| Age | Commit message (Collapse) | Author |
|---|
|
to let people use gnu 32bit compilers from e.g. Linux
|
|
This changes how we declare RPC user prototypes for device_read_inband
to use "char *data" rather than "io_buf_ptr_inband_t data". It is more
standard to pass a pointer to represent arrays compared to "char [128]". This
fixes a warning in console-client since GCC won't complain we are not
passing an exact char [128].
Also updated code to use const_io_buf_ptr_inband_t for
device_write_inband. This is a pointer to const data rather than a const
pointer.
Message-ID: <ZY7u7noOnHlyJi24@jupiter.tail36e24.ts.net>
|
|
inlined port rights.
For i686, we just change the code to use mach_port_name_inlined_t when
defining the types. This is a no-op.
For x86_64, there's a few things that are different:
- In the server code, the server handler can get inlined ports and the
array will be resized and cast as an array of mach_port_name_t. Output
parameters have a similar treatment where the inlined array in the
output is used as an array of mach_port_name_t but resized to look like
a mach_port_name_inlined_t.
- In the user side, we follow the same approach. Input ports as arrays
of mach_port_name_t are expanded into an array of mach_port_name_inlined_t.
Output ports are then converted back into an array of
mach_port_name_inlined_t so that they fit into the expected message
format.
Essentially, regardless of whether port rights are inline or out of
line, user interfaces and server stubs always receive an array of port
rights, not mach_port_name_inlined_t. However, inlined port rights will
be exchanged using mach_port_name_inlined_t.
|
|
for inlined port rights."
This reverts commit c40604042bd6e9f80e4f5fe6bc9deefb29c4b94a.
|
|
inlined port rights.
For i686, we just change the code to use mach_port_name_inlined_t when
defining the types. This is a no-op.
For x86_64, there's a few things that are different:
- In the server code, the server handler can get inlined ports and the
array will be resized and cast as an array of mach_port_name_t. Output
parameters have a similar treatment where the inlined array in the
output is used as an array of mach_port_name_t but resized to look like
a mach_port_name_inlined_t.
- In the user side, we follow the same approach. Input ports as arrays
of mach_port_name_t are expanded into an array of mach_port_name_inlined_t.
Output ports are then converted back into an array of
mach_port_name_inlined_t so that they fit into the expected message
format.
Essentially, regardless of whether port rights are inline or out of
line, user interfaces and server stubs always receive an array of port
rights, not mach_port_name_inlined_t. However, inlined port rights will
be exchanged using mach_port_name_inlined_t.
Message-ID: <20231124213041.952886-4-flaviocruz@gmail.com>
|
|
Mostly, we don't set the fields that do not exist and avoid type
mismatching (like casting unsigned short to unsigned char for
msgt_name).
We also revamp type checking to compare mach_msg_type_t to uint64_t
instead of just uint32_t as we now use the whole structure.
Message-Id: <ZGREMgn19Zptc/Pf@jupiter.tail36e24.ts.net>
|
|
Back in the wonderful old days of Mach 2, when there were no send-once
rights, dead names, or port reference counts, MIG used to have more
kinds of operations: there were functions, procedures, simpleprocedures,
routines, and simpleroutines. Routines returned an error code as their C
function return value, functions had real return values, procedures and
simpleprocedures returned void.
Functions, procedures, and simpleprocedures signalled errors by invoking
a global function. By default, a function named MsgError was invoked,
but a subsystem could specify a different function using the 'error'
directive:
error CustomErrorHandler;
In Mach 3, functions, procedures, and simpleprocedures are gone.
Routines and simpleroutines are the only remaining kinds of operations.
Some 26 years later, most of the code for generating functions,
procedures, and simpleprocedures was removed from GNU MIG in commit
7f10b4ed6a557b7a1fd1083939156a3dcf8b377e. Error directives remained,
seemingly due to an oversight. So remove them too.
Found while trying to use the word 'error' as an identifier and
receiving a cryptic syntax error from MIG.
