| Age | Commit message (Collapse) | Author |
|---|
|
In 54d68799, mtd->eb_cnt was enforced to be '1' if MTD_NO_ERASE is set.
This was done with the aim of preventing divisions by zero.
However, even if MTD_NO_ERASE is set, mtd->eb_size (eraseblock size) can
still be set to a non-zero value which would not cause a division by
zero.
Instead, enforcing an eraseblock count of '1' here even leads to
inconsistent eraseblock counting in mtd-utils and lets for example a
'flash_erase' on an mtdnand device fail:
| # flash_erase /dev/mtd0 0 0
| Erasing 32768 Kibyte @ 0 -- 0 % complete libmtd: error!: bad eraseblock number 255, mtd0 has 1 eraseblocks
| flash_erase: error!: /dev/mtd0: MTD Erase entire chip failureTrying one by one each sector.
| error 22 (Invalid argument)
| Erasing 128 Kibyte @ 0 -- 0 % complete libmtd: error!: bad eraseblock number 1, mtd0 has 1 eraseblocks
| flash_erase: error!: /dev/mtd0: MTD get bad block failed
| error 22 (Invalid argument)
Also mtdinfo would look inconsistent (eraseblock size vs amount):
| # mtdinfo /dev/mtd0
| mtd0
| Name: mtdram test device
| Type: ram
| Eraseblock size: 131072 bytes, 128.0 KiB
| Amount of eraseblocks: 1 (33554432 bytes, 32.0 MiB)
| Minimum input/output unit size: 1 byte
| Sub-page size: 1 byte
| Character device major/minor: 90:0
| Bad blocks are allowed: false
| Device is writable: true
Fix this by enforcing mtd->eb_cnt to be '1' only when mtd->eb_size is
actually zero and would lead to a division by zero otherwise.
Fixes: 54d68799 ("libmtd: avoid divide by zero")
Signed-off-by: Enrico Jorns <ejo@pengutronix.de>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Some flash types support full erase chip command which can reduce the
flash erase time. Try first to erase the entire flash and fall back
to the old method if the operation fails.
Signed-off-by: Larisa Ileana Grigore <larisa.grigore@nxp.com>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
This commit removes the C header files from the EXTRA_DIST variables
and instead assigns them to the SOURCE variable of the respective
components they belong to.
This takes care of having them distributed in the release tar ball and
helps with dependency tracking a little.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
The concept of erase blocks doesn't apply to mtd-ram devices. Such
devices set MTD_NO_ERASE to indicate this and some report 0 for the
erase block size. Avoid a divide by zero when calculating the erase
block count for such devices.
Signed-off-by: Chris Packham <chris.packham@alliedtelesis.co.nz>
Acked-by: Richard Weinberger <richard@nod.at>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
This commit fixes some uses of strncpy that could leave the destination
buffer unterminated.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
The function reads file data into a buffer and then checks if we
actually are at the end-of-file by trying to read one more byte.
For whatever reason, the code uses an int instead of a char. It's
not pretty but works. But again, this is something that every
static analysis tool barks at.
Further more, the error messages are inverted. "We aren't at EOF yet"
is printed on failure and something like "read error %m" is printed
on success.
This patch fixes all of the above.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
This patch modifies the internal helpers to read and parse integers
from sysfs files by initializing them first and removes turns an
obscure "a = open(...) if (a >= 0) {...} if (a == -1) {...}" inside
recv_image into a more straight forward if/else branch.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
This patch restructures various code parts that follow the pattern
of "stat(x, &sb) ... makes_sense(&sb) ... open(x)".
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Fix warnings abot PRIdoff_t in libmtd.c, in mtd_read (and mtd_write):
In file included from ../git/lib/libmtd.c:40:0:
../git/lib/libmtd.c: In function 'mtd_read':
../git/include/common.h:110:18: warning: format '%ld' expects argument of
type 'long int', but argument 5 has type 'off_t {aka long long int}'
[-Wformat=]
../git/include/common.h:120:2: note: in expansion of macro 'errmsg'
errmsg(fmt, ##__VA_ARGS__); \
^~~~~~
../git/lib/libmtd.c:1082:10: note: in expansion of macro 'sys_errmsg'
return sys_errmsg("cannot seek mtd%d to offset %"PRIdoff_t,
^~~~~~~~~~
/usr/lib/klibc/include/inttypes.h:28:17: note: format string is defined here
#define PRId32 "d"
Signed-off-by: Thorsten Glaser <tg@mirbsd.org>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
The libmtd library tries to obtain the available OOB size via the sysfs
with a fallback to the ECCGETLAYOUT ioctl. For some devices (e.g. plat-ram),
the fallback path is always taken and prints an error message to stderr
since the ioctl fails.
