4 Algorithm

GNU ddrescue is not a derivative of dd, nor is related to dd in any way except in that both can be used for copying data from one device to another. The key difference is that ddrescue uses a sophisticated algorithm to copy data from failing drives causing them as little additional damage as possible.

Versions of ddrescue prior to 1.19 used a divide-and-conquer strategy to rescue the difficult parts of the drive. But that caused a lot of head movement, which is bad for the drive. Therefore, newer versions try to minimize head movement to minimize drive damage.

Ddrescue manages efficiently the status of the rescue in progress and tries to rescue the good parts first, scheduling reads inside bad (or slow) areas for later. This maximizes the amount of data that can be finally recovered from a failing drive.

The standard dd utility can be used to save data from a failing drive, but it reads the data sequentially, which may wear out the drive without rescuing anything if the errors are at the beginning of the drive.

Other programs read the data sequentially but switch to small size reads when they find errors. This is a bad idea because it means spending more time at error areas, damaging the surface, the heads, and the drive mechanics, instead of getting out of them as fast as possible. This behavior reduces the chances of rescuing the remaining good data.

The algorithm of ddrescue is divided in four phases: copying, trimming, scraping, and retrying. Each phase is described below. Disc sectors are marked successively as non-tried, non-trimmed, non-scraped, and bad-sector until they are successfully read and marked as finished. The user may interrupt the process at any point, but a bad drive can block ddrescue for a long time until the kernel gives up.

The steps of the algorithm are:

1) Optionally read a mapfile describing the status of a multi-part or previously interrupted rescue. If no mapfile is specified, or is empty, or does not exist, mark all the rescue domain as non-tried.

2) (First phase; Copying) Copying is done in up to five passes. The first pass reads the non-tried parts of the input file, marking the failed blocks as non-trimmed and skipping beyond them. The second pass runs in the opposite direction as the first pass and delimits the blocks skipped by the first pass. The first two passes also skip beyond slow areas. The areas skipped are tried later in one or three additional passes (before trimming). The copying direction is reversed after each pass until all the rescue domain is tried.

The third and fourth passes read the blocks skipped due to slow areas (if any) by the first two passes, in the same direction that each block was skipped. For each block, passes 2 to 4 skip the rest of the block after finding the first error in the block. The last pass is a sweeping pass, with skipping disabled. The purpose of the multiple passes is to delimit large bad areas fast, recover the most promising areas first, keep the mapfile small, and produce good starting points for trimming.

Only non-tried areas are read in large blocks. Trimming, scraping, and retrying are done sector by sector. Each sector is tried at most two times: the first in the copying phase as part of a large block read, the second in one of the phases below as a single sector read.

3) (Second phase; Trimming) Trimming retries sector by sector the edges of the large block reads failed during the copying phase. Trimming is done in one pass as follows. For each non-trimmed block, read forwards one sector at a time from the leading edge of the block until a bad sector is found. Then read backwards one sector at a time from the trailing edge of the block until a bad sector is found. Then mark the bad sectors found (if any) as bad-sector, and mark the rest of the block as non-scraped without trying to read it. If any edge is already adjacent to a bad sector, it is considered as already trimmed and is not trimmed again.

4) (Third phase; Scraping) Scrape together, sector by sector, the data not recovered by the copying or trimming phases. Scraping is done in one pass. Each non-scraped block is read forwards, one sector at a time. Any bad sectors found are marked as bad-sector.

5) (Fourth phase; Retrying) Optionally try to read again the bad sectors until the number of retry passes specified is reached. The direction is reversed after each pass. Every bad sector is tried only once in each pass. Ddrescue can’t know if a bad sector is unrecoverable or if it will be eventually read after some retries.

6) Optionally write a mapfile for later use.

When ddrescue finishes the steps above, any areas marked as bad-sector will remain untouched in the output file. If the output file is a regular file created by ddrescue, the areas marked as bad-sector will contain zeros. If it is a device or a previously existing file, the areas marked as bad-sector will still contain the data previously present there.

The mapfile is periodically saved to disc, as well as when ddrescue finishes or is interrupted. A backup copy of the mapfile with the extension ‘.bak’ is also periodically created (if possible). So in case of a crash you can resume the rescue with little recopying. The default interval between saves varies from 30 seconds to 5 minutes depending on mapfile size (larger mapfiles are saved at longer intervals), but may be overriden. See --mapfile-interval.

The same mapfile can be used for multiple commands that copy different areas of the input file, and for multiple recovery attempts over different subsets. See this example:

Rescue the most important part of the disc first.

ddrescue -i0 -s50MiB /dev/sdc hdimage mapfile
ddrescue -i0 -s1MiB -d -r3 /dev/sdc hdimage mapfile

Then rescue some key disc areas.

ddrescue -i30GiB -s10GiB /dev/sdc hdimage mapfile
ddrescue -i230GiB -s5GiB /dev/sdc hdimage mapfile

Now rescue the rest (does not recopy what is already done).

ddrescue /dev/sdc hdimage mapfile
ddrescue -d -r3 /dev/sdc hdimage mapfile