libs (input-output)
Rust's Write
trait has the write_all
method, which is a convenience
method that writes a whole buffer, failing with ErrorKind::WriteZero
if the buffer cannot be written in full.
This RFC proposes adding its Read
counterpart: a method (here called
read_exact
) that reads a whole buffer, failing with an error (here
called ErrorKind::UnexpectedEOF
) if the buffer cannot be read in full.
When dealing with serialization formats with fixed-length fields,
reading or writing less than the field's size is an error. For the
Write
side, the write_all
method does the job; for the Read
side,
however, one has to call read
in a loop until the buffer is completely
filled, or until a premature EOF is reached.
This leads to a profusion of similar helper functions. For instance, the
byteorder
crate has a read_full
function, and the postgres
crate
has a read_all
function. However, their handling of the premature EOF
condition differs: the byteorder
crate has its own Error
enum, with
UnexpectedEOF
and Io
variants, while the postgres
crate uses an
io::Error
with an io::ErrorKind::Other
.
That can make it unnecessarily hard to mix uses of these helper
functions; for instance, if one wants to read a 20-byte tag (using one's
own helper function) followed by a big-endian integer, either the helper
function has to be written to use byteorder::Error
, or the calling
code has to deal with two different ways to represent a premature EOF,
depending on which field encountered the EOF condition.
Additionally, when reading from an in-memory buffer, looping is not
necessary; it can be replaced by a size comparison followed by a
copy_memory
(similar to write_all
for &mut [u8]
). If this
non-looping implementation is #[inline]
, and the buffer size is known
(for instance, it's a fixed-size buffer in the stack, or there was an
earlier check of the buffer size against a larger value), the compiler
could potentially turn a read from the buffer followed by an endianness
conversion into the native endianness (as can happen when using the
byteorder
crate) into a single-instruction direct load from the buffer
into a register.
First, a new variant UnexpectedEOF
is added to the io::ErrorKind
enum.
The following method is added to the Read
trait:
fn read_exact(&mut self, buf: &mut [u8]) -> Result<()>;
Aditionally, a default implementation of this method is provided:
fn read_exact(&mut self, mut buf: &mut [u8]) -> Result<()> {
while !buf.is_empty() {
match self.read(buf) {
Ok(0) => break,
Ok(n) => { let tmp = buf; buf = &mut tmp[n..]; }
Err(ref e) if e.kind() == ErrorKind::Interrupted => {}
Err(e) => return Err(e),
}
}
if !buf.is_empty() {
Err(Error::new(ErrorKind::UnexpectedEOF, "failed to fill whole buffer"))
} else {
Ok(())
}
}
And an optimized implementation of this method for &[u8]
is provided:
#[inline]
fn read_exact(&mut self, buf: &mut [u8]) -> Result<()> {
if (buf.len() > self.len()) {
return Err(Error::new(ErrorKind::UnexpectedEOF, "failed to fill whole buffer"));
}
let (a, b) = self.split_at(buf.len());
slice::bytes::copy_memory(a, buf);
*self = b;
Ok(())
}
The detailed semantics of read_exact
are as follows: read_exact
reads exactly the number of bytes needed to completely fill its buf
parameter. If that's not possible due to an "end of file" condition
(that is, the read
method would return 0 even when passed a buffer
with at least one byte), it returns an ErrorKind::UnexpectedEOF
error.
On success, the read pointer is advanced by the number of bytes read, as
if the read
method had been called repeatedly to fill the buffer. On
any failure (including an ErrorKind::UnexpectedEOF
), the read pointer
might have been advanced by any number between zero and the number of
bytes requested (inclusive), and the contents of its buf
parameter
should be treated as garbage (any part of it might or might not have
been overwritten by unspecified data).
Even if the failure was an ErrorKind::UnexpectedEOF
, the read pointer
might have been advanced by a number of bytes less than the number of
bytes which could be read before reaching an "end of file" condition.
The read_exact
method will never return an ErrorKind::Interrupted
error, similar to the read_to_end
method.
Similar to the read
method, no guarantees are provided about the
contents of buf
when this function is called; implementations cannot
rely on any property of the contents of buf
being true. It is
recommended that implementations only write data to buf
instead of
reading its contents.
Whether or not read_exact
can return an ErrorKind::Interrupted
error
is orthogonal to its semantics. One could imagine an alternative design
where read_exact
could return an ErrorKind::Interrupted
error.
The reason read_exact
should deal with ErrorKind::Interrupted
itself
is its non-idempotence. On failure, it might have already partially
advanced its read pointer an unknown number of bytes, which means it
can't be easily retried after an ErrorKind::Interrupted
error.
One could argue that it could return an ErrorKind::Interrupted
error
if it's interrupted before the read pointer is advanced. But that
introduces a non-orthogonality in the design, where it might either
return or retry depending on whether it was interrupted at the beginning
or in the middle. Therefore, the cleanest semantics is to always retry.
There's precedent for this choice in the read_to_end
method. Users who
need finer control should use the read
method directly.
