Message-ID: <20140401103617.GA10882@node.dhcp.inet.fi>
Date:	Tue, 1 Apr 2014 13:36:17 +0300
From:	"Kirill A. Shutemov" <kirill@...temov.name>
To:	Konstantin Khlebnikov <koct9i@...il.com>
Cc:	linux-mm@...ck.org, linux-kernel@...r.kernel.org,
	Hugh Dickins <hughd@...gle.com>,
	Yury Gribov <y.gribov@...sung.com>,
	Alexandr Andreev <aandreev@...allels.com>,
	Vassili Karpov <av1474@...tv.ru>,
	Andrew Morton <akpm@...ux-foundation.org>,
	"Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>
Subject: Re: [PATCH RFC] drivers/char/mem: byte generating devices and
 poisoned mappings

On Tue, Apr 01, 2014 at 01:16:07AM +0400, Konstantin Khlebnikov wrote:
> This patch adds 256 virtual character devices: /dev/byte0, ..., /dev/byte255.
> Each works like /dev/zero but generates memory filled with particular byte.

Shouldn't /dev/byte0 be an alias for /dev/zero?
I see you reuse ZERO_PAGE(0) for that, but what about all these special
cases /dev/zero has?

> Features/use cases:
> * handy source of non-zero bytes for 'dd' (dd if=/dev/byte1 ...)
> * effective way for allocating poisoned memory (just mmap, without memset)
> * /dev/byte42 is full of stars (*)
> 
> Memory filled by default with non-zero bytes might help optimize logic in some
> applications. For example (according to Yury Gribov) Address Sanitizer generates
> additional conditional jump for each memory access just to handle default zero
> byte as '0x8' to avoid memset`ing huge shadow memory map at the beginning.
> In this case allocating memory via mapping /dev/byte8 will reduce size and
> overhead of instrumented code without adding any memory usage overhead.
> 
> /dev/byteX devices have the same performance optimizations like /dev/zero.
> Shared read-only pages are allocated lazily at the first request and freed by
> the memory shrinker (design inspired by huge-zero-page). Private mappings are
> organized as normal anonymous mappings with special page-fault handler which
> allocates, initializes and installs pages like do_anonymous_page().
> 
> Unlike to /dev/zero shared ro-pages are installed into PTEs as normal pages and
> accounted into file-RSS: vm_normal_page() allows only zero-page to be installed
> as 'special'. This difference is fixable, but I don't see why it's matters.

One thing that could surprise is unexpectedly big core dump files caused
by /dev/byteX mappings. We have special handling for FOLL_DUMP && zero_page.
Not sure if /dev/byteX should be handled this way too.

> This patch also (mostly) implements effective non-zero-filled shmem/tmpfs files,
> (they are used for shared mappings) but here is no interface for the userspace.
> This feature mught be exported as ioctl or fcntl call.
> 
> Signed-off-by: Konstantin Khlebnikov <koct9i@...il.com>
> Cc: Alexandr Andreev <aandreev@...allels.com>
> Cc: Vassili Karpov <av1474@...tv.ru>
> Cc: Yury Gribov <y.gribov@...sung.com>
> Cc: Andrew Morton <akpm@...ux-foundation.org>
> Cc: Hugh Dickins <hughd@...gle.com>
> Cc: Kirill A. Shutemov <kirill.shutemov@...ux.intel.com>
> ---
>  drivers/char/Kconfig     |    7 +
>  drivers/char/mem.c       |  285 ++++++++++++++++++++++++++++++++++++++++++++++
>  include/linux/shmem_fs.h |    4 +
>  mm/shmem.c               |   58 ++++++++-
>  4 files changed, 346 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/char/Kconfig b/drivers/char/Kconfig
> index 1386749..e52cb4e 100644
> --- a/drivers/char/Kconfig
> +++ b/drivers/char/Kconfig
> @@ -15,6 +15,13 @@ config DEVKMEM
>  	  kind of kernel debugging operations.
>  	  When in doubt, say "N".
>  
> +config DEVBYTES

I don't think we want new option for this.

> +	bool "Byte generating devices"
> +	depends on SHMEM
> +	help
> +	  This option adds 256 virual devices similar to /dev/zero,
> +	  one for each byte value: /dev/byte0, /dev/byte1, ..., /dev/byte255.
> +

...

> +static ssize_t byte_read(struct file *file, char __user *buf,
> +			 size_t count, loff_t *ppos)
> +{
> +	unsigned byte = (unsigned long)file->private_data;
> +	size_t written;
> +
> +	if (!count)
> +		return 0;
> +
> +	if (!access_ok(VERIFY_WRITE, buf, count))
> +		return -EFAULT;
> +
> +	written = 0;
> +	while (count) {
> +		size_t chunk = min(count, PAGE_SIZE);
> +
> +		if (__memset_user(buf, byte, chunk))
> +			return -EFAULT;
> +		if (signal_pending(current))
> +			return written ? written : -ERESTARTSYS;
> +		written += chunk;
> +		buf += chunk;
> +		count -= chunk;
> +		cond_resched();
> +	}
> +	return written;
> +}

Shouldn't this code be shared with read_zero()?

> +
> +static ssize_t byte_aio_read(struct kiocb *iocb, const struct iovec *iov,
> +			     unsigned long nr_segs, loff_t pos)
> +{
> +	size_t written = 0;
> +	unsigned long i;
> +	ssize_t ret;
> +
> +	for (i = 0; i < nr_segs; i++) {
> +		ret = byte_read(iocb->ki_filp, iov[i].iov_base, iov[i].iov_len,
> +				&pos);
> +		if (ret < 0)
> +			break;
> +		written += ret;
> +	}
> +
> +	return written ? written : -EFAULT;
> +}

Ditto.

> +
> +static const struct file_operations byte_fops = {
> +	.llseek		= byte_lseek,
> +	.read		= byte_read,
> +	.write		= byte_write,
> +	.aio_read	= byte_aio_read,
> +	.aio_write      = byte_aio_write,
> +	.open		= byte_open,
> +	.mmap		= byte_mmap,
> +};
> +
> +static int __init byte_init(void)
> +{
> +	int major, minor;
> +
> +	major  = __register_chrdev(0, 0, 256, "byte", &byte_fops);
> +	if (major < 0) {
> +		printk("unable to get major for byte devs\n");

Hm. Can we, instead of having 256 devnodes, have one /dev/byte?
User can ask which byte it wants by write() byte to file descriptor before
using it with zero by default.

-- 
 Kirill A. Shutemov
--
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