We occasionally received reports from people getting "invalid tstate"
crashes (this is a fatal error in PyThreadState_Delete()). Finally
several people were able to reproduce it reliably and Tim Peters
discovered that there is a race condition when multiple threads are
calling this function without holding the global interpreter lock (the
function may be called without holding that).
Solved the race condition by adding a lock around the mutating uses of
interp->tstate_head. Tim and Jonathan Giddy have run tests that make
it likely that this fixes the crashes -- although Tim hasn't heard
from the person who reported the original problem.
ExtensionClasses in isinstance() and issubclass().
- abstract instance and class protocols are used *only* in those
cases that would generate errors before the patch. That is, there's
no penalty for the normal case.
- instance protocol: an object smells like an instance if it
has a __class__ attribute that smells like a class.
- class protocol: an object smells like a class if it has a
__bases__ attribute that is a tuple with elements that
smell like classes (although not all elements may actually get
sniffed ;).
man pages suggest that the proper thing to do is to add THR_NEW_LWP to
the flags on thr_create(), and that there really isn't a downside, so
I'll do that.
"""
Spec says that on success pthread_create returns 0. It does not say
that an error code will be < 0. Linux glibc2 pthread_create() returns
ENOMEM (12) when one exceed process limits. (It looks like it should
return EAGAIN, but that's another story.)
For reference, see:
http://www.opengroup.org/onlinepubs/7908799/xsh/pthread_create.html
"""
[I have a feeling that similar bugs were fixed before; perhaps someone
could check that all error checks no check for != 0?]
xrange(), especially for platforms where int and long are different
sizes (so sys.maxint isn't actually the theoretical limit for the
length of a list, but the largest C int is -- sys.maxint is the
largest Python int, which is actually a C long).
test for classes with a __complex__() method. The attribute is pulled
out of the instance with PyObject_GetAttr() but this transfers
ownership and the function object was never DECREF'd.
v temporary variable was never decref'd. Test this by starting up the
interpreter, hitting C-c, then immediately exiting.
Same potential leak can occur if error is E_NOMEM, since the return is
done in the case block. Added Py_XDECREF(v); to both blocks, just
before the return.
think we have our own DOS box (i.e. we're not started from a command
line shell), we print a message and wait for the user to hit a key
before the DOS box is closed.
The hacky heuristic for determining whether we have our *own* DOS box
(due to Mark Hammond) is to test whether we're on line zero...
The following patches (relative to 1.5.2b1) enable Python dynamic
loading to work on NetBSD platforms that use ELF (presnetly mips and
alpha systems). They automaticly determine wether the system is ELF or
a.out rather than using astatic list of platforms so that when other
NetBSD platforms move to ELF, python will continue to work without
change.
In other words, hex(sys.hexversion) == 0x010502b2 for Python 1.5.2b2.
This is derived from the new variable PY_VERSION_HEX defined in patchlevel.h.
(Cute, eh?)
Also (non-BSDI specific):
- Change the CHECK_STATUS() macro so it tests for nonzero error codes
instead of negative error codes only (this was needed for BSDI, but
appears to be correct according to the PTHREADS spec).
- use memset() to zero out the allocated lock structure. Again, this
was needed for BSDI, but can't hurt elsewhere either.
initialization of class exceptions. Specifically:
init_class_exc(): This function now returns an integer status of the
class exception initialization. No fatal errors in this method now.
Also, use PySys_WriteStderr() when writing error messages. When an
error occurs in this function, 0 is returned, but the partial creation
of the exception classes is not undone (this happens elsewhere).
Things that could trigger the fallback:
- exceptions.py fails to be imported (due to syntax error, etc.)
- one of the exception classes is missing (e.g. due to library
version mismatch)
- exception class can't be inserted into __builtin__'s dictionary
- MemoryError instance can't be pre-allocated
- some other PyErr_Occurred
newstdexception(): Changed the error message. This is still a fatal
error because if the string based exceptions can't be created, we
really can't continue.
initerrors(): Be sure to xdecref the .exc field, which might be
non-NULL if class exceptions init was aborted.
_PyBuiltin_Init_2(): If class exception init fails, print a warning
message and reinstate the string based exceptions.
that file in fact did not exist or at least was not used. Change this
so that __file__ is *only* set to the .pyc/.pyo file when it actually
read the code object from it; otherwise __file__ is set to the .py
file.
happen when you use a non-keyword argument after a keyword argument,
and in this case you also get a syntax error. I fully suspect that
the underflow is caused by the code that stops generating code when it
detects the syntax error, but I can't find the culprit right now. I
know, I know.)
