src/lib/ostream-file.c

	ostream-file.c revision 3f072ebbd4152514e4fc54ac42da535af290359b
45312f52ff3a3d4c137447be4c7556500c2f8bf2Timo Sirainen/* Copyright (c) 2002-2009 Dovecot authors, see the included COPYING file */
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen/* @UNSAFE: whole file */
345648b341f228bd7f0b89f8aa3ecb9c470d817eTimo Sirainen
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#include "lib.h"
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen#include "ioloop.h"
5a2cb3d097a2d9a9e930af997e7bf3400a8d840dTimo Sirainen#include "write-full.h"
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#include "network.h"
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#include "sendfile-util.h"
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#include "istream.h"
97c339398f1aba6f315b55a9b6ee6b020e33bea4Timo Sirainen#include "istream-internal.h"
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#include "ostream-internal.h"
16f46efe0e090fe6975acf012a61a160f4787985Andrey Panin
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#include <unistd.h>
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#include <sys/stat.h>
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#ifdef HAVE_SYS_UIO_H
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#  include <sys/uio.h>
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#endif
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen/* try to keep the buffer size within 4k..128k. ReiserFS may actually return
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen   128k as optimal size. */
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#define DEFAULT_OPTIMAL_BLOCK_SIZE 4096
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#define MAX_OPTIMAL_BLOCK_SIZE (128*1024)
16f46efe0e090fe6975acf012a61a160f4787985Andrey Panin
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#define IS_STREAM_EMPTY(fstream) \
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen    ((fstream)->head == (fstream)->tail && !(fstream)->full)
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen#define MAX_SSIZE_T(size) \
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen    ((size) < SSIZE_T_MAX ? (size_t)(size) : SSIZE_T_MAX)
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainenstruct file_ostream {
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen    struct ostream_private ostream;
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen    int fd;
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen    struct io *io;
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen    uoff_t buffer_offset;
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen    uoff_t real_offset;
7242e1ce7803b83bc82e239ef111b47c1c72dd4bAndrey Panin
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen    unsigned char *buffer; /* ring-buffer */
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen    size_t buffer_size, max_buffer_size, optimal_block_size;
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen    size_t head, tail; /* first unsent/unused byte */
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen
c57776c06ec99ba9b0dafdbf9475ea72ea8ca134Timo Sirainen    unsigned int full:1; /* if head == tail, is buffer empty or full? */
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen    unsigned int file:1;
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen    unsigned int corked:1;
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen    unsigned int flush_pending:1;
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen    unsigned int socket_cork_set:1;
d1414c09cf0d58ac983054e2f4e1a1f329272dcfTimo Sirainen    unsigned int no_socket_cork:1;
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen    unsigned int no_sendfile:1;
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen    unsigned int autoclose_fd:1;
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen};
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainenstatic void stream_send_io(struct file_ostream *fstream);
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainenstatic void stream_closed(struct file_ostream *fstream)
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen{
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen    if (fstream->io != NULL)
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen        io_remove(&fstream->io);
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen    if (fstream->autoclose_fd && fstream->fd != -1) {
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen        if (close(fstream->fd) < 0)
648d24583c1574441c4fa0331a90bd4d6e7996c5Timo Sirainen            i_error("file_ostream.close() failed: %m");
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen    }
e70d5895795732b8247ab9abb045b438e954bc46Timo Sirainen    fstream->fd = -1;
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainen
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainen    fstream->ostream.ostream.closed = TRUE;
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainen}
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainen
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainenstatic void o_stream_file_close(struct iostream_private *stream)
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainen{
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainen    struct file_ostream *fstream = (struct file_ostream *)stream;
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainen
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainen    /* flush output before really closing it */
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainen    o_stream_flush(&fstream->ostream.ostream);
0469ed17dafcc56589ce00960a23f4f06817dfb5Timo Sirainen
1ddec6312bc6882aeb17d4d46d19cbca1723b68bTimo Sirainen    stream_closed(fstream);
0469ed17dafcc56589ce00960a23f4f06817dfb5Timo Sirainen}
704fbadd78375da18dcaf2c5d93ac8cfe2c61358Timo Sirainen
704fbadd78375da18dcaf2c5d93ac8cfe2c61358Timo Sirainenstatic void o_stream_file_destroy(struct iostream_private *stream)
704fbadd78375da18dcaf2c5d93ac8cfe2c61358Timo Sirainen{
2c38504860da8a8de915f8e0f5f39d7e7bd00cf8Timo Sirainen    struct file_ostream *fstream = (struct file_ostream *)stream;
2c38504860da8a8de915f8e0f5f39d7e7bd00cf8Timo Sirainen
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen    i_free(fstream->buffer);
a64adf62fa33f2463a86f990217b0c9078531a40Timo Sirainen}
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen
97c339398f1aba6f315b55a9b6ee6b020e33bea4Timo Sirainenstatic void
97c339398f1aba6f315b55a9b6ee6b020e33bea4Timo Siraineno_stream_file_set_max_buffer_size(struct iostream_private *stream,
97c339398f1aba6f315b55a9b6ee6b020e33bea4Timo Sirainen                  size_t max_size)
97c339398f1aba6f315b55a9b6ee6b020e33bea4Timo Sirainen{
97c339398f1aba6f315b55a9b6ee6b020e33bea4Timo Sirainen    struct file_ostream *fstream = (struct file_ostream *)stream;
a64adf62fa33f2463a86f990217b0c9078531a40Timo Sirainen
2c38504860da8a8de915f8e0f5f39d7e7bd00cf8Timo Sirainen    fstream->max_buffer_size = max_size;
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen}
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen
2c38504860da8a8de915f8e0f5f39d7e7bd00cf8Timo Sirainenstatic void update_buffer(struct file_ostream *fstream, size_t size)
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen{
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen    size_t used;
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen    if (IS_STREAM_EMPTY(fstream) || size == 0)
e9e2d23e1ea5a149a7d8828d2a45b9f2313c3785Timo Sirainen        return;
e9e2d23e1ea5a149a7d8828d2a45b9f2313c3785Timo Sirainen
97c339398f1aba6f315b55a9b6ee6b020e33bea4Timo Sirainen    if (fstream->head < fstream->tail) {
97c339398f1aba6f315b55a9b6ee6b020e33bea4Timo Sirainen        /* ...HXXXT... */
f97cf1c086715db87094bc3d0a4fefdd80bd869cTimo Sirainen        used = fstream->tail - fstream->head;
0469ed17dafcc56589ce00960a23f4f06817dfb5Timo Sirainen        i_assert(size <= used);
0469ed17dafcc56589ce00960a23f4f06817dfb5Timo Sirainen        fstream->head += size;
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen    } else {
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen        /* XXXT...HXXX */
a64adf62fa33f2463a86f990217b0c9078531a40Timo Sirainen        used = fstream->buffer_size - fstream->head;
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen        if (size > used) {
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen            size -= used;
33c6d5807b449463e9b81db5ec99fe027cc1b984Timo Sirainen            i_assert(size <= fstream->tail);
8eea67470c1bd8562a62e7445d930bb2079b1a43Timo Sirainen            fstream->head = size;
94a78eb438622fa53abef1e1726714dacad4b61cTimo Sirainen        } else {
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen            fstream->head += size;
a64adf62fa33f2463a86f990217b0c9078531a40Timo Sirainen        }
2c38504860da8a8de915f8e0f5f39d7e7bd00cf8Timo Sirainen
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen        fstream->full = FALSE;
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen    }
2c38504860da8a8de915f8e0f5f39d7e7bd00cf8Timo Sirainen
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen    if (fstream->head == fstream->tail)
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen        fstream->head = fstream->tail = 0;
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen
6e873f135368bcfdd1de4458dded791d0c4d00cdTimo Sirainen    if (fstream->head == fstream->buffer_size)
e9e2d23e1ea5a149a7d8828d2a45b9f2313c3785Timo Sirainen        fstream->head = 0;
e9e2d23e1ea5a149a7d8828d2a45b9f2313c3785Timo Sirainen}
a4ac325c2802693c6b761e5a8fda961e5d7490eaTimo Sirainen
b7c2065b3f10f9ae27787a9db5aaefbfc70d4502Timo Sirainenstatic void o_stream_socket_cork(struct file_ostream *fstream)
f97cf1c086715db87094bc3d0a4fefdd80bd869cTimo Sirainen{
0469ed17dafcc56589ce00960a23f4f06817dfb5Timo Sirainen    if (fstream->corked && !fstream->socket_cork_set) {
0469ed17dafcc56589ce00960a23f4f06817dfb5Timo Sirainen        if (!fstream->no_socket_cork) {
25757faf029c369a8318349dafe952e2358df1d8Timo Sirainen            if (net_set_cork(fstream->fd, TRUE) < 0)
                fstream->no_socket_cork = TRUE;
            else
                fstream->socket_cork_set = TRUE;
        }
    }
}