Message-Id: <20230617203953.622120-1-bugaevc@gmail.com>
|
|
|
|
For the x86_64 ABI we want this to always fit into 1 byte. Even for
regular i686, msgt_name is always smaller than 25 (MACH_MSG_TYPE_LAST)
and we don't have plans to have more names.
Also throw an error if we deemed an RPC to be "TooLong" as that won't
work or work badly.
Tested by cross-compiling a basic Hurd system.
Message-Id: <ZFsuKtiLdwNpD6b1@jupiter.tail36e24.ts.net>
|
|
Message-Id: <ZAbgdBuKlIIiF6zA@jupiter.tail36e24.ts.net>
|
|
By using uintptr_t, we ensure that all 64 bit stubs are free of
undefined behavior since we support up to 8-byte alignment. This works
fine given that Mig will type check types that have higher alignment
requirements through the use of _Static_assert.
This even works for a 64 bit kernel / 32 bit user land since we can use
mig64 to build stubs for the 8-byte aligned kernel stubs and existing 32
bit mig to build the old user side stubs. Pragma is no longer required as we take
advantage of the compiler to drive the alignment and avoid undefined
behavior.
Message-Id: <Y/KsiXZKkq1q/D+3@jupiter.tail36e24.ts.net>
|
|
This was a miss in the previous patch given that mach_msg_type_long_t is
12 bytes long so won't neatly align to 8 bytes.
Message-Id: <Y+8VyADQ8+bRRUcp@jupiter.tail36e24.ts.net>
|
|
File is not needed.
Message-Id: <Y/K1bKVZXhYM7X9W@jupiter.tail36e24.ts.net>
|
|
Remove the concept of word_size since it is meaningless in some
architectures. This is also done in preparation to possibly introduce
8-byte aligned messages.
Message-Id: <Y+lkv0uMo/3+hbCb@jupiter.tail36e24.ts.net>
|
|
Since GNU Mach commit d30481122a5d24ad6b921062f93b9172ef922fc3,
i386/machine_types.defs defines types based on defined(__x86_64__).
Supressing the built-in macro definitions will now result in the wrong
type being silently selected.
-undef was initially introduced in commit
78b6a7665db7b2eae367e17102821cbdca231d19 without much of an explanation.
-ansi was introduced in commit 6940fb91859e46b2e96a331a029f2dc2a0ee51c9
"to avoid -Di386=1 and the like".
Since glibc has been using MIG with CPP set to a custom GCC invocation
which did *not* use either flag, it appears that everything works well
enough even without them. On the other hand, not having __x86_64__
defined most definetely causes issues for anything that does not set a
custom CPP when invoking MIG (i.e., most users). Other built-in
definitions could be used in the future in a similar way (e.g. on other
architectures); it's really more of a coincidence that they have not
been used until now, and things kept working with -undef.
Message-Id: <20230212111044.610942-8-bugaevc@gmail.com>
|
|
The kernel does not use these functions so we can avoid a few compiler
warnings. I think we could make the hurd servers not use these also but
currently these are still needed.
Message-Id: <Y9tpQNDG/mx84kV+@jupiter.tail36e24.ts.net>
|
|
|
|
c7654756ee19 ("Provide intptr_t and uintptr_t as default types") added
intptr_t and uintptr_t, but missed updating the testsuite.
|
|
7063f0aefd8a ("mig: replace boolean.h with stdbool.h") dropped
boolean.h, but missed dropping its installation.
|
|
Message-Id: <Y6CAwaY0IRqihOjU@mars>
|
|
During a system bootstrap, it is preferable that we don't require a full
hosted environment. For all other cases, we also do not need libc since
mach headers are self contained.
Message-Id: <Y5+WXzOwGkyvcDI0@mars>
|
|
Message-Id: <Y3/Z1CGL8D4OwT66@viriathus>
|
|
For kernel server or user subsystems we would initialize basic types
twice, once in main() and again for the subsystem declaration. Instead,
initialize basic types when the subsystem is declared and then throw
errors when types are defined multiple times.
