On 2/5/18 8:20 AM, Ming Lei wrote:

Quite a few HBAs(such as HPSA, megaraid, mpt3sas, ..) support multiple
reply queues, but tags is often HBA wide.

These HBAs have switched to use pci_alloc_irq_vectors(PCI_IRQ_AFFINITY)
for automatic affinity assignment.

Now 84676c1f21e8ff5(genirq/affinity: assign vectors to all possible CPUs)
has been merged to V4.16-rc, and it is easy to allocate all offline CPUs
for some irq vectors, this can't be avoided even though the allocation
is improved.

So all these drivers have to avoid to ask HBA to complete request in
reply queue which hasn't online CPUs assigned, and HPSA has been broken
with v4.15+:

	https://marc.info/?l=linux-kernel&m=151748144730409&w=2

This issue can be solved generically and easily via blk_mq(scsi_mq) multiple
hw queue by mapping each reply queue into hctx, but one tricky thing is
the HBA wide(instead of hw queue wide) tags.

This patch is based on the following Hannes's patch:

	https://marc.info/?l=linux-block&m=149132580511346&w=2

One big difference with Hannes's is that this patch only makes the tags sbitmap
and active_queues data structure HBA wide, and others are kept as NUMA locality,
such as request, hctx, tags, ...

The following patch will support global tags on null_blk, also the performance
data is provided, no obvious performance loss is observed when the whole
hw queue depth is same.

GLOBAL implies that it's, strangely enough, global. That isn't really the
case. Why not call this BLK_MQ_F_HOST_TAGS or something like that? I'd
welcome better names, but global doesn't seem to be a great choice.

BLK_MQ_F_SET_TAGS?

-- 
Jens Axboe