X-Git-Url: https://mattmccutchen.net/rsync/rsync.git/blobdiff_plain/c493b6b81d814db3cdd91f3bb98f5a84456756ff..ef7441669b194ba6804c2eda07beed90a29ad2b3:/io.c diff --git a/io.c b/io.c index 284e4571..bf39ff91 100644 --- a/io.c +++ b/io.c @@ -38,8 +38,8 @@ extern int bwlimit; extern size_t bwlimit_writemax; extern int io_timeout; extern int am_server; -extern int am_daemon; extern int am_sender; +extern int am_receiver; extern int am_generator; extern int msgs2stderr; extern int inc_recurse; @@ -50,6 +50,7 @@ extern int file_total; extern int file_old_total; extern int list_only; extern int read_batch; +extern int compat_flags; extern int protect_args; extern int checksum_seed; extern int protocol_version; @@ -65,10 +66,10 @@ extern iconv_t ic_send, ic_recv; int csum_length = SHORT_SUM_LENGTH; /* initial value */ int allowed_lull = 0; -int ignore_timeout = 0; int batch_fd = -1; int msgdone_cnt = 0; int forward_flist_data = 0; +BOOL flist_receiving_enabled = False; /* Ignore an EOF error if non-zero. See whine_about_eof(). */ int kluge_around_eof = 0; @@ -83,7 +84,7 @@ static struct { xbuf in, out, msg; int in_fd; int out_fd; /* Both "out" and "msg" go to this fd. */ - BOOL in_multiplexed; + int in_multiplexed; unsigned out_empty_len; size_t raw_data_header_pos; /* in the out xbuf */ size_t raw_flushing_ends_before; /* in the out xbuf */ @@ -115,7 +116,19 @@ static char int_byte_extra[64] = { 2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 5, 6, /* (C0 - FF)/4 */ }; -#define IN_MULTIPLEXED (iobuf.in_multiplexed) +/* Our I/O buffers are sized with no bits on in the lowest byte of the "size" + * (indeed, our rounding of sizes in 1024-byte units assures more than this). + * This allows the code that is storing bytes near the physical end of a + * circular buffer to temporarily reduce the buffer's size (in order to make + * some storing idioms easier), while also making it simple to restore the + * buffer's actual size when the buffer's "pos" wraps around to the start (we + * just round the buffer's size up again). */ + +#define IOBUF_WAS_REDUCED(siz) ((siz) & 0xFF) +#define IOBUF_RESTORE_SIZE(siz) (((siz) | 0xFF) + 1) + +#define IN_MULTIPLEXED (iobuf.in_multiplexed != 0) +#define IN_MULTIPLEXED_AND_READY (iobuf.in_multiplexed > 0) #define OUT_MULTIPLEXED (iobuf.out_empty_len != 0) #define PIO_NEED_INPUT (1<<0) /* The *_NEED_* flags are mutually exclusive. */ @@ -136,27 +149,47 @@ enum festatus { FES_SUCCESS, FES_REDO, FES_NO_SEND }; static flist_ndx_list redo_list, hlink_list; +static void read_a_msg(void); +static void drain_multiplex_messages(void); static void sleep_for_bwlimit(int bytes_written); -static void check_timeout(void) +static void check_timeout(BOOL allow_keepalive) { - time_t t; - - if (!io_timeout || ignore_timeout) + time_t t, chk; + + /* On the receiving side, the generator is now the one that decides + * when a timeout has occurred. When it is sifting through a lot of + * files looking for work, it will be sending keep-alive messages to + * the sender, and even though the receiver won't be sending/receiving + * anything (not even keep-alive messages), the successful writes to + * the sender will keep things going. If the receiver is actively + * receiving data, it will ensure that the generator knows that it is + * not idle by sending the generator keep-alive messages (since the + * generator might be blocked trying to send checksums, it needs to + * know that the receiver is active). Thus, as long as one or the + * other is successfully doing work, the generator will not timeout. */ + if (!io_timeout) return; - if (!last_io_in) { - last_io_in = time(NULL); - return; + t = time(NULL); + + if (allow_keepalive) { + /* This may put data into iobuf.msg w/o flushing. */ + maybe_send_keepalive(t, 0); } - t = time(NULL); + if (!last_io_in) + last_io_in = t; - if (t - last_io_in >= io_timeout) { - if (!am_server && !am_daemon) { - rprintf(FERROR, "io timeout after %d seconds -- exiting\n", - (int)(t-last_io_in)); - } + if (am_receiver) + return; + + chk = MAX(last_io_out, last_io_in); + if (t - chk >= io_timeout) { + if (am_server) + msgs2stderr = 1; + rprintf(FERROR, "[%s] io timeout after %d seconds -- exiting\n", + who_am_i(), (int)(t-chk)); exit_cleanup(RERR_TIMEOUT); } } @@ -173,9 +206,9 @@ static void check_timeout(void) * There is another case for older protocol versions (< 24) where the