update tutorial with new way of saving resume data

This commit is contained in:
arvidn 2019-12-08 11:07:05 +01:00 committed by Arvid Norberg
parent 26ec7921c3
commit 41dab18644
1 changed files with 10 additions and 10 deletions

View File

@ -186,7 +186,7 @@ file(s) are downloaded. There are two approaches to doing this:
1. read every piece of the downloaded files from disk and compare it against its
expected hash.
2. save to disk the state of which pieces (and partial pieces) are downloaded,
2. save, to disk, the state of which pieces (and partial pieces) are downloaded,
and load it back in again when resuming.
If no resume data is provided with a torrent that's added, libtorrent will
@ -212,16 +212,14 @@ The save_resume_data_alert_ looks something like this:
{
virtual std::string message() const;
// points to the resume data.
std::shared_ptr<entry> resume_data;
// the resume data
add_torrent_params params;
};
``resume_data`` points to an entry_ object. This represents a node or a tree of
nodes in a bencoded_ structure, which is the native encoding scheme in
bittorrent. It can be encoded into a byte buffer or file using `bencode()`_.
When adding a torrent with resume data, set the `add_torrent_params::resume_data`_
to contain the bencoded buffer of the resume data.
The ``params`` field is an add_torrent_params_ object containing all the state
to add the torrent back to the session again. This object can be serialized
using `write_resume_data()`_ or `write_resume_data_buf()`_, and de-serialized
with `read_resume_data()`_.
example
-------
@ -304,5 +302,7 @@ __ https://blog.libtorrent.org/2015/03/bdecode-parsers/
.. _`add_torrent_params::resume_data`: reference-Core.html#resume_data
.. _`bdecode()`: reference-Bdecoding.html#bdecode()
.. _bdecode_node: reference-Bdecoding.html#bdecode-node
.. _`write_resume_data()`: http://libtorrent.org/reference-Core.html#write_resume_data()
.. _`write_resume_data_buf()`: http://libtorrent.org/reference-Core.html#write_resume_data_buf()
.. _`read_resume_data()`: reference-Core.html#read_resume_data()