Presently The SH-4 cache flushing code uses flush_cache_4096() for most
of the real flushing work, which breaks down to a fixed 4096 unroll and
increment. Not only is this sub-optimal for larger page sizes, it's also
uncovered a bug in sh4_flush_dcache_page() when large page sizes are used
and we have no cache aliases -- resulting in only a part of the page's
D-cache lines being written back.
Signed-off-by: Valentin Sitdikov <valentin.sitdikov@siemens.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
To allow the MMU to be switched between 29bit and 32bit mode at runtime
some constants need to swapped for functions that return a runtime
value.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Replace the use of PHYSADDR() with __pa(). PHYSADDR() is based on the
idea that all addresses in P1SEG are untranslated, so we can access an
address's physical page as an offset from P1SEG. This doesn't work for
CONFIG_PMB/CONFIG_PMB_FIXED because pages in P1SEG and P2SEG are used
for PMB mappings and so can be translated to any physical address.
Likewise, replace a P1SEGADDR() use with virt_to_phys().
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Fix some callers of jump_to_uncached() and back_to_cached() that were
not annotated with __uses_jump_to_uncached.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
If PAGE_SIZE is presently over 4k we do a lot of extra flushing given
that we purge the cache 4k at a time. Make it explicitly 4k per
iteration, rather than iterating for PAGE_SIZE before looping over again.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This builds on top of the MIPS r4k code that does roughly the same thing.
This permits the use of kmap_coherent() for mapped pages with dirty
dcache lines and falls back on kmap_atomic() otherwise.
This also fixes up a problem with the alias check and defers to
shm_align_mask directly.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This kills off the unrolled segment based flushers on SH-4 and switches
over to a generic unrolled approach derived from the writethrough segment
flusher.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
PHYSADDR() runs in to issues in 32-bit mode when we do not have the
legacy P1/P2 areas mapped, as such, we need to use page_to_phys()
directly, which also happens to do the right thing in legacy 29-bit mode.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
The i-cache flush in the case of VM_EXEC was added way back when as a
sanity measure, and in practice we only care about evicting aliases from
the d-cache. As a result, it's possible to drop the i-cache flush
completely here.
After careful profiling it's also come up that all of the work associated
with hunting down aliases and doing ranged flushing ends up generating
more overhead than simply blasting away the entire dcache, particularly
if there are many mm's that need to be iterated over. As a result of
that, just move back to flush_dcache_all() in these cases, which restores
the old behaviour, and vastly simplifies the path.
Additionally, on platforms without aliases at all, this can simply be
nopped out. Presently we have the alias check in the SH-4 specific
version, but this is true for all of the platforms, so move the check up
to a generic location. This cuts down quite a bit on superfluous cacheop
IPIs.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
There was quite a lot of tab->space damage done here from a former patch,
clean it up once and for all.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This reverts commit 64a6d72213.
Unfortunately we can't use on_each_cpu() for all of the cache ops, as
some of them only require preempt disabling. This seems to be the same
issue that impacts the mips r4k caches, where this code was based on.
This fixes up a deadlock that showed up in some IRQ context cases.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This adopts the special-cased 2-way write-through dcache flusher for
N-ways and moves it in to the generic path. Assignment is done at runtime
via the check for the CCR_CACHE_WT bit in the same path as the per-way
writeback flushers.
Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Change the method used to flush the cache in write-through mode to
avoid corrupted data being written back to memory.
Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This is a pure documentation, to try to explain why the cache flushing code
for the SH4 is implemented the way it is.
Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
on_each_cpu() takes care of IRQ and preempt handling, the localized
handling in each of the called functions can be killed off.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This does a bit of rework for making the cache flushers SMP-aware. The
function pointer-based flushers are renamed to local variants with the
exported interface being commonly implemented and wrapping as necessary.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This paves the way for allowing individual CPUs to overload the
individual flushing routines that they care about without having to
depend on weak aliases. SH-4 is converted over initially, as it wires
up pretty much everything. The majority of the other CPUs will simply use
the default no-op implementation with their own region flushers wired up.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
We use flush_cache_page() outright in copy_to_user_page(), and nothing
else needs it, so just kill it off. SH-5 still defines its own version,
but that too will go away in the same fashion once it converts over.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
flush_dcache_all() is used internally by the SH-4 cache code, it is not
part of the exported cache API, so make it static and don't export it.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This migrates the alias computation and printing of probed cache
parameters from the SH-4 code to the shared cpu_cache_init().
This permits other platforms with aliases to make use of the same
probe logic without having to roll their own, and also produces
consistent output regardless of platform.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This provides a central point for CPU cache initialization routines.
This replaces the antiquated p3_cache_init() method, which the vast
majority of CPUs never cared about.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This optimizes for the cases when a CPU does not yet have a valid ASID
context associated with it, as in this case there is no work for any of
flush_cache_mm()/flush_cache_page()/flush_cache_range() to do. Based on
the the MIPS implementation.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This splits out the SH-4 __flush_xxx_region() functions and defines them
as weak symbols. This allows us to provide optimized versions without
having to ifdef cache-sh4.c to death.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This inverts the delayed dcache flush a bit to be more in line with other
platforms. At the same time this also gives us the ability to do some
more optimizations and cleanup. Now that the update_mmu_cache() callsite
only tests for the bit, the implementation can gradually be split out and
made generic, rather than relying on special implementations for each of
the peculiar CPU types.
