lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-ID: <20120328210631.GB2297@fifo99.com>
Date:	Wed, 28 Mar 2012 14:06:32 -0700
From:	Daniel Walker <dwalker@...o99.com>
To:	linux-kernel@...r.kernel.org
Cc:	Tim Bird <tim.bird@...sony.com>
Subject: [RFC] Android Logger vs. Shared Memory FIGHT!



	I went to Linaro Connect a few weeks ago. While I was there they held a
meeting with some of the Android developers and some community developers.
During the meeting Brian Swetland and Tim Bird were talking about upstreaming
Android's logger kernel changes. Brian mentioned that logger was in the kernel
for “performance reasons”. Then later Zach Pfeffer asked if Google had a test
suite for binder. They said they didn't have anything like that, which I thought
was really surprising. From that I assume there is no test suite for logger and
I doubt they have any performance analysis either.

	Since logger has a relatively simple interface I felt it would be a good
target for performance analysis (compared to something like binder). After
analyzing logger for a little while I decided it would be relatively easy to
create something similar with shared memory.

	The test case that I created uses a POSIX shared memory area as the
storage medium. This area is different than normal shared memory in that it's
visible as a file under /dev/shm/ . The shared memory area is protected with a
System V semaphore. The semaphore is really overkill since contention isn't very
likely in a real Android logging case. The writers are expected to manage the
shared area when doing writes. This allows the logging to be daemon-less.

	I tried to make the test as much apples to apples as possible. It is somewhat 
difficult since logger is doing a fair amount of stuff in kernel space which may not
be needed in userspace, or is difficult to replicate. The read side is typically used 
infrequently and I don't think performance really matters in that case. I only really 
address the write side, however my test case includes readers for both (not fully
tested).

The test case was run on both x86 and ARM. Results below all in Bytes per second,

x86:
+-----------------+----------------------+-------------------+ 
|Clocksource      |         TSC          |       ACPI_PM     |
+-----------------+----------------------+-------------------+
|Shared Memory    |     457172567.4B/s   |     28708407.6B/s |
+-----------------+----------------------+-------------------+
|Android Logger   |      77531671.5B/s   |     79575677.3B/s |
+-----------------+----------------------+-------------------+ 

ARM:
+-----------------+----------------------+
|Shared Memory    |    15336338.6B/s     |
+-----------------+----------------------+
|Android Logger   |     6615796.3B/s     |
+-----------------+----------------------+

	The x86 machine was a single core AMD 1.8Ghz machine, and the ARM was a G1
phone with an MSM7201a SoC using a v3.3 kernel in all cases. I also used chrt to
elevate the priority of the test case above all other processes in the system. The
test case was run three time and averaged. The source for the test case is below.

Noticeably absent is an hpet test. I didn't have one available.

Conclusions,

	The speed of logger isn't really dependent on the clocksource. There
is very little change between the two clocksource runs on x86 for logger. It's because
the bottle neck for logger isn't getting the time. The shared memory version does
depend heavily on the speed of the clocksource. Shared memory with an ACPI_PM
clocksource is about %64 slower than logger. In the TSC case the shared memory version
is about %490 faster. The TSC is fast, but using VDSO it's even better.

	On ARM it's more straight forward with the Shared Memory version being about
%132 faster than logger.

	As you can see there are some trade offs on x86. I should note that there are
Android phones being made with x86 platforms. So this isn't as ridiculous as it might
seem. I would guess most of those phones don't have ACPI_PM clocks however.

	I made no attempt to address any security issues in logger. Since logger has
no security built into it currently. There are at least two ways to provide security
to the shared memory interface that I know of, but I'm not planning to discuss them in
this analysis.

I welcome comments on the test case, or my analysis.
 
Here's the test case used,


/*
 * Copyright (c) 2012, Daniel Walker <dwalker@...o99.com>
 *
 * gcc -O2 logger_test.c -lrt -lpthread -o logger_test
 *
 */
#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <string.h>
#include <signal.h>

#include <fcntl.h>
#include <time.h>
#include <pthread.h>
#include <errno.h>

#include <stdint.h>
#include <semaphore.h>
#include <sys/mman.h>
#include <sys/types.h>
#include <sys/time.h>
#include <sys/stat.h>
#include <sys/uio.h>

char buffer[1024] = { [0 ... 1023] = '0' };

struct timeval tv1, tv2;

struct memory_control_area {
	sem_t mutex;
	unsigned long start, end, size, readers;
	_Bool overwrite;
	char data[0];
};

