On Fri, 2007-04-20 at 11:43 -0700, Andrew Morton wrote:
> On Fri, 20 Apr 2007 09:10:41 -0400
> James Bottomley <James.Bottomley@xxxxxxxxxxxx> wrote:
>
> > On Thu, 2007-04-19 at 22:20 -0700, Andrew Morton wrote:
> > > On Thu, 19 Apr 2007 16:27:26 -0000 "Cameron, Steve" <Steve.Cameron@xxxxxx> wrote:
> > > >
> > > > Something like
> > > >
> > > > if (sizeof(blah) > 4) {
> > > > do all the assignments with shifts
> > > > }
> > > >
> > > > might be slighly better since the CDB is already zeroed
> > > > by cmd_alloc() and doesn't need to be zeroed a 2nd time.
> > > >
> > > > -- steve
> > > >
> > > > -----Original Message-----
> > > > From: James Bottomley [mailto:James.Bottomley@xxxxxxxxxxxx]
> > > > Sent: Thu 4/19/2007 11:22 AM
> > > > To: Miller, Mike (OS Dev)
> > > > Cc: Hisashi Hifumi; akpm@xxxxxxxxxxxxxxxxxxxx; jens.axboe@xxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; linux-scsi@xxxxxxxxxxxxxxx; Cameron, Steve
> > > > Subject: RE: [PATCH] cciss: Fix warnings during compilation under 32bitenvironment
> > > >
> > > > On Thu, 2007-04-19 at 16:12 +0000, Miller, Mike (OS Dev) wrote:
> > > > > > > Nak. You still haven't told where you saw these warnings. What
> > > > > > > compiler are you using? I do not see these in my 32-bit environment.
> > > > > >
> > > > > > I think it's seen with CONFIG_LBD=n on 32 bits
> > > > > >
> > > > > > In that configuration, sector_t is a u32 (it's u64 even on 32
> > > > > > bits with CONFIG_LBD=y). The proposed code change is a
> > > > > > simple cut and paste from the sd driver.
> > > > >
> > > > > Isn't there a better way than testing each one?
> > > >
> > > > It's not such a bad option. The sizeof() test is compile time
> > > > determinable, so the compiler simply zeros the fields in the
> > > > CONFIG_LBD=n case and does the shift for CONFIG_LBD=y. It certainly
> > > > never compiles to four inline condition checks.
> > > >
> > >
> > > Boy you guys make a mess of a nice email trail :(
> > >
> > >
> > > --- linux-2.6.21-rc7.org/drivers/block/cciss.c 2007-04-17 16:36:02.000000000 +0900
> > > +++ linux-2.6.21-rc7/drivers/block/cciss.c 2007-04-17 16:25:53.000000000 +0900
> > > @@ -2552,10 +2552,10 @@ static void do_cciss_request(request_que
> > > } else {
> > > c->Request.CDBLen = 16;
> > > c->Request.CDB[1]= 0;
> > > - c->Request.CDB[2]= (start_blk >> 56) & 0xff; //MSB
> > > - c->Request.CDB[3]= (start_blk >> 48) & 0xff;
> > > - c->Request.CDB[4]= (start_blk >> 40) & 0xff;
> > > - c->Request.CDB[5]= (start_blk >> 32) & 0xff;
> > > + c->Request.CDB[2]= sizeof(start_blk) > 4 ? (start_blk >> 56) & 0xff : 0; //MSB
> > > + c->Request.CDB[3]= sizeof(start_blk) > 4 ? (start_blk >> 48) & 0xff : 0;
> > > + c->Request.CDB[4]= sizeof(start_blk) > 4 ? (start_blk >> 40) & 0xff : 0;
> > > + c->Request.CDB[5]= sizeof(start_blk) > 4 ? (start_blk >> 32) & 0xff : 0;
> > > c->Request.CDB[6]= (start_blk >> 24) & 0xff;
> > > c->Request.CDB[7]= (start_blk >> 16) & 0xff;
> > > c->Request.CDB[8]= (start_blk >> 8) & 0xff;
> > >
> > > This is not the first time we've hit this problem and presumably it won't
> > > be the last time.
> > >
> > > Could we do something like
> > >
> > > #if (BITS_PER_LONG > 32) || defined(CONFIG_LBD)
> > > #define sector_upper_32(sector) ((sector) >> 32)
> > > #else
> > > #define sector_upper_32(sector) (0)
> > > #endif
> > >
> > > and then cciss can do
> > >
> > > - c->Request.CDB[2]= start_blk >> 56;
> > > + c->Request.CDB[2]= sector_upper_32(start_blk) >> 24;
> > >
> > > which will do the right thing.
> >
> > Sure, we could do that. The slight problem is that we don't have
> > general agreement in the kernel how to handle sector_t. So, the only
> > consumer in scsi, sd, does the sizeof(block) ? thing. If you look in
> > libata you'll see that it unconditionally uses a u64 for picking up the
> > value of sector_t so the shift is never invalid ... if we're going to
> > start making macros for handling this, we probably need to ask the
> > janitors to fix all of our existing code to use them ... or we could
> > just let sleeping dogs lie ..
>
> Let's decide how we _should_ do it, implement that, then teach cciss about
> it, then try to ensure that new code uses the approved interfaces. Over
> time, people may (or may not) migrate existing code over to the new
> interfaces.
>
> They may also merge new code which uses open-coded hand-rolled stuff, too :(
>
> > Is there even a valid use for CONFIG_LBD=n anymore, anyway?
>
> Lots and lots and lots of systems don't have any disks larger than 2TB.
>
> CONFIG_LBD=y gives us an additional 3kb of instructions on i386
> allnoconfig. Other architectures might do less well. It's not a huge
> difference, but that's the way in which creeping bloatiness happens.
OK, sure, but if we really care about this saving, then unconditionally
casting to u64 is therefore wrong as well ... this is starting to open
quite a large can of worms ...
For the record, if we have to do this, I fancy sector_upper_32() ... we
should already have some similar accessor for dma_addr_t as well.
James
-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