static int o_stream_lseek(struct file_ostream *fstream)
{
    off_t ret;

    if (fstream->real_offset == fstream->buffer_offset)
        return 0;

    ret = lseek(fstream->fd, (off_t)fstream->buffer_offset, SEEK_SET);
    if (ret < 0) {
        fstream->ostream.ostream.stream_errno = errno;
        return -1;
    }

    if (ret != (off_t)fstream->buffer_offset) {
        fstream->ostream.ostream.stream_errno = EINVAL;
        return -1;
    }
    fstream->real_offset = fstream->buffer_offset;
    return 0;
}

static ssize_t o_stream_writev(struct file_ostream *fstream,
                   const struct const_iovec *iov, int iov_size)
{
    ssize_t ret, ret2;
    size_t size, sent;
    bool partial;
    int i;

    o_stream_socket_cork(fstream);
    if (iov_size == 1) {
        if (!fstream->file ||
            fstream->real_offset == fstream->buffer_offset) {
            ret = write(fstream->fd, iov->iov_base, iov->iov_len);
            if (ret > 0)
                fstream->real_offset += ret;
        } else {
            ret = pwrite(fstream->fd, iov->iov_base, iov->iov_len,
                     fstream->buffer_offset);
        }
        partial = ret != (ssize_t)iov->iov_len;
    } else {
        if (o_stream_lseek(fstream) < 0)
            return -1;

        sent = 0; partial = FALSE;
        while (iov_size > IOV_MAX) {
            size = 0;
            for (i = 0; i < IOV_MAX; i++)
                size += iov[i].iov_len;

            ret = writev(fstream->fd, (const struct iovec *)iov,
                     IOV_MAX);
            if (ret != (ssize_t)size) {
                partial = TRUE;
                break;
            }

            fstream->real_offset += ret;
            fstream->buffer_offset += ret;
            sent += ret;
            iov += IOV_MAX;
            iov_size -= IOV_MAX;
        }

        if (iov_size <= IOV_MAX) {
            size = 0;
            for (i = 0; i < iov_size; i++)
                size += iov[i].iov_len;

            ret = writev(fstream->fd, (const struct iovec *)iov,
                     iov_size);
            partial = ret != (ssize_t)size;
        }
        if (ret > 0) {
            fstream->real_offset += ret;
            ret += sent;
        } else if (!fstream->file && sent > 0) {
            /* return what we managed to get sent */
            ret = sent;
        }
    }

    if (ret < 0) {
        if (errno == EAGAIN || errno == EINTR)
            return 0;
        fstream->ostream.ostream.stream_errno = errno;
        stream_closed(fstream);
        return -1;
    }
    if (unlikely(ret == 0 && fstream->file)) {
        /* assume out of disk space */
        fstream->ostream.ostream.stream_errno = ENOSPC;
        stream_closed(fstream);
        return -1;
    }
    fstream->buffer_offset += ret;
    if (partial && fstream->file) {
        /* we failed to write everything to a file. either we ran out
           of disk space or we're writing to NFS. try to write the
           rest to resolve this. */
        size = ret;
        while (iov_size > 0 && size >= iov->iov_len) {
            size -= iov->iov_len;
            iov++;
            iov_size--;
        }
        i_assert(iov_size > 0);
        if (size == 0)
            ret2 = o_stream_writev(fstream, iov, iov_size);
        else {
            /* write the first iov separately */
            struct const_iovec new_iov;

