Strange byte alignment between windows and Linux gcc compiler - linux

I see a strange behavior of reading a bin file and mapping in to the structure in the windows compared to linux gcc compiler.
Below is my c code:
#include <inttypes.h>
#include <stdio.h>
#include <fcntl.h>
#define IV_MAX_LEN 32
#define HASH_MAX_LEN 64
#define MAX_NUM_IMGS 8
#define MAX_NUM_SRK_RECORDS 4
#define SKIP_OFFSET_1K 0x400
typedef struct {
uint8_t version;
uint16_t length;
uint8_t tag;
uint16_t srk_table_offset;
uint16_t cert_offset;
uint16_t blob_offset;
uint16_t signature_offset;
uint32_t reserved;
} __attribute__((packed)) blk_hdr_t;
typedef struct {
uint32_t offset;
uint32_t size;
uint64_t dst;
uint64_t entry;
uint32_t hab_flags;
uint32_t meta;
uint8_t hash[HASH_MAX_LEN];
uint8_t iv[IV_MAX_LEN];
} __attribute__((packed)) boot_t;
typedef struct {
uint8_t version;
uint16_t length;
uint8_t tag;
uint32_t flags;
uint16_t sw_version;
uint8_t fuse_version;
uint8_t num_images;
uint16_t sig_blk_offset;
uint16_t reserved;
boot_t img[MAX_NUM_IMGS];
blk_hdr_t blk_hdr;
uint32_t sigblk_size;
uint32_t padding;
} __attribute__((packed)) headerMain_t;
int main()
{
int ofd =1;
char filename[] = "sample.bin";
int readSizes= 0;
int headerSizes= 0;
headerMain_t header;
uint8_t *byte;
ofd = open(filename, O_RDONLY);
printf("\nOPENING File: %s!\n", filename);
printf("\n STRUCT sizes: %d, %d, %d,\n", sizeof(headerMain_t), sizeof(boot_t), sizeof(blk_hdr_t));
#if 0
if(lseek(ofd, SKIP_OFFSET_1K, SEEK_SET) < 0) {
printf("Error Read \n");
}
#endif
readSizes = read(ofd, &header, sizeof(header)) ;
headerSizes = sizeof(header);
printf("\n Read SIZE: %d / %d !",readSizes,headerSizes);
printf("\n Read Bytes: \n");
byte = (uint8_t*)&header;
for(int i=0; i<readSizes; i++)
{
printf("0x%02x, ",*byte);
byte++;
if(0 ==i%20 )
printf("\n");
}
return 0;
}
and here is the input binary file it reads. (This sample bin file is 0x01 (20 times)... 0xFF(20 times) = so 255 x 20 = 5100 bytes)
The same code is compiled and run in windows mingW-gcc and linux gcc.
Following are the strange observations seen in the windows run:
blk_hdr_t struct although 4 byte aligned and 16 bytes it total: 22 bytes
(Update: I found that this solution worked with -mno-ms-bitfields option)
Although 5100 bytes are available, read() function could read only 1000 bytes. what choked it further to read ?
read() also puts the "holes" with "0x00" bytes value in it. I don't understand this strange behaviour.
Enabling the lseek to skip first 1024 bytes shall reach the end of the file.
Everything looks perfect on the linux result (although point 3) does exist on the linux side too: although this is trivial for me, i'm curious on the behaviour)
so finally, How can make this code on windows gcc with the Exact result as in the linux gcc ?
could anyone please enlighten me ?
(The parameters in the structures cannot be reshuffled)