This patch fixes the fallback path by suppressing the error message if
errno is set to EOPNOTSUPP (i.e. the device simply doesn't support that).
Fixes: a10353584f93 ("libmtd: Add support to access OOB available size")
Reported-by: Chris Packham <Chris.Packham@alliedtelesis.co.nz>
Reviewed-by: Xiaolei Li <xiaolei.li@mediatek.com>
Tested-by: Chris Packham <Chris.Packham@alliedtelesis.co.nz>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Now that we have per-UBI volume flags (for instance for skipping CRC
check when opening it) from the Linux header, let's add it to the
ubi_mkvol_request in libubi and assign the flags to ubi_mkvol_req from
the Linux header from ubi_mkvol.
Suggested-by: Boris Brezillon <boris.brezillon@bootlin.com>
Reviewed-by: Boris Brezillon <boris.brezillon@bootlin.com>
Signed-off-by: Quentin Schulz <quentin.schulz@bootlin.com>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
This reverts commit dede98ffb706676309488d7cc660f569548d5930.
The original commit tried to fix a descrepancy between the implementation
and the documentation by making the implementation comply.
When making the change, it was overlooked, that ubinfo and ubirename were
written against the implementation instead of the behaviour specified by
the documentation. So were further internal functions like
ubi_get_vol_info1_nm which further breaks ubirmvol.
A report with an outline of a resulting problem can be read on
the mailing list:
http://lists.infradead.org/pipermail/linux-mtd/2018-June/081562.html
From the report:
steps to reproduce: have mtd-utils 2.0.1 or 2.0.2
0. make a bunch of ubi volumes in sequential order
ubimkvol /dev/ubi0 -s 64KiB -N test1
ubimkvol /dev/ubi0 -s 64KiB -N test2
ubimkvol /dev/ubi0 -s 64KiB -N test3
ubimkvol /dev/ubi0 -s 64KiB -N test4
..
1. delete the test1 volume, making a hole in the volume table
ubirmvol /dev/ubi0 -N test1
2. try an affected tool (i.e. "ubirmvol /dev/ubi0 -N test4" )
|root at mr24:/# ubirmvol /dev/ubi0 -N test4
|ubirmvol: error!: cannot find UBI volume "test4"
| error 19 (No such device)
or "ubinfo -a"
| root at mr24:/# ubinfo -a
| UBI version: 1
| Count of UBI devices: 1
| UBI control device major/minor: 10:59
| Present UBI devices: ubi0
|
| ubi0
| Volumes count: 11
| Logical eraseblock size: 15872 bytes, 15.5 KiB
| Total amount of logical eraseblocks: 1952 (30982144 bytes, 29.5 MiB)
| Amount of available logical eraseblocks: 75 (1190400 bytes, 1.1 MiB)
| Maximum count of volumes 92
| Count of bad physical eraseblocks: 0
| Count of reserved physical eraseblocks: 40
| Current maximum erase counter value: 984
| Minimum input/output unit size: 512 bytes
| Character device major/minor: 251:0
| ubinfo: error!: libubi failed to probe volume 5 on ubi0
| error 19 (No such device)
| Present volumes: 0, 1, 2, 3, 4root at mr24:/#
Reported-by: Christian Lamparter <chunkeey@gmail.com>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Signed-off-by: Xiaolei Li <xiaolei.li@mediatek.com>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
This patch exposes OOB available size to user. Then user can use
OOB free area according to OOB available size.
Steps to get OOB available size:
First, access /sys/class/mtd/mtdX/oobavail. If not exist, then
try to get ecc layout by ioctl "ECCGETLAYOUT". If none of them
work, set OOB available size to 0.
Signed-off-by: Xiaolei Li <xiaolei.li@mediatek.com>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
We use floating point just to print out KiB, MiB, GiB.
Avoid that to be klibc friendly.
Fixes compilation for aarch64 against klibc:
error: '-mgeneral-regs-only' is incompatible with floating-point argument
| printf("%s%.1f GiB", p, (double)bytes / (1024 * 1024 * 1024));
etc.