module * listing was not terminated, so we must ignore an EOF error in that case and * exit. In this situation, kluge_around_eof will be > 0. */ -static NORETURN void whine_about_eof(int fd) +static NORETURN void whine_about_eof(BOOL allow_kluge) { - if (kluge_around_eof && fd == sock_f_in) { + if (kluge_around_eof && allow_kluge) { int i; if (kluge_around_eof > 0) exit_cleanup(0); @@ -238,7 +271,8 @@ static size_t safe_read(int fd, char *buf, size_t len) who_am_i()); exit_cleanup(RERR_FILEIO); } - check_timeout(); + if (io_timeout) + maybe_send_keepalive(time(NULL), MSK_ALLOW_FLUSH); continue; } @@ -322,7 +356,8 @@ static void safe_write(int fd, const char *buf, size_t len) what_fd_is(fd), who_am_i()); exit_cleanup(RERR_FILEIO); } - check_timeout(); + if (io_timeout) + maybe_send_keepalive(time(NULL), MSK_ALLOW_FLUSH); continue; } @@ -442,23 +477,59 @@ static void forward_filesfrom_data(void) } } -/* Perform buffered input and output until specified conditions are met. When - * given a "needed" read requirement, we'll return without doing any I/O if the - * iobuf.in bytes are already available. When reading, we'll read as many - * bytes as we can into the buffer, and return as soon as we meet the minimum - * read requirement. When given a "needed" write requirement, we'll return - * without doing any I/O if that many bytes will fit in the output buffer (we - * check either iobuf.out or iobuf.msg, depending on the flags). When writing, - * we write out as much as we can, and return as soon as the given free-space - * requirement is available. +void reduce_iobuf_size(xbuf *out, size_t new_size) +{ + if (new_size < out->size) { + if (DEBUG_GTE(IO, 4)) { + const char *name = out == &iobuf.out ? "iobuf.out" + : out == &iobuf.msg ? "iobuf.msg" + : NULL; + if (name) { + rprintf(FINFO, "[%s] reduced size of %s (-%d)\n", + who_am_i(), name, (int)(out->size - new_size)); + } + } + out->size = new_size; + } +} + +void restore_iobuf_size(xbuf *out) +{ + if (IOBUF_WAS_REDUCED(out->size)) { + size_t new_size = IOBUF_RESTORE_SIZE(out->size); + if (DEBUG_GTE(IO, 4)) { + const char *name = out == &iobuf.out ? "iobuf.out" + : out == &iobuf.msg ? "iobuf.msg" + : NULL; + if (name) { + rprintf(FINFO, "[%s] restored size of %s (+%d)\n", + who_am_i(), name, (int)(new_size - out->size)); + } + } + out->size = new_size; + } +} + +/* Perform buffered input and/or output until specified conditions are met. + * When given a "needed" read or write request, this returns without doing any + * I/O if the needed input bytes or write space is already available. Once I/O + * is needed, this will try to do whatever reading and/or writing is currently + * possible, up to the maximum buffer allowances, no matter if this is a read + * or write request. However, the I/O stops as soon as the required input + * bytes or output space is available. If this is not a read request, the + * routine may also do some advantageous reading of messages from a multiplexed + * input source (which ensures that we don't jam up with everyone in their + * "need to write" code and nobody reading the accumulated data that would make + * writing possible). * - * The iobuf.out and iobuf.msg buffers are circular, so some writes into them - * will need to be split when the data needs to wrap around to the start. In - * order to help make this easier for some operations (such as the use of - * SIVAL() into the buffer) the buffers MUST have 4 bytes of overflow space at - * the end that is not not counted in the "size". The iobuf.in buffer is not - * (currently) circular. To facilitate the handling of MSG_DATA bytes as they - * are read-from/written-into the buffers, see the three raw_* iobuf vars. + * The iobuf.in, .out and .msg buffers are all circular. Callers need to be + * aware that some data copies will need to be split when the bytes wrap around + * from the end to the start. In order to help make writing into the output + * buffers easier for some operations (such as the use of SIVAL() into the + * buffer) a buffer may be temporarily shortened by a small amount, but the + * original size will be automatically restored when the .pos wraps to the + * start. See also the 3 raw_* iobuf vars that are used in the handling of + * MSG_DATA bytes as they are read-from/written-into the buffers. * * When writing, we flush data in the following priority order: * @@ -475,8 +546,6 @@ static void forward_filesfrom_data(void) * * - Make this routine able to read the generator-to-receiver batch flow? * - * - Make the input buffer circular? - * * Unlike the old routines that this replaces, it is OK to read ahead as far as * we can because the read_a_msg() routine now reads its bytes out of the input * buffer. In the old days, only raw data was in the input buffer, and any @@ -498,36 +567,17 @@ static char *perform_io(size_t needed, int flags) switch (flags & PIO_NEED_FLAGS) { case PIO_NEED_INPUT: + /* We never resize the circular input buffer. */ + if (iobuf.in.size < needed) { + rprintf(FERROR, "need to read %ld bytes, iobuf.in.buf is only %ld bytes.\n", + (long)needed, (long)iobuf.in.size); + exit_cleanup(RERR_PROTOCOL); + } + if (DEBUG_GTE(IO, 3)) { rprintf(FINFO, "[%s] perform_io(%ld, %sinput)\n", who_am_i(), (long)needed, flags & PIO_CONSUME_INPUT ? "consume&" : ""); } - - /* Make sure the input buffer is big enough to hold "needed" bytes. - * Also make sure it will fit in the free space at the end, or - * else we need to shift some bytes. */ - if (needed && iobuf.in.size < needed) { - if (!(iobuf.in.buf = realloc_array(iobuf.in.buf, char, needed))) - out_of_memory("perform_io"); - if (DEBUG_GTE(IO, 4)) { - rprintf(FINFO, "[%s] resized input buffer from %ld to %ld bytes.\n", - who_am_i(), (long)iobuf.in.size, (long)needed); - } - iobuf.in.size = needed; - } - if (iobuf.in.size - iobuf.in.pos < needed - || (iobuf.in.len < needed && iobuf.in.len < 1024 - && iobuf.in.size - (iobuf.in.pos + iobuf.in.len) < 1024)) { - memmove(iobuf.in.buf, iobuf.in.buf + iobuf.in.pos, iobuf.in.len); - if (DEBUG_GTE(IO, 4)) { - rprintf(FINFO, - "[%s] moved %ld bytes from %ld to 0 in the input buffer (size=%ld, needed=%ld).\n", - who_am_i(), (long)iobuf.in.len, (long)iobuf.in.pos, (long)iobuf.in.size, (long)needed); - } - if (iobuf.raw_input_ends_before) - iobuf.raw_input_ends_before -= iobuf.in.pos; - iobuf.in.pos = 0; - } break; case PIO_NEED_OUTROOM: @@ -593,7 +643,7 @@ static char *perform_io(size_t needed, int flags) FD_ZERO(&r_fds); FD_ZERO(&e_fds); - if (iobuf.in_fd >= 0 && iobuf.in.size - (iobuf.in.pos + iobuf.in.len)) { + if (iobuf.in_fd >= 0 && iobuf.in.size - iobuf.in.len) { if (!read_batch || batch_fd >= 0) { FD_SET(iobuf.in_fd, &r_fds); FD_SET(iobuf.in_fd, &e_fds); @@ -620,15 +670,6 @@ static char *perform_io(size_t needed, int flags) SIVAL(iobuf.out.buf + iobuf.raw_data_header_pos, 0, ((MPLEX_BASE + (int)MSG_DATA)<<24) + iobuf.out.len - 4); - if (iobuf.raw_data_header_pos + 4 > iobuf.out.size) { - int siz = (int)(iobuf.raw_data_header_pos + 4 - iobuf.out.size); - /* We used some of the overflow bytes, so move them. */ - if (DEBUG_GTE(IO, 4)) { - rprintf(FINFO, "[%s] wrap-bytes moved: %d (perform_io)\n", - who_am_i(), siz); - } - memcpy(iobuf.out.buf, iobuf.out.buf + iobuf.out.size, siz); - } if (DEBUG_GTE(IO, 1)) { rprintf(FINFO, "[%s] send_msg(%d, %ld)\n", @@ -639,6 +680,13 @@ static char *perform_io(size_t needed, int flags) iobuf.raw_data_header_pos = iobuf.raw_flushing_ends_before; if (iobuf.raw_data_header_pos >= iobuf.out.size) iobuf.raw_data_header_pos -= iobuf.out.size; + else if (iobuf.raw_data_header_pos + 4 > iobuf.out.size) { + /* The 4-byte header won't fit at the end of the buffer, + * so we'll temporarily reduce the output buffer's size + * and put the header at the start of the buffer. */ + reduce_iobuf_size(&iobuf.out, iobuf.raw_data_header_pos); + iobuf.raw_data_header_pos = 0; + } /* Yes, it is possible for this to make len > size for a while. */ iobuf.out.len += 4; } @@ -661,11 +709,19 @@ static char *perform_io(size_t needed, int flags) if (max_fd < 0) { switch (flags & PIO_NEED_FLAGS) { case PIO_NEED_INPUT: + iobuf.in.len = 0; + if (kluge_around_eof == 2) + exit_cleanup(0); + if (iobuf.in_fd == -2) + whine_about_eof(True); rprintf(FERROR, "error in perform_io: no fd for input.\n"); exit_cleanup(RERR_PROTOCOL); case PIO_NEED_OUTROOM: case PIO_NEED_MSGROOM: msgs2stderr = 1; + drain_multiplex_messages(); + if (iobuf.out_fd == -2) + whine_about_eof(True); rprintf(FERROR, "error in perform_io: no fd for output.