flushing cache after block io

flushing cache after block io

[Bahadir Balban]

Hi,

On an ARM core with write-allocate writeback cache, there can be cache
inconsistencies when an executable file is read from a block device
into memory.

On a compact flash block device driver, I got around this problem by
flushing the caches for the pages that bio buffers use on every bio
read. E.g.

cflash_read_sectors(dev, sect_begin, nsect, buf);
flush_icache_range(buf, buf + nsect*512);

Now I have the same problem with a ATA harddisk connected with a
PDC20269 controller. The question is where would be the best
high-level place to put these flushes so that it works regardless of
what block IO driver is used?


Thanks,
Bahadir

Re: flushing cache after block io

[Mark Lord]

Bahadir Balban wrote:
> On an ARM core with write-allocate writeback cache, there can be cache
> inconsistencies when an executable file is read from a block device
> into memory.
> 
> On a compact flash block device driver, I got around this problem by
> flushing the caches for the pages that bio buffers use on every bio
> read. E.g.
> 
> cflash_read_sectors(dev, sect_begin, nsect, buf);
> flush_icache_range(buf, buf + nsect*512);
> 
> Now I have the same problem with a ATA harddisk connected with a
> PDC20269 controller. The question is where would be the best
> high-level place to put these flushes so that it works regardless of
> what block IO driver is used?

Russell, didn't you go through all of this pain and fix it
a year or two ago?

Bahadir, exactly what kernel/version is this with?

Re: flushing cache after block io

[Russell King]

On Tue, May 01, 2007 at 01:46:56PM -0400, Mark Lord wrote:
> 
> Bahadir Balban wrote:
> >On an ARM core with write-allocate writeback cache, there can be cache
> >inconsistencies when an executable file is read from a block device
> >into memory.
> >
> >On a compact flash block device driver, I got around this problem by
> >flushing the caches for the pages that bio buffers use on every bio
> >read. E.g.
> >
> >cflash_read_sectors(dev, sect_begin, nsect, buf);
> >flush_icache_range(buf, buf + nsect*512);
> >
> >Now I have the same problem with a ATA harddisk connected with a
> >PDC20269 controller. The question is where would be the best
> >high-level place to put these flushes so that it works regardless of
> >what block IO driver is used?
> 
> Russell, didn't you go through all of this pain and fix it
> a year or two ago?

I thought so.

> Bahadir, exactly what kernel/version is this with?

Bahadir must be using an old kernel.

-- 
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:

Re: flushing cache after block io

[Bahadir Balban]

On 5/1/07, Mark Lord <liml@rtr.ca> wrote:
>
> Russell, didn't you go through all of this pain and fix it
> a year or two ago?
>
> Bahadir, exactly what kernel/version is this with?
>
>

This is on 2.6.19 but I'm unsure if there's a fix in the tree as I
might be one of the few to try ARM MPCore with a block device.
Although, filesystems over mtd or nfs work fine without the flushing.

If you have an idea please let me know.

Thanks,
Bahadir

Re: flushing cache after block io

[Russell King]

On Thu, May 03, 2007 at 02:25:44PM +0100, Bahadir Balban wrote:
> On 5/1/07, Mark Lord <liml@rtr.ca> wrote:
> >
> >Russell, didn't you go through all of this pain and fix it
> >a year or two ago?
> >
> >Bahadir, exactly what kernel/version is this with?
> >
> >
> 
> This is on 2.6.19 but I'm unsure if there's a fix in the tree as I
> might be one of the few to try ARM MPCore with a block device.
> Although, filesystems over mtd or nfs work fine without the flushing.

You want to be running something post 2.6.20-rc5 ideally (so 2.6.20).
There's a couple of cache flushing related patches merged between .19
and .20.

-- 
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of: