[openib-general] the valgrind support of libibverbs/libmthca is not full

[Roland Dreier]

Some of these are almost definitely due to the what the test program is
doing -- for example

 > ==10853== 72 bytes in 1 blocks are definitely lost in loss record 8 of 13
 > ==10853==    at 0x4904B7E: malloc (vg_replace_malloc.c:149)
 > ==10853==    by 0x4C3DA82: mthca_alloc_pd (verbs.c:83)
 > ==10853==    by 0x4C3BC15: mthca_alloc_context (mthca.c:179)
 > ==10853==    by 0x4A32536: ibv_open_device (device.c:126)
 > ==10853==    by 0x401FF4: open_hca (main.c:214)
 > ==10853==    by 0x40E4E7: cq_7 (test_cq.c:529)
 > ==10853==    by 0x4108A0: test_cq (test_cq.c:790)
 > ==10853==    by 0x403196: main (main.c:147)

mthca_free_pd() definitely frees the struct that is allocated in
mthca_alloc_pd(), so I think your test must just not being freeing the
PD it allocates.

others are a little more mysterious.  Out of curiousity is your test
with a mem-free or tavor mode HCA?

Are your test programs available from a git tree anywhere?

Thanks,
  Roland