Related

How to get the size of the VDSO on a Linux x86_64 system

I'd like to dump the VDSO to disk in a way that I can verify it is correct with objdump -D.
We can get the base address of the VDSO with getauxval(AT_SYSINFO_EHDR) as documented in vdso(7). But how does one get the size of the object?
I happen to know it is exactly two pages long, but I'm not certain I can rely on that.
I can't see anything in the ELF header that would indicate the size of the object as a whole, and I've also tried iterating and dumping the sections via dl_iterate_phdr(3) to no joy (presumably this would skip the ELF header?).
Any ideas? Do I really have to scrape the size out of the proc maps?
The VDSO header gives you the start of the file. To find the end, calculate the maximum of:
the end of all segments from the program header table (offset + size)
the end of the section header table
the end of the program header table
Then write everything between the start and end to disk. The following program should do the trick:
#include <stdlib.h>
#include <stdio.h>
#include <sys/auxv.h>
struct elf_fhdr_64
{
uint32_t magic;
uint8_t ei_class;
uint8_t ei_data;
uint8_t ei_version;
uint8_t ei_osabi;
uint8_t ei_abiversion;
uint8_t pad[7];
uint16_t e_type;
uint16_t e_machine;
uint32_t e_version;
uint64_t e_entry;
uint64_t e_phoff; // program header offset
uint64_t e_shoff;
uint32_t e_flags;
uint16_t e_ehsize;
uint16_t e_phentsize;
uint16_t e_phnum; // number of program headers
uint16_t e_shentsize;
uint16_t e_shnum;
// ...
};
struct elf_phdr_64
{
uint32_t p_type;
uint32_t p_flags;
uint64_t p_offset; // offset in file
uint64_t p_vaddr;
uint64_t p_paddr;
uint64_t p_filesz; // size in file
// ...
};
struct elf_shdr_64
{
uint32_t sh_name;
uint32_t sh_type;
uint64_t sh_flags;
uint64_t sh_addr; // virtual addr when loaded
uint64_t sh_offset; // offset in file
uint64_t sh_size; // size in file
// ...
};
int main(int argc, char **argv)
{
unsigned long vdso_hdr = getauxval(AT_SYSINFO_EHDR);
uint32_t elf_magic = * (uint32_t *)vdso_hdr;
if (elf_magic == 0x464C457F) {
printf("has elf magic, proceeding...\n");
}
else {
printf("no elf magic.\n");
exit(1);
}
struct elf_fhdr_64 * fhdrp = (struct elf_fhdr_64 *) vdso_hdr;
int num_phdrs = fhdrp->e_phnum;
uint16_t phentsize = fhdrp->e_phentsize;
long max_offs = 0;
for (int i = 0; i < num_phdrs; i++) {
struct elf_phdr_64 * phdr = (struct elf_phdr_64 *)(vdso_hdr
+ fhdrp->e_phoff + i * phentsize);
long file_offs = phdr->p_offset + phdr->p_filesz;
if (max_offs < file_offs) max_offs = file_offs;
}
int num_shdrs = fhdrp->e_shnum;
int shentsize = fhdrp->e_shentsize;
for (int i = 0; i < num_shdrs; i++) {
struct elf_shdr_64 * shdr = (struct elf_shdr_64 *)(vdso_hdr
+ fhdrp->e_shoff + i * shentsize);
long file_offs = shdr->sh_offset + shdr->sh_size;
if (max_offs < file_offs) max_offs = file_offs;
}
// section table:
int section_table_max = fhdrp->e_shoff + (num_shdrs * shentsize);
if (max_offs < section_table_max) max_offs = section_table_max;
// phdrs table:
int phdr_table_max = fhdrp->e_phoff + (num_phdrs * phentsize);
if (max_offs < phdr_table_max) max_offs = section_table_max;
FILE * outfile = fopen("test-vdso.so", "wb");
if (outfile) {
fwrite((void *) vdso_hdr, 1, max_offs, outfile);
fclose(outfile);
}
return 0;
}

How to test your own Linux module?

