You can use the webserver example and connect to it in a loop. At some point memory of the web-server gets exhausted. Investigate where in the LwIP stack this comes from, and provide a fix.

Email on genode-main:

Date: Wed, 21 Sep 2011 14:57:48 +0200
From: Le Bao TRUONG <[email protected]>
To: [email protected]
Subject: Error when tried to build manually Hello Demo

Hi all,

I would like build manually the 'Hello Tutorial' demo. But I have an
error to execute ./core.

I follow the tutorial :

1. Build the server side by make hello/server

2. Build the client side by

make drivers/timer
make hello/client

Then, the tutoriel told me to execute a binary ./core. But I couldn't
find core file in my bin/ directory so I build it by

make core hello/client

Then I execute ./core, my program is stopping at

int main(): --- create local services ---
int main(): --- start init ---
Could not open file "init"
ROM module "init" not present
int main(): transferred 79 MB to init


Would you like give me some advice ?

Thanks,

Le Bao
``

Right now, a server does only use the capability's unique ID to look up the server object. We should compare the unique ID to the pt_sel reported for the invokation.

Some repositories bring and use their own variants of these functions, e.g., base-fiasco.

The calls of Genode functions within the L4Linux support library have to be investigated to never throw C++ exceptions, otherwise Linux simply aborts without additional information.

When using %x in a format string on 64-bit platforms, the resulting string might get malformed when using a number equally or higher than 0xf0000000. Probably due to conversion errors in the format_string classes.

The following patch includes a test case, that triggers the error on 64-bit Genode/linux and Genode/Fiasco.OC. Just run make run/snprintf.

diff --git a/os/run/snprintf.run b/os/run/snprintf.run
new file mode 100644
--- /dev/null
+++ b/os/run/snprintf.run
@@ -0,0 +1,38 @@
+build "core init test/snprintf"
+
+create_boot_directory
+
+install_config {
+<config>
+   <parent-provides>
+       <service name="ROM"/>
+       <service name="RAM"/>
+       <service name="IRQ"/>
+       <service name="IO_MEM"/>
+       <service name="IO_PORT"/>
+       <service name="CAP"/>
+       <service name="PD"/>
+       <service name="RM"/>
+       <service name="CPU"/>
+       <service name="LOG"/>
+   </parent-provides>
+   <default-route>
+       <any-service> <parent/> <any-child/> </any-service>
+   </default-route>
+   <start name="test-snprintf">
+       <resource name="RAM" quantum="1M"/>
+   </start>
+</config>
+}
+
+build_boot_image "core init test-snprintf"
+
+append qemu_args "-m 64"
+
+run_genode_until "child exited with exit value 0" 10
+
+grep_output {^\[init -> test-snprintf}
+
+compare_output_to {
+[init -> test-snprintf] 0xf0000000
+}
\ No newline at end of file
diff --git a/os/src/test/snprintf/main.cc b/os/src/test/snprintf/main.cc
new file mode 100644
--- /dev/null
+++ b/os/src/test/snprintf/main.cc
@@ -0,0 +1,11 @@
+#include <base/printf.h>
+#include <util/arg_string.h>
+
+int main ()
+{
+   using namespace Genode;
+   static char str[1024];
+   snprintf(str, sizeof(str), "0x%x", 0xf0000000);
+   printf("%s\n", str);
+   return 0;
+}
diff --git a/os/src/test/snprintf/target.mk b/os/src/test/snprintf/target.mk
new file mode 100644
--- /dev/null
+++ b/os/src/test/snprintf/target.mk
@@ -0,0 +1,3 @@
+TARGET = test-snprintf
+LIBS   = cxx env
+SRC_CC = main.cc

The client interface should use Genode's signalling framework rather than the tricky out-of-order RPC mechansism. This way, we could unify the front-end code for all platforms.

We should make sure that the RM and ROM services provided by GDB monitor to its child limit the amount of memory that has been transferred for the session by the child.

Currently, the GPU driver is linked against the Gallium3D program. Consequently, only one program can use the GPU at a time. To allow multiple 3D-graphics application to run at the same time, the GPU driver must be executed as a separate GPU service.

Apparently, there is a race condition in the page-fault handling of the NOVA base platform. It triggers when subsequently starting and killing instances of testnit in Genode's default demo scenario.

Child::upgrade(Session_capability to_session, const char *amount) in e.g.:
base/include/base/child.h is broken for services rm, cpu, ram, and pd.
These services are not in the session pool, and thereby upgrade exists without doing anything.

Liquid framebuffer is still using the ancient XML style relying on tags only. We should change that and possibly generalize the parsing of geometry information with similar use cases (programs that take xywh as config arguments).

The thread iterator of the CPU session is an ancient relic from a profiling-tool experiment we did together with Torsten Frenzel some years back. In the meantime, we have learned better ways to observe subsystems (by virtualizing core's services) such that these functions are considered obsolete. However, unfortunately, there are a few users of these functions (e.g., OKLinux, which uses these functions to determine its own main-thread capability). We should revisit all current use cases of the thread iterator, change their implementations to use a different mechanism, and remove the thread iterator functions ('first', 'next') from the generic interface.

