Date: Tue, 24 Nov 1998 11:39:17 -0500 (EST) From: Simon Shapiro <shimon@simon-shapiro.org> To: Doug Rabson <dfr@nlsystems.com> Cc: freebsd-alpha@FreeBSD.ORG, Mike Smith <mike@smith.net.au> Subject: Re: UnAligned Access, but why? Message-ID: <XFMail.981124113917.shimon@simon-shapiro.org> In-Reply-To: <Pine.BSF.4.01.9811241131400.21711-100000@herring.nlsystems.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Doug Rabson, On 24-Nov-98 you wrote:
> On Mon, 23 Nov 1998, Mike Smith wrote:
>
> > > > >
> > > > > The bus_space_read_[14] are causing unaligned access panics.
> > > > > These
> > > > > panics
> > > > > can be eliminated by casting explicitly:
> > > >
> > > > Don't do this. Align the objects you're trying to transfer
> > > > instead.
> > >
> > > Thanx, but how? I took the CAM driver as is and just triesd to
> > > actually
> > > align the data. It makes no difference at all. I am still getting
> > > the
> > > panics. If you look again at my example (the one following the
> > > casts), you
> > > will see that it does not help.
> >
> > I saw no evidence that you were performing aligned transfers; your
> > example was, if I remember correctly, casting some arbitrary pointer to
> > a pointer to a 32-bit object. This will fail if the arbitrary
> > pointer's value is not 32-bit aligned.
> >
> > > Also (forgive my ignorance here), different PCI busss appear to have
> > > different sizes for the same thing. Or, at least this is how all
> > > this
> > > bus_rea/write stuff tries to imply.
> >
> > No. Different PCI bus implementations used different host methods to
> > achieve identical PCI bus transactions.
> >
> > > The end result is a mass confusion on my part. i am trying to fit
> > > the
> > > driver into the existing framework, not change the framework (unless
> > > obviously broken). I do not have enough information to suggest
> > > anything is
> > > broken, except the panics.
> >
> > The problem is most likely that you are attempting to access a local
> > memory data structure member of a given size that is not aligned to
> > that size, or you are attempting to access a data structure inside a
> > PCI device where the region in question is memory-mapped and the data
> > structure is not aligned in the mapping according to its size.
> >
> > > I may need some detailed help here....
> >
> > You may need to provide a more detailed example. To begin with, it'd
> > help if you can clarify whether you're sure you understand what
> > constitutes an "unaligned access", as that's pretty key to working out
> > where you're falling over here.
>
> You must not pass unaligned offsets to in[wl], out[wl], read[wl] and
> write[wl]. For 'w' sized accesses, the offset must be even aligned, for
> 'l' sized accesses, the offset must be a multiple of 4. I think NetBSD
> has code to panic if an illegal offset is used to access device memory.
> I
> may do the same (probably for kernels compiled with DIAGNOSTIC defined).
OK. This all makes lots of sense, but look at the example I provided in
the original post. It is a modification of code I did not write, but some
of you may have:
/* static __inline void */
void
bus_space_write_4(bus_space_tag_t tag, bus_space_handle_t bsh,
bus_size_t offset, u_int32_t value)
{
#if defined(_ALPHA_BUS_PIO_H_)
#if defined(_ALPHA_BUS_MEMIO_H_)
if (tag == ALPHA_BUS_SPACE_IO)
#endif
{
u_int32_t port, data;
port = (u_int32_t)bsh + (u_int32_t)offset;
data = (u_int32_t)value;
printf("%s.%d I/O handle = %x, offset = %x, value = %x\n", __FILE__,
__LINE__, b
sh, offset, value);
printf("%s.%d I/O port = %x, data = %x\n", __FILE__, __LINE__, port, data);
>>>>-------->>>> Everything is honky-dory until this point <<<<--------<<<<
outl(port, data);
>>>>-------->>>> This printf never happens. We panic before it happens
printf("%s.%d I/O DONE = %x, data = %x\n", __FILE__, __LINE__, port, data);
}
#endif
#if defined(_ALPHA_BUS_MEMIO_H_)
#if defined(_ALPHA_BUS_PIO_H_)
else
#endif
writel(bsh + offset, value);
#endif
}
You can trace back this call to the sys/dev/dpt/dpt_scsi.c file the
dpt_send_eata_command function, line 1088, where the line:
dpt_outl(dpt, HA_WDMAADDR, cmd_busaddr);
can be traced back to bus_space_write_4 above. Please trust me in saying
that I tried everything you suggest before posting my mssage. The casts at
the top of my example were necessary to avoid a similar panic. All the
data types involoved are taken from the CAM code and are either pointers or
64bit integers.
Simon
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-alpha" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?XFMail.981124113917.shimon>