_Bool stopped = 0;

#define READ_WAIT_FIFO_NAME	"/tmp/shared_memory_reader_fifo"
int read_wait_fifo;

void alarm_handler(int signal)
{
	stopped = 1;
}

struct sigaction alarm_overload = {
	.sa_handler = alarm_handler,
};

void int_handler(int signal)
{
	stopped = 1;
}

struct sigaction int_overload = {
	.sa_handler = int_handler,
};

/*
 *  This is for compatibility with Android logger
 */
#define CONTENT_LINES (64)

/*
 *  Size of the memory area. This is minus the control structure.
 */
#define CONTENT_SIZE (CONTENT_LINES*1024 - sizeof(struct memory_control_area))

/*
 * Subtracts tv1 from tv2 and prints the output.
 */
void print_diff_time(char header[], struct timeval *tv1, struct timeval *tv2, unsigned long long bytes)
{
	unsigned long long diff_useconds, diff_seconds;
	double total_time;

	diff_useconds = (tv2->tv_usec + (tv2->tv_sec*1000000)) -
			(tv1->tv_usec + (tv1->tv_sec*1000000)),

	total_time = diff_useconds/1000000.0;

	printf("%s wrote %llu bytes in %.1f seconds at a rate of %.1f Bytes/second\n",
		header, bytes, total_time, (double)(bytes/total_time));
}

char *sample_log_string(void)
{
	/*
	 *  Just some sample logs for effect.
	 */
	static char *sample_msgs[16] = {
		"New MMC card '00000' (serial 2029453699) added @ /devices/platform/msm_sdcc.2/mmc_host/mmc1/mmc1:0001\n",
		"Disk (blkdev 179:0), 7954432 secs (3884 MB) 1 partitions\n",
		"New blkdev 179.0 on media 00000, media path /devices/platform/msm_sdcc.2/mmc_host/mmc1/mmc1:0001, Dpp 1\n",
		"Partition (blkdev 179:1), 7946240 secs (3880 MB) type 0xb\n",
		"New blkdev 179.1 on media 00000, media path /devices/platform/msm_sdcc.2/mmc_host/mmc1/mmc1:0001, Dpp 0\n",
		"Evaluating dev '/devices/platform/msm_sdcc.2/mmc_host/mmc1/mmc1:0001/block/mmcblk0' for mountable filesystems for '/sdcard'\n",
		"handle_switch_event(): Ignoring switch 'No Device'\n",
		"Bootstrapping complete\n",
		"dosfsck 3.0.1 (23 Nov 2008)\n",
		"akmd 1.6.4 START \n",
		"library version: 1.2.1.1129\n",
		"dosfsck 3.0.1, 23 Nov 2008, FAT32, LFN\n",
		"Logical sector size is zero.\n",
		"/system/bin/dosfsck terminated by exit(8)\n",
		"Filesystem check failed (not a FAT filesystem)\n",
		"vfat filesystem check failed on 179:1 (No data available)\n",
	};

	static index = 0;
	
	if (index >= 16) {
		index = 0;
	}

	return sample_msgs[index++];
}

void register_signal_handler(int sig, struct sigaction *overload)
{
	int ret = sigaction(sig, overload, NULL);
	if (ret < 0) {
		perror("sigaction():");
		exit(1);
	}

}

enum fifo_direction {
	FIFO_READ,
	FIFO_WRITE,
};

int open_fifo(enum fifo_direction dir)
{
	int ret = open(READ_WAIT_FIFO_NAME, (dir == FIFO_READ) ? O_RDONLY : O_WRONLY);
	if (ret < 0) {
		perror("open():");
		exit(1);
	}
	
}

/*
 *  Test this kernels logger interface in /dev/log/
 */
void test_logger(int seconds)
{
	int fd, i, ret, prio = 4;
	unsigned long long bytes_written = 0;
	const char *tag = "";
	unsigned long long j;
    	struct iovec vec[3];

	gettimeofday(&tv1, NULL);

	fd = open("/dev/log_main", O_WRONLY);
	if (fd < 0) {
		perror("open():");
		exit(1);
	}
	
	stopped = 0;

	alarm(seconds);

	vec[0].iov_base	= (unsigned char *) &prio;
	vec[0].iov_len	= 1;
	vec[1].iov_base	= (void *) tag;
	vec[1].iov_len	= 1;

	while (!stopped) {
		char *buffer = sample_log_string();

		vec[2].iov_base	= (void *) buffer;
		vec[2].iov_len	= strlen(buffer); /* Don't include the null */