            new_iov.iov_base =
                CONST_PTR_OFFSET(iov->iov_base, size);
            new_iov.iov_len = iov->iov_len - size;
            ret2 = o_stream_writev(fstream, &new_iov, 1);
            if (ret2 > 0) {
                i_assert((size_t)ret2 == new_iov.iov_len);
                /* write the rest */
                if (iov_size > 1) {
                    ret += ret2;
                    ret2 = o_stream_writev(fstream, iov + 1,
                                   iov_size - 1);
                }
            }
        }
        if (ret2 <= 0)
            return ret2;
        ret += ret2;
    }
    return ret;
}

/* returns how much of vector was used */
static int o_stream_fill_iovec(struct file_ostream *fstream,
                   struct const_iovec iov[2])
{
    if (IS_STREAM_EMPTY(fstream))
        return 0;

    if (fstream->head < fstream->tail) {
        iov[0].iov_base = fstream->buffer + fstream->head;
        iov[0].iov_len = fstream->tail - fstream->head;
        return 1;
    } else {
        iov[0].iov_base = fstream->buffer + fstream->head;
        iov[0].iov_len = fstream->buffer_size - fstream->head;
        if (fstream->tail == 0)
            return 1;
        else {
            iov[1].iov_base = fstream->buffer;
            iov[1].iov_len = fstream->tail;
            return 2;
        }
    }
}

static int buffer_flush(struct file_ostream *fstream)
{
    struct const_iovec iov[2];
    int iov_len;
    ssize_t ret;

    iov_len = o_stream_fill_iovec(fstream, iov);
    if (iov_len > 0) {
        ret = o_stream_writev(fstream, iov, iov_len);
        if (ret < 0)
            return -1;

        update_buffer(fstream, ret);
    }

    return IS_STREAM_EMPTY(fstream) ? 1 : 0;
}

static void o_stream_file_cork(struct ostream_private *stream, bool set)
{
    struct file_ostream *fstream = (struct file_ostream *)stream;
    int ret;

    if (fstream->corked != set && !stream->ostream.closed) {
        if (set && fstream->io != NULL)
            io_remove(&fstream->io);
        else if (!set) {
            /* buffer flushing might close the stream */
            ret = buffer_flush(fstream);
            if (fstream->io == NULL &&
                (ret == 0 || fstream->flush_pending) &&
                !stream->ostream.closed) {
                fstream->io = io_add(fstream->fd, IO_WRITE,
                             stream_send_io, fstream);
            }
        }

        if (fstream->socket_cork_set) {
            i_assert(!set);
            if (net_set_cork(fstream->fd, FALSE) < 0)
                fstream->no_socket_cork = TRUE;
            fstream->socket_cork_set = FALSE;
        }
        fstream->corked = set;
    }
}

static int o_stream_file_flush(struct ostream_private *stream)
{
    struct file_ostream *fstream = (struct file_ostream *) stream;

    return buffer_flush(fstream);
}

static void
o_stream_file_flush_pending(struct ostream_private *stream, bool set)
{
    struct file_ostream *fstream = (struct file_ostream *) stream;

    fstream->flush_pending = set;
    if (set && !fstream->corked && fstream->io == NULL) {
        fstream->io = io_add(fstream->fd, IO_WRITE,
                     stream_send_io, fstream);
    }
}

static size_t get_unused_space(const struct file_ostream *fstream)
{
    if (fstream->head > fstream->tail) {
        /* XXXT...HXXX */
        return fstream->head - fstream->tail;
    } else if (fstream->head < fstream->tail) {
        /* ...HXXXT... */
        return (fstream->buffer_size - fstream->tail) + fstream->head;
    } else {
        /* either fully unused or fully used */
        return fstream->full ? 0 : fstream->buffer_size;
    }
}

static size_t o_stream_file_get_used_size(const struct ostream_private *stream)
{
    const struct file_ostream *fstream =
        (const struct file_ostream *)stream;

