docs: fix various markdownlint issues

This commit is contained in:
albertony
2025-11-01 16:47:09 +01:00
parent f361cdf1cb
commit 7672c3d586
14 changed files with 292 additions and 225 deletions

View File

@@ -707,9 +707,9 @@ the same in the new encrypted remote.
This will have the following advantages
* `rclone sync` will check the checksums while copying
* you can use `rclone check` between the encrypted remotes
* you don't decrypt and encrypt unnecessarily
- `rclone sync` will check the checksums while copying
- you can use `rclone check` between the encrypted remotes
- you don't decrypt and encrypt unnecessarily
For example, let's say you have your original remote at `remote:` with
the encrypted version at `eremote:` with path `remote:crypt`. You
@@ -719,11 +719,15 @@ as `eremote:`.
To sync the two remotes you would do
rclone sync --interactive remote:crypt remote2:crypt
```console
rclone sync --interactive remote:crypt remote2:crypt
```
And to check the integrity you would do
rclone check remote:crypt remote2:crypt
```console
rclone check remote:crypt remote2:crypt
```
## File formats
@@ -734,8 +738,8 @@ has a header and is divided into chunks.
#### Header
* 8 bytes magic string `RCLONE\x00\x00`
* 24 bytes Nonce (IV)
- 8 bytes magic string `RCLONE\x00\x00`
- 24 bytes Nonce (IV)
The initial nonce is generated from the operating systems crypto
strong random number generator. The nonce is incremented for each
@@ -753,8 +757,8 @@ authenticate messages.
Each chunk contains:
* 16 Bytes of Poly1305 authenticator
* 1 - 65536 bytes XSalsa20 encrypted data
- 16 Bytes of Poly1305 authenticator
- 1 - 65536 bytes XSalsa20 encrypted data
64k chunk size was chosen as the best performing chunk size (the
authenticator takes too much time below this and the performance drops
@@ -767,15 +771,15 @@ This uses a 32 byte (256 bit key) key derived from the user password.
1 byte file will encrypt to
* 32 bytes header
* 17 bytes data chunk
- 32 bytes header
- 17 bytes data chunk
49 bytes total
1 MiB (1048576 bytes) file will encrypt to
* 32 bytes header
* 16 chunks of 65568 bytes
- 32 bytes header
- 16 chunks of 65568 bytes
1049120 bytes total (a 0.05% overhead). This is the overhead for big
files.
@@ -798,8 +802,8 @@ it on the cloud storage system.
This means that
* filenames with the same name will encrypt the same
* filenames which start the same won't have a common prefix
- filenames with the same name will encrypt the same
- filenames which start the same won't have a common prefix
This uses a 32 byte key (256 bits) and a 16 byte (128 bits) IV both of
which are derived from the user password.
@@ -808,8 +812,8 @@ After encryption they are written out using a modified version of
standard `base32` encoding as described in RFC4648. The standard
encoding is modified in two ways:
* it becomes lower case (no-one likes upper case filenames!)
* we strip the padding character `=`
- it becomes lower case (no-one likes upper case filenames!)
- we strip the padding character `=`
`base32` is used rather than the more efficient `base64` so rclone can be
used on case insensitive remotes (e.g. Windows, Box, Dropbox, Onedrive etc).
@@ -825,6 +829,7 @@ then rclone uses an internal one.
encrypted data. For full protection against this you should always use
a salt.
## SEE ALSO
## See Also
* [rclone cryptdecode](/commands/rclone_cryptdecode/) - Show forward/reverse mapping of encrypted filenames
- [rclone cryptdecode](/commands/rclone_cryptdecode/) - Show forward/reverse
mapping of encrypted filenames.