\n"); exit_cleanup(RERR_PROTOCOL); default: @@ -676,8 +732,7 @@ static char *perform_io(size_t needed, int flags) } if (extra_flist_sending_enabled) { - if (file_total - file_old_total < MAX_FILECNT_LOOKAHEAD - && file_total - file_old_total >= MIN_FILECNT_LOOKAHEAD) + if (file_total - file_old_total < MAX_FILECNT_LOOKAHEAD) tv.tv_sec = 0; else { extra_flist_sending_enabled = False; @@ -699,19 +754,24 @@ static char *perform_io(size_t needed, int flags) send_extra_file_list(sock_f_out, -1); extra_flist_sending_enabled = !flist_eof; } else - check_timeout(); + check_timeout((flags & PIO_NEED_INPUT) != 0); FD_ZERO(&r_fds); /* Just in case... */ FD_ZERO(&w_fds); } if (iobuf.in_fd >= 0 && FD_ISSET(iobuf.in_fd, &r_fds)) { - size_t pos = iobuf.in.pos + iobuf.in.len; - size_t len = iobuf.in.size - pos; + size_t len, pos = iobuf.in.pos + iobuf.in.len; int n; + if (pos >= iobuf.in.size) { + pos -= iobuf.in.size; + len = iobuf.in.size - iobuf.in.len; + } else + len = iobuf.in.size - pos; if ((n = read(iobuf.in_fd, iobuf.in.buf + pos, len)) <= 0) { if (n == 0) { + /* Signal that input has become invalid. */ if (!read_batch || batch_fd < 0 || am_generator) - whine_about_eof(iobuf.in_fd); /* Doesn't return. */ + iobuf.in_fd = -2; batch_fd = -1; continue; } @@ -725,20 +785,23 @@ static char *perform_io(size_t needed, int flags) rsyserr(FERROR_SOCKET, errno, "read error"); } else rsyserr(FERROR, errno, "read error"); - exit_cleanup(RERR_STREAMIO); + exit_cleanup(RERR_SOCKETIO); } } if (msgs2stderr && DEBUG_GTE(IO, 2)) rprintf(FINFO, "[%s] recv=%ld\n", who_am_i(), (long)n); - if (io_timeout) + if (io_timeout) { last_io_in = time(NULL); + if (flags & PIO_NEED_INPUT) + maybe_send_keepalive(last_io_in, 0); + } stats.total_read += n; iobuf.in.len += n; } - if (iobuf.out_fd >= 0 && FD_ISSET(iobuf.out_fd, &w_fds)) { + if (out && FD_ISSET(iobuf.out_fd, &w_fds)) { size_t len = iobuf.raw_flushing_ends_before ? iobuf.raw_flushing_ends_before - out->pos : out->len; int n; @@ -753,9 +816,11 @@ static char *perform_io(size_t needed, int flags) else { /* Don't write errors on a dead socket. */ msgs2stderr = 1; - out->len = iobuf.raw_flushing_ends_before = out->pos = 0; + iobuf.out_fd = -2; + iobuf.out.len = iobuf.msg.len = iobuf.raw_flushing_ends_before = 0; rsyserr(FERROR_SOCKET, errno, "[%s] write error", who_am_i()); - exit_cleanup(RERR_STREAMIO); + drain_multiplex_messages(); + exit_cleanup(RERR_SOCKETIO); } } if (msgs2stderr && DEBUG_GTE(IO, 2)) { @@ -774,16 +839,26 @@ static char *perform_io(size_t needed, int flags) if (iobuf.raw_flushing_ends_before) iobuf.raw_flushing_ends_before -= out->size; out->pos = 0; - } - if (out->pos == iobuf.raw_flushing_ends_before) + restore_iobuf_size(out); + } else if (out->pos == iobuf.raw_flushing_ends_before) iobuf.raw_flushing_ends_before = 0; if ((out->len -= n) == empty_buf_len) { out->pos = 0; + restore_iobuf_size(out); if (empty_buf_len) iobuf.raw_data_header_pos = 0; } } + /* We need to help prevent deadlock by doing what reading + * we can whenever we are here trying to write. */ + if (IN_MULTIPLEXED_AND_READY && !(flags & PIO_NEED_INPUT)) { + while (!iobuf.raw_input_ends_before && iobuf.in.len > 512) + read_a_msg(); + if (flist_receiving_enabled && iobuf.in.len > 512) + wait_for_receiver(); /* generator only */ + } + if (ff_forward_fd >= 0 && FD_ISSET(ff_forward_fd, &r_fds)) { /* This can potentially flush all output and enable * multiplexed output, so keep this last in the loop @@ -799,27 +874,59 @@ static char *perform_io(size_t needed, int flags) if (flags & PIO_CONSUME_INPUT) { iobuf.in.len -= needed; iobuf.in.pos += needed; + if (iobuf.in.pos == iobuf.raw_input_ends_before) + iobuf.raw_input_ends_before = 0; + if (iobuf.in.pos >= iobuf.in.size) { + iobuf.in.pos -= iobuf.in.size; + if (iobuf.raw_input_ends_before) + iobuf.raw_input_ends_before -= iobuf.in.size; + } } return data; } +static void raw_read_buf(char *buf, size_t len) +{ + size_t pos = iobuf.in.pos; + char *data = perform_io(len, PIO_INPUT_AND_CONSUME); + if (iobuf.in.pos <= pos && len) { + size_t siz = len - iobuf.in.pos; + memcpy(buf, data, siz); + memcpy(buf + siz, iobuf.in.buf, iobuf.in.pos); + } else + memcpy(buf, data, len); +} + +static int32 raw_read_int(void) +{ + char *data, buf[4]; + if (iobuf.in.size - iobuf.in.pos >= 4) + data = perform_io(4, PIO_INPUT_AND_CONSUME); + else + raw_read_buf(data = buf, 4); + return IVAL(data, 0); +} + void noop_io_until_death(void) { char buf[1024]; - kluge_around_eof = 1; - set_io_timeout(protocol_version >= 31 ? 