    return fstream->buffer_size - get_unused_space(fstream);
}

static int o_stream_file_seek(struct ostream_private *stream, uoff_t offset)
{
    struct file_ostream *fstream = (struct file_ostream *)stream;

    if (offset > OFF_T_MAX || !fstream->file) {
        stream->ostream.stream_errno = EINVAL;
        return -1;
    }

    if (buffer_flush(fstream) < 0)
        return -1;

    stream->ostream.offset = offset;
    fstream->buffer_offset = offset;
    return 1;
}

static void o_stream_grow_buffer(struct file_ostream *fstream, size_t bytes)
{
    size_t size, new_size, end_size;

    size = nearest_power(fstream->buffer_size + bytes);
    if (size > fstream->max_buffer_size) {
        /* limit the size */
        size = fstream->max_buffer_size;
    } else if (fstream->corked) {
        /* try to use optimal buffer size with corking */
        new_size = I_MIN(fstream->optimal_block_size,
                 fstream->max_buffer_size);
        if (new_size > size)
            size = new_size;
    }

    if (size <= fstream->buffer_size)
        return;

    fstream->buffer = i_realloc(fstream->buffer,
                    fstream->buffer_size, size);

    if (fstream->tail <= fstream->head && !IS_STREAM_EMPTY(fstream)) {
        /* move head forward to end of buffer */
        end_size = fstream->buffer_size - fstream->head;
        memmove(fstream->buffer + size - end_size,
            fstream->buffer + fstream->head, end_size);
        fstream->head = size - end_size;
    }

    fstream->full = FALSE;
    fstream->buffer_size = size;
}

static void stream_send_io(struct file_ostream *fstream)
{
    struct ostream *ostream = &fstream->ostream.ostream;
    int ret;

    /* Set flush_pending = FALSE first before calling the flush callback,
       and change it to TRUE only if callback returns 0. That way the
       callback can call o_stream_set_flush_pending() again and we don't
       forget it even if flush callback returns 1. */
    fstream->flush_pending = FALSE;

    o_stream_ref(ostream);
    if (fstream->ostream.callback != NULL)
        ret = fstream->ostream.callback(fstream->ostream.context);
    else
        ret = o_stream_file_flush(&fstream->ostream);

    if (ret == 0)
        fstream->flush_pending = TRUE;

    if (!fstream->flush_pending && IS_STREAM_EMPTY(fstream)) {
        if (fstream->io != NULL) {
            /* all sent */
            io_remove(&fstream->io);
        }
    } else if (!fstream->ostream.ostream.closed) {
        /* Add the IO handler if it's not there already. Callback
           might have just returned 0 without there being any data
           to be sent. */
        if (fstream->io == NULL) {
            fstream->io = io_add(fstream->fd, IO_WRITE,
                         stream_send_io, fstream);
        }
    }

    o_stream_unref(&ostream);
}

static size_t o_stream_add(struct file_ostream *fstream,
               const void *data, size_t size)
{
    size_t unused, sent;
    int i;

    unused = get_unused_space(fstream);
    if (unused < size)
        o_stream_grow_buffer(fstream, size-unused);

    sent = 0;
    for (i = 0; i < 2 && sent < size && !fstream->full; i++) {
        unused = fstream->tail >= fstream->head ?
            fstream->buffer_size - fstream->tail :
            fstream->head - fstream->tail;

        if (unused > size-sent)
            unused = size-sent;
        memcpy(fstream->buffer + fstream->tail,
               CONST_PTR_OFFSET(data, sent), unused);
        sent += unused;

        fstream->tail += unused;
        if (fstream->tail == fstream->buffer_size)
            fstream->tail = 0;

        if (fstream->head == fstream->tail)
            fstream->full = TRUE;
    }

    if (sent != 0 && fstream->io == NULL &&
        !fstream->corked && !fstream->file) {
        fstream->io = io_add(fstream->fd, IO_WRITE, stream_send_io,
                     fstream);
    }

    return sent;
}

static ssize_t o_stream_file_sendv(struct ostream_private *stream,
                   const struct const_iovec *iov,
                   unsigned int iov_count)
{
    struct file_ostream *fstream = (struct file_ostream *)stream;
    size_t size, total_size, added, optimal_size;
    unsigned int i;
    ssize_t ret = 0;