Message-Id: <Y3HUt/YAKaqMMTi3@viriathus>
|
|
Basic syntax is presented below and allows users to define
nested structured types by using simpler or structure types as
members. Mig will use the C padding and alignment rules to produce
the same size as the corresponding C structures.
type timespec_t = struct {
uint32_t tv_sec;
uint32_t tv_nsec;
};
This allows us to build stubs that are more easily adaptable to other
architectures.
Message-Id: <Y2SjQSMOINY8I5Dy@viriathus>
|
|
* user.c:
- adjust comment in generated file
- set msgh_size with the same value passed to mach_msg()
Signed-off-by: Luca Dariz <luca@orpolo.org>
Message-Id: <20220628094927.442907-4-luca@orpolo.org>
|
|
* user.c: ensure fixed-length messages have the correct size. In
addition to the single-fields check, this also include padding.
Signed-off-by: Luca Dariz <luca@orpolo.org>
Message-Id: <20220628094927.442907-3-luca@orpolo.org>
|
|
On x86_64 alignment of structures is different, as the pointer size is different.
For simplicity we keep the same 4-byte alignment as used on
32-bit. This simplifies the support for 32-bit rpc on 64-bit kernels,
and also it seems not worth as an optimization, as we would need to
add more code in the ipc_kmsg* routines.
* routine.c: align both short and long descriptors
* utils.c: use a fixed alignment for data fields in structures
representing messages.
Signed-off-by: Luca Dariz <luca@orpolo.org>
Message-Id: <20220628094927.442907-2-luca@orpolo.org>
|
|
* parser.y: add information about type names
Signed-off-by: Luca Dariz <luca@orpolo.org>
Message-Id: <20220403150020.120799-3-luca@orpolo.org>
|
|
* cpu.sym: retrieve size of vm_offset_t and mach_port_name_t from
gnumach headers at compile type.
* global.{c,h}: add port size as a variable and initialize it to the
port name size.
* lexxer.l: apply port or port name size to the corresponding types,
instead of using the word size.
* parser.y: update port size if we're generating for kernel-space
(server or client). Also re-initialize default port types to account
for this change.
* type.c: use port size instead of word size in default port types and
runtime checks.
There are many assumptions about mach_port_t:
- on kernel side, its size is the same as a pointer. This allows to
replace the port name with the address of the corresponding data
structure during copyin in mach_msg()
- in mig, this is also the "word size", which is derived from gnumach
headers as the size of integer_t
- its size is also the same as natural_t, so it's possible to model
structures like mach_port_status_t as an array of integer_t in
mig. This is convenient since arrays and structures can't have
mixed types.
- its size is always the same as the port name size
This patch does not change the current behaviour on 32-bit kernels,
but allows for some of these assumptions to be broken on 64-bit
kernels. This is needed to have 32-bit port names on 64-bit kernels
and be able to support a 32-bit userspace. It still leaves the choice
for a 64-bit userspace, if all integer_t and natural_t are to be
extended to 64 bit.
However keeping 32-bit port names seems to be the right thing, based on
previous discussions [1], even for a 64-bit kernel.
The only assumption kept is that in kernel-space ports are always the
size of a pointer, as they refer to a data structure and not to a
specific port name. To ensure this is true for various user/kernel
combinations, we dynamically change the port size if we're generating
code for kernel-space server or clients, and keep the size of a port the
same of a port name for user-space servers and clients.
[1] https://lists.gnu.org/archive/html/bug-hurd/2012-04/msg00000.html
Signed-off-by: Luca Dariz <luca@orpolo.org>
Message-Id: <20220403150020.120799-2-luca@orpolo.org>
|
|
autoconf consider that 'AC_FUNC_VPRINTF' is obsolescent.
see 'info autoconf AC_FUNC_VPRINTF'
* Makefile.am: remove 'vprint.c' on src files.
configure.ac: remove 'AC_FUNC_VPRINTF' macro.
vprint.c: delete file.
