On 2/14/2018 1:30 AM, Johannes Thumshirn wrote:
On Tue, Feb 13, 2018 at 11:34:48AM -0800, James Smart wrote:
[...]
diff --git a/drivers/scsi/lpfc/lpfc_sli.c b/drivers/scsi/lpfc/lpfc_sli.c
index 3bff1f9c5df7..5e03b2c969e5 100644
--- a/drivers/scsi/lpfc/lpfc_sli.c
+++ b/drivers/scsi/lpfc/lpfc_sli.c
@@ -35,6 +35,9 @@
#include <scsi/scsi_transport_fc.h>
#include <scsi/fc/fc_fs.h>
#include <linux/aer.h>
+#ifdef CONFIG_X86
+#include <asm/set_memory.h>
+#endif
Not needed anymore now you've killed set_memory_wc(), isn't it?
Agree... but, we've done more timing and it turns out the ioremap_wc()
on X86 isn't behaving quite the same as set_memory_wc(). Works, but
it's actually slower. I think ioremap_wc() is additionally making it
cacheable, which seems to be delaying the postings to the io bus (even
if wc) until the memory barrier. While the set_memory_wc() seems to
flush as soon as the cacheline is filled.
Given everything we've seen so far - I'm going back to using
set_memory_wc() as it's the fastest latency option we've measured.
[...]
+ if (q->dpp_enable && q->phba->cfg_enable_dpp) {
+ /* write to DPP aperture taking advatage of Combined Writes */
+ tmp = (uint8_t *)wqe;
+#ifdef CONFIG_64BIT
+ for (i = 0; i < q->entry_size; i += sizeof(uint64_t))
+ writeq(*((uint64_t *)(tmp + i)), q->dpp_regaddr + i);
+#else
+ for (i = 0; i < q->entry_size; i += sizeof(uint32_t))
+ writel(*((uint32_t *)(tmp + i)), q->dpp_regaddr + i);
+#endif
+ }
+ /* ensure WQE bcopy and DPP flushed before doorbell write */
Any reason you can't use writeq() on 32 Bit as well? There's a compat version
in linux/io-64-nonatomic-hi-lo.h.
We actually ran into issues on the existence of writeq() on a 32bit
platform. Thus this code block.
-- james