    for (i = 0, size = 0; i < iov_count; i++)
        size += iov[i].iov_len;
    total_size = size;

    if (size > get_unused_space(fstream) && !IS_STREAM_EMPTY(fstream)) {
        if (o_stream_file_flush(stream) < 0)
            return -1;
    }

    optimal_size = I_MIN(fstream->optimal_block_size,
                 fstream->max_buffer_size);
    if (IS_STREAM_EMPTY(fstream) &&
        (!fstream->corked || size >= optimal_size)) {
        /* send immediately */
        ret = o_stream_writev(fstream, iov, iov_count);
        if (ret < 0)
            return -1;

        size = ret;
        while (size > 0 && iov_count > 0 && size >= iov[0].iov_len) {
            size -= iov[0].iov_len;
            iov++;
            iov_count--;
        }

        if (iov_count == 0)
            i_assert(size == 0);
        else {
            added = o_stream_add(fstream,
                    CONST_PTR_OFFSET(iov[0].iov_base, size),
                    iov[0].iov_len - size);
            ret += added;

            if (added != iov[0].iov_len - size) {
                /* buffer full */
                stream->ostream.offset += ret;
                return ret;
            }

            iov++;
            iov_count--;
        }
    }

    /* buffer it, at least partly */
    for (i = 0; i < iov_count; i++) {
        added = o_stream_add(fstream, iov[i].iov_base, iov[i].iov_len);
        ret += added;
        if (added != iov[i].iov_len)
            break;
    }
    stream->ostream.offset += ret;
    i_assert((size_t)ret <= total_size);
    i_assert((size_t)ret == total_size || !fstream->file);
    return ret;
}

static off_t io_stream_sendfile(struct ostream_private *outstream,
                struct istream *instream, int in_fd)
{
    struct file_ostream *foutstream = (struct file_ostream *)outstream;
    const struct stat *st;
    uoff_t start_offset;
    uoff_t in_size, offset, send_size, v_offset;
    ssize_t ret;

    st = i_stream_stat(instream, TRUE);
    if (st == NULL) {
        outstream->ostream.stream_errno = instream->stream_errno;
        return -1;
    }
    in_size = st->st_size;

    o_stream_socket_cork(foutstream);

    /* flush out any data in buffer */
    if ((ret = buffer_flush(foutstream)) <= 0)
        return ret;

    if (o_stream_lseek(foutstream) < 0)
        return -1;

        start_offset = v_offset = instream->v_offset;
    do {
        offset = instream->real_stream->abs_start_offset + v_offset;
        send_size = in_size - v_offset;

        ret = safe_sendfile(foutstream->fd, in_fd, &offset,
                    MAX_SSIZE_T(send_size));
        if (ret <= 0) {
            if (ret == 0 || errno == EINTR || errno == EAGAIN) {
                ret = 0;
                break;
            }

            outstream->ostream.stream_errno = errno;
            if (errno != EINVAL) {
                /* close only if error wasn't because
                   sendfile() isn't supported */
                stream_closed(foutstream);
            }
            break;
        }

        v_offset += ret;
        foutstream->real_offset += ret;
        foutstream->buffer_offset += ret;
        outstream->ostream.offset += ret;
    } while ((uoff_t)ret != send_size);

    i_stream_seek(instream, v_offset);
    return ret < 0 ? -1 : (off_t)(instream->v_offset - start_offset);
}

static off_t io_stream_copy(struct ostream_private *outstream,
                struct istream *instream)
{
    struct file_ostream *foutstream = (struct file_ostream *)outstream;
    uoff_t start_offset;
    struct const_iovec iov[3];
    int iov_len;
    const unsigned char *data;
    size_t size, skip_size;
    ssize_t ret;
    int pos;

    iov_len = o_stream_fill_iovec(foutstream, iov);

        skip_size = 0;
    for (pos = 0; pos < iov_len; pos++)
        skip_size += iov[pos].iov_len;