Some run scripts are not supported on some platforms. This condition is normally caught by some checks at the beginning of the respective run script. If such a check triggers, the run script should exit with a specific exit code that tells 'autopilot' that the script is unsupported on the platform. This way, the autopilot statistics would be even more meaningful.

When trying to compile libSDL on ARM, a compile-time assertion fails:

SDL_stdinc.h:133:1: error: size of array ‘SDL_dummy_enum’ is negative

For now, I have restricted sdl.mk to build the library only on x86 but we should better investigate the issue on ARM.

By now, capability selectors are only allocated, but never freed. Introduce smart-pointer to keep used selectors and remove not used selectors.

After this code

/* reset ownership if one thread 'lock' twice */
if (_owner == myself)
    _owner = Applicant(thread_invalid_id());

_last_applicant is invalid

The third test (stress test) of the 'os/run/signal.run' test case runs far slower on Pistachio than on all other platforms. We should find out why.

The new interfaces at 'os/include/nic/' generalize the glue code between networking drivers and Genode and have the potential to reduce code duplications among the various NIC-related services (nic_drv, nic_bridge, ...). Let's use this potential!

packet_content() returns a NULL pointer on failure, but it should throw an exception as the base API does.

All said files contain just 0..

Support for changing the size of framebuffers is highly desired. One use case is changing screen resolutions at runtime. But more importantly, we want applications to adapt their framebuffer sessions to window-size changes in our future working environment. Intuitively, I would add a way to install a signal handler to a 'Framebuffer_session'. The signal handler receives a signal each time, the framebuffer properties change. So a client can request the new framebuffer information and adapt itself. However, this might not be enough. A client should be able to specify the degree of freedom it is able to deal with. E.g., one client may be able to adapt itself to a changed color depth whereas another client does not.

Modern binutils versions 2.18, 2.19, 2.20+ allow splitting of symbols from ELF files, leaving only symbol link in target ELF.
This allows keep symbols separate with release code and makes debugging considerably easier.
E.g. you reference the elf file in gdb/insight "symbol-file foo" -> gdb knows about .debuglink sections and loads the referenced external symbol file.

Example for symbol splitting step for base\mk\prg.mk:

TARGET_DBG = $(INSTALL_DIR)/.debug/$(TARGET).dbg

$(INSTALL_DIR)/$(TARGET): $(TARGET_DBG)
$(MSG_STRIP)$(TARGET)
$(VERBOSE)ln -sf `pwd`/$(TARGET) $@
$(VERBOSE)$(CUSTOM_OBJCOPY) --add-gnu-debuglink=$(TARGET_DBG) --strip-unneeded $(abspath $(TARGET)) $@
$(VERBOSE)chmod 755 $@

$(TARGET_DBG): $(TARGET)
$(MSG_SYMBOLS)$(TARGET)
$(VERBOSE)$(CUSTOM_OBJCOPY) --only-keep-debug $(abspath $<) $@

This would produce an ELF file as usual, stripped of all symbol information "foo" and a file "foo.dbg" containing all symbol info for debugging in ".debug" subfolder.

global.mk:

MSG_SYMBOLS = @$(ECHO) " SYMBOLS "
MSG_STRIP = @$(ECHO) " STRIP+INST "

For testing, the 'eglgears.run' script must be complemented with the pl110 driver as done in 'demo.run'.

We should try to submit our Fiasco.OC patch for user-level single stepping to the mainline Fiasco.OC development. I.e., posting the patch to l4-hackers, accompanied with the following information:

• What was the reason to create the patch? (description of the desired feature / use case)
• How does the patch resolve the problem? How does it fix into the current kernel? (changed syscalls, bits, etc.)
• Which are the implications for regular Fiasco.OC users? (Fiasco.OC config)
• Copyright sign-off

GCC manpage about the Wconversion option:

Warn for implicit conversions that may alter a value. This includes conversions between real and integer, like abs (x)' whenx' is double'; conversions between signed and unsigned, likeunsigned ui = -1'; and conversions to smaller types, like sqrtf (M_PI)'. Do not warn for explicit casts likeabs ((int) x)' and ui = (unsigned) -1', or if the value is not changed by the conversion like inabs (2.0)'. Warnings about conversions between signed and unsigned integers can be disabled by using `-Wno-sign-conversion'.

In some use-cases (like QT event loop) we need a fine granularity for the timed semaphore, on the other hand the overhead of a permanently time lookup isn't desireable. Program the wake-up of the scheduler with respect of the demands of all currently active timed-semaphores instead of using a fixed wake-up schedule.

When building, we get strange warnings about array bounds in 'string.h'.

Please consider completely replacing current genode version of gdtoa (BSD) library with newer one from: http://www.netlib.org/fp/ (gdtoa.tgz)

libc\src\lib\libc\gdtoa*

The reason is that gdtoa code gets miscompiled by gcc 4.4 and later due to bad programming practices (abusing aliasing rules).
For glory details see: http://patrakov.blogspot.com/2009/03/dont-use-old-dtoac.html

Symptoms: hangs on gcc 4.4+ x86 and ARM targets that sprintf double params.

