UCT/RDMACM: fix uct_rdmacm_cm_cqs hash key

[evgeny-leksikov]

What

fix uct_rdmacm_cm_cqs hash key

Why ?

pd->handle is always 0 in container

How ?

make safer key

[yosefe]

1. can you pls add comment about container?
2. maybe the key should be both pd and handle? and not a xor of them? there is some chance that pd1 and pd2 will be considered same because xor value would be the same. so need to make sure each of them is also compared.

[evgeny-leksikov]

do you mean 86-bit key? maybe just set (or xor) the handle to hight 32 bits of 64-bit key initialized by the pd ptr?

return (uintptr_t)pd ^ ((uint64_t)pd->handle << 32);

or

uint64_t key64;
uint32_t *key32_p = &key64;

key32_p[0] = pd->handle;
key32_p[1] = (uint32_t)(uintptr_t)pd;
return key64;

[yosefe]

yes, i mean a struct of {pd, handle}
if we use xor, or reduce the pointer to 32 bit, it's not a 100% guarantee that we get the right PD, so could end up using CQ object with wrong PD

[evgeny-leksikov]

just to clarify - should not pd->handle guarantee this in normal case? I cant find any mention in docs usage examples or purposes of this handle. If yes, then it does not makes sence to me otherwise we can even do simpler:

return pd->handle ? pd->handle : pd;

[yosefe]

use GUID as key:
ibv_get_device_guid(struct ibv_device *device);

[yosefe]

bot:pipe:retest

[brminich]

@avildema, can we check some simple CS test inside container in CI somehow?