SH7705 in 32kB mode and SH-4 still need slightly different handling, but
this is something that can remain isolated in the varying page copy/clear
routines. On top of that, SH-X3 is dcache coherent, so there is no need
to bother with any of these tests in the PTEAEX version of
update_mmu_cache(), so we kill that off too.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This uses jump_to_uncached() which is now given the noinline attribute
due to the special section mapping. Kill off the inline attribute to
fix up compilation failure.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Add implementation of flush_icache_range() suitable for signal handler
and kprobes. Remove flush_cache_sigtramp() and change signal.c to use
flush_icache_range().
Signed-off-by: Chris Smith <chris.smith@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Presently most of the 29-bit physical parts do P1/P2 segmentation
with a 1:1 cached/uncached mapping, jumping between the two to
control the caching behaviour. This provides the basic infrastructure
to maintain this behaviour on 32-bit physical parts that don't map
P1/P2 at all, using a shiny new linker section and corresponding
fixmap entry.
Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Calculate the number of cache aliases on probed L2 caches, and while
we're at it, print out the detected statistics at boot time for these
also.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
There was an off-by-1 on the cache alias detection logic on SH-4,
which caused n_aliases to always be 1 even when the page size
precluded the existence of aliases.
With this corrected, 64KB pages happily reports n_aliases == 0, and
hits the appropriate fast paths in the flushing routines.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
current_cpu_data uses smp_processor_id() in order to find the
corresponding cpu_data. As the cache descs are all currently
identical, just have this look at probed results from the boot
CPU.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
The first 1MB of P3 space was reserved and used for page colouring,
as we've reworked that to use fixmaps, we can reclaim the space and
hand it back to VMALLOC_START.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This wires up kmap_coherent() and kunmap_coherent() on SH-4, and
moves away from the p3map_mutex and reserved P3 space, opting to
use fixmaps for colouring instead.
The copy_user_page()/clear_user_page() implementations are moved
to this, which fixes the nasty blowups with spinlock debugging
as a result of having some of these calls nested under the page
table lock.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
There are a lot of bogus cpu_data-> references that only end up working
for the boot CPU, convert these to current_cpu_data to fixup SMP.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This converts the lazy dcache handling to the model described in
Documentation/cachetlb.txt and drops the ptep_get_and_clear() hacks
used for the aliasing dcaches on SH-4 and SH7705 in 32kB mode. As a
bonus, this slightly cuts down on the cache flushing frequency.
With that and the PTEA handling out of the way, the update_mmu_cache()
implementations can be consolidated, and we no longer have to worry
about which configuration the cache is in for the SH7705 case.
And finally, explicitly disable the lazy writeback on SMP (SH-4A).
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
There were a number of places that made evil PAGE_SIZE == 4k
assumptions that ended up breaking when trying to play with
8k and 64k page sizes, this fixes those up.
The most significant change is the way we load THREAD_SIZE,
previously this was done via:
mov #(THREAD_SIZE >> 8), reg
shll8 reg
to avoid a memory access and allow the immediate load. With
a 64k PAGE_SIZE, we're out of range for the immediate load
size without resorting to special instructions available in
later ISAs (movi20s and so on). The "workaround" for this is
to bump up the shift to 10 and insert a shll2, which gives a
bit more flexibility while still being much cheaper than a
memory access.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
IRQs disabling in flush_cache_4096 for cache purge. Under certain
workloads we would get an IRQ in the middle of a purge operation,
and the cachelines would remain in an inconsistent state, leading
to occasional stack corruption.
Signed-off-by: Takeo Takahashi <takahashi.takeo@renesas.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
flush_cache_mm() wraps in to flush_cache_all(), which is rather
excessive given that the number of PTEs within the specified context
are generally quite low. Optimize for walking the mm's VMA list and
selectively flushing the VMA ranges from the dcache. Invalidate the
icache only if a VMA sets VM_EXEC.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
Currently when making changes to control registers, we
typically need some time for changes to take effect (8
nops, generally). However, for sh4a we simply need to
do an icbi..
This is a simple patch for implementing a general purpose
ctrl_barrier() which functions as a control register write
barrier. There's some additional documentation in the patch
itself, but it's pretty self explanatory.
There were also some places where we were not doing the
barrier, which didn't seem to have any adverse effects on
legacy parts, but certainly did on sh4a. It's safer to have
the barrier in place for legacy parts as well in these cases,
though this does make flush_tlb_all() more expensive (by an
order of 8 nops). We can ifdef around the flush_tlb_all()
case for now if it's clear that all legacy parts won't have
a problem with this.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
This reworks some of the SH-4 cache handling code to more easily
accomodate newer-style caches (particularly for the > direct-mapped
case), as well as optimizing some of the old code.
Signed-off-by: Richard Curnow <richard.curnow@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
flush_cache_range() wasn't page aligning the end of the range,
we can't assume that it will always be page aligned, and we
ended up getting unaligned faults in some rare call paths.
Additionally, we add a small optimization to just purge the
dcache entirely if the range is large enough that the page
table walking will take longer. We use an arbitrary value of
64 pages for the large range size, as per sh64.
Signed-off-by: Paul Mundt <lethal@linux-sh.org>
