NeilBrown [Wed, 12 Dec 2012 08:32:15 +0000 (19:32 +1100)]
mtd: block2mtd: throttle writes by calling balance_dirty_pages_ratelimited.
If you create a block2mtd device that is larger than main memory,
and write to all of it, then lots of pages will be dirtied but
they will never be flushed out as nothing calls any variant of
balance_dirty_pages.
It would be nice to call set_page_dirty_balance(), but that isn't exported,
so just call balance_dirty_pages_ratelimited() directly.
It could happen (1 out of 100 times) that NAND did not start up
correctly after warm rebooting, so the kernel could not find the UBI or
DMA timed out due to a stalled BCH. When resetting BCH together with
GPMI, the issue could not be observed anymore (after 10000+ reboots). We
probably need the consistent state already before sending any command to
NAND, even when no ECC is needed. I chose to keep the extra reset for
BCH when changing the flash layout to be on the safe side.
Mike Dunn [Fri, 7 Dec 2012 20:07:23 +0000 (12:07 -0800)]
mtd: nand/docg4: fix and improve read of factory bbt
This patch does two things related to reading the factory badblock table during
initialization: (1) fix error where a non-zero return code from
docg4_read_page() is assumed to be an error (it was later changed to be
max_bitflips; thanks to Brian Norris for bringing this to my attention a while
back), and (2) if there is an error reading the factory bbt, it tries reading
another (redundant) factory bbt table.
Signed-off-by: Mike Dunn <mikedunn@newsguy.com> Acked-by: Robert Jarzmik <robert.jarzmik@free.fr> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Mike Dunn [Fri, 7 Dec 2012 20:07:22 +0000 (12:07 -0800)]
mtd: nand/docg4: reserve bb marker area in ecclayout
Modify the nand_ecclayout to place the two bb marker bytes in the oob region
off-limits to the user.
Signed-off-by: Mike Dunn <mikedunn@newsguy.com> Acked-by: Robert Jarzmik <robert.jarzmik@free.fr> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Mike Dunn [Fri, 7 Dec 2012 20:07:21 +0000 (12:07 -0800)]
mtd: nand/docg4: add support for writing in reliable mode
The controller on the docg4 has a "reliable" mode, where consecutive 2k pages
are used in parallel. The initial program loader (IPL) on my Treo 680 expects
the secondary program loader (SPL) to be written in this mode. This patch adds
support for writing data in reliable mode, by way of a module parameter.
Support for reading in this mode (as the IPL does) is not supported yet, but
alternate (even-numbered) 2k pages written in reliable mode can be read normally
(odd-numbered pages will contain junk and generate ecc errors).
Signed-off-by: Mike Dunn <mikedunn@newsguy.com> Acked-by: Robert Jarzmik <robert.jarzmik@free.fr> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Lothar Waßmann [Thu, 6 Dec 2012 07:42:28 +0000 (08:42 +0100)]
mtd: mxc_nand: reorder part_probes to let cmdline override other sources
The cmdline is the easiest to change source of information. Thus
let it take precedence over 'RedBoot' and 'ofpart'. This makes the
mxc_nand driver to be in sync with all other NAND drivers that support
'cmdlinepart' partition parsing.
Also change 'const char *' to 'const char const *' as advised by checkpatch.pl
Lothar Waßmann [Thu, 6 Dec 2012 07:42:27 +0000 (08:42 +0100)]
mtd: mxc_nand: fix unbalanced clk_disable() in error path
If nand_scan_ident() or nand_scan_tail() fails, the NAND chip may have
been deselected and the clock already disabled. Thus, check 'clk_act'
in the error path to decide whether the clock still needs to be
disabled.
This fixes a:
|WARNING: at drivers/clk/clk.c:472 __clk_disable+0x3c/0x78()
Anton Prins [Tue, 27 Nov 2012 15:38:16 +0000 (16:38 +0100)]
mtd: physmap_of: error checking to prevent a NULL pointer dereference
This patch solves a NULL pointer dereference, this may occur if the tuple
is not mappable (jumps to continue in the for-loop). Out of the loop
possible results are:
- info->list_size == 0 if no of the tuples is mappable
- info->list_size == 1
- info->list_size > 1
If no one of the supplied tuples is mappable (info->list_size == 0) and
info->cmtd will not be set. But it is used in mtd_device_parse_register, OOPS!
actually it should generate an error in this case!
