| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
When searching for a suitable node that should be used for inserting a new
register, which does not fall within the range of any existing node, we not
only looks for nodes which are directly adjacent to the new register, but
for nodes within a certain proximity. This is done to avoid creating lots
of small nodes with just a few registers spacing in between, which would
increase memory usage as well as tree traversal time.
This means there might be multiple node candidates which fall within the
proximity range of the new register. If we choose the first node we
encounter, under certain register insertion patterns it is possible to end
up with overlapping ranges. This will break order in the rbtree and can
cause the cached register value to become corrupted.
E.g. take the simplified example where the proximity range is 2 and the
register insertion sequence is 1, 4, 2, 3, 5.
* Insert of register 1 creates a new node, this is the root of the rbtree
* Insert of register 4 creates a new node, which is inserted to the right
of the root.
* Insert of register 2 gets inserted to the first node
* Insert of register 3 gets inserted to the first node
* Insert of register 5 also gets inserted into the first node since
this is the first node encountered and it is within the proximity range.
Now there are two overlapping nodes.
To avoid this always choose the node that is closest to the new register.
This will ensure that nodes will not overlap. The tree traversal is still
done as a binary search, we just don't stop at the first node found. So the
complexity of the algorithm stays within the same order.
Ideally if a new register is in the range of two adjacent blocks those
blocks should be merged, but that is a much more invasive change and left
for later.
The issue was initially introduced in commit 472fdec7380c ("regmap: rbtree:
Reduce number of nodes, take 2"), but became much more exposed by commit
6399aea629b0 ("regmap: rbtree: When adding a reg do a bsearch for target
node") which changed the order in which nodes are looked-up.
Fixes: 6399aea629b0 ("regmap: rbtree: When adding a reg do a bsearch for target node")
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@kernel.org>
|
|
and 'regmap/topic/seq' into regmap-next
|
|
Replace kmalloc with specialized function kmalloc_array when the size
is a multiplication of : number * size
Signed-off-by: lixiubo <lixiubo@cmss.chinamobile.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
|
|
Replace kzalloc with specialized function kcalloc when the size is
a multiplication of : number * sizeof
Signed-off-by: lixiubo <lixiubo@cmss.chinamobile.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
|
|
A binary search is much more efficient rather than iterating
over the rbtree in ascending order which the current code is
doing.
During initialisation the reg defaults are written to the
cache in a large chunk and these are always sorted in the
ascending order so for this situation ideally we should have
iterated the rbtree in descending order.
But at runtime the drivers may write into the cache in any
random order so this patch selects to use a bsearch to give
an optimal runtime performance and also at initialisation
time when reg defaults are written the performance of binary
search would be much better than iterating in ascending order
which the current code was doing.
Signed-off-by: Nikesh Oswal <Nikesh.Oswal@wolfsonmicro.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
|
|
When inserting a new register into a block, the present bit map size is
increased using krealloc. krealloc does not clear the additionally
allocated memory, leaving it filled with random values. Result is that
some registers are considered cached even though this is not the case.
Fix the problem by clearing the additionally allocated memory. Also, if
the bitmap size does not increase, do not reallocate the bitmap at all
to reduce overhead.
Fixes: 3f4ff561bc88 ("regmap: rbtree: Make cache_present bitmap per node")
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Mark Brown <broonie@kernel.org>
Cc: stable@vger.kernel.org
|
|
When inserting a new register into a block at the lower end the present
bitmap is currently shifted into the wrong direction. The effect of this is
that the bitmap becomes corrupted and registers which are present might be
reported as not present and vice versa.
Fix this by shifting left rather than right.
Fixes: 472fdec7380c("regmap: rbtree: Reduce number of nodes, take 2")
Reported-by: Daniel Baluta <daniel.baluta@gmail.com>
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@kernel.org>
Cc: stable@vger.kernel.org
|
|
If the inlcude headers aren't sorted alphabetically, then the
logical choice is to append new ones, however that creates a
lot of potential for conflicts or duplicates because every change
will then add new includes in the same location.
Signed-off-by: Xiubo Li <Li.Xiubo@freescale.com>
Signed-off-by: Mark Brown <broonie@kernel.org>
|
|
Commit 6cfec04bcc05 ("regmap: Separate regmap dev initialization") moved the
regmap debugfs initialization after regcache initialization. This means
that the regmap debugfs directory is not created yet when the cache
initialization runs and so any debugfs files registered by the regcache are
created in the debugfs root directory rather than the debugfs directory of
the regmap instance. Fix this by adding a separate callback for the
regcache debugfs initialization which will be called after the parent
debugfs entry has been created.
Fixes: 6cfec04bcc05 (regmap: Separate regmap dev initialization)
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@linaro.org>
Cc: stable@vger.kernel.org
|
|
Change regcache_rbtree_node strcuture fields order to align the pointers on
64bits architectures.
