Address performance issue by removing use of read_avail

[shinrich]

Going through the great list of local fixes while getting ready for the 8.x sync. We had been seeing performance hot spots with these unnecessary read_avail calls.

[SolidWallOfCode]

Yes, for reasons we haven't discovered, the block list can become quite long and read_avail walks the entire list every time. In contrast, read_avail_more_than(0) will just look at the first block.

[masaori335]

I agree with that calling read_avail() should be avoided as much as possible. We observed that read_avail() made spin with long IOBufferBlock in other area in the past.

Looks this changes is doing logically same things without read_avail(). I bit worried about replacing memcpy() & consume() to read(), but it looks OK so far.

[SolidWallOfCode]

Actually, @masaori335, I think that's a bug. I want to talk with Susan before we commit this. I don't think read does the same thing as consume in this context.

[SolidWallOfCode]

Spoke with Susan, I misunderstood the IOBufferReader API. The code looks correct.

[bryancall]

Cherry picked to 8.1.0