In order to workaround Erratum A-008139, we have to invalidate the
tlb entry with tlbilx before overwriting. Due to the performance
consideration, we don't add any memory barrier when acquire/release
the tcd lock. This means the two load instructions for esel_next do
have the possibility to return different value. This is definitely
not acceptable due to the Erratum A-008139. We have two options to
fix this issue:
a) Add memory barrier when acquire/release tcd lock to order the
load/store to esel_next.
b) Just make sure to invalidate and write to the same tlb entry and
tolerate the race that we may get the wrong value and overwrite
the tlb entry just updated by the other thread.
We observe better performance using option b. So reserve an additional
register to save the value of the esel_next.
Signed-off-by: Kevin Hao <haokexin@gmail.com>
Signed-off-by: Scott Wood <scottwood@freescale.com>
It makes no sense to put the instructions for calculating the lock
value (cpu number + 1) and the clearing of eq bit of cr1 in lbarx/stbcx
loop. And when the lock is acquired by the other thread, the current
lock value has no chance to equal with the lock value used by current
cpu. So we can skip the comparing for these two lock values in the
lbz/bne loop.
Signed-off-by: Kevin Hao <haokexin@gmail.com>
Signed-off-by: Scott Wood <scottwood@freescale.com>
Some workloads take a lot of TLB misses despite using traditional
hugepages. Handle these TLB misses in the asm fastpath rather than
going through a bunch of C code. With this patch I measured around a
5x speedup in handling hugepage TLB misses.
Signed-off-by: Scott Wood <scottwood@freescale.com>
Erratum A-008139 can cause duplicate TLB entries if an indirect
entry is overwritten using tlbwe while the other thread is using it to
do a lookup. Work around this by using tlbilx to invalidate prior
to overwriting.
To avoid the need to save another register to hold MAS1 during the
workaround code, TID clearing has been moved from tlb_miss_kernel_e6500
until after the SMT section.
Signed-off-by: Scott Wood <scottwood@freescale.com>
Virtualized environments may expose a e6500 dual-threaded core
as two single-threaded e6500 cores. Take advantage of this
and get rid of the tlb lock and the trap-causing tlbsx in
the htw miss handler by guarding with CPU_FTR_SMT, as it's
already being done in the bolted tlb1 miss handler.
As seen in the results below, measurements done with lmbench
random memory access latency test running under Freescale's
Embedded Hypervisor, there is a ~34% improvement.
Memory latencies in nanoseconds - smaller is better
(WARNING - may not be correct, check graphs)
----------------------------------------------------
Host Mhz L1 $ L2 $ Main mem Rand mem
--------- --- ---- ---- -------- --------
smt 1665 1.8020 13.2 83.0 1149.7
nosmt 1665 1.8020 13.2 83.0 758.1
Signed-off-by: Laurentiu Tudor <Laurentiu.Tudor@freescale.com>
Cc: Scott Wood <scottwood@freescale.com>
[scottwood@freescale.com: commit message tweak]
Signed-off-by: Scott Wood <scottwood@freescale.com>
Commit 82d86de25b "TLB lock recursive"
introduced a bug whereby cpu 0 uses the same value for "lock held" as
is used to indicate that the lock is free. This means that cpu 1 can
acquire the lock whenever it wants, regardless of whether cpu 0 has it
locked, which in turn means we can get duplicate TLB entries.
Add one to the CPU value to ensure we do not use zero as a "lock held"
value.
Signed-off-by: Scott Wood <scottwood@freescale.com>
Reported-by: Ed Swarthout <ed.swarthout@freescale.com>
Previously TID was being cleared before the tlbsx, but not after. This
can lead to a multiway hit between a TLB entry with TID=0 (previously
inserted when PID=0) and a TLB entry with TID!=0 that matches PID.
