illumos gate

During synthetic random write benchmarks I found many secondary and tertiary block copies allocations form the same vdevs (more then 10% of all allocations!):

                1          4618459
                2           667430
                3              420

                1          4605120
                2           672441
                3              346

                1          4635516
                2           663064
                3              760

That at the end caused some allocation issues I addressed in #9738. Digging toward the source I found that the poor allocation is caused by allocation throttling. I suppose periodically under the heavy load one or another random disk becomes the only one with allocations available, and starts collecting all ditto blocks. Simple patch I made to relax allocation of secondary and tertiary blocks radically fixed the allocation issue, and seems even increased performance a bit:

                1          6222858

                1          6135604

                1          6062702