perliol - C API for Perl's implementation of IO in Layers.
/* Defining a layer ... */
#include <perliol.h>
This document describes the behavior and implementation of the PerlIO
abstraction described in perlapio when USE_PERLIO is defined (and
USE_SFIO is not).
The PerlIO abstraction was introduced in perl5.003_02 but languished as just an abstraction until perl5.7.0. However during that time a number of perl extensions switched to using it, so the API is mostly fixed to maintain (source) compatibility.
The aim of the implementation is to provide the PerlIO API in a flexible and platform neutral manner. It is also a trial of an "Object Oriented C, with vtables" approach which may be applied to perl6.
Initial discussion of the ability to modify IO streams behaviour used the term "discipline" for the entities which were added. This came (I believe) from the use of the term in "sfio", which in turn borrowed it from "line disciplines" on Unix terminals. However, this document (and the C code) uses the term "layer".
This is, I hope, a natural term given the implementation, and should avoid connotations that are inherent in earlier uses of "discipline" for things which are rather different.
The basic data structure is a PerlIOl:
typedef struct _PerlIO PerlIOl; typedef struct _PerlIO_funcs PerlIO_funcs; typedef PerlIOl *PerlIO;
struct _PerlIO
{
PerlIOl * next; /* Lower layer */
PerlIO_funcs * tab; /* Functions for this layer */
IV flags; /* Various flags for state */
};
A PerlIOl * is a pointer to the struct, and the application
level PerlIO * is a pointer to a PerlIOl * - i.e. a pointer
to a pointer to the struct. This allows the application level PerlIO *
to remain constant while the actual PerlIOl * underneath
changes. (Compare perl's SV * which remains constant while its
sv_any field changes as the scalar's type changes.) An IO stream is
then in general represented as a pointer to this linked-list of
"layers".
It should be noted that because of the double indirection in a PerlIO *,
a &(perlio->next) "is" a PerlIO *, and so to some degree
at least one layer can use the "standard" API on the next layer down.
A "layer" is composed of two parts:
The functions and attributes are accessed via the "tab" (for table)
member of PerlIOl. The functions (methods of the layer "class") are
fixed, and are defined by the PerlIO_funcs type. They are broadly the
same as the public PerlIO_xxxxx functions:
struct _PerlIO_funcs
{
Size_t fsize;
char * name;
Size_t size;
IV kind;
IV (*Pushed)(pTHX_ PerlIO *f,const char *mode,SV *arg, PerlIO_funcs *tab);
IV (*Popped)(pTHX_ PerlIO *f);
PerlIO * (*Open)(pTHX_ PerlIO_funcs *tab,
AV *layers, IV n,
const char *mode,
int fd, int imode, int perm,
PerlIO *old,
int narg, SV **args);
IV (*Binmode)(pTHX_ PerlIO *f);
SV * (*Getarg)(pTHX_ PerlIO *f, CLONE_PARAMS *param, int flags)
IV (*Fileno)(pTHX_ PerlIO *f);
PerlIO * (*Dup)(pTHX_ PerlIO *f, PerlIO *o, CLONE_PARAMS *param, int flags)
/* Unix-like functions - cf sfio line disciplines */
SSize_t (*Read)(pTHX_ PerlIO *f, void *vbuf, Size_t count);
SSize_t (*Unread)(pTHX_ PerlIO *f, const void *vbuf, Size_t count);
SSize_t (*Write)(pTHX_ PerlIO *f, const void *vbuf, Size_t count);
IV (*Seek)(pTHX_ PerlIO *f, Off_t offset, int whence);
Off_t (*Tell)(pTHX_ PerlIO *f);
IV (*Close)(pTHX_ PerlIO *f);
/* Stdio-like buffered IO functions */
IV (*Flush)(pTHX_ PerlIO *f);
IV (*Fill)(pTHX_ PerlIO *f);
IV (*Eof)(pTHX_ PerlIO *f);
IV (*Error)(pTHX_ PerlIO *f);
void (*Clearerr)(pTHX_ PerlIO *f);
void (*Setlinebuf)(pTHX_ PerlIO *f);
/* Perl's snooping functions */
STDCHAR * (*Get_base)(pTHX_ PerlIO *f);
Size_t (*Get_bufsiz)(pTHX_ PerlIO *f);
STDCHAR * (*Get_ptr)(pTHX_ PerlIO *f);
SSize_t (*Get_cnt)(pTHX_ PerlIO *f);
void (*Set_ptrcnt)(pTHX_ PerlIO *f,STDCHAR *ptr,SSize_t cnt);
};
The first few members of the struct give a function table size for
compatibility check "name" for the layer, the size to malloc for the per-instance data,
and some flags which are attributes of the class as whole (such as whether it is a buffering
layer), then follow the functions which fall into four basic groups:
A layer does not have to implement all the functions, but the whole table has to be present. Unimplemented slots can be NULL (which will result in an error when called) or can be filled in with stubs to "inherit" behaviour from a "base class". This "inheritance" is fixed for all instances of the layer, but as the layer chooses which stubs to populate the table, limited "multiple inheritance" is possible.
