connectx4.lua: fix page sizes in CREATE_{C,R}Q

Set the page size parameters for CQs and RQs so that the entire queue
fits in a single page.  The alternative would be to select a fixed
page size and add as many PAS entries as necessary.  It is unclear
whether there is a difference between the methods (e.g. for
perfomance).

src/apps/mellanox/connectx4.lua

@@ -916,7 +916,9 @@ end
 function HCA:create_cq (entries, uar_page, eqn, collapsed)
    local doorbell, doorbell_phy = memory.dma_alloc(16)
    -- Memory for completion queue entries
-   local cqe, cqe_phy = memory.dma_alloc(entries * 64, 4096)
+   local size = entries * 64
+   local cqe, cqe_phy = memory.dma_alloc(size, 4096)
+   local log_page_size = log2size(math.ceil(size/4096))
    ffi.fill(cqe, entries * 64, 0xFF)
    self:command("CREATE_CQ", 0x114, 0x0C)
       :input("opcode",        0x00,        31, 16, 0x400)
@@ -925,7 +927,7 @@ function HCA:create_cq (entries, uar_page, eqn, collapsed)
       :input("log_cq_size",   0x10 + 0x0C, 28, 24, log2size(entries))
       :input("uar_page",      0x10 + 0x0C, 23,  0, uar_page)
       :input("c_eqn",         0x10 + 0x14,  7,  0, eqn)
-      :input("log_page_size", 0x10 + 0x18, 28, 24, 4)
+      :input("log_page_size", 0x10 + 0x18, 28, 24, log_page_size)
       :input("db_addr high",  0x10 + 0x38, 31,  0, ptrbits(doorbell_phy, 63, 32))
       :input("db_addr_low",   0x10 + 0x3C, 31,  0, ptrbits(doorbell_phy, 31, 0))
       :input("pas[0] high",   0x110,       31,  0, ptrbits(cqe_phy, 63, 32))
@@ -941,6 +943,7 @@ function HCA:create_rq (cqn, pd, size, doorbell, rwq)
    local log_wq_size = log2size(size)
    local db_phy = memory.virtual_to_physical(doorbell)
    local rwq_phy = memory.virtual_to_physical(rwq)
+   local log_page_size = log2size(math.ceil(size * 64/4096))
    self:command("CREATE_RQ", 0x20 + 0x30 + 0xC4, 0x0C)
       :input("opcode",        0x00, 31, 16, 0x908)
       :input("rlkey",         0x20 + 0x00, 31, 31, 1)
@@ -951,7 +954,7 @@ function HCA:create_rq (cqn, pd, size, doorbell, rwq)
       :input("dbr_addr high", 0x20 + 0x30 + 0x10, 31,  0, ptrbits(db_phy, 63, 32))
       :input("dbr_addr low",  0x20 + 0x30 + 0x14, 31,  0, ptrbits(db_phy, 31, 0))
       :input("log_wq_stride", 0x20 + 0x30 + 0x20, 19, 16, 4)
-      :input("page_size",     0x20 + 0x30 + 0x20, 12,  8, 4) -- XXX one big page?
+      :input("log_page_size", 0x20 + 0x30 + 0x20, 12,  8, log_page_size)
       :input("log_wq_size",   0x20 + 0x30 + 0x20,  4 , 0, log_wq_size)
       :input("pas[0] high",   0x20 + 0x30 + 0xC0, 63, 32, ptrbits(rwq_phy, 63, 32))
       :input("pas[0] low",    0x20 + 0x30 + 0xC4, 31,  0, ptrbits(rwq_phy, 31, 0))