		ret = writev(fd, vec, 3);
		if (ret < 0) {
			perror("writev():");
			exit(1);
		}
		bytes_written += ret;
	}

	close(fd);

	gettimeofday(&tv2, NULL);

	print_diff_time("Android logger", &tv1, &tv2, bytes_written);	
}

void logger_reader()
{
	char buffer[1024];
	struct logger_entry {
	    uint16_t    len;    /*  length of the payload */
	    uint16_t    __pad;  /*  no matter what, we get 2 bytes of padding */
	    int32_t     pid;    /*  generating process's pid */
	    int32_t     tid;    /*  generating process's tid */
	    int32_t     sec;    /*  seconds since Epoch */
	    int32_t     nsec;   /*  nanoseconds */
	    char        msg[0]; /*  the entry's payload */
	} *casted_buffer = (struct logger_entry *)buffer;

	int i, size, fd, prio = 0;
	struct iovec vec[3]; 
	const char *tag = "";

	vec[0].iov_base = (void *) buffer,
	vec[0].iov_len = 1024,

	fd = open("/dev/log_main", O_RDONLY);
	if (fd < 0) {
		perror("open():");
		exit(1);
	}

	while (size = readv(fd, vec, 1)) { 
		for(i = 1; i < casted_buffer->len; i++) {
			printf("%c", casted_buffer->msg[i]);
		}
	}
	close(fd);	
}

/*
 *  Test write into shared memory
 */
void test_shared_memory(int seconds)
{
	struct memory_control_area *shared_area;
	unsigned long long bytes_written = 0;
	int fd, i, ret;

	stopped = 0;

	if (seconds == 0) {
		register_signal_handler(SIGINT, &int_overload);
	} else {
		alarm(seconds);
	}

	gettimeofday(&tv1, NULL);

	fd = shm_open("/shared_memory_test", O_CREAT | O_RDWR, S_IRUSR | S_IRGRP | S_IWUSR | S_IWGRP);
	if (fd < 0) {
		perror("shm_open():");
		exit(1);
	}

	ret = ftruncate(fd, CONTENT_SIZE);
	if (ret < 0) {
		perror("ftruncate():");
		exit(1);
	}

	shared_area = (struct memory_control_area *)mmap(NULL, CONTENT_SIZE, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_LOCKED, fd, 0);
	if (shared_area == MAP_FAILED) {
		perror("mmap():");
		exit(1);
	}

	while (!stopped) {
		char *buffer = sample_log_string();
		int buffer_size = strlen(buffer); /* don't write the null */
		struct timeval sample;

		sem_wait(&shared_area->mutex);
		
		/*
		 *  This isn't used but it's here for compatibility with
		 *  logger. Since logger samples the time on each write.
		 *  This can also be a delay sink depending on which clocksource
		 *  your using.
		 */
		gettimeofday(&sample, NULL);

		/*
		 *  Two cases. The first one handles when the buffer loops
		 *  around to the start, the second is just a regular write.
		 */
		if(shared_area->end + buffer_size >= CONTENT_SIZE)
		{
			unsigned long rem = CONTENT_SIZE - shared_area->end;
	
			memcpy(shared_area->data + shared_area->end, buffer, rem);

			/* Loop around to the start of the buffer */
			shared_area->end = 0;

			memcpy(shared_area->data + shared_area->end, buffer + rem, buffer_size - rem);
			
			shared_area->end = buffer_size - rem;

			shared_area->overwrite = 1;

		} else {
			memcpy(shared_area->data + shared_area->end, buffer, buffer_size);
			shared_area->end += buffer_size;
		}

		if (shared_area->overwrite)
			shared_area->start = shared_area->end + 1;

		sem_post(&shared_area->mutex);

		bytes_written += buffer_size; 
	}

	close(fd);

	gettimeofday(&tv2, NULL);

	print_diff_time("Shared memory", &tv1, &tv2, bytes_written);	

}

void init_shared_memory_file(void)
{
	struct memory_control_area *shared_area;
	int fd, ret, i;

	fd = shm_open("/shared_memory_test", O_CREAT | O_RDWR, S_IRUSR | S_IRGRP | S_IWUSR | S_IWGRP);
	if (fd < 0) {
		perror("shm_open():");
		exit(1);
	}

	ret = ftruncate(fd, CONTENT_SIZE);
	if (ret < 0) {
		perror("ftruncate():");
		exit(1);
	}

	shared_area = (struct memory_control_area *)mmap(NULL, CONTENT_SIZE, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_LOCKED, fd, 0);
	if (shared_area == MAP_FAILED) {
		perror("mmap():");
		exit(1);
	}