10 : 1); + kluge_around_eof = 2; + /* Setting an I/O timeout ensures that if something inexplicably weird + * happens, we won't hang around forever. */ + if (!io_timeout) + set_io_timeout(60); while (1) read_buf(iobuf.in_fd, buf, sizeof buf); } -/* Buffer a message for the multiplexed output stream. Is never used for MSG_DATA. */ +/* Buffer a message for the multiplexed output stream. Is not used for (normal) MSG_DATA. */ int send_msg(enum msgcode code, const char *buf, size_t len, int convert) { char *hdr; - size_t pos; + size_t needed, pos; BOOL want_debug = DEBUG_GTE(IO, 1) && convert >= 0 && (msgs2stderr || code != MSG_INFO); if (!OUT_MULTIPLEXED) @@ -828,24 +935,35 @@ int send_msg(enum msgcode code, const char *buf, size_t len, int convert) if (want_debug) rprintf(FINFO, "[%s] send_msg(%d, %ld)\n", who_am_i(), (int)code, (long)len); + /* When checking for enough free space for this message, we need to + * make sure that there is space for the 4-byte header, plus we'll + * assume that we may waste up to 3 bytes (if the header doesn't fit + * at the physical end of the buffer). */ #ifdef ICONV_OPTION if (convert > 0 && ic_send == (iconv_t)-1) convert = 0; if (convert > 0) { - /* Ensuring double-size room leaves space for a potential conversion. */ - if (iobuf.msg.len + len*2 + 4 > iobuf.msg.size) - perform_io(len*2 + 4, PIO_NEED_MSGROOM); + /* Ensuring double-size room leaves space for maximal conversion expansion. */ + needed = len*2 + 4 + 3; } else #endif - if (iobuf.msg.len + len + 4 > iobuf.msg.size) - perform_io(len + 4, PIO_NEED_MSGROOM); + needed = len + 4 + 3; + if (iobuf.msg.len + needed > iobuf.msg.size) + perform_io(needed, PIO_NEED_MSGROOM); pos = iobuf.msg.pos + iobuf.msg.len; /* Must be set after any flushing. */ if (pos >= iobuf.msg.size) pos -= iobuf.msg.size; + else if (pos + 4 > iobuf.msg.size) { + /* The 4-byte header won't fit at the end of the buffer, + * so we'll temporarily reduce the message buffer's size + * and put the header at the start of the buffer. */ + reduce_iobuf_size(&iobuf.msg, pos); + pos = 0; + } hdr = iobuf.msg.buf + pos; - iobuf.msg.len += 4; /* Leave room for the coming header bytes. */ + iobuf.msg.len += 4; /* Allocate room for the coming header bytes. */ #ifdef ICONV_OPTION if (convert > 0) { @@ -866,8 +984,8 @@ int send_msg(enum msgcode code, const char *buf, size_t len, int convert) { size_t siz; - if ((pos += 4) >= iobuf.msg.size) - pos -= iobuf.msg.size; + if ((pos += 4) == iobuf.msg.size) + pos = 0; /* Handle a split copy if we wrap around the end of the circular buffer. */ if (pos >= iobuf.msg.pos && (siz = iobuf.msg.size - pos) < len) { @@ -880,13 +998,6 @@ int send_msg(enum msgcode code, const char *buf, size_t len, int convert) } SIVAL(hdr, 0, ((MPLEX_BASE + (int)code)<<24) + len); - /* If the header used any overflow bytes, move them to the start. */ - if ((pos = hdr+4 - iobuf.msg.buf) > iobuf.msg.size) { - int siz = (int)(pos - iobuf.msg.size); - if (DEBUG_GTE(IO, 4)) - rprintf(FINFO, "[%s] wrap-bytes moved: %d (send_msg)\n", who_am_i(), siz); - memcpy(iobuf.msg.buf, iobuf.msg.buf + iobuf.msg.size, siz); - } if (want_debug && convert > 0) rprintf(FINFO, "[%s] converted msg len=%ld\n", who_am_i(), (long)len); @@ -955,13 +1066,15 @@ void io_set_sock_fds(int f_in, int f_out) void set_io_timeout(int secs) { io_timeout = secs; + allowed_lull = (io_timeout + 1) / 2; - if (!io_timeout || io_timeout > SELECT_TIMEOUT) + if (!io_timeout || allowed_lull > SELECT_TIMEOUT) select_timeout = SELECT_TIMEOUT; else - select_timeout = io_timeout; + select_timeout = allowed_lull; - allowed_lull = read_batch ? 0 : (io_timeout + 1) / 2; + if (read_batch) + allowed_lull = 0; } static void check_for_d_option_error(const char *msg) @@ -1156,25 +1269,15 @@ BOOL io_start_buffering_out(int f_out) if (msgs2stderr && DEBUG_GTE(IO, 2)) rprintf(FINFO, "[%s] io_start_buffering_out(%d)\n", who_am_i(), f_out); - if (OUT_MULTIPLEXED && !iobuf.msg.buf) { - iobuf.msg.size = IO_BUFFER_SIZE - 4; - if (!(iobuf.msg.buf = new_array(char, iobuf.msg.size + 4))) - out_of_memory("io_start_buffering_out"); - iobuf.msg.pos = iobuf.msg.len = 0; - } - if (iobuf.out.buf) { if (iobuf.out_fd == -1) iobuf.out_fd = f_out; - assert(f_out == iobuf.out_fd); + else + assert(f_out == iobuf.out_fd); return False; } - iobuf.out.size = IO_BUFFER_SIZE * 2 - 4; - /* The 4 overflow bytes makes some circular-buffer wrapping operations easier. */ - if (!