Brian Xu - Sun Microsystems - Beijing China wrote:
Garrett D'Amore wrote:
Brian Xu - Sun Microsystems - Beijing China wrote:
Garrett D'Amore wrote:
Brian Xu - Sun Microsystems - Beijing China wrote:
Hi there,
I have a question here:
Why all of the NIC drivers have to bcopy the MBLKs for transmit?
(some of them bcopy always, and some others bcopy under a
threshold of the packet length).
I think one of the reason is the overhead of the setup of dma on
the fly is greater than the overhead of bcopy for short packets. I
want to know if this is the case and if there are any other reasons.
Yes. For any packet reasonably sized bcopy (ETHERMTU or smaller)
is faster on *all* recent hardware. (This is confirmed on even an
older 300MHz Via C3.) (Hmm... I've heard that for some Niagra
systems this might not be true, however. But I've not tested it
myself.)
Even with bcopy, there is still need a pre-binded dma resource. So
the threshold of the bcopy size is based on whether the overhead for
dma bind on the fly is greater than the threshold of the bcopy to a
pre-binded dma address. For the hardware itself, it only know DMA is
needed.
The pre-bound DMA setup you pay at attach() time, and doesn't play a
role. So you have to compare the cost of bcopy() vs. the cost of
ddi_dma_addr_setup().
It is really what I meant.
There is a lot of additional complexity for tx as well, because you
have to deal with the fact that packets may cross page boundaries and
require multiple DMA cookies. This adds a lot of complexity, and not
all drivers can deal well with multiple descriptors per packet.
Just as what we do for ddi_dma_buf_bind_handle, the shadow page list
records all the mapped physical pages. so you don't have to worry
about the cross of page boundary.
Ah, but the *driver* does., because in the absence of an IOMMU you need
to be able to allocate more than one descriptor. You have some call
overhead as well... multiple ddi_dma_symc() calls per packet, probably,
and ddi_dma_nextcookie() and such.
It might not sound like much, but on hot code paths every additional
function call adds overhead. You don't have to to call many extra
function calls before you catch up to to the cost of bcopy. For
example, ignoring memory for the moment, bcopy of a 1024 byte packet
might require fewer than 150 instruction cycles.
I still don´t know if there are other reasons other than the
overhead of dma setup.
Complexity. There are various concerns, as a race with _fini() and
esballoc (for the rx path), involved.
Also you have to worry about alignment. Not all hardware can
transmit arbitrarily aligned packets. With all the work you wind up
doing to make this work correctly, you get very little performance
benefit. So its rarely worth the pain and suffering. For regular
MTU frames, it just isn't worth it, ever. On reasonably modern
hardware, anyway.
For the alignment, does how large packet transmit (dma bind on the
fly) does is OK, I think.
Packets may be aligned on *any* boundary. In fact, they are often *not*
32-bit aligned, but 16-bit aligned. Not all hardware can deal with
off-half-word alignment.
-- Garrett
Thanks,
Brian
For rx, you can eliminate a lot of the DMA costs by recycling
buffers. But the complexity to do this "well" without introducing
potential panics is high. Almost every driver that has tried has
gotten this wrong at some point. Some of them are still wrong.
-- Garrett
Thanks,
Brian
I think the situation is different with jumbo frames, though.
If what I guess is the major cause, I have a proposal and I want
to hear your advice whether it makes sense.
The most time-consuming action for the dma setup is the dma bind,
more specific, calling into the VM layer to get the PFN for the
vaddr(hat_getpfnum()), since it need to search the huge page
table. While for the MBLKs, essentially which are slab objects,
the PFN has already been determined in the slab layer, and for
most of their usage, we only touch the magazine layer, where the
PFN is a pre determined one. That is, the PFN should be considered
as a constructed state, but we don't leverage it for dma bind.
In storage, we have a field 'b_shadow' in buf(9S) to store the
pages which are recently used, through which the PFNs can be
easily got. so in
the case that b_shadow works, ddi_dma_buf_bind_handle() is much
faster than the ddi_dma_mem_bind_handle().
Another example, moving the dma bind of the HBA driver(mpt) from
Tx path to the kmem cache constrcutor, mpt driver got 26%
throughput increment. See CR6707308.
If the mblk could store the PFN info and we had a
ddi_dma_mblk_bind_handle() like interface, then I think it will
benefit the performance of the NIC drivers. I consulted the PAE,
and got a answer that the bcopy is typically about 10-15% of a
NIC TX workload.
There are things that can do to make DMA faster, better, and
simpler. In an ideal world, the GLDv3 could do most of this work,
and the mblk could just carry the ddi_dma_cookie with it.
-- Garrett
Thanks,
Brian
_______________________________________________
driver-discuss mailing list
driver-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/driver-discuss
_______________________________________________
driver-discuss mailing list
driver-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/driver-discuss
_______________________________________________
driver-discuss mailing list
driver-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/driver-discuss
