NAND/MTD support under Linux -...

NAND FlashNAND support in Linux

NAND/MTD support under Linux

Matthieu CASTET - www.parrot.com

12 July 2012

Matthieu CASTET - www.parrot.com NAND/MTD support under Linux

http://www.parrot.com



NAND Features

1 NAND Flash




NAND Features

Flash is everywhere

non-volatile computer storage chip that can be electricallyerased and reprogrammed

usb flash drivesmemory cardssolid-state drives




NAND Features

Flash is everywhere

smartphoneaudio playerdigital cameracar-kit...




NAND Features

Type of flash : NOR vs NAND

NOR NANDInterface memory Bus I/OCell Size Large SmallCell Cost High Low

Read Time Fast SlowProgram Time single Byte Fast SlowProgram Time multi Byte Slow Fast

Erase Time Slow FastPower consumption High LowCan execute code Yes No

Bit twiddling Yes 1-4 timesBad blocks at ship time No Allowed




NAND Features

1 NAND FlashNAND Features




NAND Features

BUS




NAND Features

Memory organisation




NAND Features

Memory organisation

read/write unit : pageerase unit : blocktypical block size

64 pages of 512 (32 KB)64 pages of 2,048 (128 KB)64 pages of 4,096 (256 KB)128 pages of 4,096 (512 KB)

erase : all block set to 0xFFprogram : switch bit from 1 to 0




NAND Features

SLC




NAND Features

MLC




NAND Features

Bad block

block can be bad at ship timeblock can become bad (Erase/Write failure)A marker is needed to detect them : bad block marker




NAND Features

Ecc

bitflip corruption can happen : Read/Write disturb, cell wearneed to protect data with error correcting code (ECC)done on host sidefor SLC

1 bit per 512 B (43 nm - 350 nm)4 bits per 512 B (32 nm)8 bits per 512 B (24 nm)12? bits per 512 B (19 nm)

for MLC4 bits per 512 B (43 nm)

need to do bit scrubbingbitflips can accumulate over time




NAND Features

Ecc Algorithm

Hamming1 bit correction2 bits error detectionrequires 24 bits for 512 B

Reed Solomonused in cdromexisting hardware IPwell suited to correct burstexample for ba315 controller

4 bits correction5 bits error detectionrequires 80 bits for 512 B




NAND Features

Ecc Algorithm

Hamming1 bit correction2 bits error detectionrequires 24 bits for 512 B

BCH (Bose, Ray-Chaudhuri and Hocquenghem)t bits correctionrequires t*13 bits for 512 Bextra parity bit needed for error detectionmore complex to compute than Hamming (crc like)




NAND Features

Spare area

Also called "out of band" (OOB) areaextra memory for each page (about 1/32 of page size)store BBMstore ECC checksumcan be used for some filesystem metadata




NAND Features

Spare area and ECC

512 B of dataBCH for 4 bits : 52 bitsBCH for 8 bits : 104 bitsBCH for 12 bits : 156 bitswe are out of 512/32 = 16 B = 128 bitsno space left for filesystem metadata




NAND Features

Write constraints

Number of operation (NOP) : 1-4 timescan’t rewrite ECC !

subpage : 4 * 512 Bspare area only modification

Program pages in a block sequentiallycan’t program pagei if pagen is programmed, i < ncan’t update marker in OOB area later

Minimize partial-page programming operations




NAND Features

Endurance

measured in program/erease cycle1M-100K NOR100k SLC NAND10K MLC NAND500-100 TLC NAND




NAND Features

Data retention

5-10 yearschosse between endurance / data retention

Retention required Block Cycles10 yr 10 cyc5 yr 10000 cyc1 yr 100000 cyc

for improving data retentionLimit program/eraseLimit read (read disturb)

photos won’t be readable for eternity on SDCard/SSD




NAND Features

ONFI

Open NAND Flash Interface Working Groupstandard physical interfacestandard command setmechanism for self-identificationnot followed by Samsung, Thoshiba, ... !




NAND Features

Booting on NAND

does not support execute in place (XIP)needed hardware/software loader

NAND controller with boot modeROM code

load 1 block (512 - 128 KB) in internal RAMsmall bootloader or 2 stages (x-load/u-boot)Block 0 is not bad

difficult to support future NANDidentification/geometryerror correction




NAND Features

Booting on NAND : example

s3c24124 KB of data without ECC

ba315512 B of data without ECC

omap3630ROM4 first blocks60 KB with hamming ECC32 KB with BCH ECC




MTDUBIFS

2 NAND support in Linux




MTDUBIFS

2 NAND support in LinuxMTDUBIFS




MTDUBIFS

Flash device vs block device 1/2

Block device Flash deviceConsists of sectors Consists of eraseblocksSectors are small (512, 1024 B) Eraseblocks are larger ( typically

128KB)Maintains 2 main operations:

read sector

write sector

Maintains 3 main operations:

read from eraseblock

write to eraseblock

erase eraseblock




MTDUBIFS

Flash device vs block device 2/2

Block device Flash deviceBad sectors are re-mapped andhidden by hardware (at least inmodern LBA hard drives)

Bad eraseblocks are not hiddenand should be dealt with in soft-ware

Sectors are devoid of the wear-outproperty

Eraseblocks wear-out and becomebad and unusable after about 103

(for MLC NAND) - 105 (NOR, SLCNAND) erase cycles

Flash device is more difficult to handle




MTDUBIFS

FTL devices

Flash Translation Layeremulation of block interface over flashfirmware is a black box (Good/Bad FTL)

robust to power failure ?Bad block management ?Wear leveling ok ?read disturb handling ?

optimized for FATabstraction over a standard interface

does not need to handle all aspects of NANDSSD : mix of SLC/MLC, I/O in parallel.