This RFC proposes a read_exact
function where the read pointer
(conceptually, what would be returned by Seek::seek
if the stream was
seekable) is unspecified on failure: it might not have advanced at all,
have advanced in full, or advanced partially.
Two possible alternatives could be considered: never advance the read
pointer on failure, or always advance the read pointer to the "point of
error" (in the case of ErrorKind::UnexpectedEOF
, to the end of the
stream).
Never advancing the read pointer on failure would make it impossible to
have a default implementation (which calls read
in a loop), unless the
stream was seekable. It would also impose extra costs (like creating a
temporary buffer) to allow "seeking back" for non-seekable streams.
Always advancing the read pointer to the end on failure is possible; it
happens without any extra code in the default implementation. However,
it can introduce extra costs in optimized implementations. For instance,
the implementation given above for &[u8]
would need a few more
instructions in the error case. Some implementations (for instance,
reading from a compressed stream) might have a larger extra cost.
The utility of always advancing the read pointer to the end is
questionable; for non-seekable streams, there's not much that can be
done on an "end of file" condition, so most users would discard the
stream in both an "end of file" and an ErrorKind::UnexpectedEOF
situation. For seekable streams, it's easy to seek back, but most users
would treat an ErrorKind::UnexpectedEOF
as a "corrupted file" and
discard the stream anyways.
Users who need finer control should use the read
method directly, or
when available use the Seek
trait.
This RFC proposes that the contents of the output buffer be undefined on an error return. It might be untouched, partially overwritten, or completely overwritten (even if less bytes could be read; for instance, this method might in theory use it as a scratch space).
Two possible alternatives could be considered: do not touch it on failure, or overwrite it with valid data as much as possible.
Never touching the output buffer on failure would make it much more
expensive for the default implementation (which calls read
in a loop),
since it would have to read into a temporary buffer and copy to the
output buffer on success. Any implementation which cannot do an early
return for all failure cases would have similar extra costs.
Overwriting as much as possible with valid data makes some sense; it happens without any extra cost in the default implementation. However, for optimized implementations this extra work is useless; since the caller can't know how much is valid data and how much is garbage, it can't make use of the valid data.
Users who need finer control should use the read
method directly.
It's unfortunate that write_all
used WriteZero
for its ErrorKind
;
were it named UnexpectedEOF
(which is a much more intuitive name), the
same ErrorKind
could be used for both functions.
The initial proposal for this read_exact
method called it read_all
,
for symmetry with write_all
. However, that name could also be
interpreted as "read as many bytes as you can that fit on this buffer,
and return what you could read" instead of "read enough bytes to fill
this buffer, and fail if you couldn't read them all". The previous
discussion led to read_exact
for the later meaning, and read_full
for the former meaning.
If this method fails, the buffer contents are undefined; the `read_exact' method might have partially overwritten it. If the caller requires "all-or-nothing" semantics, it must clone the buffer. In most use cases, this is not a problem; the caller will discard or overwrite the buffer in case of failure.
In the same way, if this method fails, there is no way to determine how many bytes were read before it determined it couldn't completely fill the buffer.
Situations that require lower level control can still use read
directly.
The first alternative is to do nothing. Every Rust user needing this functionality continues to write their own read_full or read_exact function, or have to track down an external crate just for one straightforward and commonly used convenience method. Additionally, unless everybody uses the same external crate, every reimplementation of this method will have slightly different error handling, complicating mixing users of multiple copies of this convenience method.
The second alternative is to just add the ErrorKind::UnexpectedEOF
or
similar. This would lead in the long run to everybody using the same
error handling for their version of this convenience method, simplifying
mixing their uses. However, it's questionable to add an ErrorKind
variant which is never used by the standard library.
Another alternative is to return the number of bytes read in the error
case. That makes the buffer contents defined also in the error case, at
the cost of increasing the size of the frequently-used io::Error
struct, for a rarely used return value. My objections to this
alternative are:
Err
return.read_exact
.The final alternative is read_full
, which returns the number of bytes
read (Result<usize>
) instead of failing. This means that every caller
has to check the return value against the size of the passed buffer, and
some are going to forget (or misimplement) the check. It also prevents
some optimizations (like the early return in case there will never be
enough data). There are, however, valid use cases for this alternative;
for instance, reading a file in fixed-size chunks, where the last chunk
(and only the last chunk) can be shorter. I believe this should be
discussed as a separate proposal; its pros and cons are distinct enough
from this proposal to merit its own arguments.
I believe that the case for read_full
is weaker than read_exact
, for
the following reasons:
read_exact
needs an extra variant in ErrorKind
, read_full
has no new error cases. This means that implementing it yourself is
easy, and multiple implementations have no drawbacks other than code
duplication.read_exact
can be optimized with an early return in cases
where the reader knows its total size (for instance, reading from a
compressed file where the uncompressed size was given in a header),
read_full
has to always write to the output buffer, so there's not
much to gain over a generic looping implementation calling read
.