On 2015-02-07 12:30PM Torsten Bögershausen wrote: >On 2015-02-07 17.45, Joachim Schmitz wrote: >> Hi there >> >> While investigating the problem with hung git-upload-pack we think to >> have found a bug in wrapper.c: >> >> #define MAX_IO_SIZE (8*1024*1024) >> >> This is then used in xread() to split read()s into suitable chunks. >> So far so good, but read() is only guaranteed to read as much as >> SSIZE_MAX bytes at a time. And on our platform that is way lower than >> those 8MB (only 52kB, POSIX allows it to be as small as 32k), and as a >> (rather strange) consequence mmap() (from compat/mmap.c) fails with >> EACCESS (why EACCESS?), because xpread() returns something > 0. >> >> How large is SSIZE_MAX on other platforms? What happens there if you >> try to >> read() more? Should't we rather use SSIZE_MAX on all platforms? If I'm >> reading the header files right, on Linux it is LONG_MAX (2TB?), so I >> guess we should really go for MIN(8*1024*1024,SSIZE_MAX)? >How about changing wrapper.c like this: >#ifndef MAX_IO_SIZE > #define MAX_IO_SIZE (8*1024*1024) >#endif >--------------------- >and to change config.mak.uname like this: >ifeq ($(uname_S),NONSTOP_KERNEL) > BASIC_CFLAGS += -DMAX_IO_SIZE=(32*1024) Does this work for you ? Yes, thank you Torsten. I have made this change in our branch (on behalf of Jojo). I think we can accept it. The (32*1024) does need to be properly quoted, however. -- To unsubscribe from this list: send the line "unsubscribe git" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html