Signed-off-by: Anton Prins <anton.prins@nl.bosch.com> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Dan Carpenter [Thu, 29 Nov 2012 14:16:51 +0000 (17:16 +0300)]
mtg: docg3: potential divide by zero in doc_write_oob()
If we set oobdelta to zero then we will either return -EINVAL or hit
a divide (modulus) by zero on the next line when we check
"(ooblen % oobdelta)". It's better to just return -EINVAL here instead.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com> Acked-by: Robert Jarzmik <robert.jarzmik@free.fr> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Josh Wu [Tue, 27 Nov 2012 10:50:31 +0000 (18:50 +0800)]
mtd: at91: atmel_nand: return bit flips for the PMECC read_page()
This patch fix pmecc's read_page() to return maximum number of bitflips, 0 if uncorrectable.
In the commit: 3f91e94f7f511de74c0d2abe08672ccdbdd1961c ("mtd: nand: read_page() returns max_bitflips ()"),
The ecc.read_page() is changed to return the maximum number of bitflips.
And when meet uncorrectable bitflips it needs to return 0.
See the comment in nand.h:
* @read_page: function to read a page according to the ECC generator
* requirements; returns maximum number of bitflips corrected in
* any single ECC step, 0 if bitflips uncorrectable, -EIO hw error
Signed-off-by: Josh Wu <josh.wu@atmel.com> Reviewed-by: Mike Dunn <mikedunn@newsguy.com> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
mtd: fix recovery after failed write-buffer operation in cfi_cmdset_0002.c
When working on a problem with some flash chips that lock up during
write-buffer operations, I think there may be a bug in the linux
handling of chips using cfi_cmdset_0002.c.
The datasheets I have found for a number of these chips all specify that
when aborting a write-buffer command, it is not enough to use the
standard reset. Rather a "write-to-buffer-reset command" is needed.
This command is quite similar for all chips, the main variance seem to
be if the final 0xF0 can go to any address or must go to addr_unlock1.
The bug is then in the recovery handling when timing out at the end of
do_write_buffer, where using the normal reset command is not sufficient.
Without this change, if the write-buffer command fails then any
following operations on the flash also fail.
Signed-off-by: Harald Nordgard-Hansen <hhansen@pvv.org> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Matthieu CASTET [Tue, 6 Nov 2012 10:51:44 +0000 (11:51 +0100)]
mtd: nand: add NAND_BUSWIDTH_AUTO to autodetect bus width
The driver call nand_scan_ident in 8 bit mode, then
readid or onfi detection are done (and detect bus width).
The driver should update its bus width before calling nand_scan_tail.
This work because readid and onfi are read work 8 byte mode.
Note that nand_scan_ident send command (NAND_CMD_RESET, NAND_CMD_READID, NAND_CMD_PARAM), address and read data
The ONFI specificication is not very clear for x16 device if high byte of address should be driven to 0,
but according to [1] it should be ok to not drive it during autodetection.
[1]
3.3.2. Target Initialization
[...]
The Read ID and Read Parameter Page commands only use the lower 8-bits of the data bus.
The host shall not issue commands that use a word data width on x16 devices until the host
determines the device supports a 16-bit data bus width in the parameter page.
Initialising ecc.strength before the call to nand_scan() fixes this.
Signed-off-by: Nathan Williams <nathan@traverse.com.au> Cc: stable@vger.kernel.org [3.4+] Acked-by: Brian Norris <computersforpeace@gmail.com> Acked-by: Mike Dunn <mikedunn@newsguy.com> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Artem Bityutskiy [Thu, 22 Nov 2012 10:16:28 +0000 (12:16 +0200)]
mtd: fix a number of checkpatch complaints
While checking the "__devinit" removal patches with checkpatch.pl, I
noticed several warnings related to a space between the function name
and '(', as well as long lines. I fixed the warnings up in this patch.
Josh Radel [Wed, 14 Nov 2012 22:11:32 +0000 (14:11 -0800)]
mtd: ofpart: Replicate mtd cmdline "lk" option with device tree "lock" property
The mtd partition command line parser already supports a "lk" option to mask
MTD_POWERUP_LOCK. This extends that same functionality to device tree
partition specifications.
Rafał Miłecki [Mon, 12 Nov 2012 12:03:21 +0000 (13:03 +0100)]
mtd: prepare place for BCMA NAND flash driver(s)
BCMA bus can contain NAND flash memory, it's registered in system as
platform device. This adds required hooks and place for controler
specific drivers.
