This documentation is for Dovecot v2.x, see wiki1 for v1.x documentation.
Differences between revisions 12 and 13
Revision 12 as of 2017-12-10 17:54:10
Size: 10723
Editor: TimoSirainen
Comment:
Revision 13 as of 2021-09-24 15:03:30
Size: 68
Editor: TimoSirainen
Comment:
Deletions are marked like this. Additions are marked like this.
Line 1: Line 1:
= 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.
 * unix: Read data from UNIX socket. Similar to file, but supports receiving file descriptors.
 * 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.

Input stream filters:
 * concat: Concatenate multiple input streams together
 * chain: Chain multiple input streams together. Similar to istream-concat, but more istreams can be added after initialization and EOF needs to be explicitly added.
 * 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.
 * sized: Require istream's length to be exactly the given size, or the last read returns error.
 * timeout: Fail the read when given timeout is reached.
 * try: Read from the first input stream that doesn't fail with EINVAL.
 * tee: Fork an input stream to multiple streams that can be read independently.
 * multiplex: Multiplex-iostreams support multiple iostream channels inside a single parent istream.
 * callback: Build an input stream by calling callback functions that return the data.
 * base64-encoder, base64-decoder: Encode/decode base64.
 * failure-at: Insert a failure at the specified offset. This can be useful for testing.
 * hash: Calculate hash of the istream while it's being read.
 * 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.
 * lib-compression/*: Read zlib/bzlib/lz4/lzma compressed data.

== 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.

Whenever i_stream_read() returns >0, all the existing pointers are potentially invalidated. v2.3+: When i_stream_read() returns <= 0, the data previously returned by i_stream_get_data() are still valid, preserved in "snapshots". (<v2.3 may or may not have invalidated them.)

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). When it needs to read from parent streams, try to use `i_stream_read_memarea(parent)` if possible so a new snapshot isn't unnecessarily created (see the snapshot discussion below).
 * {{{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.
 * {{{snapshot(prev_snapshot)}}} creates a snapshot of the data that is currently available via i_stream_get_data(), merges it with prev_snapshot (if any) and returns the merged snapshot (see below more more details).

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_try_alloc(wanted_size, size_r)` can be used when you want to store `wanted_bytes` into `w_buffer`. If the buffer isn't large enough for it, it's grown if possible. The buffer isn't grown above the stream's max buffer size. The returned `size_r` specifies how many bytes are actually available for writing at `stream->w_buffer + stream->pos`.
 * `i_stream_alloc(size) is like `i_stream_try_alloc()`, except it always succeeds allocating `size` bytes, even if it has to grow the buffer larger then the stream's max buffer size.
 * Lower-level memory allocation functions:
  * `i_stream_w_buffer_realloc(old_size)` reallocates `w_buffer` to the current `buffer_size`. If memarea's refcount is 1, this can be done with `i_realloc()`, otherwise new memory is allocated.
  * `i_stream_grow_buffer(bytes)` grows the `w_buffer` by the given number of bytes, if possible. It won't reach the stream's current max buffer size. The caller must verify from `buffer_size` how large the buffer became as a result of this call.
  * `i_stream_compress()` attempts to compress the current `w_buffer` by removing already-skipped data with `memmove()`. If `skip` is 0, it does nothing. Note that this function must not be called if `memarea` has refcount>1. Otherwise that could be modifying a snapshotted memarea.

The snapshots have made implementing slightly more complicated than earlier. There are a few different ways to implement istreams:

 * Always point {{{buffer=w_buffer}}} and use `i_stream_try_alloc()` and/or `i_stream_alloc()` to allocate the `w_buffer`. The generic code will handle all the snapshotting. Use `i_stream_read_memarea()` to read data from parent stream so multiple snapshots aren't unnecessarily created.
 * Guarantee that if `read()` returns <=0, the existing `buffer` will stay valid. Use `ISTREAM_CREATE_FLAG_NOOP_SNAPSHOT` flag in `i_stream_create()` so your filter stream isn't unnecessarily snapshotted (or causing a panic due to missing `snapshot()` implementation).
  * One way of doing this with filter streams is to read from the parent stream via `i_stream_read(parent)` and always use `buffer=i_stream_get_data(parent)`. The parent's snapshotting guarantees that the buffer will stay valid.
 * Implement the `snapshot()` yourself in the stream. You'll need to create a new memarea of the current data available via `i_stream_get_data()` and it must not change, i.e. most likely you'll need to duplicate the allocated memory. Create a new `struct istream_snapshot` and assign the allocated memarea to its `old_memarea`. Fill `prev_snapshot` field and return your new snapshot. The snapshot will be freed by the generic istream code either when the next `read()` returns >0 or when the istream is destroyed.
 * Filter streams that only pass through parent stream's contents without changes can just point to the parent stream. The default snapshotting causes the parent to be snapshotted, so the filter stream can simply use `i_stream_read_memarea()` and point to the parent's buffer.

When Dovecot is configured with `--enable-devel-checks`, `i_stream_read()` will verify that the first and the last two bytes of the buffer didn't unexpectedly change due to a `read()`. While developing istream changes you should use this to make sure the istream is working properly. Running the istream unit test also via valgrind can also be used to verify that the buffer wasn't freed.
Moved to https://doc.dovecot.org/developer_manual/design/istreams/

None: Design/InputStreams (last edited 2021-09-24 15:03:30 by TimoSirainen)