Signed-off-by: Jean-Christophe PINCE <jean-christophe.pince@intel.com>
Signed-off-by: David Cohen <david.a.cohen@linux.intel.com>
Signed-off-by: Mark Brown <broonie@linaro.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/broonie/regmap
Pull regmap updates from Mark Brown:
"A quiet release for regmap, some cleanups, fixes and:
- Improved node coalescing for rbtree, reducing memory usage and
improving performance during syncs.
- Support for registering multiple register patches.
- A quirk for handling interrupts that need to be clear when masked
in regmap-irq"
* tag 'regmap-v3.12' of git://git.kernel.org/pub/scm/linux/kernel/git/broonie/regmap:
regmap: rbtree: Make cache_present bitmap per node
regmap: rbtree: Reduce number of nodes, take 2
regmap: rbtree: Simplify adjacent node look-up
regmap: debugfs: Fix continued read from registers file
regcache-rbtree: Fix reg_stride != 1
regmap: Allow multiple patches to be registered
regmap: regcache: allow read-only regs to be cached
regmap: fix regcache_reg_present() for empty cache
regmap: core: allow a virtual range to cover its own data window
regmap: irq: document mask/wake_invert flags
regmap: irq: make flags bool and put them in a bitfield
regmap: irq: Allow to acknowledge masked interrupts during initialization
regmap: Provide __acquires/__releases annotations
|
|
With devices which have a dense and small register map but placed at a large
offset the global cache_present bitmap imposes a huge memory overhead. Making
the cache_present per rbtree node avoids the issue and easily reduces the memory
footprint by a factor of ten. For devices with a more sparse map or without a
large base register offset the memory usage might increase slightly by a few
bytes, but not significantly. E.g. for a device which has ~50 registers at
offset 0x4000 the memory footprint of the register cache goes down form 2496
bytes to 175 bytes.
Moving the bitmap to a per node basis means that the handling of the bitmap is
now cache implementation specific and can no longer be managed by the core. The
regcache_sync_block() function is extended by a additional parameter so that the
cache implementation can tell the core which registers in the block are set and
which are not. The parameter is optional and if NULL the core assumes that all
registers are set. The rbtree cache also needs to implement its own drop
callback instead of relying on the core to handle this.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@linaro.org>
|
|
Support for reducing the number of nodes and memory consumption of the rbtree
cache by allowing for small unused holes in the node's register cache block was
initially added in commit 0c7ed856 ("regmap: Cut down on the average # of nodes
in the rbtree cache"). But the commit had problems and so its effect was
reverted again in commit 4e67fb5 ("regmap: rbtree: Fix overlapping rbnodes.").
This patch brings the feature back of reducing the average number of nodes,
which will speedup node look-up, while at the same time also reducing the memory
usage of the rbtree cache. This patch takes a slightly different approach than
the original patch though. It modifies the adjacent node look-up to not only
consider nodes that are just one to the left or the right of the register but
any node that falls in a certain range around the register. The range is
calculated based on how much memory it would take to allocate a new node
compared to how much memory it takes adding a set of unused registers to an
existing node. E.g. if a node takes up 24 bytes and each register in a block
uses 1 byte the range will be from the register address - 24 to the register
address + 24. If we find a node that falls within this range it is cheaper or as
expensive to add the register to the existing node and have a couple of unused
registers in the node's cache compared to allocating a new node.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@linaro.org>
|
|
A register which is adjacent to a node will either be left to the first
register or right to the last register. It will not be within the node's range,
so there is no point in checking for each register cached by the node whether
the new register is next to it. It is sufficient to check whether the register
comes before the first register or after the last register of the node.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@linaro.org>
|
|
There are a couple of calculations, which convert between register addresses and
block indices, in regcache_rbtree_sync() and regcache_rbtree_node_alloc() which
assume that reg_stride is 1. This will break the rb cache for configurations
which do not use a reg_stride of 1.
Also rename 'base' in regcache_rbtree_sync() to 'start' to avoid confusion with
'base_reg'.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@linaro.org>
|
|
Avoid overlapping register regions by making the initial blklen of a new
node 1. If a register write occurs to a yet uncached register, that is
lower than but near an existing node's base_reg, a new node is created
and it's blklen is set to an arbitrary value (sizeof(*rbnode)). That may
cause this node to overlap with another node. Those nodes should be merged,
but this merge doesn't happen yet, so this patch at least makes the initial
blklen small enough to avoid hitting the wrong node, which may otherwise
lead to severe breakage.
Signed-off-by: David Jander <david@protonic.nl>
Signed-off-by: Mark Brown <broonie@linaro.org>
Cc: stable@vger.kernel.org
|
|
|
|
A node starting before the minimum register is no reason to reject it,
since its end could be in range. The check for the end already exists
two lines lower, so we can just remove the incorrect check.