(iobuf.out.buf = new_array(char, iobuf.out.size + 4))) - out_of_memory("io_start_buffering_out"); - iobuf.out.pos = iobuf.out.len = 0; + alloc_xbuf(&iobuf.out, ROUND_UP_1024(IO_BUFFER_SIZE * 2)); iobuf.out_fd = f_out; return True; @@ -1188,16 +1291,12 @@ BOOL io_start_buffering_in(int f_in) if (iobuf.in.buf) { if (iobuf.in_fd == -1) iobuf.in_fd = f_in; - assert(f_in == iobuf.in_fd); + else + assert(f_in == iobuf.in_fd); return False; } - iobuf.in.size = IO_BUFFER_SIZE; - if (!(iobuf.in.buf = new_array(char, iobuf.in.size))) - out_of_memory("io_start_buffering_in"); - - iobuf.in.pos = iobuf.in.len = 0; - + alloc_xbuf(&iobuf.in, ROUND_UP_1024(IO_BUFFER_SIZE)); iobuf.in_fd = f_in; return True; @@ -1242,20 +1341,29 @@ void maybe_flush_socket(int important) io_flush(NORMAL_FLUSH); } -void maybe_send_keepalive(void) +/* Older rsync versions used to send either a MSG_NOOP (protocol 30) or a + * raw-data-based keep-alive (protocol 29), both of which implied forwarding of + * the message through the sender. Since the new timeout method does not need + * any forwarding, we just send an empty MSG_DATA message, which works with all + * rsync versions. This avoids any message forwarding, and leaves the raw-data + * stream alone (since we can never be quite sure if that stream is in the + * right state for a keep-alive message). */ +void maybe_send_keepalive(time_t now, int flags) { - if (time(NULL) - last_io_out >= allowed_lull) { - if (!iobuf.msg.len && iobuf.out.len == iobuf.out_empty_len) { - if (protocol_version < 29) - return; /* there's nothing we can do */ - if (protocol_version >= 30) - send_msg(MSG_NOOP, "", 0, 0); - else { - write_int(iobuf.out_fd, cur_flist->used); - write_shortint(iobuf.out_fd, ITEM_IS_NEW); - } - } - if (iobuf.msg.len) + if (flags & MSK_ACTIVE_RECEIVER) + last_io_in = now; /* Fudge things when we're working hard on the files. */ + + if (now - last_io_out >= allowed_lull) { + /* The receiver is special: it only sends keep-alive messages if it is + * actively receiving data. Otherwise, it lets the generator timeout. */ + if (am_receiver && now - last_io_in >= io_timeout) + return; + + if (!iobuf.msg.len && iobuf.out.len == iobuf.out_empty_len) + send_msg(MSG_DATA, "", 0, 0); + if (!(flags & MSK_ALLOW_FLUSH)) { + /* Let the caller worry about writing out the data. */ + } else if (iobuf.msg.len) perform_io(iobuf.msg.size - iobuf.msg.len + 1, PIO_NEED_MSGROOM); else if (iobuf.out.len > iobuf.out_empty_len) io_flush(NORMAL_FLUSH); @@ -1276,17 +1384,22 @@ void stop_flist_forward(void) /* Read a message from a multiplexed source. */ static void read_a_msg(void) { - char *data, line[BIGPATHBUFLEN]; + char data[BIGPATHBUFLEN]; int tag, val; size_t msg_bytes; - data = perform_io(4, PIO_INPUT_AND_CONSUME); - tag = IVAL(data, 0); + /* This ensures that perform_io() does not try to do any message reading + * until we've read all of the data for this message. We should also + * try to avoid calling things that will cause data to be written via + * perform_io() prior to this being reset to 1. */ + iobuf.in_multiplexed = -1; + + tag = raw_read_int(); msg_bytes = tag & 0xFFFFFF; tag = (tag >> 24) - MPLEX_BASE; - if (DEBUG_GTE(IO, 1) && (msgs2stderr || tag != MSG_INFO)) + if (DEBUG_GTE(IO, 1) && msgs2stderr) rprintf(FINFO, "[%s] got msg=%d, len=%ld\n", who_am_i(), (int)tag, (long)msg_bytes); switch (tag) { @@ -1296,35 +1409,38 @@ static void read_a_msg(void) * the buffer the msg data will end once it is read. It is * possible that this points off the end of the buffer, in * which case the gradual reading of the input stream will - * cause this value to decrease and eventually become real. */ - iobuf.raw_input_ends_before = iobuf.in.pos + msg_bytes; + * cause this value to wrap around and eventually become real. */ + if (msg_bytes) + iobuf.raw_input_ends_before = iobuf.in.pos + msg_bytes; + iobuf.in_multiplexed = 1; break; case MSG_STATS: if (msg_bytes != sizeof stats.total_read || !am_generator) goto invalid_msg; - data = perform_io(sizeof stats.total_read, PIO_INPUT_AND_CONSUME); - memcpy((char*)&stats.total_read, data, sizeof stats.total_read); + raw_read_buf((char*)&stats.total_read, sizeof stats.total_read); + iobuf.in_multiplexed = 1; break; case MSG_REDO: if (msg_bytes != 4 || !am_generator) goto invalid_msg; - data = perform_io(4, PIO_INPUT_AND_CONSUME); - got_flist_entry_status(FES_REDO, IVAL(data, 0)); + val = raw_read_int(); + iobuf.in_multiplexed = 1; + got_flist_entry_status(FES_REDO, val); break; case