IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
the ntvfs_generic mapping functions rather than sending the exact
function asked for. This allows the generic mapping functions to be
tested by comparing the behaviour of smbtorture against two cifs
backend shares, one using "cifs:mapgeneric = true" and the other
"cifs:mapgeneric = False"
preparation for the full share modes and ntcreatex code that I am
working on.
highlights include:
- changed the way a backend determines if it is allowed to process a
request asynchronously. The previous method of looking at the
send_fn caused problems when an intermediate ntvfs module disabled
it, and the caller then wanted to finished processing using this
function. The new method is a REQ_CONTROL_MAY_ASYNC flag in
req->control_flags, which is also a bit easier to read
- fixed 2 bugs in the readbraw server code. One related to trying to
answer a readbraw with smb signing (which can't work, and crashed
our signing code), the second related to error handling, which
attempted to send a normal SMB error packet, when readbraw must
send a 0 read reply (as it has no header)
- added several more ntvfs_generic.c generic mapping functions. This
means that backends no longer need to implement such esoteric
functions as SMBwriteunlock() if they don't want to. The backend
can just request the mapping layer turn it into a write followed by
an unlock. This makes the backends considerably simpler as they
only need to implement one style of each function for lock, read,
write, open etc, rather than the full host of functions that SMB
provides. A backend can still choose to implement them
individually, of course, and the CIFS backend does that.
- simplified the generic structures to make them identical to the
principal call for several common SMB calls (such as
RAW_WRITE_GENERIC now being an alias for RAW_WRITE_WRITEX).
- started rewriting the pvfs_open() code in preparation for the full
ntcreatex semantics.
- in pvfs_open and ipc_open, initially allocate the open file
structure as a child of the request, so on error we don't need to
clean up. Then when we are going to succeed the open steal the
pointer into the long term backend context. This makes for much
simpler error handling (and fixes some bugs)
- use a destructor in the ipc backend to make sure that everthing is
cleaned up on receive error conditions.
- switched the ipc backend to using idtree for fnum allocation
- in the ntvfs_generic mapping routines, use a allocated secondary
structure not a stack structure to ensure the request pointer
remains valid even if the backend replies async.
This adds a pvfs_wait_message() routine which uses the new messaging
system, event timers and talloc destructors to give a nice generic
async event handling system with a easy to use interface. The
extensions to pvfs_lock.c are based on calls to pvfs_wait_message()
routines.
We now pass all of our smbtorture locking tests, although while
writing this code I have thought of some additonal tests that should
be added, particularly for lock cancel operations. I'll work on that
soon.
This commit also extends the smbtorture lock tests to test the rather
weird 0xEEFFFFFF locking semantics that I have discovered in
win2003. Win2003 treats the 0xEEFFFFFF boundary as special, and will
give different error codes on either side of it. Locks on both sides
are allowed, the only difference is which error code is given when a
lock is denied. Anyone like to hazard a guess as to why? It has
me stumped.
caller doesn't have to worry about the constraint of only opening a
database a single time in a process. These wrappers will ensure that
only a single open is done, and will auto-close when the last instance
is gone.
When you are finished with a database pointer, use talloc_free() to
close it.
note that this code does not take account of the threads process
model, and does not yet take account of symlinks or hard links to tdb
files.
can correctly (and quickly!) handle unaligned little endian memory
accesses, just like i386. This should reduce code size and speeds
things up quite a lot on ppc, at the expense of some inline asm code
(whcih means it only works with gcc)
on systems that have the intptr_t type, and for systems that don't
have it, they also almost certainly won't have -Wcast-qual, so we can use a
void* cast.
The motivation for this change was to avoid having to convert to/from
ucs2 strings for so many operations. Doing that was slow, used many
static buffers, and was also incorrect as it didn't cope properly with
unicode codepoints above 65536 (which could not be represented
correctly as smb_ucs2_t chars)
The two core functions that allowed this change are next_codepoint()
and push_codepoint(). These functions allow you to correctly walk a
arbitrary multi-byte string a character at a time without converting
the whole string to ucs2.
While doing this cleanup I also fixed several ucs2 string handling
bugs. See the commit for details.
The following code (which counts the number of occuraces of 'c' in a
string) shows how to use the new interface:
size_t count_chars(const char *s, char c)
{
size_t count = 0;
while (*s) {
size_t size;
codepoint_t c2 = next_codepoint(s, &size);
if (c2 == c) count++;
s += size;
}
return count;
}
which is much clearer and simpler to use. It removes a specific parent
from a pointer, no matter whether that parent is a "reference" or a
direct parent. This gives complete control over the free process.
the % coverage in terms of lines of code of a test suite. I thought a
good first place to start with gcov was the talloc test suite. When I
started the test suite covered about 60% of all lines of code in
talloc.c, and now it covers about 99%. The only lines not covered are
talloc corruption errors, as that would cause smb_panic() to fire.
It will be interesting to try gcov on the main Samba test suite for
smbd. We won't achieve 100% coverage, but it would be nice to get to
90% or more.
