this improves network fs speed by disabling stat for http entriely, and only enabling file stat for everything else.
this can be overriden in the config.
read still needs some work. opening a file should open a thread where data is read/written async.
this avoids the huge number of roundtrips per r/w.
eg, a read currently has to REST, open the data socket, send the file data, send ABOR to close and then close the data socket.
using the thread approach would just send the file over the already open data socket.
writes will also benefit from the above, just instead of REST it would be APPE.
seeks would need to do the above ABORT, close and re-open. however most reads in sphaira are not random access. so this isn't an issue.
fs path was changed to 255 in the past because that is the real path size that fs can handle.
however, this restriction does not apply to nfs, samba, http - and it is very easy to exceed this 255 path length.
to fix this, i have increased the path len to 1024. this allows fs to continue to work as the buffer is
big enough, but also gives network mounts enough space to work with.
Changelog:
- re-enable use in release build.
- remove ftpsrv and untitled from builtin ghdl options, as both packages are available in the appstore.
- add image viewer (png, jpg, bmp)
- add music player (bfstm, bfwav, mp3, wav, ogg)
- add idv3 tag parsing support for mp3.
- add "decyption" of GTA Vice City mp3.
- add usbdvd support for music playback and file browsing.
- add nsz export support (solid, block, ldm).
- add xcz export support (same as above).
- add nro fs proper mount support (romfs, nacp, icon).
- add program nca fs support.
- add bfsar fs support.
- re-write the usb protocol, still wip. replaces tinfoil protocol.
- all threads are now create with pre-emptive support with the proper affinity mask set.
- fix oob crash in libpulsar when a bfwav was opened that had more than 2 channels.
- bump yyjson version.
- bump usbhsfs version.
- disable nvjpg.
- add support for theme music of any supported playback type (bfstm, bfwav, mp3, wav, ogg).
- add support for setting background music.
- add async exit to blocking threads (download, nxlink, ftpsrv) to reduce exit time.
- add support for dumping to pc via usb.
- add null, deflate, zstd hash options, mainly used for benchmarking.
- add sidebar slider (currently unused).
- file_viwer can now be used with any filesystem.
- filebrowser will only ever stat file once. previously it would keep stat'ing until it succeeded.
- disabled themezer due to the api breaking and i am not willing to keep maintaining it.
- disable zlt handling in usbds as it's not needed for my api's because the size is always known.
- remove usbds enums and GetSpeed() as i pr'd it to libnx.
- added support for mounting nca's from any source, including files, memory, nsps, xcis etc.
- split the lru cache into it's own header as it's now used in multiple places (nsz, all mounted options).
- add support for fetching and decrypting es personalised tickets.
- fix es common ticket converting where i forgot to also convert the cert chain as well.
- remove the download default music option.
- improve performance of libpulsar when opening a bfsar by remove the large setvbuf option. instead, use the default 1k buffer and handle large buffers manually in sphaira by using a lru cache (todo: just write my own bfsar parser).
- during app init and exit, load times have been halved as i now load/exit async. timestamps have also been added to measure how long everything takes.
- download now async loads / exits the etag json file to improve init times.
- add custom zip io to dumper to support writing a zip to any dest (such as usb).
- dumper now returns a proper error if the transfer was cancelled by the user.
- fatfs mount now sets the timestamp for files.
- fatfs mount handles folders with the archive bit by reporting them as a file.
- ftpsrv config is async loaded to speed up load times.
- nxlink now tries attempt to connect/accept by handling blocking rather than just bailing out.
- added support for minini floats.
- thread_file_transfer now spawns 3 threads rather than 2, to have the middle thread be a optional processor (mainly used for compressing/decompressing).
- added spinner to progress box, taken from nvg demo.
- progress box disables sleep mode on init.
- add gamecard detection to game menu to detect a refresh.
- handle xci that have the key area prepended.
- change gamecard mount fs to use the xci mount code instead of native fs, that way we can see all the partitions rather than just secure.
- reformat the ghdl entries to show the timestamp first.
- support for exporting saves to pc via usb.
- zip fs now uses lru cache.
- fixed fs real path length not actually being 0x301, but instead 255. fixes#204
- file picker inherits from file browser now so there's a lot less duplicated code.
- file browser now saves the last highlighted file.
- fix bug in file browser where the new file path could be empty (ie not containing a /).
- added support for viewing qlaunch romfs.
- moved fs mount options to the top of the list (may revert).
from my notes:
this converts the passphrase array into an ascii string
well, turns out std::transform does not actually increase the std::string size when pushing the new ascii characters
basically, std::string (normally) has small string optimisation. For this, they have a small array which it uses instead of allocating memory.
std::transform casts the std::string itr to a void*. sooo it was basically manually writing to the class, which just so happened to be the SSO buffer. however after 32chars, it runs out...and starts overwriting the stack. so this was a stack overflow lmao
only reason i found this is because i changed some optimisation flags, so the compiler ended up inlining std::transform, which detected the buffer overflow. and then i realised how fucked it was lol.
really, these functions should be passed by && and using std::forward on assignment.
however, this means writing a lot of extra code for every single class, as well as explicitly calling move
in a lot of cases.
In the case of std::string, passing by value and calling std::move is the correct approach, especially if the
string is created as an rvalue, then it is only created once and moved into the dest.
Whereas with a const ref, the string is created and then copied into the dst, basically creating 2 copies.
The same thing happens std::function, and well any object.
However, accepting everything by value sucks if you call a constructor from within a constructor, as now you need to create 2 impls
that accept by value and the other by rvalue. All of this extra code to have a more efficent impl just isn't worth it when, going by
the benchmarks, makes no measurable difference (i count anything within >= 1ms as measurable).
