Without this patch we can get (many) kmem trace events
with call site at krealloc().
This happens because krealloc is calling __krealloc,
which performs the allocation through kmalloc_track_caller.
Since neither krealloc nor __krealloc are marked inline explicitly,
the caller can be traced as being krealloc, which clearly is not
the intended behavior.
This patch allows to get the real caller of krealloc, by creating
an always inlined function __do_krealloc, thus tracing the
call site accurately.
Acked-by: Christoph Lameter <cl@linux.com>
Cc: Glauber Costa <glommer@parallels.com>
Signed-off-by: Ezequiel Garcia <elezegarcia@gmail.com>
Signed-off-by: Pekka Enberg <penberg@kernel.org>
UBI was mistakingly using 'kfree()' instead of 'kmem_cache_free()' when
freeing "attach eraseblock" structures in vtbl.c. Thankfully, this happened
only when we were doing auto-format, so many systems were unaffected. However,
there are still many users affected.
It is strange, but the system did not crash and nothing bad happened when
the SLUB memory allocator was used. However, in case of SLOB we observed an
crash right away.
This problem was introduced in 2.6.39 by commit
"6c1e875 UBI: add slab cache for ubi_scan_leb objects"
A note for stable trees:
Because variable were renamed, this won't cleanly apply to older kernels.
Changing names like this should help:
1. ai -> si
2. aeb_slab_cache -> seb_slab_cache
3. new_aeb -> new_seb
Reported-by: Richard Genoud <richard.genoud@gmail.com>
Tested-by: Richard Genoud <richard.genoud@gmail.com>
Tested-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: stable@vger.kernel.org [v2.6.39+]
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
UBI currently prints a lot of information when it mounts a volume, which
bothers some people. Make it less chatty - print only important information
by default.
Get rid of 'dbg_msg()' macro completely.
Reported-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Without this patch, these PEB are not scrubbed until we put data in them.
Bitflip can accumulate latter and we can loose the EC header (but VID header
should be intact and allow to recover data).
Signed-off-by: Matthieu Castet <matthieu.castet@parrot.com>
Cc: stable@vger.kernel.org
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Currently UBI fails in autoresize when it is in R/O mode (e.g., because the
underlying MTD device is R/O). This patch fixes the issue - we just skip
autoresize and print a warning.
Reported-by: Pali Rohár <pali.rohar@gmail.com>
Cc: stable@vger.kernel.org
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
This patch provides a possibility to set the "maximum expected number of
bad blocks per 1024 blocks" (max_beb_per1024) for each mtd device using
the UBI_IOCATT ioctl.
Signed-off-by: Richard Genoud <richard.genoud@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
This patch provides the possibility to adjust the "maximum expected number of
bad blocks per 1024 blocks" (max_beb_per1024) for each mtd device.
The majority of NAND devices have their max_beb_per1024 equal to 20, but
sometimes it's more.
Now, we can adjust that via a kernel parameter:
ubi.mtd=<name|num|path>[,<vid_hdr_offs>[,max_beb_per1024]]
Signed-off-by: Richard Genoud <richard.genoud@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
max_beb_per1024 shouldn't be negative, and a 0 value will be treated as
the default value. For the upper bound, 768/1024 should be enough.
Signed-off-by: Richard Genoud <richard.genoud@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
This patch prepare the way for the addition of max_beb_per1024 module
parameter. There's no functional change.
Signed-off-by: Richard Genoud <richard.genoud@gmail.com>
Reviewed-by: Shmulik Ladkani <shmulik.ladkani@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
No functional changes here, just to prepare for next patch.
Signed-off-by: Richard Genoud <richard.genoud@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
UBI has changed the MTD_UBI_BEB_LIMIT semantics. It used to be a percent of
total amount of eraseblock in the partition, and now it is the maximum
amount of bad eraseblocks on the entire devise per 1024 eraseblocks. So not
only the units changed, but also the meaning.