The MS compiler doesn't call it 'long long', it uses __int64,
so a new #define, LONG_LONG, has been added and all occurrences
of 'long long' are replaced with it.
This is a patch that Bill Bummgarner did for 1.4 that hasn't made its
way into the distribution yet. This is important if you want to use
the ObjC module.
frozen packages. (I *think* this means that we can now have a
built-in module bar that's a submodule of a frozen package foo, by
registering the built-in module with a name "foo.bar" in the table of
builtin modules.)
an exception from errno, with a supplied filename (primarily used by
IOError and OSError). If class exceptions are used then the exception
is instantiated with a 3-tuple: (errno, strerror, filename). For
backwards compatibility reasons, if string exceptions are used,
filename is ignored.
PyErr_SetFromErrno(): Implement in terms of
PyErr_SetFromErrnoWithFilename().
OSError. The EnvironmentError serves primarily as the (common
implementation) base class for IOError and OSError. OSError is used
by posixmodule.c
Also added tuple definition of EnvironmentError when using string
based exceptions.
(1) If a sequence S is shorter than len(S) indicated, don't fail --
just use the shorter size. (I.e, len(S) is just a hint.)
(2) Implement the special case map(None, S) as list(S) -- it's faster.
must be enabled here, otherwise the errno we set on overflows is not
the errno that's being read by compile.c. Wonder how many other files
that do their own "#include config.h" need this too :-(
(Because of the structure of autoconf, it's not so simple to get this
into config.h...)
the filename contains at least a rudimentary pathname.
(The bad part is that we need to call getcwd() because only a prefix
of ".\\" is not enough -- we prefix the drive letter.)
and lists; if the size is negative, raise an exception. Also raise an
exception when an undefined type is found -- all this to increase the
chance that garbage input causes an exception instead of a core dump.
swapped arguments].
Also make sure that no use of a function pointer gotten from a
tp_as_sequence or tp_as_mapping structure is made without checking it
for NULL first.
the code here becomes much simpler. In particular: abs(), divmod(),
pow(), int(), long(), float(), len(), tuple(), list().
Also make sure that no use of a function pointer gotten from a
tp_as_sequence or tp_as_mapping structure is made without checking it
for NULL first.
A few other cosmetic things, such as properly reindenting slice().
old value in a temporary and XDECREF it only after then new value has
been set. This prevents the (unlikely) case where the destructor of
the member uses the containing object -- it would find it in an
undefined state.
because the path through the code would notice that sys.__path__ did
not exist and it would fall back to the default path (builtins +
sys.path) instead of failing). No longer.
Date: Thu, 14 Sep 1995 12:18:20 -0400
From: Alan Morse <alan@dvcorp.com>
To: python-list@cwi.nl
Subject: getargs bug in 1.2 and 1.3 BETA
We have found a bug in the part of the getargs code that we added
and submitted, and which was incorporated into 1.1.
The parsing of "O?" format specifiers is not handled correctly;
there is no "else" for the "if" and therefore it can never fail.
What's worse, the advancing of the varargs pointer is not
handled properly, so from then on it is out of sync, wreaking
all sorts of havoc. (If it had failed properly, then the out-of-sync
varargs would not have been an issue.)
Below is the context diff for the change.
Note that I have made a few stylistic changes beyond adding the
else case, namely:
1) Making the "O" case follow the convention established by the other
format specifiers of getting all their vararg arguments before
performing the test, rather than getting some before and some after
the test passes.
2) Making the logic of the tests parallel, so the "if" part indicates
that the format is accepted and the "else" part indicates that the
format has failed. They were inconsistent with each other and with the
the other format specifiers.
-Alan Morse (amorse@dvcorp.com)
to the table of built-in modules. This should normally be called
*before* Py_Initialize(). When the malloc() or realloc() call fails,
-1 is returned and the existing table is unchanged.
After a similar function by Just van Rossum.
int PyImport_ExtendInittab(struct _inittab *newtab);
int PyImport_AppendInittab(char *name, void (*initfunc)());
Adapted from code submitted by Just van Rossum.
PySys_WriteStdout(format, ...)
PySys_WriteStderr(format, ...)
The first function writes to sys.stdout; the second to sys.stderr. When
there is a problem, they write to the real (C level) stdout or stderr;
no exceptions are raised (but a pending exception may be cleared when a
new exception is caught).
