At 2002/7/29 11:25-0700 Wichmann, Mats D writes: > > (a) What's the right value for VSRT_ADDR_SPACE_PAGES ? > This is a configuration parameter that goes in > TESTROOT/tetexec.cfg and is used by a couple of > (currently failing) mmap tests
The current value of VSRT_ADDR_SPACE_PAGES has been tuned for 32 bit userspace. The right value to use was found mostly through a bit of experimentation. The following excerpt from the config file should help you work out what value to use: # The maximum size of a process' address space, in pages. This value # is used in mmap testing (mmap 26 and 31 and mmap_P 29 and 34) to # verify ENOMEM is returned when memory allocation is not possible # due to address space boundaries. On some systems where this value # is very large, using the actual value may cause mmap to fail with # an invalid value error. These tests are constructed so that they # completely validate their assertions using any value between 1/2 the # size of the address space and its full size can be used, and in this # circumstance a smaller value in this range should be configured here. We'll need to change the test suite setup so the value can be set on a per-architecture basis. > (b) Why does SSIZE_MAX end up set to the size of > an int, rather than a 64-bit quantity, on standard > GNU/Linux header files? This one affects the > calculation done by failing readv and writev tests > which are trying to overflow an ssize_t to test the > error code. But the don't come close, since the > calulation based on SSIZE_MAX builds 1024 iovs > of 2,098,176 bytes. The kernel readv/writev code which checks this has changed quite a bit between 2.4 and 2.5. It used to try (but I think it was wrong) to test when the total read length would overflow a 32 bit int. The 2.5 behaviour checks for overflow of ssize_t and presumably libc now has to catch up. I've asked glibc IA64 guy to look at it. Regards, Chris -- [EMAIL PROTECTED] IBM OzLabs Linux Development Group Canberra, Australia -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with subject of "unsubscribe". Trouble? Email [EMAIL PROTECTED]