Richard Genoud <richard.genoud@gmail.com> says:
"I found the board:
https://www.olimex.com/dev/sam9-L9260.html
and the nand datasheet:
http://www.rockbox.org/wiki/pub/Main/LyrePrototype/K9xxG08UXM.pdf
page 11, we can see that the max_bad_bebper1024 is 25 (100 for 4096)"
Thus, use "25" for sam9.
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
On NAND flash devices, UBI reserves some physical erase blocks (PEB) for
bad block handling. Today, the number of reserved PEB can only be set as a
percentage of the total number of PEB in each MTD partition. For example, for a
NAND flash with 128KiB PEB, 2 MTD partition of 20MiB (mtd0) and 100MiB (mtd1)
and 2% reserved PEB:
- the UBI device on mtd0 will have 2 PEB reserved
- the UBI device on mtd1 will have 16 PEB reserved
The problem with this behaviour is that NAND flash manufacturers give a
minimum number of valid block (NVB) during the endurance life of the
device, e.g.:
Parameter Symbol Min Max Unit Notes
--------------------------------------------------------------
Valid block number NVB 1004 1024 Blocks 1
From this number we can deduce the maximum number of bad PEB that a device will
contain during its endurance life: a 128MiB NAND flash (1024 PEB) will not have
less than 20 bad blocks during the flash endurance life.
But the manufacturer doesn't tell where those bad block will appear. He doesn't
say either if they will be equally disposed on the whole device (and I'm pretty
sure they won't). So, according to the datasheets, we should reserve the
maximum number of bad PEB for each UBI device (worst case scenario: 20 bad
blocks appears on the smallest MTD partition).
So this patch make UBI use the whole MTD device size to calculate the maximum
bad expected eraseblocks.
The Kconfig option is in per1024 blocks, thus it can have a default value of 20
which is *very* common for NAND devices.
Signed-off-by: Richard Genoud <richard.genoud@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
'mtd_get_device_size()' returns the size of the whole MTD device, that is the
mtd_info master size. This will be used by UBI to calculate the maximum number
of bad blocks (MBB) on a MTD device.
Artem: amended the patch a bit.
Signed-off-by: Richard Genoud <richard.genoud@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
'struct mtd_info' is not modified by 'mtd_is_partition()' so it can be marked
as "const".
Signed-off-by: Richard Genoud <richard.genoud@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
CONFIG_MTD_UBI_BEB_RESERVE has been removed and now we use
CONFIG_MTD_UBI_BEB_LIMIT instead.
Signed-off-by: Shmulik Ladkani <shmulik.ladkani@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
CONFIG_MTD_UBI_BEB_RESERVE and MIN_RESEVED_PEBS are no longer used,
since the amount of reserved eraseblocks for bad PEB handling is now
derived from 'ubi->bad_peb_limit' (ubi's maximum expected bad
eraseblocks).
Signed-off-by: Shmulik Ladkani <shmulik.ladkani@gmail.com>
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@linux.intel.com>
The existing mechanism of reserving PEBs for bad PEB handling has two
flaws:
- It is calculated as a percentage of good PEBs instead of total PEBs.
- There's no limit on the amount of PEBs UBI reserves for future bad
eraseblock handling.
This patch changes the mechanism to overcome these flaws.
The desired level of PEBs reserved for bad PEB handling (beb_rsvd_level)
is set to the maximum expected bad eraseblocks (bad_peb_limit) minus the
existing number of bad eraseblocks (bad_peb_count).
The actual amount of PEBs reserved for bad PEB handling is usually set
to the desired level (but in some circumstances may be lower than the
desired level, e.g. when attaching to a device that has too few
available PEBs to satisfy the desired level).
In the case where the device has too many bad PEBs (above the expected
limit), then the desired level, and the actual amount of PEBs reserved
are set to zero. No PEBs will be set aside for future bad eraseblock
handling - even if some PEBs are made available (e.g. by shrinking a
volume).
If another PEB goes bad, and there are available PEBs, then the
eraseblock will be marked bad (consuming one available PEB). But if
there are no available PEBs, ubi will go into readonly mode.
