From 329608d24668b044e16b54ff7a7d0ac592b2b88d Mon Sep 17 00:00:00 2001 From: Monty Brandenberg Date: Mon, 22 Sep 2014 18:49:45 -0400 Subject: Tuning and documentation. Use a fast poll frequency (0.05S) on the HTTP requests for inventory. We'll benchmark with that and see how it goes. Document some of the history of the background fetcher for future devs. Suggest some future projects to make things faster. Pointers on using LLSD with the llcorehttp library in the readme. And restructured the LLSD onCompleted() processing phases using do{}while(false) which produced a code flow that is fairly attractive. --- indra/newview/llinventorymodelbackgroundfetch.cpp | 140 +++++++++++++++------- 1 file changed, 94 insertions(+), 46 deletions(-) (limited to 'indra/newview/llinventorymodelbackgroundfetch.cpp') diff --git a/indra/newview/llinventorymodelbackgroundfetch.cpp b/indra/newview/llinventorymodelbackgroundfetch.cpp index de1d123fe5..f18832fe95 100755 --- a/indra/newview/llinventorymodelbackgroundfetch.cpp +++ b/indra/newview/llinventorymodelbackgroundfetch.cpp @@ -42,6 +42,56 @@ #include "bufferstream.h" #include "llcorehttputil.h" +// History (may be apocryphal) +// +// Around V2, an HTTP inventory download mechanism was added +// along with inventory LINK items referencing other inventory +// items. As part of this, at login, the entire inventory +// structure is downloaded 'in the background' using the +// backgroundFetch()/bulkFetch() methods. The UDP path can +// still be used and is found in the 'DEPRECATED OLD CODE' +// section. +// +// The old UDP path implemented a throttle that adapted +// itself during running. The mechanism survived info HTTP +// somewhat but was pinned to poll the HTTP plumbing at +// 0.5S intervals. The reasons for this particular value +// have been lost. It's possible to switch between UDP +// and HTTP while this is happening but there may be +// surprises in what happens in that case. +// +// Conversion to llcorehttp reduced the number of connections +// used but batches more data and queues more requests (but +// doesn't due pipelining due to libcurl restrictions). The +// poll interval above was re-examined and reduced to get +// inventory into the viewer more quickly. +// +// Possible future work: +// +// * Don't download the entire heirarchy in one go (which +// might have been how V1 worked). Implications for +// links (which may not have a valid target) and search +// which would then be missing data. +// +// * Review the download rate throttling. Slow then fast? +// Detect bandwidth usage and speed up when it drops? +// +// * A lot of calls to notifyObservers(). It looks like +// these could be collapsed by maintaining a 'dirty' +// bit and there appears to be an attempt to do this. +// But it isn't used or is used in a limited fashion. +// Are there semanic issues requiring a call after certain +// updateItem() calls? +// +// * An error on a fetch could be due to one item in the batch. +// If the batch were broken up, perhaps more of the inventory +// would download. (Handwave here, not certain this is an +// issue in practice.) +// +// * Conversion to AISv3. +// + + namespace { @@ -488,11 +538,11 @@ void LLInventoryModelBackgroundFetch::bulkFetch() // a fast/slow fetch throttle. Once login is complete and the scene // is mostly loaded, we could turn up the throttle and fill missing // inventory more quickly. + static const U32 max_batch_size(10); static const S32 max_concurrent_fetches(12); // Outstanding requests, not connections static const F32 new_min_time(0.05f); // *HACK: Clean this up when old code goes away entirely. - static const U32 max_batch_size(10); - mMinTimeBetweenFetches = 0.01f; + mMinTimeBetweenFetches = new_min_time; if (mMinTimeBetweenFetches < new_min_time) { mMinTimeBetweenFetches = new_min_time; // *HACK: See above. @@ -702,63 +752,61 @@ namespace void BGFolderHttpHandler::onCompleted(LLCore::HttpHandle handle, LLCore::HttpResponse * response) { - // Single-pass do-while used for common exit handling - do + do // Single-pass do-while used for common exit handling { LLCore::HttpStatus status(response->getStatus()); // status = LLCore::HttpStatus(404); // Dev tool to force error handling if (! status) { processFailure(status, response); + break; // Goto common exit } - else - { - // Response body should be present. - LLCore::BufferArray * body(response->getBody()); - // body = NULL; // Dev tool to force error handling - if (! body || ! body->size()) - { - LL_WARNS(LOG_INV) << "Missing data in inventory folder query." << LL_ENDL; - processFailure("HTTP response missing expected body", response); - break; // Goto common exit - } - // Could test 'Content-Type' header but probably unreliable. + // Response body should be present. + LLCore::BufferArray * body(response->getBody()); + // body = NULL; // Dev tool to force error handling + if (! body || ! body->size()) + { + LL_WARNS(LOG_INV) << "Missing data in inventory folder query." << LL_ENDL; + processFailure("HTTP response missing expected body", response); + break; // Goto common exit + } - // Convert response to LLSD - // body->write(0, "Garbage Response", 16); // Dev tool to force error handling - LLSD body_llsd; - if (! LLCoreHttpUtil::responseToLLSD(response, true, body_llsd)) - { - // INFOS-level logging will occur on the parsed failure - processFailure("HTTP response contained malformed LLSD", response); - break; // goto common exit - } + // Could test 'Content-Type' header but probably unreliable. - // Expect top-level structure to be a map - // body_llsd = LLSD::emptyArray(); // Dev tool to force error handling - if (! body_llsd.isMap()) - { - processFailure("LLSD response not a map", response); - break; // goto common exit - } + // Convert response to LLSD + // body->write(0, "Garbage Response", 16); // Dev tool to force error handling + LLSD body_llsd; + if (! LLCoreHttpUtil::responseToLLSD(response, true, body_llsd)) + { + // INFOS-level logging will occur on the parsed failure + processFailure("HTTP response contained malformed LLSD", response); + break; // goto common exit + } - // Check for 200-with-error failures - // - // See comments in llinventorymodel.cpp about this mode of error. - // - // body_llsd["error"] = LLSD::emptyMap(); // Dev tool to force error handling - // body_llsd["error"]["identifier"] = "Development"; - // body_llsd["error"]["message"] = "You left development code in the viewer"; - if (body_llsd.has("error")) - { - processFailure("Inventory application error (200-with-error)", response); - break; // goto common exit - } + // Expect top-level structure to be a map + // body_llsd = LLSD::emptyArray(); // Dev tool to force error handling + if (! body_llsd.isMap()) + { + processFailure("LLSD response not a map", response); + break; // goto common exit + } - // Okay, process data if possible - processData(body_llsd, response); + // Check for 200-with-error failures + // + // See comments in llinventorymodel.cpp about this mode of error. + // + // body_llsd["error"] = LLSD::emptyMap(); // Dev tool to force error handling + // body_llsd["error"]["identifier"] = "Development"; + // body_llsd["error"]["message"] = "You left development code in the viewer"; + if (body_llsd.has("error")) + { + processFailure("Inventory application error (200-with-error)", response); + break; // goto common exit } + + // Okay, process data if possible + processData(body_llsd, response); } while (false); -- cgit v1.2.3