Signed-off-by: Andrea Adami <andrea.adami@gmail.com>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Signed-off-by: Rock Lee <rockdotlee@gmail.com>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Use casts to void instead. Clang generates warnings about that by
default.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
This patch eliminates warnings generated by the -Wmissing-prototypes
option. With this flag set, we are now forced to have prototypes for
all global, exported functions, that have to be made visible to the
definitions and we are forced to mark all local functions as static.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
AC_CHECK_HEADERS already makes sure our config header contains a
HAVE_$FOO_H macro if a header was found. There is no need to
awkwardly set our own Automake conditionals and check for it all
over the place in the Automake files.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Fixes:
| LD ubi-utils/ubiformat
| .../git/ubi-utils/libiniparser.a(libiniparser.o): In function
| ` LD ubi-utils/ubirename
| iniparser_getdouble':
| .../git/ubi-utils/libiniparser.c:336: undefined reference to `atof'
Grep doesn't reveal any occurrence of iniparser_getdouble(), using atof() so
remove it: floating-point is not supported in klibc
Upstream-Status: Pending
Signed-off-by: Andrea Adami <andrea.adami@gmail.com>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
First issue is that ioctl() in klibc doesn't expect a constant as arg3.
Second issue is that arg3 in klibc ioctl() implementation is not optional.
Fixes:
| ubi-utils/libubi.c: In function 'do_attach':
| ubi-utils/libubi.c:698:8: warning: passing argument 3 of 'ioctl' discards
| 'const' qualifier from pointer target type
| ret = ioctl(fd, UBI_IOCATT, r);
| ^
| In file included from ubi-utils/libubi.c:32:0:
| .../lib/klibc/include/sys/ioctl.h:15:14: note: expected 'void *' but argument
| is of type 'const struct ubi_attach_req *'
| __extern int ioctl(int, int, void *);
| ^
| ubi-utils/libubi.c: In function 'ubi_vol_block_create':
| ubi-utils/libubi.c:1118:9: error: too few arguments to function 'ioctl'
| return ioctl(fd, UBI_IOCVOLCRBLK);
| ^
| In file included from ubi-utils/libubi.c:32:0:
| .../lib/klibc/include/sys/ioctl.h:15:14: note: declared here
| __extern int ioctl(int, int, void *);
| ^
| ubi-utils/libubi.c: In function 'ubi_vol_block_remove':
| ubi-utils/libubi.c:1123:9: error: too few arguments to function 'ioctl'
| return ioctl(fd, UBI_IOCVOLRMBLK);
| ^
| In file included from ubi-utils/libubi.c:32:0:
| .../usr/lib/klibc/include/sys/ioctl.h:15:14: note: declared here
| __extern int ioctl(int, int, void *);
| ^
Signed-off-by: Andrea Adami <andrea.adami@gmail.com>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
If the specified UBI device or volume does not exist, the function
is supposed to set errno to ENODEV.
This patch adds a check to ubi_get_vol_info1 to change the errno
to ENODEV if vol_get_major cannot access the underlying sysfs file,
so the function propperly returns that the device or volume does
not exist, instead of failing with errno set to ENOENT.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
The documentation of libmtd_open says, if it returns NULL and errno is
zero, MTD is not present. However, the current version always returns
a libmtd_t object. The function internally checks, if it can access the
MTD sysfs files and, if not, sets a flag to use the procfs fallback.
This patch adds an additional check to libmtd_open, to test if the
MTD procfs file can be read and fails with errno cleared if it does
not exist.
Furhtermore, mtd_get_info is documented to fail with errno set to ENODEV
if MTD is not present. First of all, this was broken in the original
version. It was implemented to specification for the sysfs code path,
but if MTD is not present, that won't be executed, because of the flag
set by libmtd_open. This makes the check not only redundant, but masks
an actual error (the sysfs paths suddenly not being readable anymore).
The legacy path that was used if the sysfs files are not avaible fails
with ENOENT if it cannot read the procfs file. With the above changes
in addition, we don't have a libmtd_t object if neither sysfs nor
procfs is readable, so this error status no longer makes sense.
This patch removes the documentation on the ENODEV errno, and
makes sure that mtd_get_info always returns with apropriate errno
on failure.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
UDEV_SETTLE_HACK addresses a problem which does no longer exist on Linux.
These days we have devtmpfs. New devices will automatically created on
the kernel side and user space has no longer to wait for udev.
As udev has a hard dependency on devtmpfs we can depend on it too.
People which don't use udev nor plain devtmpfs are anyways on their own.
Android, I'm looking at you...
Signed-off-by: Richard Weinberger <richard@nod.at>
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
Historically, the mtd-utils and ubi-utils were seperate packages. The
ubi-utils were at some point merged into the mtd-utils. They first
appeared in the release tar-ball in version 1.1.0 in their own
sub-hirarchy with their own buildsystem, readme, documentation, etc.
A lot of the duplicated stuff got centralized/removed over time.
This patch further cleans up the directory hirarchy duplication by
moving common libraries from the ubi-utils/ into the central lib/
and include/ directories in the top directory of the mtd-utils package.