I also modified the talloc.c sources to be able to be build standalone, using:
gcc -c -D_STANDALONE_ -Iinlcude lib/talloc.c
that should make it much easier to re-use talloc in other projects
a const pointer really means that "the data pointed to by this pointer
won't change", and that is certainly true of talloc(). The fact that
some behind-the-scenes meta-data can change doesn't matter from the
point of view of const.
this fixes a number of const warnings caused by const data structures
being passed as talloc contexts. That will no longer generate a
warning.
also changed the talloc leak reporting option from --leak-check to
--leak-report, as all it does is generate a report on exit. A new
--leak-report-full option has been added that shows the complete tree
of memory allocations, which is is quite useful in tracking things down.
NOTE: I find it quite useful to insert talloc_report_full(ptr, stderr)
calls at strategic points in the code while debugging memory
allocation problems, particularly before freeing a major context (such
as the connection context). This allows you to see if that context has
been accumulating too much data, such as per-request data, which
should have been freed when the request finished.
taking a context (so when you pass a NULL pointer you end up with
memory in a top level context). Fixed it by changing the API to take a
context. The context is only used if the pointer you are reallocing is
NULL.
rather than manual reference counts
- properly support SMBexit in the cifs and posix backends
- added a logoff method to all backends
With these changes the RAW-CONTEXT test now passes against the posix backend
to get auto-naming of pointers very cheaply.
- fixed a couple of memory leaks found with the new tricks
A typical exit report for smbd is now:
talloc report on 'null_context' (total 811 bytes in 54 blocks)
auth/auth_sam.c:334 contains 20 bytes in 1 blocks
struct auth_serversupplied_info contains 498 bytes in 33 blocks
UNNAMED contains 8 bytes in 1 blocks
lib/data_blob.c:40 contains 16 bytes in 1 blocks
iconv(CP850,UTF8) contains 61 bytes in 4 blocks
iconv(UTF8,CP850) contains 61 bytes in 4 blocks
iconv(UTF8,UTF-16LE) contains 67 bytes in 4 blocks
iconv(UTF-16LE,UTF8) contains 67 bytes in 4 blocks
UNNAMED contains 13 bytes in 1 blocks
which is much better than before
auto-derived from the type you are allocating. This is done with
basically zero overhead by relying on the stringify operator in cpp
producing string constants.
the result is that --leak-check nicely names all pointers that come
from talloc_p()
const warnings for a long time, and no real way to approach a
solution. Some of them are unavoidable due to the way the C standard
works (for example, any function that provides strchr() like
functionality _must_ produce a const warning)
I will be converting a bunch of places that currently produce const
warnings to use the discard_const_p(). Some of these will be
unavoidable const problems, some of them will be ones we will fix up
over time. At least this change means we will no longer be swamped
with const warnings, and we will easily be able to see when new
problems emerge.
server side request structure to prevent a structing being freed in
some circumstances. This change replaces this with the much more
robust mechanism of talloc_increase_ref_count().
- Samba4-style code in lib/registry (struct registry_key instead of REG_KEY, etc)
- Use hives (like Windows has drives) instead of one root key (like a Unix FS)
- usability fixes in the GTK utilities (autodetect the username,
enable/disable options, etc)
- fix gwsam compile
- several bugfixes in the registry rpc code
- do charset conversion in nt4 registry backend
original core level calls). The old code was completely wrong in many respects.
also fixed the EA_SIZE level in the server
extended the RAW-SEARCH test suite to test the new code properly
I had previously thought this was unnecessary, as windows doesn't use
standards compliant UTF-16, and for filesystem operations treats bytes
as UCS-2, but Bjoern Jacke has pointed out to me that this means we
don't correctly store extended UTF-16 characters as UTF-8 on
disk. This can be seen with (for example) the gothic characters with
codepoints above 64k.
This commit also adds a LOCAL-ICONV torture test that tests the first
1 million codepoints against the system iconv library, and tests 5
million random UTF-16LE buffers for identical error handling to the
system iconv library.
the lib/iconv.c changes need backporting to samba3
I plan on replacing the concept by adding a generic destructor in all talloc ptrs, so you can do:
talloc_set_destructor(ptr, my_destructor);
to setup a function that will be called on free.
It simplifies our structure handling a lot, making the code shorter
and easier to understand. Look at the diff carefully and see if you
can understand it. If you're still confused then please ask.
This version does the following:
1) talloc_free(), talloc_realloc() and talloc_steal() lose their
(redundent) first arguments
2) you can use _any_ talloc pointer as a talloc context to allocate
more memory. This allows you to create complex data structures
where the top level structure is the logical parent of the next
level down, and those are the parents of the level below
that. Then destroy either the lot with a single talloc_free() or
destroy any sub-part with a talloc_free() of that part
3) you can name any pointer. Use talloc_named() which is just like
talloc() but takes the printf style name argument as well as the
parent context and the size.
The whole thing ends up being a very simple piece of code, although
some of the pointer walking gets hairy.
So far, I'm just using the new talloc() like the old one. The next
step is to actually take advantage of the new interface
properly. Expect some new commits soon that simplify some common
coding styles in samba4 by using the new talloc().