Jonas Gorski [Mon, 12 Nov 2012 09:52:49 +0000 (10:52 +0100)]
mtd: bcm63xxpart: make fixed part length calculation more generic
The CFE does not use 4K sectors even if the flash supports it, so
for the fixed partitions like CFE itself or NVRAM the erase block
size is always 64k or bigger.
Ezequiel Garcia [Sat, 10 Nov 2012 16:08:20 +0000 (13:08 -0300)]
mtd: mtd_blkdevs: Replace request handler kthread with a workqueue
By replacing a kthread with a workqueue, the code is now a bit clearer.
There's also a slight reduction of code size (numbers apply for x86):
Before:
text data bss dec hex filename
3248 36 0 3284 cd4 drivers/mtd/mtd_blkdevs.o
After:
text data bss dec hex filename
3150 36 0 3186 c72 drivers/mtd/mtd_blkdevs.o
Due to lack of real hardware, tests have been performed on an emulated
environment with mtdswap and mtdblock over nandsim devices.
Some real testing should be done, before merging this patch.
Ezequiel Garcia [Fri, 9 Nov 2012 15:36:35 +0000 (12:36 -0300)]
mtd: mtd_blkdev: Use a different name for block_device_operations variable
struct mtd_blktrans_ops is a type, and mtd_blktrans_ops is a variable.
To improve code clarity it's better to not use the same names,
so we just change the latter.
Huang Shijie [Fri, 9 Nov 2012 08:23:45 +0000 (16:23 +0800)]
mtd: de-select the chip when it is not used
When we scan several nand chips with nand_scan(), such as
.......................
nand_scan(*, 2);
.......................
In nand_scan_ident(), the maxchips will become 2, so the current code
will select chip 1 to read the device ID. But the chip 0 is still
selected in this case.
To make the logic clear, we'd better de-select the chip when it is not used.
This patch de-select the nand chip if it is not used any more.
Users of jffs2_do_reserve_space() expect they still held
erase_completion_lock after call to it. But there is a path
where jffs2_do_reserve_space() leaves erase_completion_lock unlocked.
The patch fixes it.
Found by Linux Driver Verification project (linuxtesting.org).
mtd: physmap_of: allow to specify the mtd name for retro compatiblity
linux,mtd-name allow to specify the mtd name for retro capability with
physmap-flash drivers as boot loader pass the mtd partition via the old
device name physmap-flash.
mtd: fsl_ifc_nand: wait tWB time, poll R/B before command execution
IFC_FIR_OP_CMD0 issues command for execution without checking flash
readiness. It may cause problem if flash is not ready. Instead use
IFC_FIR_OP_CW0 which Wait for tWB time and poll R/B to return high or
time-out, before issuing command.
NAND_CMD_READID command implemention does not fulfill above requirement. So
update its programming.
Arnd Bergmann [Tue, 6 Nov 2012 21:55:28 +0000 (22:55 +0100)]
mtd: uninitialized variable warning in map.h
The map_word_load() function initializes exactly
as many words in the buffer as required, but gcc
cannot figure this out and gives a misleading
warning. Marking the local variable as
uninitialized_var shuts up that warning.
Without this patch, building acs5k_defconfig results in:
drivers/mtd/chips/cfi_cmdset_0002.c: In function 'cfi_amdstd_panic_write':
include/linux/mtd/map.h:331:11: warning: 'r.x[0]' may be used uninitialized in this function [-Wuninitialized]
drivers/mtd/chips/cfi_cmdset_0002.c: In function 'cfi_amdstd_write_words':
include/linux/mtd/map.h:331:11: warning: 'r.x[0]' may be used uninitialized in this function [-Wuninitialized]
drivers/mtd/chips/cfi_cmdset_0001.c: In function 'cfi_intelext_write_words':
include/linux/mtd/map.h:331:11: warning: 'r.x[0]' may be used uninitialized in this function [-Wuninitialized]
Arnd Bergmann [Tue, 6 Nov 2012 21:55:27 +0000 (22:55 +0100)]
mtd: diskonchip: use inline functions for DocRead/DocWrite
The diskonchip drivers traditionally use home-grown macros for
doing MMIO accesses, which cause a lot of warnings, at least
on ARM machines:
drivers/mtd/devices/doc2000.c: In function 'doc_write':
drivers/mtd/devices/doc2000.c:854:5: warning: value computed is not used [-Wunused-value]
drivers/mtd/devices/doc2000.c: In function 'doc_erase':
drivers/mtd/devices/doc2000.c:1123:5: warning: value computed is not used [-Wunused-value
drivers/mtd/nand/diskonchip.c: In function 'doc2000_read_byte':
drivers/mtd/nand/diskonchip.c:318:3: warning: value computed is not used [-Wunused-value]
A nicer solution is to use the architecture-defined I/O accessors.