Signed-off-by: Shmulik Ladkani <shmulik.ladkani@gmail.com>
Introduce 'ubi->bad_peb_limit', which specifies an upper limit of PEBs
UBI expects to go bad. Currently, it is initialized to a fixed percentage
of total PEBs in the UBI device (configurable via CONFIG_MTD_UBI_BEB_LIMIT).
The 'bad_peb_limit' is intended to be used for calculating the amount of PEBs
UBI needs to reserve for bad eraseblock handling.
Artem: minor amendments.
Signed-off-by: Shmulik Ladkani <shmulik.ladkani@gmail.com>
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@linux.intel.com>
We are going to kill the CONFIG_MTD_UBI_BEB_RESERVE configuration option soon
and use the CONFIG_MTD_UBI_BEB_LIMIT instead. In order to do this smoothly,
we now introduce the new configuration option to sam9_l9260_defconfig, and
will kill the old one after the corresponding UBI changes are done.
Signed-off-by: Shmulik Ladkani <shmulik.ladkani@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Use hash table to store ports of datapath. Allow 64K ports per switch.
Signed-off-by: Pravin B Shelar <pshelar@nicira.com>
Signed-off-by: Jesse Gross <jesse@nicira.com>
The vlan encapsulation fields in the maximum flow defintion were
never updated when the representation changed before upstreaming.
In theory this could cause a kernel panic when a maximum length
flow is used. In practice this has never happened (to my knowledge)
because skb allocations are padded out to a cache line so you would
need the right combination of flow and packet being sent to userspace.
Signed-off-by: Jesse Gross <jesse@nicira.com>
Add a "C5" cpuidle state to the SH7372 SoC connected to the A4S power
domain in such a way that A4S may be turned off by cpuidle if all
I/O devices in that domain have been suspended (or do not have
attached drivers).
This requires some reorganization of the initialization of SH7372
power management which affects the the boards based on it, Mackerel
and AP4EVB.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
The sh7372 cpuidle code uses the same artificially designed routine
shmobile_cpuidle_enter() as the .enter() callback for all of its
cpuidle states. However, shmobile_cpuidle_enter() calls a different
"enter" function for each state using an array of function pointers
populated by the sh7372 PM initialization code. Moreover, the
states[] array of the shmobile cpuidle driver is populated by that
code as well, although in principle it just might have been filled
with static data.
All of that complexity goes away if the sh7372 cpuidle code is
allowed to define its own cpuidle driver structure that can be passed
for registration to the common shmobile cpuidle initialization
routine, so modify the code accordingly.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
The role of the only function in the common.c file in
arch/arm/mach-shmobile, shmobile_init_late(), is to call two
initializers whose definitions depend on kernel configuration
options. Those initializers may very well be called from a static
inline function in arm/mach-shmobile/include/mach/common.h,
though, in which makes the code a bit easier to read. Moreover,
the common.c may be dropped entirely then.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
The !console_suspend_enabled check in sh7372_enter_suspend() seems
to be reversed and the condition it is supposed to catch (console
clock enabled) should be detected by the sh7372_sysc_valid() check
anyway, so remove it.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
Use the function rmobile_add_devices_to_domains() introduced
previously for adding devices to PM domains during the AP4EVB
initialization instead of a series of rmobile_add_device_to_domain*()
calls. This also causes the default device PM QoS latencies to be
used on that board in analogy with Mackerel.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
On SH7372 and Mackerel devices are added to PM domains through a
series of rmobile_add_device_to_domain_td() calls where the last
argument is always the same. This is quite inefficient, so add
a common function for adding devices to PM domains that reads the
domain-device pairs information from a table and use it during SH7372
and Mackerel initialization.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
The results of adaptive latency computations in
GENPD_DEV_TIMED_CALLBACK() show that the start/stop and save/restore
state latencies of all devices on the Mackerel board I have tried are
a little below 250 us. Therefore, if the 250 us is used as the
common initial value of the latency fields in struct gpd_timing_data
for all devices on Mackerel, the latency values will never have to
change at run time and there won't be any overhead related to
re-computation of the corresponding PM QoS data.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
The results of adaptive latency computations in
GENPD_DEV_TIMED_CALLBACK() show that the start/stop and save/restore
state latencies of all devices on SH7372 I have tried are a little