	/* Warm up the shared memory area */
	for (i = 0; i < CONTENT_SIZE - sizeof(struct memory_control_area); i++)
		shared_area->data[i] = 0;

	sem_init(&shared_area->mutex, 1, 1);

	shared_area->readers = 0;
	shared_area->start = 0;
	shared_area->end = 0;

	close(fd);
}

void shared_memory_reader()
{
	int fd, i, ret, bytes_to_read, bytes_read = 0;
	struct memory_control_area *shared_area;
	char output_buf[1024];

	fd = shm_open("/shared_memory_test", O_CREAT | O_RDWR, S_IRUSR | S_IRGRP | S_IWUSR | S_IWGRP);
	if (fd < 0) {
		perror("shm_open():");
		exit(1);
	}

	ret = ftruncate(fd, CONTENT_SIZE);
	if (ret < 0) {
		perror("ftruncate():");
		exit(1);
	}

	shared_area = (struct memory_control_area *)mmap(NULL, CONTENT_SIZE, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_LOCKED, fd, 0);
	if (shared_area == MAP_FAILED) {
		perror("mmap():");
		exit(1);
	}
	
	bytes_to_read = CONTENT_SIZE - shared_area->end;

	while (bytes_to_read) {
		if (bytes_to_read >= 1024) {
			memcpy(output_buf, shared_area->data + shared_area->end + bytes_read, 1024);
			bytes_read += 1024;
			bytes_to_read -= 1024;
		} else {
			memcpy(output_buf, shared_area->data + shared_area->end + bytes_read, bytes_to_read);
			bytes_read += bytes_to_read;
			bytes_to_read = 0;
		}
		
		for(i = 0; i < 1024; i++)
			printf("%c", output_buf[i]);	
	}
	
	bytes_to_read = shared_area->end;

	while (bytes_to_read) {
		if (bytes_to_read >= 1024) {
			memcpy(output_buf, shared_area->data + bytes_read, 1024);
			bytes_read += 1024;
			bytes_to_read -= 1024;
		} else {
			memcpy(output_buf, shared_area->data + bytes_read, bytes_to_read);
			bytes_read += bytes_to_read;
			bytes_to_read = 0;
		}

		for(i = 0; i < 1024; i++)
			printf("%c", output_buf[i]);
	}
}

int main(int argc, char *argv[])
{
	char help[] =	"Android Logger Performance Tester\n"
			"\n"
			"Options:\n"
			"-l:	Test android logger in /dev/log/ (default)\n"
			"-s:	Test shared memory in /dev/shm\n"
			"	Must be mount with a command similar to \"mount -t tmpfs none /dev/shm\"\n"
			"-n X:	Number of seconds to run the test (default 60, 0 == unlimited)\n"
			"-i:	Initialize shared memory file\n"
			"-r:	Run shared memory reader\n"
			"-t:	Run logger reader\n"
			"\n";			

	int c, ret;
	int seconds = 60;

	ret = mlockall(MCL_CURRENT | MCL_FUTURE);
	if (ret < 0) {
		perror("mlockall():");
		exit(1);
	}

	register_signal_handler(SIGALRM, &alarm_overload);

	if (access(READ_WAIT_FIFO_NAME, F_OK) < 0) {
		ret = mkfifo(READ_WAIT_FIFO_NAME, S_IWUSR | S_IWGRP | S_IRUSR | S_IRGRP);
		if (ret < 0) {
			perror("mkfifo():");
			exit(1);
		}
	}

	while ((c = getopt(argc, argv, "hiutlsrn:")) != -1)  {

		switch (c) {
			case 'h':
				printf("%s\n", help);
				break;	
			case 'u':
				if (shm_unlink("/shared_memory_test") != 0) {
					perror("shm_unlink():");
					exit(1);
				}
			case 'i':
				init_shared_memory_file();
				break;
			case 'l':
				test_logger(seconds);
				break;
			case 's':
				test_shared_memory(seconds);
				break;
			case 'r':
				shared_memory_reader();
				break;
			case 't':
				logger_reader();
				break;
			case 'n':
				errno = 0;
				seconds = strtoll(optarg, NULL, 10);
				if (errno != 0) {
					perror("strtol():");
					exit(1);
				}
		}
	}

	return 0;
}

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists

Powered by Openwall GNU/*/Linux Powered by OpenVZ