This can theoretically result in undefined behavior, though we probably
get lucky due to the details of the overlap. It also results in the
inability to use multihit detection to detect other conflicting TLB
entries, as well as poorer TLB utilization due to duplicating kernel
TLB entries.
Rather than try to patch up MAS1 after tlbsx, the entire value is
saved/restored as with MAS2.
I observed a slight improvement in TLB miss performance with this patch
applied.
Signed-off-by: Scott Wood <scottwood@freescale.com>
Reported-by: Ed Swarthout <ed.swarthout@freescale.com>
While bolted handlers (including e6500) do not need to deal with a TLB
miss recursively causing another TLB miss, nested TLB misses can still
happen with crit/mc/debug exceptions -- so we still need to honor
SPRG_TLB_EXFRAME.
We don't need to spend time modifying it in the TLB miss fastpath,
though -- the special level exception will handle that.
Signed-off-by: Scott Wood <scottwood@freescale.com>
Cc: Mihai Caraman <mihai.caraman@freescale.com>
Cc: kvm-ppc@vger.kernel.org
Once special level interrupts are supported, we may take nested TLB
misses -- so allow the same thread to acquire the lock recursively.
The lock will not be effective against the nested TLB miss handler
trying to write the same entry as the interrupted TLB miss handler, but
that's also a problem on non-threaded CPUs that lack TLB write
conditional. This will be addressed in the patch that enables crit/mc
support by invalidating the TLB on return from level exceptions.
Signed-off-by: Scott Wood <scottwood@freescale.com>
The previous code added wrong TLBs and causes machine check errors if
a driver accessed passed the end of the linear mapping instead of
a clean page fault.
Signed-off-by: Ralph E. Bellofatto <ralphbel@us.ibm.com>
Signed-off-by: Benjamin Krill <ben@codiert.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
...and make CONFIG_PPC_FSL_BOOK3E conflict with CONFIG_PPC_64K_PAGES.
This fixes a build break with CONFIG_PPC_64K_PAGES on 64-bit book3e,
that was introduced by commit 28efc35fe6
("powerpc/e6500: TLB miss handler with hardware tablewalk support").
Signed-off-by: Scott Wood <scottwood@freescale.com>
It was branching to the cleanup part of the non-bolted handler,
which would have been bad if there were any chips with tlbsrx.
that use the bolted handler.
Signed-off-by: Scott Wood <scottwood@freescale.com>
There are a few things that make the existing hw tablewalk handlers
unsuitable for e6500:
- Indirect entries go in TLB1 (though the resulting direct entries go in
TLB0).
- It has threads, but no "tlbsrx." -- so we need a spinlock and
a normal "tlbsx". Because we need this lock, hardware tablewalk
is mandatory on e6500 unless we want to add spinlock+tlbsx to
the normal bolted TLB miss handler.
- TLB1 has no HES (nor next-victim hint) so we need software round robin
(TODO: integrate this round robin data with hugetlb/KVM)
- The existing tablewalk handlers map half of a page table at a time,
because IBM hardware has a fixed 1MiB indirect page size. e6500
has variable size indirect entries, with a minimum of 2MiB.
So we can't do the half-page indirect mapping, and even if we
could it would be less efficient than mapping the full page.
- Like on e5500, the linear mapping is bolted, so we don't need the
overhead of supporting nested tlb misses.
Note that hardware tablewalk does not work in rev1 of e6500.
We do not expect to support e6500 rev1 in mainline Linux.
Signed-off-by: Scott Wood <scottwood@freescale.com>
Cc: Mihai Caraman <mihai.caraman@freescale.com>
Embedded.Hypervisor category defines GSPRG0..3 physical registers for guests.
Avoid SPRG4-7 usage as scratch in host exception handlers, otherwise guest
SPRG4-7 registers will be clobbered.
For bolted TLB miss exception handlers, which is the version currently
supported by KVM, use SPRN_SPRG_GEN_SCRATCH aka SPRG0 instead of
SPRN_SPRG_TLB_SCRATCH aka SPRG6. Keep using TLB PACA slots to fit in one
64-byte cache line.
