: Unknown reason for short read: %s\n", fio->fn);
        }

        gmx_fseek(fio->fp, 0, SEEK_END);
        return -1;
    }
    // Return the file position to the end of the file.
    gmx_fseek(fio->fp, 0, SEEK_END);

    if (debug)
    {
        fprintf(debug, "chksum %s readlen %ld\n", fio->fn, static_cast<long int>(readLength));
    }

    gmx_md5_init(&state);
    gmx_md5_append(&state, buf.data(), readLength);
    *checksum = gmx_md5_finish(&state);
    return readLength;
}


/*
 * fio: file to compute md5 for
 * offset: starting pointer of region to use for md5
 * digest: return array of md5 sum
 */
int gmx_fio_get_file_md5(t_fileio* fio, gmx_off_t offset, std::array<unsigned char, 16>* checksum)
{
    int ret;

    gmx_fio_lock(fio);
    ret = gmx_fio_int_get_file_md5(fio, offset, checksum);
    gmx_fio_unlock(fio);

    return ret;
}

/* The fio_mutex should ALWAYS be locked when this function is called */
static int gmx_fio_int_get_file_position(t_fileio* fio, gmx_off_t* offset)
{
    /* Flush the file, so we are sure it is written */
    if (gmx_fio_int_flush(fio))
    {
        char buf[STRLEN];
        sprintf(buf, "Cannot write file '%s'; maybe you are out of disk space?", fio->fn);
        gmx_file(buf);
    }

    /* We cannot count on XDR being able to write 64-bit integers,
       so separate into high/low 32-bit values.
       In case the filesystem has 128-bit offsets we only care
       about the first 64 bits - we'll have to fix
       this when exabyte-size output files are common...
     */
    *offset = gmx_ftell(fio->fp);

    return 0;
}

std::vector<gmx_file_position_t> gmx_fio_get_output_file_positions()
{
    std::vector<gmx_file_position_t> outputfiles;
    t_fileio*                        cur;

    Lock openFilesLock(open_file_mutex);
    cur = gmx_fio_get_first();
    while (cur)
    {
        /* Skip the checkpoint files themselves, since they could be open when
           we call this routine... */
        if (!cur->bRead && cur->iFTP != efCPT)
        {
            outputfiles.emplace_back();

            std::strncpy(outputfiles.back().filename, cur->fn, STRLEN - 1);

            /* Get the file position */
            gmx_fio_int_get_file_position(cur, &outputfiles.back().offset);
            if (!GMX_FAHCORE)
            {
                outputfiles.back().checksumSize = gmx_fio_int_get_file_md5(
                        cur, outputfiles.back().offset, &outputfiles.back().checksum);
            }
        }

        cur = gmx_fio_get_next(cur);
    }

    return outputfiles;
}


char* gmx_fio_getname(t_fileio* fio)
{
    char* ret;
    gmx_fio_lock(fio);
    ret = fio->fn;
    gmx_fio_unlock(fio);

    return ret;
}

int gmx_fio_getftp(t_fileio* fio)
{
    int ret;

    gmx_fio_lock(fio);
    ret = fio->iFTP;
    gmx_fio_unlock(fio);

    return ret;
}

void gmx_fio_rewind(t_fileio* fio)
{
    gmx_fio_lock(fio);

    if (fio->xdr)
    {
        xdr_destroy(fio->xdr);
        frewind(fio->fp);
        xdrstdio_create(fio->xdr, fio->fp, fio->xdrmode);
    }
    else
    {
        frewind(fio->fp);
    }
    gmx_fio_unlock(fio);
}


int gmx_fio_flush(t_fileio* fio)
{
    int ret;

    gmx_fio_lock(fio);
    ret = gmx_fio_int_flush(fio);
    gmx_fio_unlock(fio);

    return ret;
}


static int gmx_fio_int_fsync(t_fileio* fio)
{
    int rc = 0;

    if (fio->fp)
    {
        rc = gmx_fsync(fio->fp);
    }
    return rc;
}


int gmx_fio_fsync(t_fileio* fio)
{
    int rc;

    gmx_fio_lock(fio);
    rc = gmx_fio_int_fsync(fio);
    gmx_fio_unlock(fio);

    return rc;
}


t_fileio* gmx_fio_all_output_fsync()
{
    t_fileio* ret = nullptr;
    t_fileio* cur;

    Lock openFilesLock(open_file_mutex);
    cur = gmx_fio_get_first();
    while (cur)
    {
        if (!cur->bRead)
        {
            /* if any of them fails, return failure code */
            int rc = gmx_fio_int_fsync(cur);
            if (rc != 0 && !ret)
            {
                ret = cur;
            }
        }
        cur = gmx_fio_get_next(cur);
    }

    /* in addition, we force these to be written out too, if they're being
       redirected. We don't check for errors because errors most likely mean
       that they're not redirected. */
    fflush(stdout);
    fflush(stderr);
#if HAVE_FSYNC
    /* again, fahcore defines HAVE_FSYNC and fsync() */
    fsync(STDOUT_FILENO);
    fsync(STDERR_FILENO);
#endif

    return ret;
}


gmx_off_t gmx_fio_ftell(t_fileio* fio)
{
    gmx_off_t ret = 0;

    gmx_fio_lock(fio);
    if (fio->fp)
    {
        ret = gmx_ftell(fio->fp);
    }
    gmx_fio_unlock(fio);
    return ret;
}

int gmx_fio_seek(t_fileio* fio, gmx_off_t fpos)
{
    int rc;

    gmx_fio_lock(fio);
    if (fio->fp)
    {
        rc = gmx_fseek(fio->fp, fpos, SEEK_SET);
    }
    else
    {
        gmx_file(fio->fn);
    }
    gmx_fio_unlock(fio);
    return rc;
}

FILE* gmx_fio_getfp(t_fileio* fio)
{
    FILE* ret = nullptr;

    gmx_fio_lock(fio);
    if (fio->fp)
    {
        ret = fio->fp;
    }
    gmx_fio_unlock(fio);
    return ret;
}

gmx_bool gmx_fio_getread(t_fileio* fio)
{
    gmx_bool ret;

    gmx_fio_lock(fio);
    ret = fio->bRead;
    gmx_fio_unlock(fio);

    return ret;
}

int xtc_seek_time(t_fileio* fio, real time, int natoms, gmx_bool bSeekForwardOnly)
{
    int ret;

    gmx_fio_lock(fio);
    ret = xdr_xtc_seek_time(time, fio->fp, fio->xdr, natoms, bSeekForwardOnly);
    gmx_fio_unlock(fio);

    return ret;
}