In the quest to remove all stack VLA usage from the kernel[1], this
allocates a fixed size stack array to cover the range needed for
bch. This was done instead of a preallocation on the SLAB due to
performance reasons, shown by Ivan Djelic:
little-endian, type sizes: int=4 long=8 longlong=8
cpu: Intel(R) Core(TM) i5 CPU 650 @ 3.20GHz
calibration: iter=4.9143µs niter=2034 nsamples=200 m=13 t=4
Buffer allocation | Encoding throughput (Mbit/s)
---------------------------------------------------
on-stack, VLA | 3988
on-stack, fixed | 4494
kmalloc | 1967
So this change actually improves performance too, it seems.
The resulting stack allocation can get rather large; without
CONFIG_BCH_CONST_PARAMS, it will allocate 4096 bytes, which
trips the stack size checking:
lib/bch.c: In function ‘encode_bch’:
lib/bch.c:261:1: warning: the frame size of 4432 bytes is larger than 2048 bytes [-Wframe-larger-than=]
Even the default case for "allmodconfig" (with CONFIG_BCH_CONST_M=14 and
CONFIG_BCH_CONST_T=4) would have started throwing a warning:
lib/bch.c: In function ‘encode_bch’:
lib/bch.c:261:1: warning: the frame size of 2288 bytes is larger than 2048 bytes [-Wframe-larger-than=]
But this is how large it's always been; it was just hidden from
the checker because it was a VLA. So the Makefile has been adjusted to
silence this warning for anything smaller than 4500 bytes, which should
provide room for normal cases, but still low enough to catch any future
pathological situations.
[1] https://lkml.kernel.org/r/CA+55aFzCG-zNmZwX4A2FQpadafLfEzK6CC=qPXydAacU1RqZWA@mail.gmail.com
Signed-off-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Ivan Djelic <ivan.djelic@parrot.com>
Tested-by: Ivan Djelic <ivan.djelic@parrot.com>
Acked-by: Boris Brezillon <boris.brezillon@bootlin.com>
Signed-off-by: Boris Brezillon <boris.brezillon@bootlin.com>
This is a new software BCH encoding/decoding library, similar to the shared
Reed-Solomon library.
Binary BCH (Bose-Chaudhuri-Hocquenghem) codes are widely used to correct
errors in NAND flash devices requiring more than 1-bit ecc correction; they
are generally better suited for NAND flash than RS codes because NAND bit
errors do not occur in bursts. Latest SLC NAND devices typically require at
least 4-bit ecc protection per 512 bytes block.
This library provides software encoding/decoding, but may also be used with
ASIC/SoC hardware BCH engines to perform error correction. It is being
currently used for this purpose on an OMAP3630 board (4bit/8bit HW BCH). It
has also been used to decode raw dumps of NAND devices with on-die BCH ecc
engines (e.g. Micron 4bit ecc SLC devices).
Latest NAND devices (including SLC) can exhibit high error rates (typically
a dozen or more bitflips per hour during stress tests); in order to
minimize the performance impact of error correction, this library
implements recently developed algorithms for fast polynomial root finding
(see bch.c header for details) instead of the traditional exhaustive Chien
root search; a few performance figures are provided below:
Platform: arm926ejs @ 468 MHz, 32 KiB icache, 16 KiB dcache
BCH ecc : 4-bit per 512 bytes
Encoding average throughput: 250 Mbits/s
Error correction time (compared with Chien search):
average worst average (Chien) worst (Chien)
----------------------------------------------------------
1 bit 8.5 µs 11 µs 200 µs 383 µs
2 bit 9.7 µs 12.5 µs 477 µs 728 µs
3 bit 18.1 µs 20.6 µs 758 µs 1010 µs
4 bit 19.5 µs 23 µs 1028 µs 1280 µs
In the above figures, "worst" is meant in terms of error pattern, not in
terms of cache miss / page faults effects (not taken into account here).
The library has been extensively tested on the following platforms: x86,
x86_64, arm926ejs, omap3630, qemu-ppc64, qemu-mips.
Signed-off-by: Ivan Djelic <ivan.djelic@parrot.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