For critical exception handlers use SPRG3 instead of SPRG7. Provide a routine
to store and restore user-visible SPRGs. This will be subsequently used
to restore VDSO information in SPRG3. Add EX_R13 to paca slots to free up
SPRG3 and change the critical exception epilog to use it.
Signed-off-by: Mihai Caraman <mihai.caraman@freescale.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Hook DO_KVM macro into 64-bit booke for KVM integration. Extend interrupt
handlers' parameter list with interrupt vector numbers to accomodate the macro.
Only the bolted version of tlb miss handers is addressed now.
Signed-off-by: Mihai Caraman <mihai.caraman@freescale.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Some macros use RA where when RA=R0 the values is 0, so make this
the enforced mnemonic in the macro.
Idea suggested by Andreas Schwab.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Anything that uses a constructed instruction (ie. from ppc-opcode.h),
need to use the new R0 macro, as %r0 is not going to work.
Also convert usages of macros where we are just determining an offset
(usually for a load/store), like:
std r14,STK_REG(r14)(r1)
Can't use STK_REG(r14) as %r14 doesn't work in the STK_REG macro since
it's just calculating an offset.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Before hugetlb, at each level of the table, we test for
!0 to determine if we have a valid table entry. With hugetlb, this
compare becomes:
< 0 is a normal entry
0 is an invalid entry
> 0 is huge
This works because the hugepage code pulls the top bit off the entry
(which for non-huge entries always has the top bit set) as an
indicator that we have a hugepage.
Signed-off-by: Becky Bruce <beckyb@kernel.crashing.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
I happened to comment this code while I was digging through it;
we might as well commit that. I also made some whitespace
changes - the existing code had a lot of unnecessary newlines
that I found annoying when I was working on my tiny laptop.
No functional changes.
Signed-off-by: Becky Bruce <beckyb@kernel.crashing.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Enable hugepages on Freescale BookE processors. This allows the kernel to
use huge TLB entries to map pages, which can greatly reduce the number of
TLB misses and the amount of TLB thrashing experienced by applications with
large memory footprints. Care should be taken when using this on FSL
processors, as the number of large TLB entries supported by the core is low
(16-64) on current processors.
The supported set of hugepage sizes include 4m, 16m, 64m, 256m, and 1g.
Page sizes larger than the max zone size are called "gigantic" pages and
must be allocated on the command line (and cannot be deallocated).
This is currently only fully implemented for Freescale 32-bit BookE
processors, but there is some infrastructure in the code for
64-bit BooKE.
Signed-off-by: Becky Bruce <beckyb@kernel.crashing.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
On MMUs such as FSL where we can guarantee the entire linear mapping is
bolted, we don't need to worry about linear TLB misses. If on top of
that we do a full table walk, we get rid of all recursive TLB faults, and
can dispense with some state saving. This gains a few percent on
TLB-miss-heavy workloads, and around 50% on a benchmark that had a high
rate of virtual page table faults under the normal handler.
While touching the EX_TLB layout, remove EX_TLB_MMUCR0, EX_TLB_SRR0, and
EX_TLB_SRR1 as they're not used.
