Re: [PATCH v3] MIPS: mm: Prevent a TLB shutdown on initial uniquification

[Thomas Bogendoerfer <tsbogend@alpha.franken.de>]

On Thu, Nov 13, 2025 at 05:21:10AM +0000, Maciej W. Rozycki wrote:
> Depending on the particular CPU implementation a TLB shutdown may occur 
> if multiple matching entries are detected upon the execution of a TLBP 
> or the TLBWI/TLBWR instructions.  Given that we don't know what entries 
> we have been handed we need to be very careful with the initial TLB 
> setup and avoid all these instructions.
> 
> Therefore read all the TLB entries one by one with the TLBR instruction, 
> bypassing the content addressing logic, and truncate any large pages in 
> place so as to avoid a case in the second step where an incoming entry 
> for a large page at a lower address overlaps with a replacement entry 
> chosen at another index.  Then preinitialize the TLB using addresses 
> outside our usual unique range and avoiding clashes with any entries 
> received, before making the usual call to local_flush_tlb_all().
> 
> This fixes (at least) R4x00 cores if TLBP hits multiple matching TLB 
> entries (SGI IP22 PROM for examples sets up all TLBs to the same virtual 
> address).
> 
> Signed-off-by: Maciej W. Rozycki <macro@orcam.me.uk>
> Fixes: 35ad7e181541 ("MIPS: mm: tlb-r4k: Uniquify TLB entries on init")
> Cc: stable@vger.kernel.org # v6.17+
> ---
> Hi,
> 
>  On second thoughts I decided against including wired entries in our VPN 
> matching discovery for clash avoidance.  The reason is as I wrote before 
> it makes no sense to have wired entries for KSEG0 addresses, so it should 
> be safe to assume we won't ever make one, and then if a wired entry maps a 
> large page, which is quite likely, then our clash avoidance logic won't 
> handle an overlap where the two VPN values of a clashing pair don't match, 
> so it makes no sense to pretend we can handle wired entries with the code 
> as proposed.
> 
>  Pasting v2 discussion below verbatim as it still applies.
> 
>  Verified the same way as before, also with some diagnostics added so as 
> to make sure things get set up correctly, with my Malta/74Kf system for a 
> 32-bit configuration and with my SWARM/BCM1250 system for a 64-bit one.
> 
>  In addition to the Wired register setup discussed with v1 I have realised 
> the incoming entries may include large pages, possibly exceeding the size 
> of KSEG0 even.  Such entries may overlap with our temporary entries added 
> in the second step, so truncate any large pages in place as this ensures 
> no clash happens with the received contents of the TLB.
> 
>  NB this doesn't handle incoming PageGrain.ESP having been set, but it's 
> an unrelated preexisting issue that would have to be handled elsewhere.  
> Possibly it doesn't matter in reality.
> 
>  Additionally PageMask is left set at what has been retrieved from the 
> last incoming TLB entry in the first step and has to be reset to our page 
> size before proceeding with the second step.
> 
>  And last but not least the comparator function returned 0 incorrectly 
> when the difference between 64-bit elements was positive but with none of 
> the high-order 32 bits set.  Fixed with a branchless sequence of 3 machine 
> instructions, which I think is the minimum here (only the sign and zero 
> matter here, but this sequence actually produces -1/0/1, because why not).
> No change for the 32-bit case, the difference is returned as is.
> 
>   Maciej
> 
> Changes from v2 (at 
> <https://lore.kernel.org/r/alpine.DEB.2.21.2511122032400.25436@angie.orcam.me.uk/>):
> 
> - Revert the v2 update to include wired entries while reading original 
>   contents of TLB.
> 
> Changes from v1 (at 
> <https://lore.kernel.org/r/alpine.DEB.2.21.2511110547430.25436@angie.orcam.me.uk/>):
> 
> - Also include wired entries while reading original contents of TLB.
> 
> - Truncate any large pages in place while reading original TLB entries.
> 
> - Reset PageMask to PM_DEFAULT_MASK after reading in TLB entries.
> 
> - Fix the 64-bit case for the sort comparator.
> ---
>  arch/mips/mm/tlb-r4k.c |  100 ++++++++++++++++++++++++++++++-------------------
>  1 file changed, 63 insertions(+), 37 deletions(-)

applied to mips-fixes.

Thomas.

-- 
Crap can work. Given enough thrust pigs will fly, but it's not necessarily a
good idea.                                                [ RFC1925, 2.3 ]