This includes:
- libuib.a & libubigen.a used by the ubi utilities
- libscan.a currently only used by ubiformat
- libiniparser.a used by ubinize
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
This patch moves the remaining 3 functions from ubiutils-common.{c,h}
into libmtd common.{c,h}.
The functions are only generic utility functions that other mtd-utils
programs may also find usefull and every program that uses libubi links
against libmtd anyway so there is no real reason for keeping around a
seperate ubiutils-common with only generic helper functions.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
|
|
In order to use test files, allow sysfs root
to be set during compile time
Signed-off-by: Daniel Walter <dwalter@sigma-star.at>
|
|
Uninitialized buffers lead to failing
unittests, since padding was not set to 0.
Additionally this stops valgrind from complaining as well.
Signed-off-by: Daniel Walter <dwalter@sigma-star.at>
|
|
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Signed-off-by: Richard Weinberger <richard@nod.at>
|
|
This patch adds a libmissing library to mtd-utils, containing
implementations of functionality found in glibc but typically
missing from embedded C libraries such as uclibc ot musl.
For now, the library only contains stub implementations of
the backtrace*() family of functions.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Signed-off-by: Richard Weinberger <richard@nod.at>
|
|
This patch is largely based on Richards original RFC.
The major differences to the RFC patch are:
- Add missing sumtools & mtdpart targets
- Fix name of mkfs.jffs2 target
- Add missing subdir-objects option for non-recursive make
- Move all automake options to configure.ac
- Add manpages to install target
- Make XATTR & LZO support configurable
- Install binaries to sbin directory like in the old build system
- Install flash_erase wrapper script
- Add files missing from distribution target
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Signed-off-by: Richard Weinberger <richard@nod.at>
|
|
The function _tries_ to support short reads but doesn't adjust the
pointer into the buffer. If a short read happens, we scrambles the
flash contents. Interrupted reads aren't handled. Short or
interrupted writes aren't handled at all. Either a write succeeds
writing the entire buffer or the function gives up.
During an attempt at fixing it, it was discovered, that no mtd-utils
program uses this function. Furthermore, its highly specific nature
makes it more of a "feature looking for use case".
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Signed-off-by: Richard Weinberger <richard@nod.at>
|
|
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Signed-off-by: Richard Weinberger <richard@nod.at>
|
|
The mtd_get_dev_info1 function reads (among other things) name and type
string into coresponding struct mtd_dev_info fields.
The struct mtd_dev_info has the string fields marked const, requiring
them to be cast to non-const version during initialization.
This cast was previously omitted from the dev_read_data calls,
triggering warnings during compilation.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Signed-off-by: Richard Weinberger <richard@nod.at>
|
|
This patch fixes the return status of the mtd_torture function
in libmtd.
The torture test function is currently only used by the ubiformat
utility to check if a block is bad after a write fails (blocks are
marked bad if the function returns an error status). However, the
way the function was written, it ALWAYS returns an error value
regardless of whether it failed or not.
Signed-off-by: David Oberhollenzer <david.oberhollenzer@sigma-star.at>
Signed-off-by: Richard Weinberger <richard@nod.at>
Reviewed-by: Boris Brezillon <boris.brezillon@free-electrons.com>
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
|
|
Assuming the read() call does not return zero and the result is less
than len, the current implementation will overwrite the data already
read in buf which doesn't seem correct.
With this patch, subsequent calls to read() within the loop will now no
longer overwrite the existing contents of buf.
Signed-off-by: Marcus Prebble <marcus.prebble@axis.com>
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
|
|
Signed-off-by: Baruch Siach <baruch@tkos.co.il>
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
|
|
Fixes compilation on hosts with the musl C library.
Also drops the unused u_short typedef.
Signed-off-by: Imre Kaloz <kaloz@openwrt.org>
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
|
|
ioctl(MEMGETREGIONINFO) has one input parameter (regionindex) and three
output parameters (info about the erase region). There are two problems
in mtdinfo/libmtd here:
1. mtdinfo.c doesn't initialize its region_info_user struct, instead
passing uninitialized data to mtd_regioninfo()
2. mtd_regioninfo() fails to utilize the 'regidx' parameter to fill out
the regionindex parameter properly, so the garbage from mtdinfo.c is
propagated to the ioctl()
This means that mtdinfo will continuously probe the same (possibly
out-of-range) erase region, instead of looping over the valid regions.
Let's fix this in the mtd_regioninfo() helper, and at the same time,
let's zero out the mtdinfo.c buffer, as an additional precaution to keep
from using uninitialized data.