With this update you also need to update:

libc\src\lib\libc\stdio\vfprintf.c
libc\src\lib\libc\stdio\vfwprintf.c
libc\src\lib\libc\stdio\xprintf_float.c

and remove following redefines:

define dtoa __dtoa

define freedtoa __freedtoa

Instead of having a 'Rm_session_component::dissolve' function, we should better extend the RM session interface via a 'remove_client' function and handle the protocol thread destruction analogously to how a thread is created.

Port of the OKLinux nitpicker/fb stub driver to L4Linux.

In 'genode/main.cc', we still use the IPC framework directly for triggering the input handling. We should change this code to use the RPC framework instead.

Some contrib code depends on trylock() to implement lock-state-dependent policies. To now, I've not found a way to safely implement this feature outside of Genode::Lock. A prominent example using trylock() is Madwifi.

On platforms using APIC and not PIC (NOVA, Fiasco.OC) to use certain drivers (like SATA) we need the concrete IRQ -> Device mapping. Parse ACPI, use DSDT in there to find out IRQ -> Device mapping.

The RM fault handling on Pistachio does not work yet. The problem seems to be that the pager thread receives page faults as short IPC messages whereas core-local wake-up notifications are sent to the pager as long IPC messages (via the Genode IPC framework). Hence, the pager thread wrongly considers wakeup messages as IPC errors. My proposed solution is to replace the use of the IPC framework (in 'include/base/pager.h') for the core-internal wake-up call by a kernel-specific implementation of 'Pager_object::wake_up'.

For testing, the 'base/run/rm_fault.run' script can be used.

Currently, only x86_32 is supported.

Currently, a global background can be assigned by a client. This way, the backdrop program can supply a background image. However, there is no proper handling of the case that this session gets closed. Right now, it never happens (for backdrop started by init) but in general, this should be fixed.

The current POSIX-like exit() handling does not fit Genode very well. Additionally, it interferes (historically) with atexit handling and static destructors. By the best the child must assume to be killed right after calling Parent::exit(), which means no finalizers will be called.

The current input session interface relies on polling, which is pretty lame and introduces unwanted latencies when chaining input-session implementations (e.g., running nested nitpicker instances). The use of signals would improve the situation.

Handling flexpages is not easy (e.g., look at the unmap code in OKL4). Right now, such code appears at several places. We could improve the situation with a generic flexpage iterator, which takes an address range as constructor argument and generates log2-sized flexpages spanning the whole range.

Linux supports delegation of file descriptors via Unix-domain sockets called "SCM_RIGHTS via ancillary messages" (man 7 unix):

SCM_RIGHTS
  Send or receive a set of open file descriptors from another process.
  The data portion contains an integer array of the file descriptors.
  The passed file descriptors behave as though they have been created
  with dup(2).

A good read seems to be http://www.lst.de/~okir/blackhats/node121.html

t seems the tcpip thread wakes up with a bogus message from sys_mbox_fetch(). This can easily reproduced under Linux:

terminal 1> make run/lwip_lx
terminal 2> ab -c 2 -n 5000 http://10.0.2.16/

Currently, Cpu_session::set_pager() needs to be called before Cpu_session::exception_handler() for the exception handler to be set on Genode/Fiasco.OC. This calling order requirement should be documented somewhere.

The Fiasco.OC debugger capability shall be used in future to print out information (in core) over the kernel debugger. This might also solve some issues on multi-processor platforms (serializing the output).

The verbose flags in the ps2 mouse and keyboard implementation must be defined at compile time. However, the ps2 driver is often used for diagnostic purposes. In such a case, the driver must be recompiled. It would be convenient to have 'verbose_mouse' 'verbose_keyboard' arguments to the config node of the PS/2 driver.

We need to run Genode subsystems within dedicated chroot environments. Such an environment must use only one socket to communicate to the outside world.

Currently, the Gallium3D/GEM stack is using msleep to "synchronize". This should be replaced by proper interrupt handling.

The set of functionality that we always link to Genode programs becomes pretty clear now. It consists of the libraries listed in 'base/mk/base-libs.mk'. To simplify the directory layout and make the distinction between base and non-base libraries more clear, those libraries could be merged into a single 'base' library. In contrast to other static libraries, this base library is guaranteed to be linked only once (to 'ld.lib.so' for dynamic binaries). Core is the only program that does not link to this library. It assembles the needed parts of the library by itself. Eventually, we could use two "base" libraries, namely 'cxx' and 'base' because the library description file for cxx is rather complicated.

Please consider adding "void usleep(unsigned us)" method to base interface because some platforms can actually implement such feature.
Opcode "USLEEP" .
On platforms where impl is not possible/feasible you could add default iml that just:

(us < 500) ? msleep(1) : msleep((us + 500)/1000)

To keep abstract interface class, consider something like Session_base with two pure virtual members: msleep, usleep.
class Session : public Session_base which provides default impl of usleep. This way all the current timer drivers/servers code doesnt need to be changed to implement usleep() (even if its dummy).

Port of the OKLinux audio_out stub driver to L4Linux.