DiskdIOStrategy.cc

Go to the documentation of this file.

/*
 * Copyright (C) 1996-2025 The Squid Software Foundation and contributors
 *
 * Squid software is distributed under GPLv2+ license and includes
 * contributions from numerous individuals and organizations.
 * Please see the COPYING and CONTRIBUTORS files for details.
 */
 
/* DEBUG: section 79    Squid-side DISKD I/O functions. */
 
#include "squid.h"
#include "comm.h"
#include "comm/Loops.h"
#include "compat/select.h"
#include "ConfigOption.h"
#include "diomsg.h"
#include "DiskdFile.h"
#include "DiskdIOStrategy.h"
#include "DiskIO/DiskFile.h"
#include "fd.h"
#include "SquidConfig.h"
#include "SquidIpc.h"
#include "StatCounters.h"
#include "Store.h"
#include "unlinkd.h"
 
#include <cerrno>
#if HAVE_SYS_IPC_H
#include <sys/ipc.h>
#endif
#if HAVE_SYS_MSG_H
#include <sys/msg.h>
#endif
#if HAVE_SYS_SHM_H
#include <sys/shm.h>
#endif
 
diskd_stats_t diskd_stats;
 
size_t DiskdIOStrategy::nextInstanceID (0);
const int diomsg::msg_snd_rcv_sz = sizeof(diomsg) - sizeof(mtyp_t);
 
size_t
DiskdIOStrategy::newInstance()
{
    return ++nextInstanceID;
}
 
bool
DiskdIOStrategy::shedLoad()
{
    /*
     * Fail on open() if there are too many requests queued.
     */
 
    if (away > magic1) {
        debugs(79, 3, "storeDiskdIO::shedLoad: Shedding, too many requests away");
 
        return true;
    }
 
    return false;
}
 
int
DiskdIOStrategy::load()
{
    /* Calculate the storedir load relative to magic2 on a scale of 0 .. 1000 */
    /* the parse function guarantees magic2 is positive */
    return away * 1000 / magic2;
}
 
void
DiskdIOStrategy::openFailed()
{
    ++diskd_stats.open_fail_queue_len;
}
 
DiskFile::Pointer
DiskdIOStrategy::newFile(char const *path)
{
    if (shedLoad()) {
        openFailed();
        return nullptr;
    }
 
    return new DiskdFile (path, this);
}
 
DiskdIOStrategy::DiskdIOStrategy() : magic1(64), magic2(72), away(0), smsgid(-1), rmsgid(-1), wfd(-1), instanceID(newInstance())
{}
 
bool
DiskdIOStrategy::unlinkdUseful() const
{
    return true;
}
 
void
DiskdIOStrategy::unlinkFile(char const *path)
{
    if (shedLoad()) {
        /* Damn, we need to issue a sync unlink here :( */
        debugs(79, 2, "storeDiskUnlink: Out of queue space, sync unlink");
        unlinkdUnlink(path);
        return;
    }
 
    /* We can attempt a diskd unlink */
    int x;
 
    ssize_t shm_offset;
 
    char *buf;
 
    buf = (char *)shm.get(&shm_offset);
 
    if (!buf) {
        unlinkdUnlink(path);
        return;
    }
 
    xstrncpy(buf, path, SHMBUF_BLKSZ);
 
    x = send(_MQD_UNLINK,
             0,
             (StoreIOState::Pointer )nullptr,
             0,
             0,
             shm_offset);
 
    if (x < 0) {
        int xerrno = errno;
        debugs(79, DBG_IMPORTANT, "storeDiskdSend UNLINK: " << xstrerr(xerrno));
        ::unlink(buf);      /* XXX EWW! */
        //        shm.put (shm_offset);
    }
 