    start_offset = instream->v_offset;
    for (;;) {
        (void)i_stream_read_data(instream, &data, &size,
                     foutstream->optimal_block_size-1);
        if (size == 0) {
            /* all sent */
            break;
        }

        pos = iov_len++;
        iov[pos].iov_base = (void *) data;
        iov[pos].iov_len = size;

        ret = o_stream_writev(foutstream, iov, iov_len);
        if (ret < 0)
            return -1;

        if (skip_size > 0) {
            if ((size_t)ret < skip_size) {
                update_buffer(foutstream, ret);
                skip_size -= ret;
                ret = 0;
            } else {
                update_buffer(foutstream, skip_size);
                ret -= skip_size;
                skip_size = 0;
            }
        }
        outstream->ostream.offset += ret;
        i_stream_skip(instream, ret);

        if ((size_t)ret != iov[pos].iov_len)
            break;

        i_assert(skip_size == 0);
        iov_len = 0;
    }

    return (off_t) (instream->v_offset - start_offset);
}

static off_t io_stream_copy_backwards(struct ostream_private *outstream,
                      struct istream *instream, uoff_t in_size)
{
    struct file_ostream *foutstream = (struct file_ostream *)outstream;
    uoff_t in_start_offset, in_offset, in_limit, out_offset;
    const unsigned char *data;
    size_t buffer_size, size, read_size;
    ssize_t ret;

    i_assert(IS_STREAM_EMPTY(foutstream));

    /* figure out optimal buffer size */
    buffer_size = instream->real_stream->buffer_size;
    if (buffer_size == 0 || buffer_size > foutstream->buffer_size) {
        if (foutstream->optimal_block_size > foutstream->buffer_size) {
            o_stream_grow_buffer(foutstream,
                         foutstream->optimal_block_size -
                         foutstream->buffer_size);
        }

        buffer_size = foutstream->buffer_size;
    }

    in_start_offset = instream->v_offset;
    in_offset = in_limit = in_size;
    out_offset = outstream->ostream.offset + (in_offset - in_start_offset);

    while (in_offset > in_start_offset) {
        if (in_offset - in_start_offset <= buffer_size)
            read_size = in_offset - in_start_offset;
        else
            read_size = buffer_size;
        in_offset -= read_size;
        out_offset -= read_size;

        for (;;) {
            i_assert(in_offset <= in_limit);

            i_stream_seek(instream, in_offset);
            read_size = in_limit - in_offset;

            (void)i_stream_read_data(instream, &data, &size,
                         read_size-1);
            if (size >= read_size) {
                size = read_size;
                if (instream->mmaped) {
                    /* we'll have to write it through
                       buffer or the file gets corrupted */
                    i_assert(size <=
                         foutstream->buffer_size);
                    memcpy(foutstream->buffer, data, size);
                    data = foutstream->buffer;
                }
                break;
            }

            /* buffer too large probably, try with smaller */
            read_size -= size;
            in_offset += read_size;
            out_offset += read_size;
            buffer_size -= read_size;
        }
        in_limit -= size;

        ret = pwrite_full(foutstream->fd, data, size, out_offset);
        if (ret < 0) {
            /* error */
            outstream->ostream.stream_errno = errno;
            return -1;
        }
        i_stream_skip(instream, size);
    }

    outstream->ostream.offset += in_size - in_start_offset;
    return (off_t) (in_size - in_start_offset);
}

static off_t o_stream_file_send_istream(struct ostream_private *outstream,
                    struct istream *instream)
{
    struct file_ostream *foutstream = (struct file_ostream *)outstream;
    const struct stat *st;
    off_t ret;
    int in_fd;

    in_fd = i_stream_get_fd(instream);

    if (in_fd == foutstream->fd) {
        /* copying data within same fd. we'll have to be careful with
           seeks and overlapping writes. */
        st = i_stream_stat(instream, TRUE);
        if (st == NULL) {
            outstream->ostream.stream_errno = instream->stream_errno;
            return -1;
        }
        i_assert(instream->v_offset <= (uoff_t)st->st_size);

        ret = (off_t)outstream->ostream.offset -
            (off_t)(instream->real_stream->abs_start_offset +
                instream->v_offset);
        if (ret == 0) {
            /* copying data over itself. we don't really
               need to do that, just fake it. */
            return st->st_size - instream->v_offset;
        }
        if (ret > 0) {
            /* overlapping */
            return io_stream_copy_backwards(outstream, instream,
                            st->st_size);
        }
    }

    if (!foutstream->no_sendfile && in_fd != -1 && instream->seekable) {
        ret = io_stream_sendfile(outstream, instream, in_fd);
        if (ret >= 0 || outstream->ostream.stream_errno != EINVAL)
            return ret;