Today I am getting started with developing Linux modules. It was rather hard to write, compile and work with Helloworld, but I've done it.
My second module with open, write, read functions is ready, but I really dont know how to test it. Write method just makes printk(). My module is loaded, its name is iamnoob. How to test this write(...) function and to find smth in var/log/syslog?
cat > iamnoob just writes a file to the dir. Same with cp and other.
Sorry for noob question, i've googled, but no answer has been found. Sorry for poor English.
A basic kernel module would normally include registering a character device.
Simple imlementation requires:
Register chrdev region with specific major & minor.
Allocate file operations structure and implement the basic read / write APIs.
Initialize and register character device with the file operations structure to the major / minor region.
See the following code snippet as a template of a module (only read / write APIs are imlemented):
#include <linux/module.h>
#include <linux/fs.h>
#include <linux/cdev.h>
#include <linux/init.h>
#include <linux/slab.h>
#include <asm-generic/uaccess.h>
#define MY_BUFFER_SIZE (1024 * 10)
#define MY_CHRDEV_MAJOR 217
#define MY_CHRDEV_MINOR 0
static struct cdev my_cdev;
static unsigned char *my_buf;
static dev_t my_dev = MKDEV(MY_CHRDEV_MAJOR, MY_CHRDEV_MINOR);
ssize_t my_read(struct file *file, char __user * buf, size_t count, loff_t * ppos)
{
int size;
size = MY_BUFFER_SIZE - 100 - (int)*ppos;
if (size > count)
size = count;
if (copy_to_user(buf, my_buf + *ppos, count))
return -EFAULT;
*ppos += size;
return size;
}
ssize_t my_write(struct file *file, const char __user *buf, size_t count, loff_t *ppos)
{
int size;
size = MY_BUFFER_SIZE - 100 - (int)*ppos;
if (size > count)
size = count;
if (copy_from_user(my_buf + *ppos, buf, count))
return -EFAULT;
*ppos += size;
return size;
}
long my_unlocked_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
{
printk ("%s!\n", __FUNCTION__);
return 0;
}
int my_mmap(struct file *f, struct vm_area_struct *vma)
{
printk ("%s!\n", __FUNCTION__);
return 0;
}
int my_open(struct inode *i, struct file *f)
{
printk ("%s!\n", __FUNCTION__);
return 0;
}
int my_release(struct inode *i, struct file *f)
{
printk ("%s!\n", __FUNCTION__);
return 0;
}
struct file_operations my_fops =
{
.owner = THIS_MODULE,
.read = &my_read,
.write = &my_write,
.unlocked_ioctl = &my_unlocked_ioctl,
.mmap = &my_mmap,
.open = &my_open,
.release = &my_release,
};
static int __init my_module_init(void)
{
int line = 0;
unsigned char *pos;
printk ("%s!\n", __FUNCTION__);
my_buf = (unsigned char *)kzalloc(MY_BUFFER_SIZE, 0);
if (my_buf == NULL) {
printk("%s - failed to kzallocate buf!\n", __FUNCTION__);
return -1;
}
pos = my_buf;
while (pos - my_buf < MY_BUFFER_SIZE - 100) {
sprintf(pos, "Line #%d\n", line++);
pos += strlen(pos);
}
cdev_init(&my_cdev, &my_fops);
if (register_chrdev_region(my_dev, 1, "my_dev")) {
pr_err("Failed to allocate device number\n");
}
cdev_add(&my_cdev, my_dev, 1);
printk ("%s - registered chrdev\n", __FUNCTION__);
return 0;
}
static void __exit my_module_exit(void)
{
printk ("my_module_exit.\n");
unregister_chrdev_region(my_dev, 1);
return;
}
module_init(my_module_init);
module_exit(my_module_exit);
MODULE_LICENSE("GPL");
This module uses a buffer for file operations, therefore can be tested on any machine, regardless of its HW. Make sure you avoid unnecessary printk's as loops may harm your kernel stability.
Once this is done, in user-space shell you should create a /dev node to represent your character device:
sudo mknod /dev/[dev_name] c [major] [minor]
for example:
sudo mknod /dev/my_dev c 217 0
Then you can test your read / write APIs with:
sudo insmod my_modult.ko
cat /dev/my_dev
less -f /dev/my_dev
sudo su
root> echo "This is a test" > /dev/my_dev
root> exit
cat /dev/my_dev
The shell commands listed above perform read, then login as root (to allow writing to device), write to the char dev, then exit and read again to see the changes.
Now you'd normally implement ioctl and mmap if needed.

