strange segfaults with CompactFlash in "true-ide" mode on PXA270

[sshtylyov@mvista.com (Sergei Shtylyov)]

Hello.

Juergen Schindele wrote:

>>> we have a strange problem when booting from a Compact Flash card
>>> as root-device. In one of 10 cases while booting an application from
>>> root-filesystem dies with "Segmentation fault" without a visible error.
>>> After booting the same application runs hundred times without error.
>>> (we're talking about standard appl. like chat, pppd, hwclock and so on)
>>>
>>> For debugging purposes we put printk's in kernels IDE driver 
>>> and the problem disappeared. But replacing the printk by a usleep
>>> or so (which is not a solution !) brings the problem back.
>>> It seems that the application received faulty data from CF.
>>> But how and why ?????????????????????????????
>>>
>>> We tested with a custom PXA270 board with a CompactFlash in 
>>> "true ide" on processor bus driven by "generic ide / pata_platform"
>>> driver from linux-2.6.27.x.
>>> (by the way linux-2.6.20.x showed the same problem).
>>>
>>> Any ideas, hints, experiences, patches are very welcome :-)
>>> --------------------------------------------------------------
>> I guess that is the old cache problem that apparently was never really
>> fixed:
>> http://marc.info/?t=108537828400002&r=1&w=2

> Hello
> this hint was very helpful for understanding the problem.
> In the mail from Russell King at 
> <http://marc.info/?l=linux-arm-kernel&m=108611676807909&w=2>
> he said : .... the rule is if the CPU writes to a page cache page,
> it must call flush_dcache_page afterwards to ensure cache coherency with user space".

> so i made the following patch in drivers/ide/ide-taskfile.c
> in the function ide_pio_sector() where blocks from disk are read
> into a cached page to apply the rule from Russell. 
> ===================================================================
> --- ide-taskfile.c
> +++ ide-taskfile.c.new
> @@ -278,6 +278,7 @@
>                 hwif->tp_ops->input_data(drive, rq, buf, SECTOR_SIZE);
> 
>         kunmap_atomic(buf, KM_BIO_SRC_IRQ);
> +       flush_dcache_page(page);
>  #ifdef CONFIG_HIGHMEM
>         local_irq_restore(flags);
>  #endif

> And from there on the problem disappeared :-))
> What are you think about ???

    Flushing is only needed for reads and shouldn't be done for the slab pages 
(judging on libata's code).

> Isn't this missing for years in IDE PIO driver ???

    It's actually missing from the ARM port of IDE. E.g. the MIPS port 
provides for that -- see arch/mips/include/asm/mach-generic/ide.h.

WBR, Sergei