From 33649e848b34dcaa5f4be49862fcbf0a49b46df3 Mon Sep 17 00:00:00 2001 From: Nat Goodspeed Date: Mon, 12 Jan 2015 12:11:37 -0500 Subject: Experimentally try full-size LLLeap stream test on Windows. The new toolchain may (!) have fixed a longstanding bug in LLLeap / APR when we try to pump large volumes of data through a Windows named pipe using APR nonblocking I/O. This used to fail pretty consistently because the APR nonblocking write call would sometimes spuriously return "would block" when in fact the data buffer was completely written; the caller would later retry, which of course would duplicate some of the data in the pipe. Preliminary experiments with VS 2013 suggest this may have been resolved. This changeset is to propagate the experiment to a wider range of Windows systems; we may need to revert it if in fact the bug persists. --- indra/llcommon/tests/llleap_test.cpp | 17 ----------------- 1 file changed, 17 deletions(-) (limited to 'indra') diff --git a/indra/llcommon/tests/llleap_test.cpp b/indra/llcommon/tests/llleap_test.cpp index 653b043de4..2d88e2c676 100755 --- a/indra/llcommon/tests/llleap_test.cpp +++ b/indra/llcommon/tests/llleap_test.cpp @@ -38,24 +38,7 @@ StringVec sv(const StringVec& listof) { return listof; } #define sleep(secs) _sleep((secs) * 1000) #endif -#if ! LL_WINDOWS const size_t BUFFERED_LENGTH = 1023*1024; // try wrangling just under a megabyte of data -#else -// "Then there's Windows... sigh." The "very large message" test is flaky in a -// way that seems to point to either the OS (nonblocking writes to pipes) or -// possibly the apr_file_write() function. Poring over log messages reveals -// that at some point along the way apr_file_write() returns 11 (Resource -// temporarily unavailable, i.e. EAGAIN) and says it wrote 0 bytes -- even -// though it did write the chunk! Our next write attempt retries the same -// chunk, resulting in the chunk being duplicated at the child end, corrupting -// the data stream. Much as I would love to be able to fix it for real, such a -// fix would appear to require distinguishing bogus EAGAIN returns from real -// ones -- how?? Empirically this behavior is only observed when writing a -// "very large message". To be able to move forward at all, try to bypass this -// particular failure by adjusting the size of a "very large message" on -// Windows. -const size_t BUFFERED_LENGTH = 65336; -#endif // LL_WINDOWS void waitfor(const std::vector& instances, int timeout=60) { -- cgit v1.2.3