pci_alloc_consistent for small allocations?

From: Adam J. Richter (adam@yggdrasil.com)
Date: Sun Sep 02 2001 - 10:08:00 EST


        In looking at the ieee1394 OHCI driver, I noticed that it
appears to make 104 calls to pci_alloc_consistent for data structures
that are 16 or 64 bytes. Currently, on x86, pci_alloc_consistent
allocates at least one full page per call, so it looks like the
ohci1394 driver allocates 416kB per controller as a result of these
data structures.

        It is easy enough to change the ohci driver to just
do a few pci_alloc_consistent calls, but, in grepping through the
kernel, I see that there are lots of calls to pci_alloc_consistent
calls requesting small amounts of memory. So, I think it might reduce
kernel memory consumption to have pci_alloc_consistent do something
slightly smarter for allocations of less than a page. Two pretty
simple approaches come to mind:

        1. If it is the case that a side effect of the slab memory allocator
is that allocations of less than a page (or a certain size) never cross
page boundaries, then the x86 version of pci_alloc_consistent can just
use kmalloc/kfree when the size is below a certain amount. Could someone
tell me if this is the case?

        2. Assuming that is not the case, I have written, but not yet tested,
a change to pci_alloc_consistent for x86 where sub-page allocations can share
a single page. The first four bytes of a shared page are used as a
reference counter. This would be two bytes were it not for alignment
considerations. This change can only aggregate small allocations when
they occur in succession. I have attached the patch for illustration, but,
it could use some better variable naming and, I repeat, I have not tested
it at all yet. In the middle of working on it, I thought of option #1 and
figured I should ask on linux-kernel before investing more time in this idea.

Adam J. Richter __ ______________ 4880 Stevens Creek Blvd, Suite 104
adam@yggdrasil.com \ / San Jose, California 95129-1034
+1 408 261-6630 | g g d r a s i l United States of America
fax +1 408 261-6631 "Free Software For The Rest Of Us."

--- linux-2.4.10-pre3/arch/i386/kernel/pci-dma.c Mon Feb 14 15:34:21 2000
+++ linux/arch/i386/kernel/pci-dma.c Sun Sep 2 07:50:33 2001
@@ -13,15 +13,40 @@
 #include <linux/pci.h>
 #include <asm/io.h>
 
+static void *reg_partial; /* = NULL */
+static void *dma_partial; /* = NULL */
+
 void *pci_alloc_consistent(struct pci_dev *hwdev, size_t size,
                            dma_addr_t *dma_handle)
 {
         void *ret;
         int gfp = GFP_ATOMIC;
+ void **partial;
+ int avail;
 
         if (hwdev == NULL || hwdev->dma_mask != 0xffffffff)
                 gfp |= GFP_DMA;
- ret = (void *)__get_free_pages(gfp, get_order(size));
+
+ partial = (gfp & GFP_DMA) ? &dma_partial : &reg_partial;
+
+ /* if *partial is NULL or has been incremented to the beginning
+ of a new page, avail will be zero, and *partial will not be
+ used. */
+ avail = (-((unsigned long)*partial)) & PAGE_MASK;
+
+ if (avail >= size) {
+ ret = *partial;
+ *(int*)(((unsigned long)*partial) & ~PAGE_MASK) += 1;
+ (*partial) += size;
+ } else {
+ ret = (void *)__get_free_pages(gfp, get_order(size));
+
+ if (ret != NULL && (avail + size + sizeof(int) < PAGE_SIZE)) {
+ *(int*)ret = 1;
+ ret += sizeof(int);
+ *partial = ret + size;
+ }
+ }
 
         if (ret != NULL) {
                 memset(ret, 0, size);
@@ -33,5 +58,9 @@
 void pci_free_consistent(struct pci_dev *hwdev, size_t size,
                          void *vaddr, dma_addr_t dma_handle)
 {
+ if (((unsigned long)vaddr & PAGE_MASK) != 0) {
+ if (--*(int*)(((unsigned long)vaddr) & ~PAGE_MASK) != 0)
+ return; /* shared page */
+ }
         free_pages((unsigned long)vaddr, get_order(size));
 }

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/



This archive was generated by hypermail 2b29 : Fri Sep 07 2001 - 21:00:14 EST