| Age | Commit message (Collapse) | Author |
|---|
|
|
|
This doesn't really address 3325 directly but it is the result of research
done while hunting it down. First, this is a thread safety improvement for
canceled requests that have gone into resource wait state. I don't think
we've seen a failure there but there was a window. It also cleans the
resource wait queue earlier which lets us do less work and get requests
more quickly into llcorehttp by bypassing the resource wait state. With
this, I finally feel comfortable about rundown of requests.
|
|
A/B comparison with original code showed the newer issuing lower-priority
requests of the cache reader and some other minor changes. Brought them
into agreement (this is cargo-cult programming). Made the HTTP resource
semaphore an atomic int for rigorous correctness across threads. I
swear I'm going to tear down this code someday.
|
|
Dropped an argument during integration which made the total byte count read
lower than expected. Everything else is fine, however.
|
|
cleaned up LLStat and removed unnecessary includes
|
|
Isolate llcorehttp initialization into a utility class (LLAppCoreHttp)
that provides glue between app and library (sets up policies, handles
notifications). Introduce 'TextureFetchConcurrency' debug setting to
provide some field control when absolutely necessary.
|
|
Big delta was converting the new texture debugger support code
to the new library. Viewer manifest should probably get an eyeball
before release.
|
|
|
|
|
|
Seems to be working correctly. Not certain this is the fastest possible way
to provide a std::streambuf interface but it's visually acceptable.
|
|
log on exit.
With much trial-and-error, cleaned up the banner on the texture console and made everything
mostly fit. Added global cache read, cache write and resource wait count events to the
console display to show if cache is working. On clean exit, emit a log line to report
stats to log file (intended for automated tests, maybe):
LLTextureFetch::endThread: CacheReads: 2618, CacheWrites: 117, ResWaits: 0, TotalHTTPReq: 117
|
|
LLProxy support, HttpOptions starting to work, HTTP resource waiting fixed.
Non-LLThread-based threads need to do some registration or LLMutex locks taken out in these
threads will not work as expected (SH-3154). We'll get a better solution later, this fixes
some things for now. Tracing of operations now supported. Global and per-request (via
HttpOptions) tracing levels of [0..3]. The 2 and 3 levels use libcurl's VERBOSE mode
combined with CURLOPT_DEBUGFUNCTION to stream high levels of detail into the log. *Very*
laggy but useful. Simple GET request supported (no Range: header). Really just a
degenrate case of a ranged get but supplied an API anyway. Global option to use the
LLProxy interface to setup CURL handles for either socks5 or http proxy usage. This
isn't really the most encapsulated way to do this but a better solution will have to
come later. The wantHeaders and tracing options are now supported in HttpOptions giving
per-request controls. Big refactoring of the HTTP resource waiter in lltexturefetch.
What I was doing before wasn't correct. Instead, I'm implementing the resource wait
after the Semaphore model (though not using system semaphores). So instead of having
a sequence like: SEND_HTTP_REQ -> WAIT_HTTP_RESOURCE -> SEND_HTTP_REQ, we now
do WAIT_HTTP_RESOURCE -> WAIT_HTTP_RESOURCE2 (actual wait) -> SEND_HTTP_REQ. Works
well but the prioritized filling of the corehttp library needs some performance
work later.
|
|
surprised me. Added a retry queue similar to ready queue to the
policy object which is sorted by retry time. Currently do five
retries (after the initial try) delayed by .25, .5, 1, 2 and 5
seconds. Removed the retry logic from the lltexturefetch module.
Upped the waiting time in the unit test for the retries. People
won't like this but tough, need tests.
|
|
Went through all the code and tried to document lock and thread usage
in the module. There's a huge comment block introducing all of this
at the beginning and I believe it's correct (though not quite complete).
Keep it updated, people. Added a new state, WAIT_HTTP_RESOURCE, that's
sort of a side-state of SEND_HTTP_REQ. If we hit a high-water mark
for HTTP requests, the extra are shunted to the new state once. Once
levels fall to a low-water mark, we run through a wait list of UUIDs,
sort the valid ones by priority and release them for service. This
keeps the HTTP layer busy while leaving the active queue shallow enough
that requests can still be re-prioritzed cheaply. Priority model
changed. The new state uses the PRIORITY_LOW mask, the old users
of _LOW are now at PRIORITY_NORMAL and sleepers woken up after an
external event are kicked off at PRIORITY_HIGH. This combination
along with the new state should avoid priority inversion and keep
things running without resorting to an infinite pipeline. New
state displays as "HTW" with green text in the texture console.