I2C EEPROM Read/Write Cubieboard 2 Arch Linux

I am trying to read and write to the AT24MAC402 EEPROM over i2c on the Cubieboard 2 using Arch Linux. I am using the i2c-dev library and i2c-tools.
Datasheet:
http://www.atmel.com/images/atmel-8807-seeprom-at24mac402-602-datasheet.pdf
I can successfully write (kind of...) to a chosen address and sequentially write many bites starting at that address. The issues are:
Cannot re-select another address to write once the first address has been selected.
Cannot point the the EEPROM to the location I wish to read from (by dummy-writing), and therefore have almost no real control over the EEPROM.
Upon looking at the datasheet (for hours on end), it looks as if I don't have as much control over the I2C communications as I may need using the i2c-dev library.. It would be great if I could just write X bits or X bytes directly to the EEPROM.
In short, I would like advice on how I can read and write properly to this EEPROM.
char buf[10];
int com_serial;
int failcount;
int i2c_init(char filename[40], int addr)
{
int file;
if ((file = open(filename,O_RDWR)) < 0)
{
printf("Failed to open the bus.");
/* ERROR HANDLING; you can check errno to see what went wrong */
com_serial=0;
exit(1);
}
if (ioctl(file,I2C_SLAVE,addr) < 0)
{
printf("Failed to acquire bus access and/or talk to slave.\n");
/* ERROR HANDLING; you can check errno to see what went wrong */
com_serial=0;
exit(1);
}
return file;
}
int main (int argc, char *argv[]) {
char read_buf[16];
char write_buf[17];
int i;
int file;
file=i2c_init("/dev/i2c-1",0x50); //dev,slavei2caddr
write_buf[0] = 0x00;
write_buf[1] = 'H';
write_buf[2] = 'i';
write_buf[3] = '!';
write(file, write_buf, 4);
//Successfully prints "Hi!" to bytes 0x00 -> 0x02
//Setting EEPROM to point to address 0xA0 to start reading (arbitrary address with known values: all 0xFF)
write_buf[0] = 0xA0;
write(file, write_buf, 1);
//Reading 1 byte from EEPROM, even though there is a '2'; 2 bytes would be '3'
read(file, read_buf, 2);
for (i=1; i<3; i++){
printf("%X", read_buf[i]);
}
//Prints out from address 0x04 to 0x05 instead of 0xA0 to 0xA1
printf("\n");
}
I did work properly using the functions from the linux/i2c-dev.h.
To test the code I get the output generated by i2cdump and put as input to i2c-stub-from-dump tool, it lets you setup one or more fake I2C chips on the i2c-stub bus based on dumps of the chips you want to emulate.
#include <stdio.h>
#include <fcntl.h>
#include <unistd.h>
#include <linux/i2c-dev.h>
int i2c_init(const char * i2c_device, const int chip_address)
{
int file;
if ((file = open(i2c_device, O_RDWR)) < 0) {
return -1;
}
if (ioctl(file, I2C_SLAVE, chip_address) < 0) {
close(file);
return -1;
}
return file;
}
int i2c_write(int file, const int data_address, const unsigned char * data, size_t size)
{
return i2c_smbus_write_i2c_block_data(file, data_address, size, data);
}
void i2c_read(int file, const int data_address, unsigned char * data_vector, size_t size)
{
unsigned char reg = data_address;
unsigned int i;
for(i = 0; i < size; ++i, ++reg) {
data_vector[i] = i2c_smbus_read_byte_data(file, reg);
}
}
int main(void) {
char device[] = "/dev/i2c-6";
int address = 0x50;
unsigned char buffer_before[30] = {0};
unsigned char buffer_after[30] = {0};
unsigned char data[] = "Hello World!";
int file;
file = i2c_init(device, address);
if (file > 0) {
i2c_read(file, 0x00, buffer_before, sizeof(data));
i2c_write(file, 0x00, data, sizeof(data));
i2c_read(file, 0x00, buffer_after, sizeof(data));
close (file);
}
printf("data read before write: %s\n", buffer_before);
printf("data read after write: %s\n", buffer_after);
return 0;
}

