Boot Linux faster!

Check our new training course

Boot Linux faster!

Check our new training course
and Creative Commons CC-BY-SA
lecture and lab materials

Bootlin logo

Elixir Cross Referencer

/* main.c - load/store portion of FPU sharing test */

/*
 * Copyright (c) 2011-2014 Wind River Systems, Inc.
 *
 * SPDX-License-Identifier: Apache-2.0
 */

/*
DESCRIPTION
This module implements the load/store portion of the FPU sharing test. This
microkernel version of this test utilizes a pair of tasks, while the nanokernel
verions utilizes a task and a fiber.

The load/store test validates the nanokernel's floating point unit context
save/restore mechanism. This test utilizes a pair of threads of different
priorities that each use the floating point registers. The context
switching that occurs exercises the kernel's ability to properly preserve the
floating point registers. The test also exercises the kernel's ability to
automatically enable floating point support for a task, if supported.

FUTURE IMPROVEMENTS
On architectures where the non-integer capabilities are provided in a hierarchy,
for example on IA-32 the USE_FP and USE_SSE options are provided, this test
should be enhanced to ensure that the architectures' _Swap() routine doesn't
context switch more registers that it needs to (which would represent a
performance issue).  For example, on the IA-32, the test should issue
a fiber_fp_disable() from main(), and then indicate that only x87 FPU
registers will be utilized (fiber_fp_enable()).  The fiber should continue
to load ALL non-integer registers, but main() should validate that only the
x87 FPU registers are being saved/restored.
 */

#ifndef CONFIG_FLOAT
#error Rebuild with the FLOAT config option enabled
#endif

#ifndef CONFIG_FP_SHARING
#error Rebuild with the FP_SHARING config option enabled
#endif

#if defined(CONFIG_ISA_IA32)
#ifndef CONFIG_SSE
#error Rebuild with the SSE config option enabled
#endif
#endif /* CONFIG_ISA_IA32 */

#include <zephyr.h>

#if defined(CONFIG_ISA_IA32)
  #if defined(__GNUC__)
    #include <float_regs_x86_gcc.h>
  #else
    #include <float_regs_x86_other.h>
  #endif /* __GNUC__ */
#elif defined(CONFIG_CPU_CORTEX_M4)
  #if defined(__GNUC__)
    #include <float_regs_arm_gcc.h>
  #else
    #include <float_regs_arm_other.h>
  #endif /* __GNUC__ */
#endif

#include <arch/cpu.h>
#include <tc_util.h>
#include "float_context.h"
#include <stddef.h>
#include <string.h>

#define MAX_TESTS 500

/* space for float register load/store area used by low priority task */

static struct fp_register_set float_reg_set_load;
static struct fp_register_set float_reg_set_store;

/* space for float register load/store area used by high priority thread */

static struct fp_register_set float_reg_set;

/* stack for high priority fiber (also use .bss for float_reg_set) */

static char __stack fiber_stack[1024];

static struct nano_timer fiber_timer;
static void *dummy_timer_data;	/* allocate just enough room for a pointer */

/* flag indicating that an error has occurred */

int fpu_sharing_error;

/*
 * Test counters are "volatile" because GCC may not update them properly
 * otherwise. (See description of pi calculation test for more details.)
 */

static volatile unsigned int load_store_low_count = 0;
static volatile unsigned int load_store_high_count = 0;

static void load_store_high(int, int);

/**
 *
 * @brief Low priority FPU load/store thread
 *
 * @return N/A
 */

void main(void)
{
	unsigned int i;
	unsigned char init_byte;
	unsigned char *store_ptr = (unsigned char *)&float_reg_set_store;
	unsigned char *load_ptr = (unsigned char *)&float_reg_set_load;

	volatile char volatile_stack_var = 0;

	PRINT_DATA("Floating point sharing tests started\n");
	PRINT_LINE;

	/*
	 * Start a single fiber which will regularly preempt the background
	 * task, and perform similiar floating point register manipulations
	 * that the background task performs; except that a different constant
	 * is loaded into the floating point registers.
	 */

	task_fiber_start(fiber_stack,
			 sizeof(fiber_stack),
			 load_store_high,
			 0,	/* arg1 */
			 0,	/* arg2 */
			 5,	/* priority */
			 FP_OPTION /* options */
			 );

	/*
	 * Initialize floating point load buffer to known values;
	 * these values must be different than the value used in other threads.
	 */

	init_byte = MAIN_FLOAT_REG_CHECK_BYTE;
	for (i = 0; i < SIZEOF_FP_REGISTER_SET; i++) {
		load_ptr[i] = init_byte++;
	}

	/* Keep cranking forever, or until an error is detected. */

	for (load_store_low_count = 0; ; load_store_low_count++) {

		/*
		 * Clear store buffer to erase all traces of any previous
		 * floating point values that have been saved.
		 */

		memset(&float_reg_set_store, 0, SIZEOF_FP_REGISTER_SET);

		/*
		 * Utilize an architecture specific function to load all the
		 * floating point registers with known values.
		 */

		_load_all_float_registers(&float_reg_set_load);

		/*
		 * Waste some cycles to give the high priority load/store
		 * thread an opportunity to run when the low priority thread is
		 * using the floating point registers.
		 *
		 * IMPORTANT: This logic requires that sys_tick_get_32() not
		 * perform any floating point operations!
		 */

		while ((sys_tick_get_32() % 5) != 0) {
			/*
			 * Use a volatile variable to prevent compiler
			 * optimizing out the spin loop.
			 */
			++volatile_stack_var;
		}