The per-instance data are held in memory beyond the basic PerlIOl struct, by making a PerlIOl the first member of the layer's struct thus:
typedef struct
{
struct _PerlIO base; /* Base "class" info */
STDCHAR * buf; /* Start of buffer */
STDCHAR * end; /* End of valid part of buffer */
STDCHAR * ptr; /* Current position in buffer */
Off_t posn; /* Offset of buf into the file */
Size_t bufsiz; /* Real size of buffer */
IV oneword; /* Emergency buffer */
} PerlIOBuf;
In this way (as for perl's scalars) a pointer to a PerlIOBuf can be treated as a pointer to a PerlIOl.
table perlio unix
| |
+-----------+ +----------+ +--------+
PerlIO ->| |--->| next |--->| NULL |
+-----------+ +----------+ +--------+
| | | buffer | | fd |
+-----------+ | | +--------+
| | +----------+
The above attempts to show how the layer scheme works in a simple case.
The application's PerlIO * points to an entry in the table(s)
representing open (allocated) handles. For example the first three slots
in the table correspond to stdin,stdout and stderr. The table
in turn points to the current "top" layer for the handle - in this case
an instance of the generic buffering layer "perlio". That layer in turn
points to the next layer down - in this case the lowlevel "unix" layer.
The above is roughly equivalent to a "stdio" buffered stream, but with much more flexibility:
read/write/lseek is not appropriate for (say)
sockets then the "unix" layer can be replaced (at open time or even
dynamically) with a "socket" layer.
mmap than read. An "unbuffered" stream can be implemented
simply by not having a buffer layer.
The generic flag bits are a hybrid of O_XXXXX style flags deduced
from the mode string passed to PerlIO_open(), and state bits for
typical buffer layers.
PerlIO_error()).
PerlIO_binmode() will mess with this
flag rather than add/remove layers if the PERLIO_K_CANCRLF bit is set
for the layers class.
unlink()ed, or should be deleted on close().
gets" interface.
Normally set based on PERLIO_K_FASTGETS for the class and by the
existence of the function(s) in the table. However a class that
normally provides that interface may need to avoid it on a
particular instance. The "pending" layer needs to do this when
it is pushed above a layer which does not support the interface.
(Perl's sv_gets() does not expect the streams fast gets behaviour
to change during one "get".)
Size_t fsize;
Size of the function table. This is compared against the value PerlIO code "knows" as a compatibility check. Future versions may be able to tolerate layers compiled against an old version of the headers.
char * name;
The name of the layer whose open() method Perl should invoke on open(). For example if the layer is called APR, you will call:
open $fh, ">:APR", ...
and Perl knows that it has to invoke the PerlIOAPR_open() method implemented by the APR layer.
Size_t size;
The size of the per-instance data structure, e.g.:
sizeof(PerlIOAPR)
If this field is zero then PerlIO_pushed does not malloc anything
and assumes layer's Pushed function will do any required layer stack
manipulation - used to avoid malloc/free overhead for dummy layers.
If the field is non-zero it must be at least the size of PerlIOl,
PerlIO_pushed will allocate memory for the layer's data structures
and link new layer onto the stream's stack. (If the layer's Pushed
method returns an error indication the layer is popped again.)