kernel driver reading ok from user space, but writing back is always 0

So I'm working my way through kernel driver programming, and currently I'm trying to build a simple data transfer between application and kernel driver.
I am using simple character device as a link between these two, and I have succeeded to transfer data to driver, but I can't get meaningful data back to user space.
Kernel driver looks like this:
#include <linux/init.h>
#include <linux/module.h>
#include <linux/kernel.h> /* printk() */
#include <linux/errno.h> /* error codes */
#include <linux/types.h> /* size_t */
#include <linux/proc_fs.h>
#include <asm/uaccess.h> /* copy_from/to_user */
MODULE_LICENSE("GPL");
//Declarations
int memory_open(struct inode *inode, struct file *filp);
int memory_release(struct inode *inode, struct file *filp);
ssize_t memory_read(struct file *filp, char *buf, size_t count, loff_t *f_pos);
ssize_t memory_write(struct file *filp, char *buf, size_t count, loff_t *f_pos);
void memory_exit(void);
int memory_init(void);
/* Structure that declares the usual file access functions */
struct file_operations memory_fops = {
read: memory_read,
write: memory_write,
open: memory_open,
release: memory_release
};
//Default functions
module_init(memory_init);
module_exit(memory_exit);
/* Global variables of the driver */
/* Major number */
int memory_major = 60;
/* Buffer to store data */
char* tx_buffer;
char* rx_buffer;
int BUFFER_SIZE=64;
int actual_rx_size=0;
int memory_init(void) {
int result;
/* Registering device */
result = register_chrdev(memory_major, "move_data", &memory_fops);
if (result < 0) {
printk(
"<1>move_data: cannot obtain major number %d\n", memory_major);
return result;
}
/* Allocating memory for the buffers */
//Allocate buffers
tx_buffer = kmalloc(BUFFER_SIZE, GFP_KERNEL);
rx_buffer = kmalloc(BUFFER_SIZE, GFP_KERNEL);
//Check allocation was ok
if (!tx_buffer || !rx_buffer) {
result = -ENOMEM;
goto fail;
}
//Reset the buffers
memset(tx_buffer,0, BUFFER_SIZE);
memset(rx_buffer,0, BUFFER_SIZE);
printk("<1>Inserting memory module\n");
return 0;
fail:
memory_exit();
return result;
}
void memory_exit(void) {
/* Freeing the major number */
unregister_chrdev(memory_major, "memory");
/* Freeing buffers */
if (tx_buffer) {
kfree(tx_buffer); //Note kfree
}
if (rx_buffer) {
kfree(rx_buffer); //Note kfree
}
printk("<1>Removing memory module\n");
}
//Read function
ssize_t memory_read(struct file *filp, char *buf, size_t count, loff_t *f_pos) {
printk("user requesting data, our buffer has (%d) \n", actual_rx_size);
/* Transfering data to user space */
int retval = copy_to_user(buf,rx_buffer,actual_rx_size);
printk("copy_to_user returned (%d)", retval);
return retval;
}
ssize_t memory_write( struct file *filp, char *buf,
size_t count, loff_t *f_pos) {
//zero the input buffer
memset(tx_buffer,0,BUFFER_SIZE);
memset(rx_buffer,0,BUFFER_SIZE);
printk("New message from userspace - count:%d\n",count);
int retval = copy_from_user(tx_buffer,buf,count);
printk("copy_from_user returned (%d) we read [%s]\n",retval , tx_buffer);
printk("initialize rx buffer..\n");
memcpy(rx_buffer,tx_buffer, count);
printk("content of rx buffer [%s]\n", rx_buffer);
actual_rx_size = count;
return count; //inform that we read all (fixme?)
}
//Always successfull
int memory_open(struct inode *inode, struct file *filp) { return 0; }
int memory_release(struct inode *inode, struct file *filp) { return 0; }
And the userspace application is simple as well:
#include <unistd.h> //open, close | always first, defines compliance
#include <fcntl.h> //O_RDONLY
#include <stdio.h>
#include <stdlib.h> //printf
#include <string.h>
int main(int args, char *argv[])
{
int BUFFER_SIZE = 20;
char internal_buf[BUFFER_SIZE];
int to_read = 0;
memset(internal_buf,0,BUFFER_SIZE);
if (args < 3) {