Signed-off-by: Maarten ter Huurne <maarten@treewalker.org>
Signed-off-by: Mark Brown <broonie@linaro.org>
|
|
The parameter passed to the regmap lock/unlock callbacks needs to be
map->lock_arg, regcache passes just map. This works fine in the case that no
custom locking callbacks are used since in this case map->lock_arg equals map,
but will break when custom locking callbacks are used. The issue was introduced
in commit 0d4529c5("regmap: make lock/unlock functions customizable") and is
fixed by this patch.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
The parameter passed to the regmap lock/unlock callbacks needs to be
map->lock_arg, regcache passes just map. This works fine in the case that no
custom locking callbacks are used, since in this case map->lock_arg equals map,
but will break when custom locking callbacks are used. The issue was introduced
in commit 0d4529c5 ("regmap: make lock/unlock functions customizable") and is
fixed by this patch.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
If range information has been provided then when we allocate a rbnode
within a range allocate the entire range. The goal is to minimise the
number of reallocations done when combining or extending blocks. At
present only readability and yes_ranges are taken into account, this is
expected to cover most cases efficiently.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
In preparation for being slightly smarter about how we allocate memory
factor out the node allocation.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
Linux 3.9-rc7
|
|
The idea of holding blocks of registers in device format is shared between
at least rbtree and lzo cache formats so split out the loop that does the
sync from the rbtree code so optimisations on it can be reused.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
Reviewed-by: Dimitris Papastamos <dp@opensource.wolfsonmicro.com>
|
|
The idea of maintaining a bitmap of present registers is something that
can usefully be used by other cache types that maintain blocks of cached
registers so move the code out of the rbtree cache and into the generic
regcache code.
Refactor the interface slightly as we go to wrap the set bit and enlarge
bitmap operations (since we never do one without the other) and make it
more robust for reads of uncached registers by bounds checking before we
look at the bitmap.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
Reviewed-by: Dimitris Papastamos <dp@opensource.wolfsonmicro.com>
|
|
This will bring no meaningful benefit by itself, it is done as a separate
commit to aid bisection if there are problems with the following commits
adding support for coalescing adjacent writes.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
This patch aims to bring down the average number of nodes
in the rbtree cache and increase the average number of registers
per node. This should improve general lookup and traversal times.
This is achieved by setting the minimum size of a block within the
rbnode to the size of the rbnode itself. This will essentially
cache possibly non-existent registers so to combat this scenario,
we keep a separate bitmap in memory which keeps track of which register
exists. The memory overhead of this change is likely in the order of
~5-10%, possibly less depending on the register file layout. On my test
system with a bitmap of ~4300 bits and a relatively sparse register
layout, the memory requirements for the entire cache did not increase
(the cutting down of nodes which was about 50% of the original number
compensated the situation).
A second patch that can be built on top of this can look at the
ratio `sizeof(*rbnode) / map->cache_word_size' in order to suitably
adjust the block length of each block.
Signed-off-by: Dimitris Papastamos <dp@opensource.wolfsonmicro.com>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
The last register block, which falls into the specified range, is not handled
correctly. The formula which calculates the number of register which should be
synced is inverse (and off by one). E.g. if all registers in that block should
be synced only one is synced, and if only one should be synced all (but one) are
synced. To calculate the number of registers that need to be synced we need to
subtract the number of the first register in the block from the max register
number and add one. This patch updates the code accordingly.
The issue was introduced in commit ac8d91c ("regmap: Supply ranges to the sync
operations").
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
Cc: stable@vger.kernel.org
|
|
Provide a feel of how much overhead the rbtree cache adds to
the game.
[Slightly reworded output in debugfs -- broonie]
Signed-off-by: Dimitris Papastamos <dp@opensource.wolfsonmicro.com>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
It's more idiomatic to pass the map structure around and this means we
can use other bits of information from the map.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
If we're updating a value in place it's more work to read the value and
compare the value with what we're about to set than it is to just write
the value into the cache; there are no further operations after writing
in the code even though there's an early return here.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
regmap_config.reg_stride is introduced. All extant register addresses
are a multiple of this value. Users of serial-oriented regmap busses will
typically set this to 1. Users of the MMIO regmap bus will typically set
this based on the value size of their registers, in bytes, so 4 for a
32-bit register.
Throughout the regmap code, actual register addresses are used. Wherever
the register address is used to index some array of values, the address
is divided by the stride to determine the index, or vice-versa. Error-
checking is added to all entry-points for register address data to ensure
that register addresses actually satisfy the specified stride. The MMIO
bus ensures that the specified stride is large enough for the register
size.