IV kind;
MODE argument. When this
flag is used it's up to the layer to validate the args.
IV (*Pushed)(pTHX_ PerlIO *f,const char *mode, SV *arg);
The only absolutely mandatory method. Called when the layer is pushed
onto the stack. The mode argument may be NULL if this occurs
post-open. The arg will be non-NULL if an argument string was
passed. In most cases this should call PerlIOBase_pushed() to
convert mode into the appropriate PERLIO_F_XXXXX flags in
addition to any actions the layer itself takes. If a layer is not
expecting an argument it need neither save the one passed to it, nor
provide Getarg() (it could perhaps Perl_warn that the argument
was un-expected).
Returns 0 on success. On failure returns -1 and should set errno.
IV (*Popped)(pTHX_ PerlIO *f);
Called when the layer is popped from the stack. A layer will normally
be popped after Close() is called. But a layer can be popped
without being closed if the program is dynamically managing layers on
the stream. In such cases Popped() should free any resources
(buffers, translation tables, ...) not held directly in the layer's
struct. It should also Unread() any unconsumed data that has been
read and buffered from the layer below back to that layer, so that it
can be re-provided to what ever is now above.
Returns 0 on success and failure.
PerlIO * (*Open)(...);
The Open() method has lots of arguments because it combines the
functions of perl's open, PerlIO_open, perl's sysopen,
PerlIO_fdopen and PerlIO_reopen. The full prototype is as
follows:
PerlIO * (*Open)(pTHX_ PerlIO_funcs *tab, AV *layers, IV n, const char *mode, int fd, int imode, int perm, PerlIO *old, int narg, SV **args);
Open should (perhaps indirectly) call PerlIO_allocate() to allocate
a slot in the table and associate it with the layers information for
the opened file, by calling PerlIO_push. The layers AV is an
array of all the layers destined for the PerlIO *, and any
arguments passed to them, n is the index into that array of the
layer being called. The macro PerlIOArg will return a (possibly
NULL) SV * for the argument passed to the layer.
The mode string is an "fopen()-like" string which would match
the regular expression /^[I#]?[rwa]\+?[bt]?$/.
The 'I' prefix is used during creation of stdin..stderr via
special PerlIO_fdopen calls; the '#' prefix means that this is
sysopen and that imode and perm should be passed to
PerlLIO_open3; 'r' means read, 'w' means write and
'a' means append. The '+' suffix means that both reading and
writing/appending are permitted. The 'b' suffix means file should
be binary, and 't' means it is text. (Almost all layers should do
the IO in binary mode, and ignore the b/t bits. The :crlf layer
should be pushed to handle the distinction.)
If old is not NULL then this is a PerlIO_reopen. Perl itself
does not use this (yet?) and semantics are a little vague.
If fd not negative then it is the numeric file descriptor fd,
which will be open in a manner compatible with the supplied mode
string, the call is thus equivalent to PerlIO_fdopen. In this case
nargs will be zero.
If nargs is greater than zero then it gives the number of arguments
passed to open, otherwise it will be 1 if for example
PerlIO_open was called. In simple cases SvPV_nolen(*args) is the
pathname to open.
Having said all that translation-only layers do not need to provide
Open() at all, but rather leave the opening to a lower level layer
and wait to be "pushed". If a layer does provide Open() it should
normally call the Open() method of next layer down (if any) and
then push itself on top if that succeeds.
Returns NULL on failure.
IV (*Binmode)(pTHX_ PerlIO *f);
Optional. Used when :raw layer is pushed (explicitly or as a result
of binmode(FH)). If not present layer will be popped. If present
should configure layer as binary (or pop itself) and return 0.
If it returns -1 for error binmode will fail with layer
still on the stack.
SV * (*Getarg)(pTHX_ PerlIO *f, CLONE_PARAMS *param, int flags);
Optional. If present should return an SV * representing the string argument passed to the layer when it was pushed. e.g. ":encoding(ascii)" would return an SvPV with value "ascii". (param and flags arguments can be ignored in most cases)
IV (*Fileno)(pTHX_ PerlIO *f);
Returns the Unix/Posix numeric file descriptor for the handle. Normally
PerlIOBase_fileno() (which just asks next layer down) will suffice
for this.