Here, we use the __raw_readl/__raw_writel style, instead of the
proper readl/writel ones, in order to preserve the odd semantics
of the existing macros that have their own barrier implementation
and no byte swap. It would be nice to fix this properly and use
the correct accessors as well as make the word size independent
from the architecture, but I guess the hardware is obsolete
enough that we should better not mess the driver an more than
necessary.
Arnd Bergmann [Tue, 6 Nov 2012 21:55:26 +0000 (22:55 +0100)]
mtd: diskonchip: don't warn about ARM architecture
Enabling the diskonchip drivers on most architectures results
in a pointless warning "#warning Unknown architecture for
DiskOnChip. No default probe locations defined". The driver
can in fact handle the default location already through the
CONFIG_MTD_DOCPROBE_ADDRESS, which gets set on the platforms
that need it, and we get a run-time error if this is not
set correctly.
Matthieu CASTET [Mon, 5 Nov 2012 14:00:44 +0000 (15:00 +0100)]
mtd: nand_wait: warn if the nand is busy on exit
This patch allow to detect buggy driver/hardware with
bad RnB (dev_ready) management or when timeout occurs in polling mode.
This works when dev_ready is set or not set.
There are 2 methods to wait for an erase/program command completion:
1. Wait until nand RnB pin goes high (that's what chip->dev_ready usually does)
2. Poll the device: send a status (0x70) command and read status byte in a loop
until bit NAND_STATUS_READY is set
In all cases, you should send a status command after completion, to check if
the operation was successful. And if the operation completed, the status should
have bit NAND_STATUS_READY set.
for a big nand chip Micron MT29F64G08AFAAAWP(8GB), we got the following wrong
result:
.............................................
"mtd: partition size too small (0)"
.............................................
We can not get any partition.
The "4g(test)" partition triggers a overflow of the "size". The memparse()
returns 4g to the "size", but the size is "unsigned long" type, so a overflow
occurs, the "size" becomes zero in the end.
This patch changes the "size"/"offset" to "unsigned long long" type,
and replaces the UINT_MAX with ULLONG_MAX for macros SIZE_REMAINING and
OFFSET_CONTINUOUS.
Viresh Kumar [Mon, 29 Oct 2012 17:17:26 +0000 (22:47 +0530)]
mtd: map: Fix compilation warning
This patch is an attempt to fix following compilation warning.
In file included from drivers/mtd/chips/cfi_cmdset_0001.c:35:0:
drivers/mtd/chips/cfi_cmdset_0001.c: In function 'cfi_intelext_write_words':
include/linux/mtd/map.h:331:11: warning: 'r.x[0]' may be used uninitialized in this function [-Wmaybe-uninitialized]
I could have used uninitialized_var() too, but didn't used it as the final else
part of map_word_load() is missing. So there is a chance that it might be passed
uninitialized. Better initialize to zero.
Matthieu CASTET [Fri, 19 Oct 2012 15:29:33 +0000 (17:29 +0200)]
mtdoops: don't erase flash at each boot
The current version on mtdoops erase first block of mtdoops partition at each
boot if there is no oops stored in flash. This can wear the flash.
When mtdoops start, find_next_position is called to find the next free entry in
the circular buffer. But if the flash is erased, find_next_position don't find
anything (maxcount == 0xffffffff) and start with the first entry after erasing it.
The scanning that is done in find_next_position already track free/used entries.
So if at the end of the scanning we don't find anything, we can start at the
first entry and erased the entry only if it is marked as used.
Most of this is implemented in mtdoops_inc_counter, so to avoid duplicating
code, if we don't find anything we set position to -1. mtdoops_inc_counter with
increment it, erase the entry if needed and start as before with nextpage = 0
and nextcount = 1).
