Message-ID: <Pine.LNX.4.64.0703140712080.18655@williams.orchestra.cse.unsw.EDU.AU>
Date:	Wed, 14 Mar 2007 07:21:41 +1100 (EST)
From:	Kandan Venkataraman <kven709@....unsw.EDU.AU>
To:	linux-kernel@...r.kernel.org
cc:	akpm@...ux-foundation.org
Subject: [PATCH] Loop device - Tracking page writes made to a loop device
 through mmap

All comments have been taken care of.

Description:

A  file_operations structure variable called loop_fops is initialised with
the default block device file operations (def_blk_fops).
The mmap operation is overriden with a new function called loop_file_mmap.

A vm_operations structure variable called loop_file_vm_ops is initialised
with the default operations for a disk file.
The page_mkwrite operation in this variable is initialised to a new
function called loop_track_pgwrites.

In the function lo_open, the file operations pointer of the device file is
initialised with the address of loop_fops.

The function loop_file_mmap simply calls generic_file_mmap and then
initialises the vm_ops of the vma with address of loop_file_vm_ops.

The function loop_track_pgwrites stores the page offset of the page that
is being written to, in a red-black tree within the loop device.

A flag lo_track_pgwrite has been added to the struct loop_device to turn
on/off tracking of page writes.

Three new ioctls have been added.

The ioctl cmd LOOP_SET_TRACK_PGWRITE sets/unsets the flag lo_track_pgwrite
in the struct loop_device.
The ioctl cmd LOOP_GET_PGWRITES retrieves the page offsets of pages that
have been written to.
The ioctl cmd LOOP_CLR_PGWRITES empties the red-black tree.

The existing interface has been left untouched. Thus applications such as 
losetup would work as is.

The functionality added would allow us to have a read only version and a 
write version of memory by doing the following:
Associate a normal file as backing storage for  the loop device and mmap
to the loop device. Call this mmapped address space as area1.
Mmap to a normal file of identical size. Call this mmapped address space
as area2.

Changes made to area1 can be periodically copied to area2 using the ioctl
cmds (retreive dirty page offsets and copy the dirty pages from area1 to
area2). This facility would provide a quick way of updating the read only
version.

Motivation for new ioctls:

Imagine a business server application which processes messages from
clients as they come in (say over a TCP connection).
Some of those messages may be transactions, i.e. they cause data changes
in the application.
Rest of those messages may be queries i.e. they get information from the
application.
The application can consist of two processes. One process will handle the
transactions.
The other process will handle the queries. Each process will have its own
copy of the business data.
The process handling transactions can mmap to the loop device for its copy
of the memory. The loop device must have a normal file for its backing
storage.
The process handling queries can mmap to another normal file for its copy
of the memory.  Both these memories have identical data at the beginning.
Queries and transactions can now be handled simultaneously by the
respective processes.
The query process can update its memory periodically by obtaining the
changes that have have happened to the loop device.
By using the ioctl call to retrieve the dirty page offsets, only the dirty
pages need to be copied over to the query process's copy of memory. We can
infact have multiple processes to handle queries sharing the same memory.
During this copy over, the transaction process will hold off processing
transactions till the update is complete.

This would be very useful for high speed in-memory transaction systems,
where the query load can be passed of to other processes. Example of such
systems would be a stock trading system, where clients buy and sell
stock(equity, options etc).
At the same time lot of clients would be downloading market data and this
can be done independently of the transactions.

This new facility will provide a way of tracking changes made to business
data, independent of the application domain.


Test program:

Before you run the test program, please create the backing storage file
for the loop device as follows

dd if=/dev/zero of=/root/file bs=4K count=10

Set bs to be whatever pagesize is in your machine. In my machine it was
4K.


#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
#include <unistd.h>
#include <sys/wait.h>
#include <sys/stat.h>
#include <sys/mman.h>
#include <fcntl.h>
#include <string.h>
#include <assert.h>
#include <signal.h>
#include <errno.h>
#include <linux/loop.h>

int main()
{
        int maxPages = 10;
        char* start = 0;
        int fd;
        int dfd;
        int *array = 0;
        int pageSize;
        int elemsPerPage;
        struct loop_info64 info;
        struct loop_pgoff_array pgarray;

        pgarray.max = maxPages;
        pgarray.pgoff = calloc(maxPages, sizeof(long));

        if (pgarray.pgoff == NULL) {
                fprintf(stderr, "can't create pgarray\n");
                exit(1);
        }

        pageSize = getpagesize();
        elemsPerPage = pageSize/sizeof(int);

        /* open the device file */
        if ((fd = open ("/dev/loop0", O_RDWR, S_IRWXU)) < 0) {
                fprintf(stderr, "can't create device file for writing\n");
                goto out5;
        }

        /* open the disk file  to set as backing storage*/
        if ((dfd = open ("/root/file", O_RDWR, S_IRWXU)) < 0) {
                fprintf(stderr, "can't create device file for writing\n");
                goto out4;
        }

        if (ioctl(fd, LOOP_SET_FD, dfd) < 0) {
                perror("ioctl: LOOP_SET_FD");
                goto out3;
        }

        if ((start = mmap(0, maxPages * pageSize, PROT_READ | PROT_WRITE,
                          MAP_SHARED, fd, 0)) == MAP_FAILED) {
                perror("mmap error");
                goto out2;
        }


        if (ioctl(fd, LOOP_SET_TRACK_PGWRITE, 1) < 0) {
                perror("ioctl: LOOP_SET_TRACK_PGWRITE");
                goto out1;
        }

        if (ioctl(fd, LOOP_CLR_PGWRITES, 0) < 0) {
                perror("ioctl: LOOP_CLR_PGWRITES");
                goto out1;
        }

        array = (int *)start;

        array[0] = 5;

        fprintf(stderr, "value = %d\n", array[0]);

        array[1] = 9;

        fprintf(stderr, "value = %d\n", array[1]);

        array[elemsPerPage] = 14;

        fprintf(stderr, "value = %d\n", array[elemsPerPage]);

        array[3*elemsPerPage+60] = 35;

        fprintf(stderr, "value = %d\n", array[3*elemsPerPage+60]);

        if (ioctl(fd, LOOP_GET_PGWRITES, &pgarray) < 0) {
                perror("ioctl: LOOP_GET_PGWRITES");
                goto out1;
        }

        int i;
        for (i= 0; i < pgarray.num; i++)
                fprintf(stderr, "offset %ld\n", pgarray.pgoff[i]);

out1:
        munmap(start, maxPages * pageSize);
out2:
        ioctl(fd, LOOP_CLR_FD, 0);
out3:
        close(dfd);
out4:
        close(fd);
out5:
        return 0;
}



Signed-off-by: Kandan Venkataraman kandan.venkataraman@...group.com



View attachment "patch2" of type "TEXT/PLAIN" (8883 bytes)

Powered by blists - more mailing lists