Both take a printf-style format string as their first argument followed
by a variable length argument list determined by the format string.
*** WARNING ***
The format should limit the total size of the formatted output string to
1000 bytes. In particular, this means that no unrestricted "%s" formats
should occur; these should be limited using "%.<N>s where <N> is a
decimal number calculated so that <N> plus the maximum size of other
formatted text does not exceed 1000 bytes. Also watch out for "%f",
which can print hundreds of digits for very large numbers.
PyThreadState_GetDict() returns a dictionary that can be used to hold such
state; the caller should pick a unique key and store its state there. If
PyThreadState_GetDict() returns NULL, an exception has been raised (most
likely MemoryError) and the caller should pass on the exception. */
PyObject *
PyThreadState_GetDict()
Frozen packages are indicated by a negative size (the code string
is the __import__.py file). A frozen package module has its __path__
set to a string, the package name.
time can be in PyImport_ImportModuleEx(). Recursive calls from the
same thread are okay.
Potential problems:
- The lock should really be part of the interpreter state rather than
global, but that would require modifying more files, and I first want
to figure out whether this works at all.
- One could argue that the lock should be per module -- however that
would be complicated to implement. We would have to have a linked
list of locks per module name, *or* invent a new object type to
represent a lock, so we can store the locks in the module or in a
separate dictionary. Both seem unwarranted. The one situation where
this can cause problems is when loading a module takes a long time,
e.g. when the module's initialization code interacts with the user --
during that time, no other threads can run. I say, "too bad."
(modified) and use that.
Some differences in the cleanup algorithm:
- Clear __main__ before the other modules.
- Delete more sys variables: including ps1, ps2, exitfunc, argv, and
even path -- this will prevent new imports!
- Restore stdin, stdout, stderr from __stdin__, __stdout__,
__stderr__, effectively deleting hooks that the user might have
installed -- so their (the hooks') destructors will run.
This is an option for OS-es with case-insensitive but case-preserving
filesystems. It is currently supported for Win32 and MacOS. To
enable it, #define CHECK_IMPORT_CASE in your platform specific
config.h. It is enabled by default on those systems where it is
supported. On Win32, it can be disabled at runtime by setting the
environment variable PYTHONCASEOK (to any value).
When enabled, the feature checks that the case of the requested module
name matches that of the filename found in the filesystem, and raises
a NameError exception when they don't match.
pass it the true file. This is used to set __file__ properly, instead
of believing what the code object carries with it. (If the pointer
is NULL, the code object's co_filename is still used.)
- Add Py_FrozenFlag, intended to suppress error messages fron
getpath.c in frozen binaries.
- Add Py_GetPythonHome() and Py_SetPythonHome(), intended to allow
embedders to force a different PYTHONHOME.
- Add new interface PyErr_PrintEx(flag); same as PyErr_Print() but
flag determines whether sys.last_* are set or not. PyErr_Print()
now simply calls PyErr_PrintEx(1).
(1) Explicitly clear __builtin__._ and sys.{last,exc}_* before
clearing anything else. These are common places where user values
hide and people complain when their destructors fail. Since the
modules containing them are deleted *last* of all, they would come too
late in the normal destruction order. Sigh.
(2) Add some debugging aid to cleanup (after a suggestion by Marc
Lemburg) -- print the names of the modules being cleaned, and (when
-vv is used) print the names of the variables being cleared.
now implement the following finalization strategy.
1. Whenever this code deletes a module, its directory is cleared
carefully, as follows:
- set all names to None that begin with exactly one underscore
- set all names to None that don't begin with two underscores
- clear the directory
2. Modules are deleted in the following order:
- modules with a reference count of 1, except __builtin__ or __sys__
- repeat until no more are found with a reference count of 1
- __main__ if it's still there
- all remaining modules except __builtin__ or sys
- sys
_ __builtin__
This is a bit of a hack: when the shared library is loaded, the module
name is "package.module", but the module calls Py_InitModule*() with just
"module" for the name. The shared library loader squirrels away the true
name of the module in _Py_PackageContext, and Py_InitModule*() will
substitute this (if the name actually matches).
1) The __builtins__ variable in the __main__ module is set to the
__builtin__ module instead of its __dict__.
2) Get rid of the SIGHUP and SIGTERM handlers. They can't be made to
work reliably when threads may be in use, they are Unix specific, and
Python programmers can now program this functionality is a safer way
using the signal module.