Also during the scan phase, we use the MTDOOPS_KERNMSG_MAGIC to detect corruped
entries.
Bastian Hecht [Fri, 19 Oct 2012 10:15:36 +0000 (12:15 +0200)]
mtd: sh_flctl: Add device tree support
The flctl can now be probed via device tree setup in addition to the
existing platform data way.
SoC specific setup data is set in the .data member of the OF match, so
kept within the driver itself, while board/user specific setup - like
partitioning - is taken from the device tree.
Bastian Hecht [Fri, 19 Oct 2012 10:15:35 +0000 (12:15 +0200)]
mtd: sh_flctl: Add DMA capabilty
The code probes if DMA channels can get allocated and tears them down at
removal/failure if needed.
If available it uses them to transfer the data part (not ECC). On
failure we fall back to PIO mode.
Based on Guennadi Liakhovetski's code from the sh_mmcif driver.
Bastian Hecht [Fri, 19 Oct 2012 10:15:34 +0000 (12:15 +0200)]
mtd: sh_flctl: Minor cleanups
Some small fixes to avoid sparse and smatch complain. Other cosmetic fixes
as well.
- Change of the type of the member index in struct sh_flctl from signed
to unsigned. We use index by addressing array members, so unsigned is more
concise here. Adapt functions relying on sh_flctl::index.
- Remove a blurring cast in write_fiforeg().
- Apply consistent naming scheme when refering to the data buffer.
- Shorten some unnecessarily verbose functions.
- Remove spaces at start of lines.
Huang Shijie [Sat, 13 Oct 2012 17:03:29 +0000 (13:03 -0400)]
mtd: gpmi: remove unneccessary header
The whole gpmi-nand driver has turned to pure devicetree supported.
So the linux/mtd/gpmi-nand.h is not neccessary now. Just remove it,
and move some macros to the gpmi-nand driver itself.
Alexander Shiyan [Wed, 17 Oct 2012 06:08:27 +0000 (10:08 +0400)]
mtd: gpio-nand: Ability to use driver for configurations without RDY-pin
In some configurations of "gpio-nand" RDY-pin may be not connected.
This patch allow to use driver for these configurations. In this case
we are assume that device always ready.
Signed-off-by: Alexander Shiyan <shc_work@mail.ru> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Vikram Narayanan [Wed, 10 Oct 2012 17:42:02 +0000 (23:12 +0530)]
mtd: mtd_oobtest: printk -> pr_{info,err,crit}
Use pr_info() and pr_err() while defining pr_fmt(). This saves a few
characters, joins a few lines, and makes the code a little more readable
(and grep-able).
Vipin Kumar [Tue, 9 Oct 2012 10:44:53 +0000 (16:14 +0530)]
mtd: nand: Increase the ecc placement locations to 640
Few devices like H27UBG8T2CTR have a writesize/oobsize of 8KB/640B.
This means that the maximum oobsize has gone up to 640 bytes and consequently
the maximum ecc placement locations have also gone up to 640.
Vipin Kumar [Tue, 9 Oct 2012 10:44:50 +0000 (16:14 +0530)]
mtd: fsmc_nand: use relaxed variants of io accessors
Use relaxed variants of readl/writel accessors. readl/writel io accessors use
explicit dsb instruction which causes stalls in the processor core resulting
several cycles of delay for each access
Use relaxed variants where ever possible. This also results in an improved
read/write performance.
Vipin Kumar [Tue, 9 Oct 2012 10:44:48 +0000 (16:14 +0530)]
mtd: fsmc_nand: modify the wait to uninterruptible
Interruptible wait caused trouble in fsmc hardware state machine if the
application was killed abruptly. To make fsmc operation safe turn wait in to
un-interruptible.
Linus Walleij [Thu, 11 Oct 2012 22:23:02 +0000 (00:23 +0200)]
ARM: nomadik: fixup some FSMC merge problems
Due to a clash between refactoring and due to loss of a header
file that remained in my working tree the Nomadik stopped
compiling after switching to the FSMC driver. This patch fixes
it up.
