Input Streams
=============
'lib/istream.h' describes Dovecot's input streams. Input streams can be stacked
on top of each others as many times as wanted.
Input streams actually reading data:
* file: Read data from fd using 'pread()' for files and 'read()' for
non-files.
* mmap: Read data from file using 'mmap()'. This usually seems to be slower
than just using it with 'read()', so this input stream is probably quite
unnecessary.
* data: Read data from memory.
* plugins/zlib/zlib, bzlib: Read zlib/bzlib compressed data from fd. Some day
this should be changed to become a filter instead.
Input stream filters:
* concat: Concatenate multiple input streams together
* seekable: Make a number of (possibly non-seekable) input streams into a
single seekable input stream. If all of the input streams are already
seekable, a concat stream is created instead.
* Usually the only non-seekable input streams are non-file fds, such as
pipes or sockets.
* crlf: Change all newlines to either LFs or CRLFs, by adding or removing CRs
as necessary.
* limit: Limit input stream's length, so after reading a given number of bytes
it returns EOF.
* tee: Fork an input stream to multiple streams that can be read
independently.
* lib-mail/dot: Read SMTP-style DATA input where the input ends with an empty
"." line.
* lib-mail/header-filter: Add/remove/modify email headers.
Reading
-------
'i_stream_read()' tries to read more data into the stream's buffer. It returns:
* -2: Nothing was read, because buffer is full.
* -1: Either input reached EOF, or read failed and stream_errno was set.
* 0: Input stream is non-blocking, and no more input is available now.
* >0: Number of bytes read.
Reading from a stream doesn't actually go forward in the stream, that needs to
be done manually with 'i_stream_skip()'. This makes it easy to read full data
records into the stream directly instead of creating separate buffers. For
example when reading line-based input you can keep reading input into the
stream until you find LF and then just access the string directly from the
input buffer. There are actually helper functions for
this:'i_stream_next_line()' attempts to return the next line if available,
'i_stream_read_next_line()' does the same but does a read to try to get the
data.
Because more and more data can be read into the buffer, the buffer size is
typically limited, and once this limit is reached read returns -2. The buffer
size is usually given as parameter to the 'i_stream_create_*()', filters use
their parent stream's buffer size. The buffer size can be also changed with
'i_stream_set_max_buffer_size()'. Figuring out what the buffer size should be
depends on the situation. It should be large enough to contain all valid input,
but small enough that users can't cause a DoS by sending a too large record and
having Dovecot eat up all the memory.
Once read returns -1, the stream has reached EOF. 'stream->eof=TRUE' is also
set. In this situation it's important to remember that there may still be data
available in the buffer. If 'i_stream_have_bytes_left()' returns FALSE, there
really isn't anything left to read.
Example:
---%<-------------------------------------------------------------------------
/* read line-based data from file_fd, buffer size has no limits */
struct istream *input = i_stream_create_fd(file_fd, (size_t)-1, FALSE);
const char *line;
/* return the last line also even if it doesn't end with LF.
this is generally a good idea when reading files (but not a good idea
when reading commands from e.g. socket). */
i_stream_set_return_partial_line(input, TRUE);
while ((line = i_stream_read_next_line(input)) != NULL) {
/* handle line */
}
i_stream_destroy(&input);
---%<-------------------------------------------------------------------------
Internals
---------
'lib/istream-internal.h' describes the internal API that input streams need to
implement. The methods that need to be implemented are:
* 'read()' is the most important function. It can also be tricky to get it
completely bug-free. See the existing unit tests for other istreams and try
to test the edge cases as well (such as ability to read one byte at a time
and also with max buffer size of 1).
* 'seek(v_offset, mark)' seeks to given offset. The 'mark' parameter is
necessary only when it's difficult to seek backwards in the stream, such as
when reading compressed input.
* 'sync()' removes everything from internal buffers, so that if the underlying
file has changed the changes get noticed immediately after sync.
* 'get_size(exact)' returns the size of the input stream, if it's known. If
'exact=TRUE', the returned size must be the same how many bytes can be read
from the input. If 'exact=FALSE', the size is mainly used to compare against
another stat to see if the underlying input had changed. For example with
compressed input the size could be the compressed size.
* 'stat(exact)' stats the file, filling as much of the fields as makes sense.
'st_size' field is filled the same way as with 'get_size()', or set to -1 if
it's unknown.
There are some variables available:
* 'buffer' contains pointer to the data.
* First 'skip' bytes of the buffer are already skipped over (with
'i_stream_skip()' or seeking).
* Data up to 'pos' bytes (beginning after 'skip') in the buffer are available
with 'i_stream_get_data()'. If pos=skip, it means there is no available data
in the buffer.
If your input stream needs a write buffer, you can use some of the common
helper functions and variables:
* 'w_buffer' contain the pointer where you can write data. It should be kept
in sync with 'buffer'.
* 'buffer_size' specifies the buffer's size, and 'max_buffer_size' the max.
size the buffer can be grown to.
* 'i_stream_get_buffer_space(wanted_bytes)' can be used to request wanted
number of bytes from the buffer. The returned available buffer size may be
less (or more), and calling the function again may or may not return a
larger size.
(This file was created from the wiki on 2017-05-11 04:42)
