Function wrapping is unreliable when using drwrap_skip_call

Created by: kholia

Function wrapping in drwrap extension seems to be unreliable, possibly when combined with drwrap_skip_call feature.

Here is my drwrap client code (contained in nop.c),

#include "dr_api.h"
#include "utils.h"
#include "drwrap.h"
#include "drmgr.h"
#include <stdio.h>

void wrap_pre_sleep(void *wrapcxt, OUT void **user_data)
{
    unsigned long int status = 0;

    drwrap_skip_call(wrapcxt, (void *) status, 0);
}

static void module_load_event(void *drcontext, const module_data_t * mod, bool loaded)
{
    bool ok;

    app_pc towrap = (app_pc) dr_get_proc_address(mod->handle, "sleep");
    if (towrap != NULL) {
        ok = drwrap_wrap(towrap, wrap_pre_sleep, NULL);
    }
}

DR_EXPORT void
dr_init(client_id_t id)
{
    drmgr_init();
    drwrap_init();
    drmgr_register_module_load_event(module_load_event);
}

Here is my application code (contained in application.c),

#include <stdio.h>
#include <unistd.h>

int main()
{
	while (1) {
		fprintf(stderr, ".");
		sleep(256); // we will try to nop this
	}

	return 0;
}

Expected behavior,

$ ./bin64/drrun -c api/bin/libnop.so -- ../application
......... (this should keep printing without any long pauses)

sleep from application.c should never be called.

Actual behavior,

$ ./bin64/drrun -c api/bin/libnop.so -- ../application
.................................................................(long pause occurs here)

Sometimes the original wrapped function sleep is called, which is not expected!

This outcome is very puzzling to me, and I can't explain it. Hopefully, this is not a bug / limitation in DynamoRIO.

All this code is available at https://github.com/kholia/dynamorio/tree/missed-call-wrapping for testing.