[BenH: Fixed build with 64K pages (wsp config)]
Signed-off-by: Scott Wood <scottwood@freescale.com>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Fix some minor typos:
* informations => information
* there own => their own
* these => this
Signed-off-by: Sylvestre Ledru <sylvestre.ledru@scilab.org>
Signed-off-by: Randy Dunlap <randy.dunlap@oracle.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We were seeing oops like the following when we did an rmmod on a module:
Unable to handle kernel paging request for instruction fetch
Faulting instruction address: 0x8000000000008010
Oops: Kernel access of bad area, sig: 11 [#1]
SMP NR_CPUS=2 P5020 DS
last sysfs file: /sys/devices/qman-portals.2/qman-pool.9/uevent
Modules linked in: qman_tester(-)
NIP: 8000000000008010 LR: c000000000074858 CTR: 8000000000008010
REGS: c00000002e29bab0 TRAP: 0400 Not tainted
(2.6.34.6-00744-g2d21f14)
MSR: 0000000080029000 <EE,ME,CE> CR: 24000448 XER: 00000000
TASK = c00000007a8be600[4987] 'rmmod' THREAD: c00000002e298000 CPU: 1
GPR00: 8000000000008010 c00000002e29bd30 8000000000012798 c00000000035fb28
GPR04: 0000000000000002 0000000000000002 0000000024022428 c000000000009108
GPR08: fffffffffffffffe 800000000000a618 c0000000003c13c8 0000000000000000
GPR12: 0000000022000444 c00000000fffed00 0000000000000000 0000000000000000
GPR16: 00000000100c0000 0000000000000000 00000000100dabc8 0000000010099688
GPR20: 0000000000000000 00000000100cfc28 0000000000000000 0000000010011a44
GPR24: 00000000100017b2 0000000000000000 0000000000000000 0000000000000880
GPR28: c00000000035fb28 800000000000a7b8 c000000000376d80 c0000000003cce50
NIP [8000000000008010] .test_exit+0x0/0x10 [qman_tester]
LR [c000000000074858] .SyS_delete_module+0x1f8/0x2f0
Call Trace:
[c00000002e29bd30] [c0000000000748b4] .SyS_delete_module+0x254/0x2f0 (unreliable)
[c00000002e29be30] [c000000000000580] syscall_exit+0x0/0x2c
Instruction dump:
XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
38600000 4e800020 60000000 60000000 <4e800020> 60000000 60000000 60000000
---[ end trace 4f57124939a84dc8 ]---
This appears to be due to checking the wrong permission bits in the
instruction_tlb_miss handling if the address that faulted was in vmalloc
space. We need to look at the supervisor execute (_PAGE_BAP_SX) bit and
not the user bit (_PAGE_BAP_UX/_PAGE_EXEC).
Also removed a branch level since it did not appear to be used.
Reported-by: Jeffrey Ladouceur <Jeffrey.Ladouceur@freescale.com>
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Support for TLB reservation (or TLB Write Conditional) and Paired MAS
registers are optional for a processor implementation so we handle
them via MMU feature sections.
We currently only used paired MAS registers to access the full RPN + perm
bits that are kept in MAS7||MAS3. We assume that if an implementation has
hardware page table at this time it also implements in TLB reservations.
Signed-off-by: Kumar Gala <galak@kernel.crashing.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This is an attempt at cleaning up a bit the way we handle execute
permission on powerpc. _PAGE_HWEXEC is gone, _PAGE_EXEC is now only
defined by CPUs that can do something with it, and the myriad of
#ifdef's in the I$/D$ coherency code is reduced to 2 cases that
hopefully should cover everything.
The logic on BookE is a little bit different than what it was though
not by much. Since now, _PAGE_EXEC will be set by the generic code
for executable pages, we need to filter out if they are unclean and
recover it. However, I don't expect the code to be more bloated than
it already was in that area due to that change.
I could boast that this brings proper enforcing of per-page execute
permissions to all BookE and 40x but in fact, we've had that now for
some time as a side effect of my previous rework in that area (and
I didn't even know it :-) We would only enable execute permission if
the page was cache clean and we would only cache clean it if we took
and exec fault. Since we now enforce that the later only work if
VM_EXEC is part of the VMA flags, we de-fact already enforce per-page
execute permissions... Unless I missed something
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
This adds the TLB miss handler assembly, the low level TLB flush routines
along with the necessary hook for dealing with our virtual page tables
or indirect TLB entries that need to be flushes when PTE pages are freed.
There is currently no support for hugetlbfs
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