		/*
		 * Utilize an architecture specific function to dump the
		 * contents of all floating point registers to memory.
		 */

		_store_all_float_registers(&float_reg_set_store);

		/*
		 * Compare each byte of buffer to ensure the expected value is
		 * present, indicating that the floating point registers weren't
		 * impacted by the operation of the high priority thread(s).
		 *
		 * Display error message and terminate if discrepancies are
		 * detected.
		 */

		init_byte = MAIN_FLOAT_REG_CHECK_BYTE;

		for (i = 0; i < SIZEOF_FP_REGISTER_SET; i++) {
			if (store_ptr[i] != init_byte) {
				TC_ERROR("load_store_low found 0x%x instead of 0x%x @ offset 0x%x\n",
						store_ptr[i],
						init_byte, i);
				TC_ERROR("Discrepancy found during iteration %d\n",
						load_store_low_count);
				fpu_sharing_error = 1;
			}
			init_byte++;
		}

		/*
		 * Terminate if a test error has been reported.
		 */

		if (fpu_sharing_error) {
			TC_END_RESULT(TC_FAIL);
			TC_END_REPORT(TC_FAIL);
			return;
		}

#if defined(CONFIG_ISA_IA32)
		/*
		 * After every 1000 iterations (arbitrarily chosen), explicitly
		 * disable floating point operations for the task. The
		 * subsequent execution of _load_all_float_registers() will result
		 * in an exception to automatically re-enable floating point
		 * support for the task.
		 *
		 * The purpose of this part of the test is to exercise the
		 * task_float_disable() API, and to also continue exercising
		 * the (exception based) floating enabling mechanism.
		 */
		if ((load_store_low_count % 1000) == 0) {
			task_float_disable(sys_thread_self_get());
		}
#elif defined(CONFIG_CPU_CORTEX_M4)
		/*
		 * The routine task_float_disable() allows for thread-level
		 * granularity for disabling floating point. Furthermore, it
		 * is useful for testing on the fly thread enabling of floating
		 * point. Neither of these capabilities are currently supported
		 * for ARM.
		 */
#endif
	}
}

/**
 *
 * @brief High priority FPU load/store thread
 *
 * @return N/A
 */

void load_store_high(int unused1, int unused2)
{
	unsigned int i;
	unsigned char init_byte;
	unsigned char *reg_set_ptr = (unsigned char *)&float_reg_set;

	ARG_UNUSED(unused1);
	ARG_UNUSED(unused2);

	/* initialize timer; data field is not used */

	nano_timer_init(&fiber_timer, (void *)dummy_timer_data);

	/* test until the specified time limit, or until an error is detected */

	while (1) {
		/*
		 * Initialize the float_reg_set structure by treating it as
		 * a simple array of bytes (the arrangement and actual number
		 * of registers is not important for this generic C code).  The
		 * structure is initialized by using the byte value specified
		 * by the constant FIBER_FLOAT_REG_CHECK_BYTE, and then
		 * incrementing the value for each successive location in the
		 * float_reg_set structure.
		 *
		 * The initial byte value, and thus the contents of the entire
		 * float_reg_set structure, must be different for each
		 * thread to effectively test the nanokernel's ability to
		 * properly save/restore the floating point values during a
		 * context switch.
		 */

		init_byte = FIBER_FLOAT_REG_CHECK_BYTE;

		for (i = 0; i < SIZEOF_FP_REGISTER_SET; i++) {
			reg_set_ptr[i] = init_byte++;
		}

		/*
		 * Utilize an architecture specific function to load all the
		 * floating point registers with the contents of the
		 * float_reg_set structure.
		 *
		 * The goal of the loading all floating point registers with
		 * values that differ from the values used in other threads is
		 * to help determine whether the floating point register
		 * save/restore mechanism in the nanokernel's context switcher
		 * is operating correctly.
		 *
		 * When a subsequent nano_fiber_timer_test() invocation is
		 * performed, a (cooperative) context switch back to the
		 * preempted task will occur. This context switch should result
		 * in restoring the state of the task's floating point
		 * registers when the task was swapped out due to the
		 * occurrence of the timer tick.
		 */

		_load_then_store_all_float_registers(&float_reg_set);

		/*
		 * Relinquish the processor for the remainder of the current
		 * system clock tick, so that lower priority threads get a
		 * chance to run.
		 *
		 * This exercises the ability of the nanokernel to restore the
		 * FPU state of a low priority thread _and_ the ability of the
		 * nanokernel to provide a "clean" FPU state to this thread
		 * once the sleep ends.
		 */

		nano_fiber_timer_start(&fiber_timer, 1);
		nano_fiber_timer_test(&fiber_timer, TICKS_UNLIMITED);

		/* periodically issue progress report */

		if ((++load_store_high_count % 100) == 0) {
			PRINT_DATA("Load and store OK after %u (high) + %u (low) tests\n",
					load_store_high_count,
					load_store_low_count);
		}

		/* terminate testing if specified limit has been reached */

		if (load_store_high_count == MAX_TESTS) {
			TC_END_RESULT(TC_PASS);
			TC_END_REPORT(TC_PASS);
			return;
		}
	}
}