Patented




MTDUBIFS

FTL devices




MTDUBIFS

MTD

MTD stands for "Memory Technology Devices"Provides an abstraction of flash devicesHides many aspects specific to particular flashtechnologiesProvides uniform API to access various types of flashesE.g., MTD supports NAND, NOR, ECC-ed NOR,DataFlash, OneNAND, etcProvides partitioning capabilities




MTDUBIFS

MTD API

In-kernel API (struct mdt_device) and user-space API(/dev/mtd0)

Information (device size, min. I/O unit size, etc)Read from and write to eraseblocksErase an eraseblockMark an eraseblock as badCheck if an eraseblock is bad

Does not hide bad eraseblocksDoes not do any wear-leveling




MTDUBIFS

MTD NAND support

MTD started with NOR supportAPI was limited to 4GB flash (32 bits offset)first NAND controller were dummy/pio

new controllers are smarter and need a high level interfacedmawork at page/block level




MTDUBIFS

MTD evolution : NAND ECC

bitflips happen quicklyresults in new, previously untested cases/situationscorruption in bad block marker/bad block table

protect bad block table with ECCbe robust to bit-flip in bad block marker

corruption of fs data in OOB areaavoid using OOB area or protect data

empty page not all 0xFFuse ECC that works on empty page or empty marker




MTDUBIFS

MTD evolution : NAND ECC

bitflips are more frequentneed stronger (software) ECC algorithm

BCHfrequent bit scrubbing

bitflip_thresholdmtd report bitflip only if nb_bitflip >= bitflip_threshold




MTDUBIFS

MTD evolution : ONFI

some new flash devices do not support legacyidentification

add ONFI identificationsave complex heuristic

speed optimization : auto-detect NAND timings




MTDUBIFS

Smart NAND

support of new generation NAND is complexneed software modificationneed hardware modification (software BCH ECC is slow)

NAND with internal ECC managementstatus command report ECC statusoptional enable/disablefree or used OOB area




MTDUBIFS

2 NAND support in LinuxMTDUBIFS




MTDUBIFS

Linux flash filesystem

JFFS2Linux 2.4.10 (2001-09)RAM usageslow mountsummary

YAFFS/YAFFS2not mainlineuse OOB areaused by android

UBIFSLinux 2.6.27 (2008-10)next generation of the JFFS2




MTDUBIFS

UBI/UBIFS stack




MTDUBIFS

UBI

Stands for "Unsorted Block Images"Provides an abstraction of "UBI volume"Has kernel API (include/mtd/ubi-user.h) and user spaceAPI (/dev/ubi0)Provides wear-levelingHides bad eraseblocksAllows run-time volume creation, deletion, and re-sizeIs somewhat similar to LVM, but for MTD devices




MTDUBIFS

UBI volume vs MTD device

MTD device UBI deviceConsists of physical eraseblocks(PEB), typically 128 KB

Consists of logical eraseblocks (LEB),slightly smaller than PEB (e.g 126/124KB)

Has 3 main operations

read from PEB

write to PEB

erase PEB

Has 3 main operations

read from LEB

write to LEB

erase LEB

May have bad PEBs Does not have bad LEB (handle a cer-tain amount of bad PEB)

PEBs wear out LEBs do not wear out - UBI spread theI/O load across the whole flash

MTD devices are static UBI volumes are dynamic




MTDUBIFS

Main idea behind UBI

Maps LEBs to PEBsAny LEB may be mapped to any PEBEraseblock headers store mapping information and erasecount




MTDUBIFS

Other

Handle bit-flips by moving data to a different PEBConfigurable wear-leveling thresholdAtomic LEB changeVolume update/rename operationSuitable for MLC NANDPerforms operations in backgroundWorks on NAND, NOR and other flash typesTolerant to power cutsSimple and robust designeasy support in bootloader




MTDUBIFS

UBIFS relies on UBI

UBIFS does not care about bad eraseblocks and relies onUBIUBIFS does not care about wear-leveling and relies on UBIUBIFS exploits the atomic LEB change feature




MTDUBIFS

Requirements

Good scalabilityData structures are treesOnly journal has to be replayed

High performanceWrite-backBackground commitRead/write is allowed during commitMulti-head journal minimizes amount of GCTNC makes look-ups fastLPT caches make LEB searches fast




MTDUBIFS

Requirements

On-the-flight compressionPower-cut tolerance

All updates are out-of-placeWas extensively tested

High reliabilityAll data is protected by CRC32 checksumChecksum may be disabled for data

RecoverabilityAll nodes have a header which fully describes the nodeThe index may be fully re-constructed from the headersNo tool to rebuild it at the moment




MTDUBIFS

TODO

UBI scales linearlyRead all eraseblock headers on attach1 or 2 page per eraseblock

UBI fastmap patch in progress




MTDUBIFS

TODO

UBIFS do not handle MLCneed background flash scanning for bitflipspaired pages problem




MTDUBIFS

TODO

UBI/UBIFS do not handle unstable bitspower cut just before the end of page programpower cut just after the start of page erase

page can be read, but randomly fails after a few times

power cut just after the start of page programpower cut just before the end of page erase

not a full-erased page. Contains 0xFF, but data written tothis page randomly fails

need to track program/erase operation and restartinterrupted one




MTDUBIFS

Questions ?

Merci pour votre attention !Thanks for your attention!Questions?



NAND/MTD support under Linux -...

Documents

Transcript of NAND/MTD support under Linux -...