Returns -1 on error, which is considered to include the case where the layer cannot provide such a file descriptor.
PerlIO * (*Dup)(pTHX_ PerlIO *f, PerlIO *o, CLONE_PARAMS *param, int flags);
XXX: Needs more docs.
Used as part of the "clone" process when a thread is spawned (in which
case param will be non-NULL) and when a stream is being duplicated via
'&' in the open.
Similar to Open, returns PerlIO* on success, NULL on failure.
SSize_t (*Read)(pTHX_ PerlIO *f, void *vbuf, Size_t count);
Basic read operation.
Typically will call Fill and manipulate pointers (possibly via the
API). PerlIOBuf_read() may be suitable for derived classes which
provide "fast gets" methods.
Returns actual bytes read, or -1 on an error.
SSize_t (*Unread)(pTHX_ PerlIO *f, const void *vbuf, Size_t count);
A superset of stdio's ungetc(). Should arrange for future reads to
see the bytes in vbuf. If there is no obviously better implementation
then PerlIOBase_unread() provides the function by pushing a "fake"
"pending" layer above the calling layer.
Returns the number of unread chars.
SSize_t (*Write)(PerlIO *f, const void *vbuf, Size_t count);
Basic write operation.
Returns bytes written or -1 on an error.
IV (*Seek)(pTHX_ PerlIO *f, Off_t offset, int whence);
Position the file pointer. Should normally call its own Flush
method and then the Seek method of next layer down.
Returns 0 on success, -1 on failure.
Off_t (*Tell)(pTHX_ PerlIO *f);
Return the file pointer. May be based on layers cached concept of position to avoid overhead.
Returns -1 on failure to get the file pointer.
IV (*Close)(pTHX_ PerlIO *f);
Close the stream. Should normally call PerlIOBase_close() to flush
itself and close layers below, and then deallocate any data structures
(buffers, translation tables, ...) not held directly in the data
structure.
Returns 0 on success, -1 on failure.
IV (*Flush)(pTHX_ PerlIO *f);
Should make stream's state consistent with layers below. That is, any
buffered write data should be written, and file position of lower layers
adjusted for data read from below but not actually consumed.
(Should perhaps Unread() such data to the lower layer.)
Returns 0 on success, -1 on failure.
IV (*Fill)(pTHX_ PerlIO *f);
The buffer for this layer should be filled (for read) from layer below. When you "subclass" PerlIOBuf layer, you want to use its _read method and to supply your own fill method, which fills the PerlIOBuf's buffer.
Returns 0 on success, -1 on failure.
IV (*Eof)(pTHX_ PerlIO *f);
Return end-of-file indicator. PerlIOBase_eof() is normally sufficient.
Returns 0 on end-of-file, 1 if not end-of-file, -1 on error.
IV (*Error)(pTHX_ PerlIO *f);
Return error indicator. PerlIOBase_error() is normally sufficient.
Returns 1 if there is an error (usually when PERLIO_F_ERROR is set,
0 otherwise.
void (*Clearerr)(pTHX_ PerlIO *f);
Clear end-of-file and error indicators. Should call PerlIOBase_clearerr()
to set the PERLIO_F_XXXXX flags, which may suffice.
void (*Setlinebuf)(pTHX_ PerlIO *f);
Mark the stream as line buffered. PerlIOBase_setlinebuf() sets the
PERLIO_F_LINEBUF flag and is normally sufficient.
STDCHAR * (*Get_base)(pTHX_ PerlIO *f);
Allocate (if not already done so) the read buffer for this layer and return pointer to it. Return NULL on failure.
Size_t (*Get_bufsiz)(pTHX_ PerlIO *f);
Return the number of bytes that last Fill() put in the buffer.
STDCHAR * (*Get_ptr)(pTHX_ PerlIO *f);
Return the current read pointer relative to this layer's buffer.