Initial error report from Yang, when running "mtdinfo /dev/mtd0" on a
Cavium 6100 board:
root@CN61XX:~# mtdinfo /dev/mtd0
mtd0
Name: phys_mapped_flash
Type: nor
Eraseblock size: 65536 bytes, 64.0 KiB
Amount of eraseblocks: 128 (8388608 bytes, 8.0 MiB)
Minimum input/output unit size: 1 byte
Sub-page size: 1 byte
Additional erase regions: 0
Character device major/minor: 90:0
Bad blocks are allowed: false
Device is writable: true
libmtd: error!: MEMGETREGIONINFO ioctl failed for erase region 0
error 22 (Invalid argument)
Eraseblock region 0: info is unavailable
libmtd: error!: MEMGETREGIONINFO ioctl failed for erase region 1
error 22 (Invalid argument)
Eraseblock region 1: info is unavailable
Reported-by: Yang Wei <Wei.Yang@windriver.com>
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
|
|
On legacy systems, if "/proc/mtd" doesn't exist or gives a read error,
mtd_dev_present returns -1 (since it calls legacy_dev_present), contrary
to what's specified in the header file.
This causes checks like
if (mtd_dev_present(n)) {
...
}
to give false positives. Fix this by comparing the return value to 1.
Signed-off-by: Guido MartÃnez <guido@vanguardiasur.com.ar>
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
|
|
In the current code, the MTD_NANDFLASH stands for both the SLC and MLC.
In the kernel, the MTD_NANDFLASH only stands for the SLC now,
so in order to keep the logic unchanged, we should also check the MLC
NAND by MTD_MLCNANDFLASH.
Signed-off-by: Huang Shijie <b32955@freescale.com>
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
|
|
The MTD_MLCNANDFLASH case is missed in the current code.
This patch adds it.
Signed-off-by: Huang Shijie <b32955@freescale.com>
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
|
|
We should use the off_t type instead of off64_t or u_int32_t as its
length is controlled by the WITHOUT_LARGEFILE flag.
Signed-off-by: Richard Genoud <richard.genoud@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
|
|
On legacy kernels with ROM devices, we can get mtdinfo errors like:
libmtd: error!: cannot open "/dev/mtd4"
error 13 (Permission denied)
mtdinfo: error!: libmtd failed get MTD device 4 information
error 13 (Permission denied)
We don't need O_RDRW access for informational ioctls(), so make this
O_RDONLY.
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
|
|
If we don't check for the MTD before calling `legacy_get_dev_info1', we may
get errors like:
libmtd: MTD subsystem is old and does not support sysfs, so MTD character device nodes have to exist
libmtd: error!: "/dev/mtd2" is not a character device
mtdinfo: error!: libmtd failed get MTD device 2 information
error 22 (Invalid argument)
So reverse the order of these two checks.
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
|
|
ioctl(MEMWRITE) is implemented with memdup_user(), and so it allocates
kernel memory in contiguous regions. This limits its usefulness for large
amounts of data, since contiguous kernel memory can become scarce. I have
experienced "out of memory" problems with ubiformat, for instance, which
writes in eraseblock-sized regions:
...
ubiformat: flashing eraseblock 12 -- 72 % complete
ubiformat: page allocation failure.
order:8, mode:0xd0
Call Trace:
[<8043fa7c>] dump_stack+0x8/0x34
[<8008c940>] __alloc_pages_nodemask+0x408/0x618
[<800bd748>] cache_alloc_refill+0x400/0x730
[<800bdbbc>] __kmalloc+0x144/0x154
[<8009cae4>] memdup_user+0x24/0x94
[<802d04e4>] mtd_ioctl+0xba8/0xbd0
[<802d0544>] mtd_unlocked_ioctl+0x38/0x5c
[<800d43c0>] do_vfs_ioctl+0xa4/0x6e4
[<800d4a44>] sys_ioctl+0x44/0xa0
[<8000f95c>] stack_done+0x20/0x40
...
libmtd: error!: MEMWRITE ioctl failed for eraseblock 12 (mtd0)
error 12 (Cannot allocate memory)
ubiformat: error!: cannot write eraseblock 12
error 12 (Cannot allocate memory)
This error can be mitigated for now by only using ioctl(MEMWRITE) when we
need to write OOB data, since we can only do this in small transactions
anyway. Then, data-only transactions (like those originating from
ubiformat) can be carried out with write() calls.
This issue can also be solved within the kernel ioctl(), but either way,
this patch is still useful, since write() is more straightforward (and
efficient?) than ioctl() for data-only writes.
Signed-off-by: Brian Norris <computersforpeace@gmail.com>
Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
|