printf("2 Input arguments needed\nTo read 10 bytes: \"%s read 10\" \
\nTo write string \"hello\": \"%s write hello\"\nExiting..\n", argv[0], argv[0]);
return 1;
}
//Check the operation
if (strcmp(argv[1],"write") == 0) {
printf("input lenght:%d", strlen(argv[2]));
//Make sure our write fits to the internal buffer
if(strlen(argv[2]) >= BUFFER_SIZE) {
printf("too long input string, max buffer[%d]\nExiting..", BUFFER_SIZE);
return 2;
}
printf("write op\n");
memcpy(internal_buf,argv[2], strlen(argv[2]));
printf("Writing [%s]\n", internal_buf);
FILE * filepointer;
filepointer = fopen("/dev/move_data", "w");
fwrite(internal_buf, sizeof(char) , strlen(argv[2]), filepointer);
fclose(filepointer);
} else if (strcmp(argv[1],"read") == 0) {
printf("read op\n");
to_read = atoi(argv[2]);
FILE * filepointer;
filepointer = fopen("/dev/move_data", "r");
int retval = fread(internal_buf, sizeof(char) , to_read, filepointer);
fclose(filepointer);
printf("Read %d bytes from driver string[%s]\n", retval, internal_buf);
} else {
printf("first argument has to be 'read' or 'write'\nExiting..\n");
return 1;
}
return 0;
}
When I execute my application, this is what happens:
./rw write "testing testing"
kernel side:
[ 2696.607586] New message from userspace - count:15
[ 2696.607591] copy_from_user returned (0) we read [testing testing]
[ 2696.607593] initialize rx buffer..
[ 2696.607594] content of rx buffer [testing testing]
So all look correct. But when I try to read:
./rw read 15
read op
Read 0 bytes from driver string[]
Kernel
[ 617.096521] user requesting data, our buffer has (15)
[ 575.797668] copy_to_user returned (0)
[ 617.096528] copy_to_user returned (0)
I guess it's quite simple what I'm doing wrong, since if I don't return 0, I can get some data back, but for example if I read with cat, it will continue looping endlessly.
I would like to understand what mistakes I have made in my thinking.
Is there a way that kernel driver would just spit out it's buffer, and then return 0, so that I wouldn't have to build some protocol there in between to take care of how much data has been read etc.
Thanks for your suggestions!
Edit: corrected the printk statement in memory_write function, and added memory_read function trace
Your read function always returns 0 because you are returning retval, and not the count of bytes read. As long as the copy_to_user() call always succeeds, retval will always be 0. Instead, as long as copy_to_user() succeeds, you should return the number of bytes actually written to user space. This documentation states that copy_to_user() returns the total number of bytes that it was unable to copy.
As an aside, you are ignoring the value of count. It is very possible that the user is requesting less data than you have available in your buffer. You should never ignore count.
Now you have the problem where your function never returns a 0. Returning a 0 is important because is tells the user application that there is no more data available for reading and the user application should close the device file.
You need to keep track in your driver how many bytes have been read vs. how many bytes have been written. This may be implemented using your actual_rx_size.
Try this:
//Read function
ssize_t memory_read(struct file *filp, char *buf, size_t count, loff_t *f_pos) {
ssize_t bytes;
if (actual_rx_size < count)
bytes = actual_rx_size;
else
bytes = count;
printk("user requesting data, our buffer has (%d) \n", actual_rx_size);
/* Check to see if there is data to transfer */
if (bytes == 0)
return 0;
/* Transfering data to user space */
int retval = copy_to_user(buf,rx_buffer,bytes);
if (retval) {
printk("copy_to_user() could not copy %d bytes.\n", retval);
return -EFAULT;
} else {
printk("copy_to_user() succeeded!\n");
actual_rx_size -= bytes;
return bytes;
}
}