    ++diskd_stats.unlink.ops;
}
 
void
DiskdIOStrategy::init()
{
    int pid;
    void * hIpc;
    int rfd;
    int ikey;
    const char *args[5];
    char skey1[32];
    char skey2[32];
    char skey3[32];
    Ip::Address localhost;
 
    ikey = (getpid() << 10) + (instanceID << 2);
    ikey &= 0x7fffffff;
    smsgid = msgget((key_t) ikey, 0700 | IPC_CREAT);
 
    if (smsgid < 0) {
        int xerrno = errno;
        debugs(50, DBG_CRITICAL, MYNAME << "msgget: " << xstrerr(xerrno));
        fatal("msgget failed");
    }
 
    rmsgid = msgget((key_t) (ikey + 1), 0700 | IPC_CREAT);
 
    if (rmsgid < 0) {
        int xerrno = errno;
        debugs(50, DBG_CRITICAL, MYNAME << "msgget: " << xstrerr(xerrno));
        fatal("msgget failed");
    }
 
    shm.init(ikey, magic2);
    snprintf(skey1, 32, "%d", ikey);
    snprintf(skey2, 32, "%d", ikey + 1);
    snprintf(skey3, 32, "%d", ikey + 2);
    args[0] = "diskd";
    args[1] = skey1;
    args[2] = skey2;
    args[3] = skey3;
    args[4] = nullptr;
    localhost.setLocalhost();
    pid = ipcCreate(IPC_STREAM,
                    Config.Program.diskd,
                    args,
                    "diskd",
                    localhost,
                    &rfd,
                    &wfd,
                    &hIpc);
 
    if (pid < 0)
        fatalf("execl: %s", Config.Program.diskd);
 
    if (rfd != wfd)
        comm_close(rfd);
 
    fd_note(wfd, "squid -> diskd");
 
    commUnsetFdTimeout(wfd);
    commSetNonBlocking(wfd);
    Comm::QuickPollRequired();
}
 
/*
 * SHM manipulation routines
 */
void
SharedMemory::put(ssize_t offset)
{
    int i;
    assert(offset >= 0);
    assert(offset < nbufs * SHMBUF_BLKSZ);
    i = offset / SHMBUF_BLKSZ;
    assert(i < nbufs);
    assert(CBIT_TEST(inuse_map, i));
    CBIT_CLR(inuse_map, i);
    --diskd_stats.shmbuf_count;
}
 
void *
 
SharedMemory::get(ssize_t * shm_offset)
{
    char *aBuf = nullptr;
    int i;
 
    for (i = 0; i < nbufs; ++i) {
        if (CBIT_TEST(inuse_map, i))
            continue;
 
        CBIT_SET(inuse_map, i);
 
        *shm_offset = i * SHMBUF_BLKSZ;
 
        aBuf = buf + (*shm_offset);
 
        break;
    }
 
    if (!aBuf) {
        debugs(79, DBG_IMPORTANT, "ERROR: out of shared-memory buffers");
        return nullptr;
    }
 
    assert(aBuf);
    assert(aBuf >= buf);
    assert(aBuf < buf + (nbufs * SHMBUF_BLKSZ));
    ++diskd_stats.shmbuf_count;
 
    if (diskd_stats.max_shmuse < diskd_stats.shmbuf_count)
        diskd_stats.max_shmuse = diskd_stats.shmbuf_count;
 
    return aBuf;
}
 
void
SharedMemory::init(int ikey, int magic2)
{
    nbufs = (int)(magic2 * 1.3);
    id = shmget((key_t) (ikey + 2),
                nbufs * SHMBUF_BLKSZ, 0600 | IPC_CREAT);
 
    if (id < 0) {
        int xerrno = errno;
        debugs(50, DBG_CRITICAL, MYNAME << "shmget: " << xstrerr(xerrno));
        fatal("shmget failed");
    }
 
    buf = (char *)shmat(id, nullptr, 0);
 
    if (buf == (void *) -1) {
        int xerrno = errno;
        debugs(50, DBG_CRITICAL, MYNAME << "shmat: " << xstrerr(xerrno));
        fatal("shmat failed");
    }
 
    inuse_map = (char *)xcalloc((nbufs + 7) / 8, 1);
    diskd_stats.shmbuf_count += nbufs;
 
    for (int i = 0; i < nbufs; ++i) {
        CBIT_SET(inuse_map, i);
        put (i * SHMBUF_BLKSZ);
    }
}
 
void
DiskdIOStrategy::unlinkDone(diomsg * M)
{
    debugs(79, 3, "storeDiskdUnlinkDone: file " << shm.buf + M->shm_offset << " status " << M->status);
    ++statCounter.syscalls.disk.unlinks;
 
    if (M->status < 0)
        ++diskd_stats.unlink.fail;
    else
        ++diskd_stats.unlink.success;
}
 
void
DiskdIOStrategy::handle(diomsg * M)
{
    if (!cbdataReferenceValid (M->callback_data)) {
        /* I.e. already closed file
         * - say when we have a error opening after
         *   a read was already queued
         */
        debugs(79, 3, "storeDiskdHandle: Invalid callback_data " << M->callback_data);
        cbdataReferenceDone (M->callback_data);
        return;
    }
 
