Re: [PATCH] Introduce nodemask_t ADT [0/7]

[Jesse Barnes <jbarnes@sgi.com>]

On Thursday 18 March 2004 4:58 pm, Zwane Mwaikambo wrote:
> > It probably shouldn't have anything to do with PCI directly either,
> > so .... ;-) My former thought was that you might just want the most
> > local memory for DMAing into.
> 
> That knowledge should be in the dma api thing shouldn't it? But in it's
> current incarnation i don't see how that's possible.

Couldn't we mostly hide it under the covers (though obviously not in
the intentionally remote case I mentioned earlier):


===== arch/ia64/sn/io/machvec/pci_dma.c 1.28 vs edited =====
--- 1.28/arch/ia64/sn/io/machvec/pci_dma.c	Sun Mar 14 11:17:06 2004
+++ edited/arch/ia64/sn/io/machvec/pci_dma.c	Thu Mar 18 17:08:13 2004
@@ -130,13 +130,11 @@
 	device_sysdata = SN_DEVICE_SYSDATA(hwdev);
 	vhdl = device_sysdata->vhdl;
 
-	/*
-	 * Allocate the memory.
-	 * FIXME: We should be doing alloc_pages_node for the node closest
-	 *        to the PCI device.
-	 */
-	if (!(cpuaddr = (void *)__get_free_pages(GFP_ATOMIC, get_order(size))))
+	cpuaddr = alloc_pages_node(pci_to_node(hwdev), GFP_ATOMIC, (get_order(size)));
+	if (!cpuaddr)
 		return NULL;
+
+	cpuaddr = page_to_virt(cpuaddr);
 
 	memset(cpuaddr, 0x0, size);