Linus Walleij [Thu, 4 Oct 2012 07:28:29 +0000 (09:28 +0200)]
ARM: nomadik: switch over to using the FSMC driver
The Nomadik NAND driver is really just a subset of the existing
FSMC driver, so let's switch over to using that driver instead,
since it handles more variants of this chip. The callbacks for
setting up the chip is doing stuff now handled by the FSMC
driver.
mtd: denali: split the generic driver and PCI layer
The Denali controller can also be found in SoC devices attached to a
simple bus. Move the PCI specific parts into denali_pci so that we can
add a denali_dt that uses the same driver but for a device tree driver
instead of a PCI based device.
nand_chip.badblock_pattern will be dynamically calculated to the same
1-byte-length pattern that is required by fsl_elbc_nand. This custom
badblock_pattern is no longer needed, then, and its removal may help
facilitate further nand_bbt.c/nand_base.c cleanup in the future (one
down, many to go?)
Anyway, with nand_bbt.c fixed, this effectively reverts:
mtd: ofpart: Fix incorrect NULL check in parse_ofoldpart_partitions()
The pointer returned by kzalloc should be tested for NULL
to avoid potential NULL pointer dereference later. Incorrect
pointer was being tested for NULL. Bug introduced by commit fbcf62a3
(mtd: physmap_of: move parse_obsolete_partitions to become separate
parser).
This patch fixes this bug.
Simply 'parse_cmdline_partitions': the outer loop iterating over
'partitions' is actually a search loop, it does not execute the inner
loop for each partition, only for the matched partition.
Let's break when search is successful, and move all inner code (relevant
only for the matched partition) outside of the outer loop.
Resulting code is much more readable, and makes the indent level sane.
Some Samsung NAND with "5-byte" ID really appear to have 6-byte IDs, with
wraparound like:
Samsung K9K8G08U0D
ec d3 51 95 58 ec ec d3
Samsung K9F1G08U0C
ec f1 00 95 40 ec ec f1
Samsung K9F2G08U0B
ec da 10 95 44 00 ec da
This bad wraparound makes it hard to reliably detect the difference
between Samsung SLC with 5-byte ID and Samsung SLC with 6-byte ID.
The fix is to, for now, only use the new Samsung table for MLC. We
cannot support the new SLC (K9FAG08U0M) until Samsung gives better ID
decode information.
Note that this applies in addition to the previous regression fix:
Together, these patches completely restore the previous detection
behavior so that we cannot see any more regressions in Samsung SLC NAND
(finger crossed). With luck, I can get a hold of a Samsung
representative and stop having to cross my fingers eventually.
A combination of the following two commits caused a regression in 3.7-rc1
when identifying some Samsung NAND, so that some previously working NAND
were no longer detected properly:
Particularly, a regression was seen on Samsung K9F2G08U0B, with the
following full 8-byte READ ID string:
ec da 10 95 44 00 ec da
The basic problem is that Samsung manufactures both SLC and MLC NAND
that use a non-standard decoding table for deriving information from
their IDs. I have heuristically determined that all the chips that use
the new table have ID strings which wrap around after the 6th byte.
Unfortunately, I overlooked the fact that some older Samsung SLC (which
use a different decoding table) have "5 byte ID strings" which also wrap
around after the 6th byte.
This patch re-introduces a distinction between these old and new Samsung
NAND by checking that the 6th byte is non-zero, allowing both old and
new Samsung NAND to be detected properly.
Signed-off-by: Brian Norris <computersforpeace@gmail.com> Tested-by: Brian Norris <computersforpeace@gmail.com> Reported-by: Marek Vasut <marex@denx.de> Tested-by: Marek Vasut <marex@denx.de> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Thomas Betker [Wed, 17 Oct 2012 20:59:30 +0000 (22:59 +0200)]
jffs2: Fix lock acquisition order bug in jffs2_write_begin
jffs2_write_begin() first acquires the page lock, then f->sem. This
causes an AB-BA deadlock with jffs2_garbage_collect_live(), which first
acquires f->sem, then the page lock:
We fix this by restructuring jffs2_write_begin() to take f->sem before
the page lock. However, we make sure that f->sem is not held when
calling jffs2_reserve_space(), as this is not permitted by the locking
rules.
The deadlock above was observed multiple times on an SoC with a dual
ARMv7 (Cortex-A9), running the long-term 3.4.11 kernel; it occurred
when using scp to copy files from a host system to the ARM target
system. The fix was heavily tested on the same target system.
mtd: onenand: Make flexonenand_set_boundary static
Fixes the following sparse warning:
drivers/mtd/onenand/onenand_base.c:3697:5: warning:
symbol 'flexonenand_set_boundary' was not declared. Should it be static?