If we run on a CPU which doesn't implement a particular cache then we
would previously get stuck in an infinite loop, executing a cache op on
the first "line" of the missing cache & then incrementing the address by
0. This was being avoided for the L2 caches, but not for the L1s. Fix
this by generalising the check for a zero line size & avoiding the cache
op loop when this is the case.
Signed-off-by: Paul Burton <paul.burton@mips.com>
Cc: Daniel Schwierzeck <daniel.schwierzeck@gmail.com>
Cc: u-boot@lists.denx.de
When writing code, for example during relocation, we ensure that the
icache has a coherent view of the new instructions with a call to
flush_cache(). This handles the bulk of the work to ensure the new
instructions will execute as expected, however it does not ensure that
the CPU pipeline doesn't already contain instructions taken from a stale
view of the affected memory. This could theoretically be a problem for
relocation, but in practice typically isn't because we sync caches for
enough code after the entry point of the newly written code that by the
time the CPU pipeline might possibly fetch any of it we'll have long ago
written it back & invalidated any stale icache entries. This is however
a problem for shorter regions of code.
In preparation for later patches which write shorter segments of code,
ensure any instruction hazards are cleared by flush_cache() by
introducing & using a new instruction_hazard_barrier() function which
makes use of the jr.hb instruction to clear the hazard.
Signed-off-by: Paul Burton <paul.burton@mips.com>
Cc: Daniel Schwierzeck <daniel.schwierzeck@gmail.com>
Cc: u-boot@lists.denx.de
A typical use of cache maintenance functions is to force writeback of
data which a device is about to read using DMA - for example a
descriptor or command structure. Such users of cache maintenance
functions require that operations on the cache have completed before
they proceed to instruct a device to read memory. This requires that we
place a completion barrier (ie. sync instruction) between the cache ops
and whatever write informs the device to perform DMA.
Whilst strictly speaking this isn't all users of the cache maintenance
functions & we could instead place the barriers in the drivers that
require them, it would be much more invasive to do so than to just have
the barrier be the default by placing it in the cache functions
themselves. The cost is low enough that it shouldn't matter to us in any
rare cases that we use the cache functions when not performing DMA.
Signed-off-by: Paul Burton <paul.burton@mips.com>
Cc: Daniel Schwierzeck <daniel.schwierzeck@gmail.com>
Cc: u-boot@lists.denx.de
Without adding a prompt for CONFIG_MIPS_CM_BASE, Kconfig doesn't allow
defconfigs to set it. Provide the prompt in order to allow for that.
Signed-off-by: Paul Burton <paul.burton@imgtec.com>
Signed-off-by: Daniel Schwierzeck <daniel.schwierzeck@gmail.com>
Reviewed-by: Simon Glass <sjg@chromium.org>
This patch adds support for initialising & maintaining L2 caches on MIPS
systems. The L2 cache configuration may be advertised through either
coprocessor 0 or the MIPS Coherence Manager depending upon the system,
and support for both is included.
If the L2 can be bypassed then we bypass it early in boot & initialise
the L1 caches first, such that we can start making use of the L1
instruction cache as early as possible. Otherwise we initialise the L2
first such that the L1s have no opportunity to generate access to the
uninitialised L2.
Signed-off-by: Paul Burton <paul.burton@imgtec.com>
Rather than probing the cache line sizes on every call of any cache
maintenance function, probe them once during boot & store the values in
the global data structure for later use. This will reduce the overhead
of the cache maintenance functions, which isn't a big deal yet but
becomes more important once L2 caches which may expose their properties
via coprocessor 2 or the CM are supported.
Signed-off-by: Paul Burton <paul.burton@imgtec.com>
Commit fb64cda579 ("MIPS: Abstract cache op loops with a macro")
accidentally modified invalidate_dcache_range to operate on the L1
Icache instead of the Dcache. Fix the cache op used to operate on the
Dcache.
Signed-off-by: Paul Burton <paul.burton@imgtec.com>
Fixes: fb64cda579 ("MIPS: Abstract cache op loops with a macro")
The various cache maintenance routines perform a number of loops over
cache lines. Rather than duplicate the code for performing such loops,
abstract it out into a new cache_loop macro which performs an arbitrary
number of cache ops on a range of addresses. This reduces duplication in
the existing L1 cache maintenance code & will allow for not adding
further duplication when introducing L2 cache support.
Signed-off-by: Paul Burton <paul.burton@imgtec.com>
Allow L1 Icache & L1 Dcache line size to be specified separately, since
there's no architectural mandate that they be the same. The
[id]cache_line_size functions are tidied up to take advantage of the
fact that the Kconfig entries are always present to simply check them
for zero rather than needing to #ifdef on their presence.
Signed-off-by: Paul Burton <paul.burton@imgtec.com>
[removed CONFIG_SYS_CACHELINE_SIZE in include/configs/pic32mzdask.h]
Signed-off-by: Daniel Schwierzeck <daniel.schwierzeck@gmail.com>
Move details of the L1 cache line sizes & total sizes into Kconfig,
defaulting to 0. A new CONFIG_SYS_CACHE_SIZE_AUTO Kconfig entry is
introduced to allow platforms to select auto-detection of cache sizes,
and it defaults to being enabled if none of the cache sizes are set by
the configuration (ie. sizes are all the default 0), and code is
adjusted to #ifdef on that rather than on the definition of the sizes
(which will always be defined even if 0).
Signed-off-by: Paul Burton <paul.burton@imgtec.com>
This patch makes sure that the flush/invalidate_dcache_range() functions
can handle corner-case calls like this -- invalidate_dcache_range(0, 0, 0);
This call is valid and is happily produced by USB EHCI code for example.
The expected behavior of the cache function(s) in this case is that they
will do no operation, since the size is zero.
The current implementation though has a problem where such invocation will
result in a hard CPU hang. This is because under such conditions, where the
start_addr = 0 and stop = 0, the addr = 0 and aend = 0xffffffe0 . The loop
will then try to iterate over the entire address space, which in itself is
wrong. But iterating over the entire address space might also hit some odd
address which will cause bus hang. The later happens on the Atheros MIPS.
Signed-off-by: Marek Vasut <marex@denx.de>
Cc: Daniel Schwierzeck <daniel.schwierzeck@gmail.com>
Cc: Hans de Goede <hdegoede@redhat.com>
Move the more developed mips32 version of the cache maintenance
functions to a common arch/mips/lib/cache.c, in order to reduce
duplication between mips32 & mips64.
Signed-off-by: Paul Burton <paul.burton@imgtec.com>
Cc: Daniel Schwierzeck <daniel.schwierzeck@gmail.com>