Setting interp->builtins to the __builtin__ module instead of to its
dictionary had the unfortunate side effect of always running in
restricted execution mode :-(
I will check in a different way of setting __main__.__builtins__ to
the __builtin__ module later.
Also, there was a typo -- a comment was unfinished, and as a result
some finalizations were not being executed.
In Bart Simpson style,
I Will Not Check In Untested Changes.
I Will Not Check In Untested Changes.
I Will Not Check In Untested Changes.
I Will Not Check In Untested Changes.
I Will Not Check In Untested Changes.
I Will Not Check In Untested Changes.
I Will Not Check In Untested Changes.
I Will Not Check In Untested Changes.
I Will Not Check In Untested Changes.
I Will Not Check In Untested Changes.
- The interp->builtins variable (and hence, __main__.__builtins__) is
once again initialized to the built-in *module* instead of its
dictionary.
- The finalization order is once again changed. Signals are finalized
relatively early, because (1) it DECREF's the signal handlers, and if
a signal handler happens to be a bound method, deleting it could cause
problems when there's no current thread around, and (2) we don't want
to risk executing signal handlers during finalization.
__init__.py (or __init__.pyc/.pyo, whichever applies) is considered a
package. All other subdirectories are left alone. Should make Konrad
Hinsen happy!
tstate swapping. Only the acquiring and releasing of the lock is
conditional (twice, under ``#ifdef WITH_THREAD'' and inside ``if
(interpreter_lock)'').
but annoying memory leak. This was introduced when PyExc_Exception
was added; the loop above populating the PyExc_StandardError exception
tuple started at index 1 in bltin_exc, but PyExc_Exception was added
at index 0, so PyExc_StandardError was getting inserted in itself!
How else can a tuple include itself?!
Change the loop to start at index 2.
This was a *fun* one! :-)
dummy entry to sys.modules, marking the absence of a submodule by the
same name.
Thus, if module foo.bar executes the statement "import time",
sys.modules['foo.time'] will be set to None, once the absence of a
module foo.time is confirmed (by looking for it in foo's path).
The next time when foo.bar (or any other submodule of foo) executes
"import time", no I/O is necessary to determine that there is no
module foo.time.
(Justification: It may seem strange to pollute sys.modules. However,
since we're doing the lookup anyway it's definitely the fastest
solution. This is the same convention that 'ni' uses and I haven't
heard any complaints.)
right thing.
Still to do:
- Make reload() of a submodule work.
- Performance tweaks -- currently, a submodule that tries to import a
global module *always* searches the package directory first, even if
the global module was already imported. Not sure how to solve this
one; probably need to record misses per package.
- Documentation!
This doesn't yet support "import a.b.c" or "from a.b.c import x", but
it does recognize directories. When importing a directory, it
initializes __path__ to a list containing the directory name, and
loads the __init__ module if found.
The (internal) find_module() and load_module() functions are
restructured so that they both also handle built-in and frozen modules
and Mac resources (and directories of course). The imp module's
find_module() and (new) load_module() also have this functionality.
Moreover, imp unconditionally defines constants for all module types,
and has two more new functions: find_module_in_package() and
find_module_in_directory().
There's also a new API function, PyImport_ImportModuleEx(), which
takes all four __import__ arguments (name, globals, locals, fromlist).
The last three may be NULL. This is currently the same as
PyImport_ImportModule() but in the future it will be able to do
relative dotted-path imports.
Other changes:
- bltinmodule.c: in __import__, call PyImport_ImportModuleEx().
- ceval.c: always pass the fromlist to __import__, even if it is a C
function, so PyImport_ImportModuleEx() is useful.
- getmtime.c: the function has a second argument, the FILE*, on which
it applies fstat(). According to Sjoerd this is much faster. The
first (pathname) argument is ignored, but remains for backward
compatibility (so the Mac version still works without changes).
By cleverly combining the new imp functionality, the full support for
dotted names in Python (mini.py, not checked in) is now about 7K,
lavishly commented (vs. 14K for ni plus 11K for ihooks, also lavishly
commented).
Good night!
- Changed semantics for initialized flag (again); forget the ref
counting, forget the fatal errors -- redundant calls to
Py_Initialize() or Py_Finalize() calls are simply ignored.
- Automatically import site.py on initialization, unless a flag is set
not to do this by main().
Added PyErr_MemoryErrorInst to hold the pre-instantiated instance when
using class based exceptions.