MSG_IO_ERROR: - if (msg_bytes != 4 || am_sender) + if (msg_bytes != 4) goto invalid_msg; - data = perform_io(4, PIO_INPUT_AND_CONSUME); - val = IVAL(data, 0); + val = raw_read_int(); + iobuf.in_multiplexed = 1; io_error |= val; - if (!am_generator) + if (am_receiver) send_msg_int(MSG_IO_ERROR, val); break; case MSG_IO_TIMEOUT: if (msg_bytes != 4 || am_server || am_generator) goto invalid_msg; - data = perform_io(4, PIO_INPUT_AND_CONSUME); - val = IVAL(data, 0); + val = raw_read_int(); + iobuf.in_multiplexed = 1; if (!io_timeout || io_timeout > val) { if (INFO_GTE(MISC, 2)) rprintf(FINFO, "Setting --timeout=%d to match server\n", val); @@ -1332,15 +1448,20 @@ static void read_a_msg(void) } break; case MSG_NOOP: + /* Support protocol-30 keep-alive method. */ + if (msg_bytes != 0) + goto invalid_msg; + iobuf.in_multiplexed = 1; if (am_sender) - maybe_send_keepalive(); + maybe_send_keepalive(time(NULL), MSK_ALLOW_FLUSH); break; case MSG_DELETED: - if (msg_bytes >= sizeof line) + if (msg_bytes >= sizeof data) goto overflow; if (am_generator) { - memcpy(line, perform_io(msg_bytes, PIO_INPUT_AND_CONSUME), msg_bytes); - send_msg(MSG_DELETED, line, msg_bytes, 1); + raw_read_buf(data, msg_bytes); + iobuf.in_multiplexed = 1; + send_msg(MSG_DELETED, data, msg_bytes, 1); break; } #ifdef ICONV_OPTION @@ -1350,12 +1471,12 @@ static void read_a_msg(void) int add_null = 0; int flags = ICB_INCLUDE_BAD | ICB_INIT; - INIT_CONST_XBUF(outbuf, line); + INIT_CONST_XBUF(outbuf, data); INIT_XBUF(inbuf, ibuf, 0, (size_t)-1); while (msg_bytes) { size_t len = msg_bytes > sizeof ibuf - inbuf.len ? sizeof ibuf - inbuf.len : msg_bytes; - memcpy(ibuf + inbuf.len, perform_io(len, PIO_INPUT_AND_CONSUME), len); + raw_read_buf(ibuf + inbuf.len, len); inbuf.pos = 0; inbuf.len += len; if (!(msg_bytes -= len) && !ibuf[inbuf.len-1]) @@ -1377,13 +1498,14 @@ static void read_a_msg(void) msg_bytes = outbuf.len; } else #endif - memcpy(line, perform_io(msg_bytes, PIO_INPUT_AND_CONSUME), msg_bytes); + raw_read_buf(data, msg_bytes); + iobuf.in_multiplexed = 1; /* A directory name was sent with the trailing null */ - if (msg_bytes > 0 && !line[msg_bytes-1]) - log_delete(line, S_IFDIR); + if (msg_bytes > 0 && !data[msg_bytes-1]) + log_delete(data, S_IFDIR); else { - line[msg_bytes] = '\0'; - log_delete(line, S_IFREG); + data[msg_bytes] = '\0'; + log_delete(data, S_IFREG); } break; case MSG_SUCCESS: @@ -1394,8 +1516,8 @@ static void read_a_msg(void) inc_recurse ? "/inc" : ""); exit_cleanup(RERR_STREAMIO); } - data = perform_io(4, PIO_INPUT_AND_CONSUME); - val = IVAL(data, 0); + val = raw_read_int(); + iobuf.in_multiplexed = 1; if (am_generator) got_flist_entry_status(FES_SUCCESS, val); else @@ -1404,8 +1526,8 @@ static void read_a_msg(void) case MSG_NO_SEND: if (msg_bytes != 4) goto invalid_msg; - data = perform_io(4, PIO_INPUT_AND_CONSUME); - val = IVAL(data, 0); + val = raw_read_int(); + iobuf.in_multiplexed = 1; if (am_generator) got_flist_entry_status(FES_NO_SEND, val); else @@ -1424,7 +1546,7 @@ static void read_a_msg(void) case MSG_ERROR: case MSG_ERROR_XFER: case MSG_WARNING: - if (msg_bytes >= sizeof line) { + if (msg_bytes >= sizeof data) { overflow: rprintf(FERROR, "multiplexing overflow %d:%lu [%s%s]\n", @@ -1432,34 +1554,51 @@ static void read_a_msg(void) inc_recurse ? "/inc" : ""); exit_cleanup(RERR_STREAMIO); } - memcpy(line, perform_io(msg_bytes, PIO_INPUT_AND_CONSUME), msg_bytes); - rwrite((enum logcode)tag, line, msg_bytes, !am_generator); + raw_read_buf(data, msg_bytes); + iobuf.in_multiplexed = 1; + rwrite((enum logcode)tag, data, msg_bytes, !am_generator); if (first_message) { - if (list_only && !am_sender && tag == 1 && msg_bytes < sizeof line) { - line[msg_bytes] = '\0'; - check_for_d_option_error(line); + if (list_only && !am_sender && tag == 1 && msg_bytes < sizeof data) { + data[msg_bytes] = '\0'; + check_for_d_option_error(data); } first_message = 0; } break; case MSG_ERROR_EXIT: + if (msg_bytes == 4) + val = raw_read_int(); + else if (msg_bytes == 0) + val = 0; + else + goto invalid_msg; + iobuf.in_multiplexed = 1; + if (DEBUG_GTE(EXIT, 3)) + rprintf(FINFO, "[%s] got MSG_ERROR_EXIT with %ld bytes\n", who_am_i(), (long)msg_bytes); if (msg_bytes == 0) { if (!am_sender && !am_generator) { + if (DEBUG_GTE(EXIT, 3)) { + rprintf(FINFO, "[%s] sending MSG_ERROR_EXIT (len 0)\n", + who_am_i()); + } send_msg(MSG_ERROR_EXIT, "", 0, 