linux-mips
[Top] [All Lists]

Re: [PATCH 1/4] MIPS: Replace _PAGE_READ with _PAGE_NO_READ

To: Ralf Baechle <ralf@linux-mips.org>
Subject: Re: [PATCH 1/4] MIPS: Replace _PAGE_READ with _PAGE_NO_READ
From: David Daney <ddaney@caviumnetworks.com>
Date: Fri, 13 May 2011 10:17:46 -0700
Cc: Kevin Cernekee <cernekee@gmail.com>, linux-mips@linux-mips.org, linux-kernel@vger.kernel.org
In-reply-to: <20110513155605.GA30674@linux-mips.org>
Original-recipient: rfc822;linux-mips@linux-mips.org
References: <7aa38c32b7748a95e814e5bb0583f967@localhost> <20110513150707.GA26389@linux-mips.org> <BANLkTikcyEzjOWt9pWToE=89dySSEYbw_g@mail.gmail.com> <20110513155605.GA30674@linux-mips.org>
Sender: linux-mips-bounce@linux-mips.org
User-agent: Mozilla/5.0 (X11; U; Linux x86_64; en-US; rv:1.9.1.15) Gecko/20101027 Fedora/3.0.10-1.fc12 Thunderbird/3.0.10
On 05/13/2011 08:56 AM, Ralf Baechle wrote:
On Fri, May 13, 2011 at 08:46:18AM -0700, Kevin Cernekee wrote:

On hardware that does not support RI/XI, EntryLo bits 31:30 / 63:62 will
remain unset and RI/XI permissions will not be enforced.

Nice idea but it breaks on 64-bit hardware running 32-bit kernels.  On
those the RI/XI bits written to c0_entrylo0/1 31:30 will be interpreted as
physical address bits 37:36.

Hmm, are you sure?  (Unfortunately I do not have a 64-bit machine to
test it on.)

I did not touch David's existing build_update_entries(), which makes a
point not to set the RI/XI bits when the RIXI feature is disabled:

         if (kernel_uses_smartmips_rixi) {
                 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_NO_EXEC));
                 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_NO_EXEC));
                 UASM_i_ROTR(p, tmp, tmp, ilog2(_PAGE_GLOBAL) -
ilog2(_PAGE_NO_EXEC));
                 if (r4k_250MHZhwbug())
                         UASM_i_MTC0(p, 0, C0_ENTRYLO0);
                 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
                 UASM_i_ROTR(p, ptep, ptep, ilog2(_PAGE_GLOBAL) -
ilog2(_PAGE_NO_EXEC));
         } else {
                 UASM_i_SRL(p, tmp, tmp, ilog2(_PAGE_GLOBAL)); /*
convert to entrylo0 */
                 if (r4k_250MHZhwbug())
                         UASM_i_MTC0(p, 0, C0_ENTRYLO0);
                 UASM_i_MTC0(p, tmp, C0_ENTRYLO0); /* load it */
                 UASM_i_SRL(p, ptep, ptep, ilog2(_PAGE_GLOBAL)); /*
convert to entrylo1 */
                 if (r45k_bvahwbug())
                         uasm_i_mfc0(p, tmp, C0_INDEX);
         }

If RIXI is enabled, it shifts the SW bits off the end of the register,
then rotates the RI/XI bits into place.

If RIXI is disabled, it shifts the SW bits + RI/XI bits off the end of
the register.  It should not be setting bits 31:30 or 63:62, ever.

(A side issue here is that ROTR is a MIPS R2 instruction, so we could
never remove the old handler and use the RIXI version of the TLB
handler on an R1 machine.)

If setting EntryLo bits 31:30 for RI/XI is illegal on a 64-bit system
running a 32-bit kernel, I suspect we will have a problem with the
existing RIXI TLB update code, regardless of whether my changes are
applied.

I'm not totally certain with my explanation but it seemed like a good
working hypothesis.  Jayachandran C. bisected this morning's linux-queue
on his Netlogic XLR which is MIPS64 R1 and found this comment causing
the problem.

He should dump out the TLB Refill handler with the patch set applied. Perhaps it is erroneously getting in RI/XI mode.

Put #define DEBUG 1 as the first line of tlbex.c, then boot the kernel with 'debug' on the command line. Also it is best to have early printk enabled so you can see the dump.

David Daney.

<Prev in Thread] Current Thread [Next in Thread>