ioatdma: Push pending transactions to hardware more frequently
Every 20 descriptors turns out to be to few append commands with newer/faster CPUs. Pushing every 4 still cuts down on MMIO writes to an acceptable level without letting the DMA engine run out of work. Signed-off-by: Chris Leech <christopher.leech@intel.com>
This commit is contained in:
parent
7dcca30a32
commit
000725d56a
|
@ -310,7 +310,7 @@ static dma_cookie_t do_ioat_dma_memcpy(struct ioat_dma_chan *ioat_chan,
|
|||
list_splice_init(&new_chain, ioat_chan->used_desc.prev);
|
||||
|
||||
ioat_chan->pending += desc_count;
|
||||
if (ioat_chan->pending >= 20) {
|
||||
if (ioat_chan->pending >= 4) {
|
||||
append = 1;
|
||||
ioat_chan->pending = 0;
|
||||
}
|
||||
|
@ -818,7 +818,7 @@ static void __devexit ioat_remove(struct pci_dev *pdev)
|
|||
}
|
||||
|
||||
/* MODULE API */
|
||||
MODULE_VERSION("1.7");
|
||||
MODULE_VERSION("1.9");
|
||||
MODULE_LICENSE("GPL");
|
||||
MODULE_AUTHOR("Intel Corporation");
|
||||
|
||||
|
|
Loading…
Reference in New Issue