    /* set errno passed from diskd.  makes debugging more meaningful */
    if (M->status < 0)
        errno = -M->status;
 
    if (M->newstyle) {
        DiskdFile *theFile = (DiskdFile *)M->callback_data;
        theFile->unlock();
        theFile->completed (M);
    } else
        switch (M->mtype) {
 
        case _MQD_OPEN:
 
        case _MQD_CREATE:
 
        case _MQD_CLOSE:
 
        case _MQD_READ:
 
        case _MQD_WRITE:
            assert (0);
            break;
 
        case _MQD_UNLINK:
            unlinkDone(M);
            break;
 
        default:
            assert(0);
            break;
        }
 
    cbdataReferenceDone (M->callback_data);
}
 
int
DiskdIOStrategy::send(int mtype, int id, DiskdFile *theFile, size_t size, off_t offset, ssize_t shm_offset, Lock *requestor)
{
    diomsg M;
    M.callback_data = cbdataReference(theFile);
    theFile->lock();
    M.requestor = requestor;
    M.newstyle = true;
 
    if (requestor)
        requestor->lock();
 
    return SEND(&M, mtype, id, size, offset, shm_offset);
}
 
int
DiskdIOStrategy::send(int mtype, int id, RefCount<StoreIOState> sio, size_t size, off_t offset, ssize_t shm_offset)
{
    diomsg M;
    M.callback_data = cbdataReference(sio.getRaw());
    M.newstyle = false;
 
    return SEND(&M, mtype, id, size, offset, shm_offset);
}
 
int
DiskdIOStrategy::SEND(diomsg *M, int mtype, int id, size_t size, off_t offset, ssize_t shm_offset)
{
    static int send_errors = 0;
    static int last_seq_no = 0;
    static int seq_no = 0;
    int x;
 
    M->mtype = mtype;
    M->size = size;
    M->offset = offset;
    M->status = -1;
    M->shm_offset = (int) shm_offset;
    M->id = id;
    M->seq_no = ++seq_no;
 
    if (M->seq_no < last_seq_no)
        debugs(79, DBG_IMPORTANT, "WARNING: sequencing out of order");
 
    x = msgsnd(smsgid, M, diomsg::msg_snd_rcv_sz, IPC_NOWAIT);
 
    last_seq_no = M->seq_no;
 
    if (0 == x) {
        ++diskd_stats.sent_count;
        ++away;
    } else {
        int xerrno = errno;
        debugs(79, DBG_IMPORTANT, MYNAME << "msgsnd: " << xstrerr(xerrno));
        cbdataReferenceDone(M->callback_data);
        ++send_errors;
        assert(send_errors < 100);
        if (shm_offset > -1)
            shm.put(shm_offset);
    }
 
    /*
     * We have to drain the queue here if necessary.  If we don't,
     * then we can have a lot of messages in the queue (probably
     * up to 2*magic1) and we can run out of shared memory buffers.
     */
    /*
     * Note that we call Store::Root().callbackk (for all SDs), rather
     * than callback for just this SD, so that while
     * we're "blocking" on this SD we can also handle callbacks
     * from other SDs that might be ready.
     */
 
    struct timeval delay = {0, 1};
 
    while (away > magic2) {
        xselect(0, nullptr, nullptr, nullptr, &delay);
        Store::Root().callback();
 
        if (delay.tv_usec < 1000000)
            delay.tv_usec <<= 1;
    }
 
    return x;
}
 
ConfigOption *
DiskdIOStrategy::getOptionTree() const
{
    ConfigOptionVector *result = new ConfigOptionVector;
    result->options.push_back(new ConfigOptionAdapter<DiskdIOStrategy>(*const_cast<DiskdIOStrategy *>(this), &DiskdIOStrategy::optionQ1Parse, &DiskdIOStrategy::optionQ1Dump));
    result->options.push_back(new ConfigOptionAdapter<DiskdIOStrategy>(*const_cast<DiskdIOStrategy *>(this), &DiskdIOStrategy::optionQ2Parse, &DiskdIOStrategy::optionQ2Dump));
    return result;
}
 