Simplified the creation of all built-in exceptions, both class based
and string based. Actually, for class based exceptions, the string
ones are still created just in case there's a problem creating the
class based ones (so you still get *some* exception handling!). Now
the init and fini functions run through a list of structure elements,
creating the strings (and optionally classes) for every entry.
initerrors(): the new base class exceptions StandardError,
LookupError, and NumberError are initialized when using string
exceptions, to tuples containing the list of derived string
exceptions. This GvR trick enables forward compatibility! One bit of
nastiness is that the C code has to know the inheritance tree embodied
in exceptions.py.
Added the two phase init and fini functions.
the -X command line option.
Py_Initialize(): Handle the two phase initialization of the built-in
module.
Py_Finalize(): Handle the two phase finalization of the built-in
module.
parse_syntax_error(): New function which parses syntax errors that
PyErr_Print() will catch. This correctly parses such errors
regardless of whether PyExc_SyntaxError is an old-style string
exception or new-fangled class exception.
PyErr_Print(): Many changes:
1. Normalize the exception.
2. Handle SystemExit exceptions which might be class based. Digs
the exit code out of the "code" attribute. String based
SystemExit is handled the same as before.
3. Handle SyntaxError exceptions which might be class based. Digs
the various information bits out of the instance's attributes
(see parse_syntax_error() for details). String based
SyntaxError still works too.
4. Don't write the `:' after the exception if the exception is
class based and has an empty string str() value.
(PyExc_MemoryErrorInst) raise this instead of PyExc_MemoryError. This
only happens when exception classes are enabled (e.g. when Python is
started with -X).
former rather than the latter, since PyErr_NormalizeException takes
PyObject** and I didn't want to change the interface for set_exc_info
(but I did want the changes propagated to eval_code2!).
UNPACK_LIST byte codes and added a third code path that allows
generalized sequence unpacking. Now both syntaxes:
a, b, c = seq
[a, b, c] = seq
can be used to unpack any sequence with the exact right number of
items.
unpack_sequence(): out-lined implementation of generalized sequence
unpacking. tuple and list unpacking are still inlined.
PyErr_GivenExceptionMatches().
set_exc_info(): make sure to normalize exceptions.
do_raise(): Use PyErr_NormalizeException() if type is a class.
loop_subscript(): Use PyErr_ExceptionMatches() instead of raw pointer
compare for PyExc_IndexError.
- int PyErr_GivenExceptionMatches(obj1, obj2)
Returns 1 if obj1 and obj2 are the same object, or if obj1 is an
instance of type obj2, or of a class derived from obj2
- int PyErr_ExceptionMatches(obj)
Higher level wrapper around PyErr_GivenExceptionMatches() which uses
PyErr_Occurred() as obj1. This will be the more commonly called
function.
- void PyErr_NormalizeException(typeptr, valptr, tbptr)
Normalizes exceptions, and places the normalized values in the
arguments. If type is not a class, this does nothing. If type is a
class, then it makes sure that value is an instance of the class by:
1. if instance is of the type, or a class derived from type, it does
nothing.
2. otherwise it instantiates the class, using the value as an
argument. If value is None, it uses an empty arg tuple, and if
the value is a tuple, it uses just that.
classes as their second arguments. The former takes a class as the
first argument and returns true iff first is second, or is a subclass
of second.
The latter takes any object as the first argument and returns true iff
first is an instance of the second, or any subclass of second.
Also, change all occurances of pointer compares against
PyExc_IndexError with PyErr_ExceptionMatches() calls.
ExitThread(). As discussed in c.l.p, this takes care of
initialization and finalization of thread-local storage allocated by
the C runtime system. Not sure whether non-MS compilers grok this
though (but who cares :-).
scheme based on object's types, have a simple two-phase scheme based
on object's *names*:
/* To make the execution order of destructors for global
objects a bit more predictable, we first zap all objects
whose name starts with a single underscore, before we clear
the entire dictionary. We zap them by replacing them with
None, rather than deleting them from the dictionary, to
avoid rehashing the dictionary (to some extent). */
Py_Initmodule(), which is a macro wrapper around it).
The return value is now a NULL pointer if the initialization failed.
This may make old modules fail with a SEGFAULT, since they don't
expect this kind of failure. That's OK, since (a) it "never" happens,
and (b) they would fail with a fatal error otherwise, anyway.
Tons of extension modules should now check the return value of
Py_Initmodule*() -- that's on my TODO list.
Guido van Rossum
Barry Warsaw
Fred Drake