What's the max file mapping size in 64bits machine

I'm new to 64-bits architecture. Could you tell me what's MAX file size supported by file mapping in 64 bits linux machine. I want to open more than 20GB files by file mapping, is it available?
I write a sample code. But it causes Bus Error when I get the value of the pointer in GBSIZE offset:
unsigned char* pCur = pBegin + GBSIZE;
//pBegin is the pointer returned by mmap
printf("%c",*pCur);
BTW, printf("%c",*pBegin ); works fine. and my address sizes : 38 bits physical, 48 bits virtual
Here is the full code:
#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/mman.h>
//#define FILEPATH "smallfile"
#define FILEPATH "bigfile"
#define GBSIZE (1024L*1024L*1024L)
#define TBSIZE (1024L*GBSIZE)
#define NUMSIZE (20L * GBSIZE)
//#define NUMSIZE (10)
#define FILESIZE (NUMINTS * sizeof(int))
int main(int argc, char *argv[])
{
int i;
int fd;
unsigned char *pBegin;
fd = open(FILEPATH, O_RDONLY);
if (fd == -1) {
perror("Error opening file for reading");
exit(EXIT_FAILURE);
}
pBegin = mmap(0, NUMSIZE, PROT_READ, MAP_SHARED, fd, 0);
if (pBegin == MAP_FAILED) {
close(fd);
perror("Error mmapping the file");
exit(EXIT_FAILURE);
}
/** ERROR happens here!!! **/
unsigned char* pCur = pBegin + GBSIZE;
printf("%c",*pCur);
if (munmap(pBegin, NUMSIZE) == -1) {
perror("Error un-mmapping the file");
}
close(fd);
return 0;
}
Although pointers are 64-bit wide, most processors do not actually support virtual addresses using the full 64 bits. To see what size virtual addresses your processor supports, look in /proc/cpuinfo (48 bits is typical).
grep "address sizes" /proc/cpuinfo
Additionally, half of the virtual address space is used by the kernel and not available to userspace - leaving 47 bits in the current Linux implementation.
However, even taking this into account, you will still have plenty of room for a 20GB file. 47 bits in theory means a virtual address space of 128TB.
From the mmap(2) man page:
void *mmap(void *addr, size_t length, int prot, int flags,
int fd, off_t offset);
length is a size_t, which on 64-bit machines is 64 bits in length. Therefore yes, you can theoretically map a 20GB file.
64-bit addresses allow for many orders of magnitude more than 20 GB.
(This answer was originally edited into the question by OP)
You have requested a 20GB map onto a file which was only 50MB in size.
As described by the mmap man page, mmap succeeds when you request the length too big, however it will give SIGBUS or SIGSEGV when you actually try to read beyond the end of the underlying file.
Agree with MarkR, you are dereference an invalid address.
// A bug in these lines.
unsigned char* pCur = pBegin + GBSIZE;
printf("%c",*pCur);
unsigned char* pEnd = pBegin + NUMSIZE;
unsigned char* pLast = pEnd - 1;
unsigned char* pCur = pLast;
I modified your code to use HUGE TLB flags as the following.
#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <unistd.h>
#include <fcntl.h>
#include <sys/mman.h>
#define MAP_HUGETLB 0x40000 /* create a huge page mapping */
#define MAP_HUGE_SHIFT 26
#define MAP_HUGE_1GB (30 << MAP_HUGE_SHIFT)
#define KSIZE 1024L
#define MSIZE (1024L*1024L)
#define GSIZE (1024L*1024L*1024L)
#define TSIZE (1024L*GSIZE)
#define INIT_MEM 0
// Fail on my MacBook Pro (Retina, 13-inch, Early 2015)
// Darwin Kernel Version 16.5.0:x86_64
// #define NUMSIZE (16L * TSIZE)
// mmap ok; init: got killed; signal 9
// #define NUMSIZE (8L * TSIZE)
// Got killed signal 9
// #define NUMSIZE (1L * TSIZE)
// OK
// #define NUMSIZE (200L * GSIZE)
// OK
#define NUMSIZE (20L * GSIZE)
typedef unsigned long long ETYPE;
#define MEMSIZE (NUMSIZE*sizeof(ETYPE))
#define PGSIZE (16*KSIZE)
void init(ETYPE* ptr) {
*ptr = (ETYPE)ptr;
}
int verify(ETYPE* ptr) {
if (*ptr != (ETYPE)ptr) {
fprintf(stderr, "ERROR: 0x%016llx != %p.\n", *ptr, ptr);
return -1;
}
else {
fprintf(stdout, "OK: 0x%016llx = %p.\n", *ptr, ptr);
}
return 0;
}
int main(int argc, char *argv[])
{
int i;
int fd;
ETYPE *pBegin;
int flags = MAP_SHARED | MAP_ANONYMOUS | MAP_HUGETLB | MAP_HUGE_1GB;
printf("mmap memory size:%lu GB\n", MEMSIZE/GSIZE);
pBegin = (ETYPE*) mmap(0, MEMSIZE, PROT_READ | PROT_WRITE, flags, -1, 0);
if (pBegin == MAP_FAILED) {
perror("Error mmapping the file");
exit(EXIT_FAILURE);
}
ETYPE* pEnd = pBegin + NUMSIZE;
ETYPE* pCur = pBegin;
#if INIT_MEM
while (pCur < pEnd) {
init(pCur);
// ++pCur; //slow if init all addresses.
pCur += (PGSIZE/sizeof(ETYPE));
}
#endif
init(&pBegin[0]);
init(&pBegin[NUMSIZE-1]);
verify(&pBegin[0]);
verify(&pBegin[NUMSIZE-1]);
if (munmap(pBegin, MEMSIZE) == -1) {
perror("Error un-mmapping the file");
}
return 0;
}

Resources