support for "owa", audio-only webm, was introduced in iOS 17.5
owa is a more compliant alternative to opus-caf from iOS 11,
which was technically limited to CBR opus, a limitation which
we ignored since it worked mostly fine for regular opus too
being the new officially-recommended way to do things,
we'll default to owa for iOS 18 and later, even though
iOS still has some bugs affecting our use specifically:
if a weba file is preloaded into a 2nd audio object,
safari will throw a spurious exception as playback is
initiated, even as the file is playing just fine
the `.ld` stuff is an attempt at catching and ignoring this
spurious error without eating any actual network exceptions
previously, the `?zip` url-suffix would create a cp437 zipfile,
and `?zip=utf` would use utf-8, which is now generally expected
now, both `?zip=utf` and `?zip` will produce a utf8 zipfile,
and `?zip=dos` provides the old behavior
fixes a bug reported on discord:
a sha512 checksum does not cleanly encode to base64, and the
padding runs afoul of the safety-check added in 988a7223f4
as there is not a single reason to use a filekey that long,
fix it by setting an upper limit (which is still ridiculous)
if an untrusted x-forwarded-for is received, then disable
some features which assume the client-ip to be correct:
* listing dotfiles recently uploaded from own ip
* listing ongoing uploads from own ip
* unpost recently uploaded files
this is in addition to the existing vivid warning in
the serverlogs, which empirically is possible to miss
may improve upload performance in some particular uncommon scenarios,
for example if hdd-writes are uncached, and/or the hdd is drastically
slower than the network throughput
one particular usecase where nosparse *might* improve performance
is when the upload destination is cloud-storage provided by FUSE
(for example an s3 bucket) but this is educated guesswork
try to decode some malicious xml on startup; if this succeeds,
then force-disable all xml-based features (primarily WebDAV)
this is paranoid future-proofing against unanticipated changes
in future versions of python, specifically if the importlib or
xml.etree.ET behavior changes in a way that somehow reenables
entity expansion, which (still hypothetically) would probably
be caused by failing to unload the `_elementtree` c-module
no past or present python versions are affected by this change
when loading up2k snaps, entries are forgotten if
the relevant file has been deleted since last run
when the entry is an unfinished upload, the file that should
be asserted is the .PARTIAL, and not the placeholder / final
filename (which, unintentionally, was the case until now)
if .PARTIAL is missing but the placeholder still exists,
the only safe alternative is to forget/disown the file,
since its state is obviously wrong and unknown
also includes a slight tweak to the json upload info:
when exactly one file is uploaded, the json-response has a
new top-level property, `fileurl` -- this is just a copy of
`files[0].url` as a workaround for castdrian/ishare#107
("only toplevel json properties can be referenced")
when hashing files on android-chrome, read a contiguous range of
several chunks at a time, ensuring each read is at least 48 MiB
and then slice that cache into the correct chunksizes for hashing
especially on GrapheneOS Vanadium (where webworkers are forbidden),
improves worst-case speed (filesize <= 256 MiB) from 13 to 139 MiB/s
48M was chosen wrt RAM usage (48*4 MiB); a target read-size of
16M would have given 76 MiB/s, 32M = 117 MiB/s, and 64M = 154 MiB/s
additionally, on all platforms (not just android and/or chrome),
allow async hashing of <= 3 chunks in parallel on main-thread
when chunksize <= 48 MiB, and <= 2 at <= 96 MiB; this gives
decent speeds approaching that of webworkers (around 50%)
this is a new take on c06d928bb5
which was removed in 184af0c603
when a chrome-beta temporarily fixed the poor file-read performance
(afaict the fix was reverted and never made it to chrome stable)
as for why any of this is necessary,
the security features in android have the unfortunate side-effect
of making file-reads from web-browsers extremely expensive;
this is especially noticeable in android-chrome, where
file-hashing is painfully slow, around 9 MiB/s worst-case
this is due to a fixed-time overhead for each read operation;
reading 1 MiB takes 60 msec, while reading 16 MiB takes 112 msec
fixes a bug reported on discord;
1. run with `--idp-h-usr=iu -v=srv::A`
2. upload a file with up2k; this succeeds
3. announce an idp user: `curl -Hiu:a 127.1:3923`
4. upload another file; fails with "fs-reload"
the idp announce would `up2k.reload` which raises the
`reload_flag` and `rescan_cond`, but there is nothing
listening on `rescan_cond` because `have_e2d` was false
must assume e2d if idp is enabled, because `have_e2d` will
only be true if there are non-idp volumes with e2d enabled