Signed-off-by: Stephen Warren <swarren@nvidia.com>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
regmap-stride
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/broonie/regmap
Pull two more small regmap fixes from Mark Brown:
- Now we have users for it that aren't running Android it turns out
that regcache_sync_region() is much more useful to drivers if it's
exported for use by modules. Who knew?
- Make sure we don't divide by zero when doing debugfs dumps of
rbtrees, not visible up until now because everything was providing at
least some cache on startup.
* tag 'regmap-3.4-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/broonie/regmap:
regmap: prevent division by zero in rbtree_show
regmap: Export regcache_sync_region()
|
|
Some bus types have very fast IO. For these, acquiring a mutex for every
IO operation is a significant overhead. Allow busses to indicate their IO
is fast, and enhance regmap to use a spinlock for those busses.
[Currently limited to native endian registers -- broonie]
Signed-off-by: Stephen Warren <swarren@nvidia.com>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
If there are no nodes in the cache, nodes will be 0, so calculating
"registers / nodes" will cause division by zero.
Signed-off-by: Stephen Warren <swarren@nvidia.com>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
Cc: stable@vger.kernel.org
|
|
The code currently passes the register offset in the current block to
regcache_lookup_reg. This works fine as long as there is only one block and with
base register of 0, but in all other cases it will look-up the default for a
wrong register, which can cause unnecessary register writes. This patch fixes
it by passing the actual register number to regcache_lookup_reg.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
Cc: <stable@vger.kernel.org>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/paulg/linux
Pull <linux/device.h> avoidance patches from Paul Gortmaker:
"Nearly every subsystem has some kind of header with a proto like:
void foo(struct device *dev);
and yet there is no reason for most of these guys to care about the
sub fields within the device struct. This allows us to significantly
reduce the scope of headers including headers. For this instance, a
reduction of about 40% is achieved by replacing the include with the
simple fact that the device is some kind of a struct.
Unlike the much larger module.h cleanup, this one is simply two
commits. One to fix the implicit <linux/device.h> users, and then one
to delete the device.h includes from the linux/include/ dir wherever
possible."
* tag 'device-for-3.4' of git://git.kernel.org/pub/scm/linux/kernel/git/paulg/linux:
device.h: audit and cleanup users in main include dir
device.h: cleanup users outside of linux/include (C files)
|
|
into regmap-next
|
|
For files that are actively using linux/device.h, make sure
that they call it out. This will allow us to clean up some
of the implicit uses of linux/device.h within include/*
without introducing build regressions.
Yes, this was created by "cheating" -- i.e. the headers were
cleaned up, and then the fallout was found and fixed, and then
the two commits were reordered. This ensures we don't introduce
build regressions into the git history.
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
|
|
Otherwise we'll end up running with bogus register numbers.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
Most of the current users have register 0 as a volatile register or don't
have a register 0 so it's not been apparent that it's not getting synced.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
In order to allow us to support partial sync operations add minimum and
maximum register arguments to the sync operation and update the rbtree
and lzo caches to use this new information. The LZO implementation is
obviously not good, we could exit the iteration earlier, but there may
be room for more wide reaching optimisation there.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
The debugfs functions don't stub themselves out quite so well as might
be desirable so provide functions which do do this stubbing.
Reported-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
Show the register ranges we have in each rbtree node in debugfs, plus
some statistics on how big each node is and the total number of nodes.
It may also be worth collecting data on the ranges of dirty registers
to see if there's much mileage in trying to coalesce writes on sync.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
Calling regcache_exit from regcache_rbtree_init is first of all a layering
violation and secondly will cause double frees. regcache_exit will free buffers
allocated by the core, but the core will also free the same buffers when the
cacheops init callback returns an error. Thus we end up with a double free.
Fix this by not calling regcache_exit but only free those buffers which, have
been allocated in this function.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Acked-by: Dimitris Papastamos <dp@opensource.wolfsonmicro.com>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
|
Simplify the check for registers set at their default value by avoiding
picking a default value in the case where we don't have one. Instead we
only compare the current value to the current value when we looked one
up. This fixes the case where we don't have a default stored but the value
was set to zero when that isn't the chip default.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
Acked-by: Dimitris Papastamos <dp@opensource.wolfsonmicro.com>
|
|
Ensure that when we start up in cache only mode we can store defaults of
zero, otherwise if the hardware is unavailable we won't be able to read.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
Acked-by: Dimitris Papastamos <dp@opensource.wolfsonmicro.com>
|
|
If a register isn't cached then let callers know that so they can fall
back or error handle appropriately.
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
Acked-by: Dimitris Papastamos <dp@opensource.wolfsonmicro.com>
|
|
Signed-off-by: Dimitris Papastamos <dp@opensource.wolfsonmicro.com>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>
|