Request cancelation and worker run-down should now be more
correct but this edge case may need more attention.
|
|
Identified and reacted to the priority inversion problem we
have in texturefetch. Includes the introduction of a priority_queue
for the requests that are ready. Start some parameterization in
anticipation of having policy_class everywhere. Removed _assert.h
which isn't really needed in indra codebase. Implemented async
setPriority request (which I hope I can get rid of eventually along
with all priorities in this library). Converted to using unsigned
int for priority rather than float. Implemented POST and did
groundwork for PUT.
|
|
|
|
|
|
This is the first functional viewer pass with the HTTP work of the texture fetch
code performed by the llcorehttp library. Not exactly a 'drop-in' replacement
but a work-alike with some changes (e.g. handler notification in consumer
thread versus responder notification in worker thread).
This also includes some temporary changes in the priority scheme to prevent
the kind of priority inversion found in VWR-28996. Scheme used here does
provide liveness if not optimal responsiveness or order-of-operation.
The llcorehttp library at this point is far from optimally performing.
Its worker thread is making relatively poor use of cycles it gets and
it doesn't idle or sleep intelligently yet. This early integration step
helps shake out the interfaces, implementation niceties will be covered
soon.
|
|
|
|
|
|
|
|
|
|
from cache
|
|
on and off.
|
|
|
|
|
|
|
|
HTTP when all objects loading are done.
|
|
As a viewer architect, I would like to understand how fast each of the components of the texture pipeline can run in isolation
|
|
read latency
|
|
occlusion queries from previous frame are still pending and perform texture decode work.
|
|
Viewer attempting to load precached images in file types that are not being used.)
|
|
Linux startup crash appears to be due to static/global C++ init of LLAtomic
types. The initializer with explicit value makes some runtime calls and it
looks like these assume, at least on Linux, that apr_initialize() has been
called. So move the static POST count to a member and provide accessors
and increment/decrement. Command queue was built on a pointer to a class
in anonymous namespace and that's not quite valid. Made it a nested
class (really a nested forward declaration) while keeping the derived
classes in anonymous.
|
|
|
|
representation transferring ownership, doing data aggregation
in a very pedantic way. That's just adding unneeded cost and
complication. Used the same objects to transport data as are
collecting it and everything got simpler, faster, easier to
read with fewer gotchas. Bit myself *again* doing the min/max/mean
merges but the unittests where there to pick me up again. Added
a per-region FPS metric while I was at it. This is much asked
for and there was a convenient place to sample the value.
|
|
The TextureFetch thread was still stalling out due to
a different path that determines whether there is work
or not in the thread (uses getPending()) and that had
to be harmonized with the changes to runCondition().
I'm not happy with this solution but a refactor of the
LLThread tree isn't in the cards right now.
|
|
In the viewer, the region's UUID is acquired very late and isn't generally
used as the canonical region identifier. The U64 region hash is a better
and more consistently used identifier so I'm moving over to using that as
the region key. Don't have a proper reserved invalid region hash which is
unfortunate, but then, so much is.
|
|
Using unpause() method in derived class rather than wake() in furthest base class
solved the stalling problem. I still think too many levels of the
LLTextureFetch hierarchy are keeping thread state, however. The
LLViewerRegion instance an agent enters doesn't necessarily have its
region_id yet, that only comes after the handshake, if any. So add
a few more metrics insertion points to propagate region into metrics.
Finally, try to launch a final metrics report when a quit is initiated.
|
|
After discussions, renamed 'QAModeMetricsSubmetrics' to 'QAModeMetrics'
and confirmed that LLCachedControl<> is the way to go. Moved the
resulting flag out of LLViewerAssetStats (where it didn't belong)
and it lives in both LLAppViewer and LLTextureFetch where it does
belong.
|
|
Removed declared but undefined interfaces from LLTextureFetch family.
Had inserted the cross-thread command processor into some of
LLTextureFetchWorker's processing which was unnatural and
probably wrong. Moved it to LLTextureFetch which turned out to
be far, far more natural. Better documentation on the asLLSD()
format. Refined LLSD stats merger logic and enhanced unit
tests to verify same.
|
|
Asset fetch requests wrapped to allow for measurements.
Asset fetch enqueues, dequeues and completion times recorded
to asset stats collector. Texture fetch operations (http
and udp) recorded to asset stats collector. Stats collector
time vallue switched from F32 to U64 which is the more common
type in the viewer. Cross-thread mechanism introduced to
communicate region changes and generate global statistics
messages. Facility to deliver metrics via Capabilities
sketched in but needs additional work. Documentation and
diagrams added.
|
|
back-out the back-out for this branch. yay.
|
|
Backing out this merge that I pushed (prematurely) to the wrong place.
|
|
|
|
|
|
EXT-7397: decoding textures get stuck in REQ status
|
|
|
|
and EXT-7397: decoding textures get stuck in REQ status
|
|
either direction) when HTTP Textures is enabled
Fixed baked textures downloading. Uploading will be fixed on the server side.
|
|
LLWorkerClass::checkWork: ASSERT(workreq).
|