SSize_t (*Get_cnt)(pTHX_ PerlIO *f);
Return the number of bytes left to be read in the current buffer.
void (*Set_ptrcnt)(pTHX_ PerlIO *f, STDCHAR *ptr, SSize_t cnt);
Adjust the read pointer and count of bytes to match ptr and/or cnt.
The application (or layer above) must ensure they are consistent.
(Checking is allowed by the paranoid.)
The file perlio.c provides the following layers:
read(), write(),
lseek(), close(). No buffering. Even on platforms that distinguish
between O_TEXT and O_BINARY this layer is always O_BINARY.
A very complete generic buffering layer which provides the whole of
PerlIO API. It is also intended to be used as a "base class" for other
layers. (For example its Read() method is implemented in terms of
the Get_cnt()/Get_ptr()/Set_ptrcnt() methods).
"perlio" over "unix" provides a complete replacement for stdio as seen
via PerlIO API. This is the default for USE_PERLIO when system's stdio
does not permit perl's "fast gets" access, and which do not
distinguish between O_TEXT and O_BINARY.
O_TEXT and O_BINARY.
O_TEXT and O_BINARY opens. (At some point
"unix" will be replaced by a "native" Win32 IO layer on that platform,
as Win32's read/write layer has various drawbacks.) The "crlf" layer is
a reasonable model for a layer which transforms data in some way.
mmap() functions this layer is provided (with
"perlio" as a "base") which does "read" operations by mmap()ing the
file. Performance improvement is marginal on modern systems, so it is
mainly there as a proof of concept. It is likely to be unbundled from
the core at some point. The "mmap" layer is a reasonable model for a
minimalist "derived" layer.
Fill() pops itself off the stack
and so resumes reading from layer below.)
PERLIO_K_RAW bit set. Layers can modify that behaviour by defining
their own Binmode entry.
PERLIO_F_UTF8 flag on the layer which was (and now is once more)
the top of the stack.
In addition perlio.c also provides a number of PerlIOBase_xxxx()
functions which are intended to be used in the table slots of classes
which do not need to do anything special for a particular method.
Layers can made available by extension modules. When an unknown layer is encountered the PerlIO code will perform the equivalent of :
use PerlIO 'layer';
Where layer is the unknown layer. PerlIO.pm will then attempt to:
require PerlIO::layer;
If after that process the layer is still not defined then the open
will fail.
The following extension layers are bundled with perl:
use Encoding;
makes this layer available, although PerlIO.pm "knows" where to find it. It is an example of a layer which takes an argument as it is called thus:
open( $fh, "<:encoding(iso-8859-7)", $pathname );
Provides support for reading data from and writing data to a scalar.
open( $fh, "+<:scalar", \$scalar );
When a handle is so opened, then reads get bytes from the string value
of $scalar, and writes change the value. In both cases the position
in $scalar starts as zero but can be altered via seek, and
determined via tell.
Please note that this layer is implied when calling open() thus:
open( $fh, "+<", \$scalar );
Provided to allow layers to be implemented as Perl code. For instance:
use PerlIO::via::StripHTML; open( my $fh, "<:via(StripHTML)", "index.html" );
See PerlIO::via for details.
Things that need to be done to improve this document.
Explain how to make a valid fh without going through open()(i.e. apply a layer). For example if the file is not opened through perl, but we want to get back a fh, like it was opened by Perl.
How PerlIO_apply_layera fits in, where its docs, was it made public?
Currently the example could be something like this:
PerlIO *foo_to_PerlIO(pTHX_ char *mode, ...)
{
char *mode; /* "w", "r", etc */
const char *layers = ":APR"; /* the layer name */
PerlIO *f = PerlIO_allocate(aTHX);
if (!f) {
return NULL;
}
PerlIO_apply_layers(aTHX_ f, mode, layers);
if (f) {
PerlIOAPR *st = PerlIOSelf(f, PerlIOAPR);
/* fill in the st struct, as in _open() */
st->file = file;
PerlIOBase(f)->flags |= PERLIO_F_OPEN;
return f;
}
return NULL;
}
The handling of errors by the layer is not specified. e.g. when $! should be set explicitly, when the error handling should be just delegated to the top layer.
Probably give some hints on using SETERRNO() or pointers to where they can be found.