[PARISC] Document that we tolerate "Relaxed Ordering"

This means "DMA Read returns" can bypass "MMIO Writes".
Violating the PCI specs in this case improves outbound DMA "flows"
and is currently not required by any drivers.

This is NOT a new behavior. Previous chipsets did this
already and I believe ZX1 PDC was already setting this
for hpux. I just want to further document the behavior.

Signed-off-by: Grant Grundler <grundler@parisc-linux.org>
Signed-off-by: Kyle McMartin <kyle@parisc-linux.org>

diff --git a/drivers/parisc/sba_iommu.c b/drivers/parisc/sba_iommu.c
index 42b32ff..278f325 100644 (file)
--- a/drivers/parisc/sba_iommu.c
+++ b/drivers/parisc/sba_iommu.c
@@ -178,6 +178,11 @@ extern struct proc_dir_entry * proc_mckinley_root;
 #define ROPE6_CTL      0x230
 #define ROPE7_CTL      0x238
 
+#define IOC_ROPE0_CFG  0x500   /* pluto only */
+#define   IOC_ROPE_AO    0x10  /* Allow "Relaxed Ordering" */
+
+
+
 #define HF_ENABLE      0x40
 
 
@@ -1759,19 +1764,33 @@ printk("sba_hw_init(): mem_boot 0x%x 0x%x 0x%x 0x%x\n", PAGE0->mem_boot.hpa,
 
        sba_dev->num_ioc = num_ioc;
        for (i = 0; i < num_ioc; i++) {
-               /*
-               ** Make sure the box crashes if we get any errors on a rope.
-               */
-               WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE0_CTL);
-               WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE1_CTL);
-               WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE2_CTL);
-               WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE3_CTL);
-               WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE4_CTL);
-               WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE5_CTL);
-               WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE6_CTL);
-               WRITE_REG(HF_ENABLE, sba_dev->ioc[i].ioc_hpa + ROPE7_CTL);
-
-               /* flush out the writes */
+               unsigned long ioc_hpa = sba_dev->ioc[i].ioc_hpa;
+               unsigned int j;
+
+               for (j=0; j < sizeof(u64) * ROPES_PER_IOC; j+=sizeof(u64)) {
+
+                       /*
+                        * Clear ROPE(N)_CONFIG AO bit.
+                        * Disables "NT Ordering" (~= !"Relaxed Ordering")
+                        * Overrides bit 1 in DMA Hint Sets.
+                        * Improves netperf UDP_STREAM by ~10% for bcm5701.
+                        */
+                       if (IS_PLUTO(sba_dev->iodc)) {
+                               unsigned long rope_cfg, cfg_val;
+
+                               rope_cfg = ioc_hpa + IOC_ROPE0_CFG + j;
+                               cfg_val = READ_REG(rope_cfg);
+                               cfg_val &= ~IOC_ROPE_AO;
+                               WRITE_REG(cfg_val, rope_cfg);
+                       }
+
+                       /*
+                       ** Make sure the box crashes on rope errors.
+                       */
+                       WRITE_REG(HF_ENABLE, ioc_hpa + ROPE0_CTL + j);
+               }
+
+               /* flush out the last writes */
                READ_REG(sba_dev->ioc[i].ioc_hpa + ROPE7_CTL);
 
                DBG_INIT("      ioc[%d] ROPE_CFG 0x%Lx  ROPE_DBG 0x%Lx\n",