bool
DiskdIOStrategy::optionQ1Parse(const char *name, const char *value, int isaReconfig)
{
    if (strcmp(name, "Q1") != 0)
        return false;
 
    int old_magic1 = magic1;
 
    magic1 = atoi(value);
 
    if (!isaReconfig)
        return true;
 
    if (old_magic1 < magic1) {
        /*
        * This is because shm.nbufs is computed at startup, when
        * we call shmget().  We can't increase the Q1/Q2 parameters
        * beyond their initial values because then we might have
        * more "Q2 messages" than shared memory chunks, and this
        * will cause an assertion in storeDiskdShmGet().
        */
        /* TODO: have DiskdIO hold a link to the swapdir, to allow detailed reporting again */
        debugs(3, DBG_IMPORTANT, "WARNING: cannot increase cache_dir Q1 value while Squid is running.");
        magic1 = old_magic1;
        return true;
    }
 
    if (old_magic1 != magic1)
        debugs(3, DBG_IMPORTANT, "cache_dir new Q1 value '" << magic1 << "'");
 
    return true;
}
 
void
DiskdIOStrategy::optionQ1Dump(StoreEntry * e) const
{
    storeAppendPrintf(e, " Q1=%d", magic1);
}
 
bool
DiskdIOStrategy::optionQ2Parse(const char *name, const char *value, int isaReconfig)
{
    if (strcmp(name, "Q2") != 0)
        return false;
 
    int old_magic2 = magic2;
 
    magic2 = atoi(value);
 
    if (!isaReconfig)
        return true;
 
    if (old_magic2 < magic2) {
        /* See comments in Q1 function above */
        debugs(3, DBG_IMPORTANT, "WARNING: cannot increase cache_dir Q2 value while Squid is running.");
        magic2 = old_magic2;
        return true;
    }
 
    if (old_magic2 != magic2)
        debugs(3, DBG_IMPORTANT, "cache_dir new Q2 value '" << magic2 << "'");
 
    return true;
}
 
void
DiskdIOStrategy::optionQ2Dump(StoreEntry * e) const
{
    storeAppendPrintf(e, " Q2=%d", magic2);
}
 
/*
 * Sync any pending data. We just sit around and read the queue
 * until the data has finished writing.
 */
void
DiskdIOStrategy::sync()
{
    static time_t lastmsg = 0;
 
    while (away > 0) {
        if (squid_curtime > lastmsg) {
            debugs(47, DBG_IMPORTANT, "storeDiskdDirSync: " << away << " messages away");
            lastmsg = squid_curtime;
        }
 
        callback();
    }
}
 
/*
 * Handle callbacks. If we have more than magic2 requests away, we block
 * until the queue is below magic2. Otherwise, we simply return when we
 * don't get a message.
 */
 
int
DiskdIOStrategy::callback()
{
    diomsg M;
    int x;
    int retval = 0;
 
    if (away >= magic2) {
        ++diskd_stats.block_queue_len;
        retval = 1;
        /* We might not have anything to do, but our queue
         * is full.. */
    }
 
    if (diskd_stats.sent_count - diskd_stats.recv_count >
            diskd_stats.max_away) {
        diskd_stats.max_away = diskd_stats.sent_count - diskd_stats.recv_count;
    }
 
    while (1) {
        x = msgrcv(rmsgid, &M, diomsg::msg_snd_rcv_sz, 0, IPC_NOWAIT);
 
        if (x < 0)
            break;
        else if (x != diomsg::msg_snd_rcv_sz) {
            debugs(47, DBG_IMPORTANT, "storeDiskdDirCallback: msgget returns " << x);
            break;
        }
 
        ++diskd_stats.recv_count;
        --away;
        handle(&M);
        retval = 1;     /* Return that we've actually done some work */
 
        if (M.shm_offset > -1)
            shm.put ((off_t) M.shm_offset);
    }
 
    return retval;
}
 
void
DiskdIOStrategy::statfs(StoreEntry & sentry)const
{
    storeAppendPrintf(&sentry, "Pending operations: %d\n", away);
}