Hello everyone,
Thank you very much for the time you've spent trying to help. I'm not 100% sure, but I've probably found the solution for the "H:0x0 D:0x8 P:0x0" errors and weird HCA behaviour, where errors occurred under no significant storage and target load. Stress testing with fio never showed any problems, that's why I suspected that the hangs in the BUSY states were caused by something else. I'm using SCST target stack and played with ibdump recently to find out, that the cause was indeed limited to the initiator side. What I did wrong is that I've setup the blades to use OS Control for the power management and even setting it to High Performance in the ESXi - servers were throttling in C1/C1E states and probably messing something with the PCI-Express power too. It suppose to have an impact only on the latency, but explicitly disabling these features in the BIOS made my logs clean. It's passed only 7 days since the change, so it's still to early to be certain, but I've a good feeling.
Thanks again!