0); io_flush(FULL_FLUSH); } - val = 0; - } else if (msg_bytes == 4) { - data = perform_io(4, PIO_INPUT_AND_CONSUME); - val = IVAL(data, 0); - if (protocol_version >= 31) { - if (am_generator) - send_msg_int(MSG_ERROR_EXIT, val); - else - send_msg(MSG_ERROR_EXIT, "", 0, 0); + } else if (protocol_version >= 31) { + if (am_generator) { + if (DEBUG_GTE(EXIT, 3)) { + rprintf(FINFO, "[%s] sending MSG_ERROR_EXIT with exit_code %d\n", + who_am_i(), val); + } + send_msg_int(MSG_ERROR_EXIT, val); + } else { + if (DEBUG_GTE(EXIT, 3)) { + rprintf(FINFO, "[%s] sending MSG_ERROR_EXIT (len 0)\n", + who_am_i()); + } + send_msg(MSG_ERROR_EXIT, "", 0, 0); } - } else - goto invalid_msg; + } /* Send a negative linenum so that we don't end up * with a duplicate exit message. */ _exit_cleanup(val, __FILE__, 0 - __LINE__); @@ -1468,6 +1607,26 @@ static void read_a_msg(void) tag, who_am_i(), inc_recurse ? "/inc" : ""); exit_cleanup(RERR_STREAMIO); } + + assert(iobuf.in_multiplexed > 0); +} + +static void drain_multiplex_messages(void) +{ + while (IN_MULTIPLEXED_AND_READY && iobuf.in.len) { + if (iobuf.raw_input_ends_before) { + size_t raw_len = iobuf.raw_input_ends_before - iobuf.in.pos; + iobuf.raw_input_ends_before = 0; + if (raw_len >= iobuf.in.len) { + iobuf.in.len = 0; + break; + } + iobuf.in.len -= raw_len; + if ((iobuf.in.pos += raw_len) >= iobuf.in.size) + iobuf.in.pos -= iobuf.in.size; + } + read_a_msg(); + } } void wait_for_receiver(void) @@ -1492,6 +1651,7 @@ void wait_for_receiver(void) } } else { struct file_list *flist; + flist_receiving_enabled = False; if (DEBUG_GTE(FLIST, 2)) { rprintf(FINFO, "[%s] receiving flist for dir %d\n", who_am_i(), ndx); @@ -1502,6 +1662,7 @@ void wait_for_receiver(void) if (preserve_hard_links) match_hard_links(flist); #endif + flist_receiving_enabled = True; } } } @@ -1623,12 +1784,12 @@ void read_buf(int f, char *buf, size_t len) { if (f != iobuf.in_fd) { if (safe_read(f, buf, len) != len) - whine_about_eof(f); /* Doesn't return. */ + whine_about_eof(False); /* Doesn't return. */ goto batch_copy; } if (!IN_MULTIPLEXED) { - memcpy(buf, perform_io(len, PIO_INPUT_AND_CONSUME), len); + raw_read_buf(buf, len); total_data_read += len; if (forward_flist_data) write_buf(iobuf.out_fd, buf, len); @@ -1639,20 +1800,15 @@ void read_buf(int f, char *buf, size_t len) } while (1) { - char *data; size_t siz; while (!iobuf.raw_input_ends_before) read_a_msg(); siz = MIN(len, iobuf.raw_input_ends_before - iobuf.in.pos); - data = perform_io(siz, PIO_INPUT_AND_CONSUME); - if (iobuf.in.pos == iobuf.raw_input_ends_before) - iobuf.raw_input_ends_before = 0; - - /* The bytes at the "data" pointer will survive long - * enough to make a copy, but not past future I/O. */ - memcpy(buf, data, siz); + if (siz >= iobuf.in.size) + siz = iobuf.in.size; + raw_read_buf(buf, siz); total_data_read += siz; if (forward_flist_data) @@ -2080,11 +2236,11 @@ void io_printf(int fd, const char *format, ...) va_end(ap); if (len < 0) - exit_cleanup(RERR_STREAMIO); + exit_cleanup(RERR_PROTOCOL); if (len > (int)sizeof buf) { rprintf(FERROR, "io_printf() was too long for the buffer.\n"); - exit_cleanup(RERR_STREAMIO); + exit_cleanup(RERR_PROTOCOL); } write_sbuf(fd, buf); @@ -2098,6 +2254,9 @@ void io_start_multiplex_out(int fd) if (msgs2stderr && DEBUG_GTE(IO, 2)) rprintf(FINFO, "[%s] io_start_multiplex_out(%d)\n", who_am_i(), fd); + if (!iobuf.msg.buf) + alloc_xbuf(&iobuf.msg, ROUND_UP_1024(IO_BUFFER_SIZE)); + iobuf.out_empty_len = 4; /* See also OUT_MULTIPLEXED */ io_start_buffering_out(fd); @@ -2111,7 +2270,7 @@ void io_start_multiplex_in(int fd) if (msgs2stderr && DEBUG_GTE(IO, 2)) rprintf(FINFO, "[%s] io_start_multiplex_in(%d)\n", who_am_i(), fd); - iobuf.in_multiplexed = True; /* See also IN_MULTIPLEXED */ + iobuf.in_multiplexed = 1; /* See also IN_MULTIPLEXED */ io_start_buffering_in(fd); } @@ -2122,7 +2281,7 @@ int io_end_multiplex_in(int mode) if (msgs2stderr && DEBUG_GTE(IO, 2)) rprintf(FINFO, "[%s] io_end_multiplex_in(mode=%d)\n", who_am_i(), mode); - iobuf.in_multiplexed = False; + iobuf.in_multiplexed = 0; if (mode == MPLX_SWITCHING) iobuf.raw_input_ends_before = 0; else @@ -2160,7 +2319,7 @@ void start_write_batch(int fd) * is involved. */ write_int(batch_fd, protocol_version); if (protocol_version >= 30) - write_byte(batch_fd, inc_recurse); + write_byte(batch_fd, compat_flags); write_int(batch_fd, checksum_seed); if (am_sender)