        /* sendfile() not supported (with this fd), fallback to
           regular sending. */
        outstream->ostream.stream_errno = 0;
        foutstream->no_sendfile = TRUE;
    }

    return io_stream_copy(outstream, instream);
}

static struct file_ostream *
o_stream_create_fd_common(int fd, bool autoclose_fd)
{
    struct file_ostream *fstream;

    fstream = i_new(struct file_ostream, 1);
    fstream->fd = fd;
    fstream->autoclose_fd = autoclose_fd;
    fstream->optimal_block_size = DEFAULT_OPTIMAL_BLOCK_SIZE;

    fstream->ostream.iostream.close = o_stream_file_close;
    fstream->ostream.iostream.destroy = o_stream_file_destroy;
    fstream->ostream.iostream.set_max_buffer_size =
        o_stream_file_set_max_buffer_size;

    fstream->ostream.cork = o_stream_file_cork;
    fstream->ostream.flush = o_stream_file_flush;
    fstream->ostream.flush_pending = o_stream_file_flush_pending;
    fstream->ostream.get_used_size = o_stream_file_get_used_size;
    fstream->ostream.seek = o_stream_file_seek;
    fstream->ostream.sendv = o_stream_file_sendv;
    fstream->ostream.send_istream = o_stream_file_send_istream;

    return fstream;
}

static void fstream_init_file(struct file_ostream *fstream)
{
    struct stat st;

    fstream->no_sendfile = TRUE;
    if (fstat(fstream->fd, &st) < 0)
        return;

    if ((uoff_t)st.st_blksize > fstream->optimal_block_size) {
        /* use the optimal block size, but with a reasonable limit */
        fstream->optimal_block_size =
            I_MIN(st.st_blksize, MAX_OPTIMAL_BLOCK_SIZE);
    }

    if (S_ISREG(st.st_mode)) {
        fstream->no_socket_cork = TRUE;
        fstream->file = TRUE;
    }
}

struct ostream *
o_stream_create_fd(int fd, size_t max_buffer_size, bool autoclose_fd)
{
    struct file_ostream *fstream;
    struct ostream *ostream;
    off_t offset;

    fstream = o_stream_create_fd_common(fd, autoclose_fd);
    fstream->max_buffer_size = max_buffer_size;
    ostream = o_stream_create(&fstream->ostream);

    offset = lseek(fd, 0, SEEK_CUR);
    if (offset >= 0) {
        ostream->offset = offset;
        fstream->real_offset = offset;
        fstream->buffer_offset = offset;
        fstream_init_file(fstream);
    } else {
        if (net_getsockname(fd, NULL, NULL) < 0) {
            fstream->no_sendfile = TRUE;
            fstream->no_socket_cork = TRUE;
        }
    }

    if (max_buffer_size == 0)
        fstream->max_buffer_size = fstream->optimal_block_size;

    return ostream;
}

struct ostream *
o_stream_create_fd_file(int fd, uoff_t offset, bool autoclose_fd)
{
    struct file_ostream *fstream;
    struct ostream *ostream;

    if (offset == (uoff_t)-1)
        offset = lseek(fd, 0, SEEK_CUR);

    fstream = o_stream_create_fd_common(fd, autoclose_fd);
    fstream_init_file(fstream);
    fstream->max_buffer_size = fstream->optimal_block_size;
    fstream->real_offset = offset;
    fstream->buffer_offset = offset;

    ostream = o_stream_create(&fstream->ostream);
    ostream->offset = offset;
    return ostream;
}