Message-Id: <Ym+H/OM0PnucKO4w@debian>
|
|
This is only done when data is inlined with a concrete size. It
ensures the C and Mig types have the same size in the target arch.
Tested by building the hurd package. No assertions were triggered.
Message-Id: <YekIQaxvs+4FrHyw@viriathus>
|
|
Although in practice the buffers can be modified since they are from the
message, it leads to missing const where it would otherwise make sense.
|
|
to avoid forcing the caller to respect the definite string size.
|
|
We do not actually use yywrap, and this allows to avoid requiring libflex,
making cross-compilations simpler.
* lexxer.l: Add noyywrap option.
|
|
This will allow to relieve constraints in callers, e.g. dir_lookup("") would
otherwise produce a warning with gcc-11 since char[1024] would mean that
dir_lookup would read all 1024 characters while it is not.
* utils.c (UserVarQualifier): Also use const_ qualifier when type is
string_t.
* tests/includes/types.h (const_string_t): New type.
|
|
* cpu.sym: Add sizeof_long, sizeof_float, sizeof_double.
|
|
This replaces the implementation from the University of Utah, covered by the
advertising clause, with the implementation from CMU, picked up from the GNU
Mach source, which is free from the advertising clause.
This includes the addition of the sizeof as seen used by mig with
git grep word_size
git grep sizeof_
|
|
This replaces the implementation from the University of Utah, covered by the
advertising clause, with the implementation from CMU, picked up from the GNU
Mach source, which is free from the advertising clause.
|
|
This replaces the implementation from the University of Utah, covered by the
advertising clause, with the implementation from CMU, picked up from the GNU
Mach source, which is free from the advertising clause.
|
|
Also removed itMakeIntType which was not used anymore.
Users can use char, int, and short types without having to define them.
These types are defined using the builtin MACH_MSG_TYPE_* types and are
architecture independent since they have the same size as the C char,
short and int.
If these basic types are redefined, MIG will still produce stub code but
will produce a warning.
* cpu.sym: Define sizeof_int, char, short.
* tests/base_types.defs: Remove int.
* tests/good/complex-types.defs: Use byte instead of char.
* tests/good/directions: Remove char and int.
* tests/good/types.defs: Remove char and also use short as a parameter
in 'alltypes'.
* type.c: Define itCIntTypeDecl. Remove itMakeIntType. Call itInsert for
char, short and int.
Message-Id: <20160419070513.GA12642@debian>
|
|
* configure.ac (AM_INIT_AUTOMAKE): Change dist-bzip2 to dist-xz.
|
|
* header.c (WriteServerHeader): Print __inline instead of inline.
|
|
Users of RPCs want to be able to pass pointers to const data, so add
const qualifiers to RPCs as appropriate.
* utils.c (UserVarConst, UserVarQualifier): New functions.
(WriteUserVarDecl): Use UserVarQualifier to qualify function parameter.
(WriteFieldDeclPrim): Use UserVarConst to qualify pointer to user
variable.
|
|
Nowadays' compilers are able to recognize memcpy and replace it
appropriately without having to tell them so through a structure
assignment. That also avoids warnings about type puning.
* utils.c (WriteCopyType): Emit memcpy call instead of type-puned
assignement.
|
|
* utils.c (WriteStaticLongDecl): Explicitly cast name to unsigned short to
ignore truncation of MACH_MSG_TYPE_POLYMORPHIC i.e. -1.
|
|
|
|
* configure.ac (AC_INIT): Set version to 1.8.
* NEWS: Finalize for 1.8.
|
|
... to match the GNU Mach and Hurd NEWS files.
* NEWS: Remove bullet points.
|
|
|
|
* utils.c (WriteCheckDecl, WriteStaticShortDecl): Explicitly cast name to
unsigned char to ignore truncation of MACH_MSG_TYPE_POLYMORPHIC i.e. -1.
|