below 250 us. Therefore, if the 250 us is used as the common initial
value of the latency fields in struct gpd_timing_data for all devices
on SH7372, the latency values will never have to change at run time
and there won't be any overhead related to re-computation of the
corresponding PM QoS data.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
Make it possible to specify device start/stop and save/restore
state latencies directy when adding devices to PM domains. For
this purpose, introduce rmobile_add_device_to_domain_td() whose
third argument is a pointer to a struct gpd_timing_data object
containing device latency data.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
The results of adaptive latency computations in __pm_genpd_poweron()
and pm_genpd_poweroff() show that the power on/power off latencies
of all power domains in SH7372 are a little below 250 us. Therefore,
if 250 us is used as the common initial value of the latency fields
in struct generic_pm_domain for all domains, the latency values
will never have to change at run time and there won't be any overhead
related to re-computation of the corresponding PM QoS data.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
Since rmobile_init_pm_domain() is not called anywhere outside of
arch/arm/mach-shmobile/pm-rmobile.c any more, it can be made static
and its header may be removed from pm-rmobile.h. Modify the code
accordingly.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
Instead of giving a name to every r8a7779's PM domain object, put
them all into a table and initialize them all together in a loop.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
Instead of giving a name to every r8a7740's PM domain object, put
them all into a table and use rmobile_init_domains(), introduced by a
previous patch, for initializing them all altogether. Also, use
pm_genpd_add_subdomain_names() for adding A3SP as a subdomain of A4S.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
Instead of giving a name to every sh7372's PM domain object, put them
all into a table and use rmobile_init_domains(), introduced by a
previous patch, for initializing them all altogether. Also, use
pm_genpd_add_subdomain_names() for adding subdomains to the PM
domains and pm_genpd_poweron_name() for turning on the A4S domain
when preparing for system suspend.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
The sh7372_enter_suspend() routine checks the status field of the
generic PM domain object corresponding to the A4S domain in order to
check if it can turn that domain off when entering system sleep.
However, it shouldn't rely on the specific values of the generic
data structures this way, so make it use its own mechanism to
recognize when it is safe to turn that domain off.
For this purpos, introduce a boolean variable a4s_suspend_ready
that will be set by the A4S' suspend routine and unset by its
resume routine executed by rmobile_pd_power_down() and
__rmobile_pd_power_up(), respectively.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
Add a new routine, rmobile_init_domains(), allowing the caller
to initialize all generic PM objects stored in a table in one
operation.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
Make the power management code under arch/arm/mach-shmobile/
use pm_genpd_add_subdomain_names() for adding subdomains to power
domains, which makes it possible to drop
rmobile_pm_add_subdomain() and will allow us to carry out those
operations for domain objects stored in tables in a straightforward
way.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
If the r8a7779's PM domains are given names, this SoC and its boards
will be able to use rmobile_add_device_to_domain() for adding devices
to those domains and r8a7779_add_device_to_domain(), which is not
used anywhere at the moment anyway, may be dopped.
Accordingly, give names to the r8a7779's PM domains and drop
r8a7779_add_device_to_domain().
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
Make the power management code under arch/arm/mach-shmobile/ use
names of power domains instead of pointers to domain objects for
adding devices to the domains. This will allow us to put the
domain objects into tables and register them all in one shot
going forward.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Magnus Damm <damm@opensource.se>
This change initialises the cpu id field of cs_cpu_dbs_info structure in
conservative governor and keep this consistent with other governors.
Similar initialisation is present in ondemand governor.
Signed-off-by: Amit Daniel Kachhap <amit.daniel@samsung.com>
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Make it possible to use domain names in operations connecting cpuidle
to and disconnecting it from a PM domain. This is useful on
platforms where PM domain objects are organized in such a way that
the names of the domains are easier to use than the addresses of
those objects.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
