Some bus_dma(9) fixes for Xen:

- Attempt to gracefully recover from a failed decrease_reservation or
  increase_reservation, by avoiding physical memory loss.
- always store a machine address in ds_addr; this avoids some mistakes
  where machine address would in some case be freed at physical address, or
  mapped as physical address.

sys/arch/x86/include/bus_private.h

@@ -1,4 +1,4 @@
-/*	$NetBSD: bus_private.h,v 1.5 2006/02/16 20:17:15 perry Exp $	*/
+/*	$NetBSD: bus_private.h,v 1.6 2006/08/28 19:58:56 bouyer Exp $	*/
 /*	NetBSD: bus.h,v 1.8 2005/03/09 19:04:46 matt Exp	*/
 
 /*-
@@ -150,10 +150,23 @@ struct x86_bus_dma_cookie {
 #define _BUS_PHYS_TO_BUS(pa)	((bus_addr_t)(pa))
 #endif /* !defined(_BUS_PHYS_TO_BUS) */
 
+#if !defined(_BUS_BUS_TO_PHYS)
+#define _BUS_BUS_TO_PHYS(ba)	((paddr_t)(ba))
+#endif /* !defined(_BUS_BUS_TO_PHYS) */
+
 #if !defined(_BUS_VM_PAGE_TO_BUS)
 #define	_BUS_VM_PAGE_TO_BUS(pg)	_BUS_PHYS_TO_BUS(VM_PAGE_TO_PHYS(pg))
 #endif /* !defined(_BUS_VM_PAGE_TO_BUS) */
 
+#if !defined(_BUS_BUS_TO_VM_PAGE)
+#define	_BUS_BUS_TO_VM_PAGE(ba)	PHYS_TO_VM_PAGE(ba)
+#endif /* !defined(_BUS_BUS_TO_VM_PAGE) */
+
+#if !defined(_BUS_PMAP_ENTER)
+#define _BUS_PMAP_ENTER(pmap, va, ba, prot, flags) \
+    pmap_enter(pmap, va, ba, prot, flags)
+#endif /* _BUS_PMAP_ENTER */
+
 #if !defined(_BUS_VIRT_TO_BUS)
 #include <uvm/uvm_extern.h>
 

sys/arch/x86/x86/bus_dma.c

@@ -1,4 +1,4 @@
-/*	$NetBSD: bus_dma.c,v 1.29 2006/03/01 12:38:12 yamt Exp $	*/
+/*	$NetBSD: bus_dma.c,v 1.30 2006/08/28 19:58:57 bouyer Exp $	*/
 
 /*-
  * Copyright (c) 1996, 1997, 1998 The NetBSD Foundation, Inc.
@@ -38,7 +38,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: bus_dma.c,v 1.29 2006/03/01 12:38:12 yamt Exp $");
+__KERNEL_RCSID(0, "$NetBSD: bus_dma.c,v 1.30 2006/08/28 19:58:57 bouyer Exp $");
 
 /*
  * The following is included because _bus_dma_uiomove is derived from
@@ -994,7 +994,7 @@ _bus_dmamem_free(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs)
 		for (addr = segs[curseg].ds_addr;
 		    addr < (segs[curseg].ds_addr + segs[curseg].ds_len);
 		    addr += PAGE_SIZE) {
-			m = PHYS_TO_VM_PAGE(addr);
+			m = _BUS_BUS_TO_VM_PAGE(addr);
 			TAILQ_INSERT_TAIL(&mlist, m, pageq);
 		}
 	}
@@ -1039,7 +1039,7 @@ _bus_dmamem_map(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
 		    addr += PAGE_SIZE, va += PAGE_SIZE, size -= PAGE_SIZE) {
 			if (size == 0)
 				panic("_bus_dmamem_map: size botch");
-			pmap_enter(pmap_kernel(), va, addr,
+			_BUS_PMAP_ENTER(pmap_kernel(), va, addr,
 			    VM_PROT_READ | VM_PROT_WRITE,
 			    PMAP_WIRED | VM_PROT_READ | VM_PROT_WRITE);
 			/*
@@ -1129,7 +1129,7 @@ _bus_dmamem_mmap(bus_dma_tag_t t, bus_dma_segment_t *segs, int nsegs,
 			continue;
 		}
 
-		return (x86_btop((caddr_t)segs[i].ds_addr + off));
+		return (x86_btop(_BUS_BUS_TO_PHYS(segs[i].ds_addr + off)));
 	}
 
 	/* Page not found. */

sys/arch/xen/include/bus_private.h

@@ -1,4 +1,4 @@
-/*	$NetBSD: bus_private.h,v 1.6 2006/02/16 20:17:15 perry Exp $	*/
+/*	$NetBSD: bus_private.h,v 1.7 2006/08/28 19:58:57 bouyer Exp $	*/
 
 /*-
  * Copyright (c)2005 YAMAMOTO Takashi,
@@ -29,7 +29,11 @@
 #include <uvm/uvm_extern.h>
 
 #define	_BUS_PHYS_TO_BUS(pa)	((bus_addr_t)xpmap_ptom(pa))
+#define	_BUS_BUS_TO_PHYS(ba)	((paddr_t)xpmap_mtop(ba))
 #define	_BUS_VIRT_TO_BUS(pm, va) _bus_virt_to_bus((pm), (va))
+#define _BUS_BUS_TO_VM_PAGE(ba) (PHYS_TO_VM_PAGE(xpmap_mtop(ba)))
+#define _BUS_PMAP_ENTER(pmap, va, ba, prot, flags) \
+    pmap_enter(pmap, va, xpmap_mtop(ba), prot, flags)
 
 static __inline bus_addr_t _bus_virt_to_bus(struct pmap *, vaddr_t);
 

sys/arch/xen/x86/xen_bus_dma.c

@@ -1,4 +1,4 @@
-/*	$NetBSD: xen_bus_dma.c,v 1.6 2006/01/15 22:09:52 bouyer Exp $	*/
+/*	$NetBSD: xen_bus_dma.c,v 1.7 2006/08/28 19:58:57 bouyer Exp $	*/
 /*	NetBSD bus_dma.c,v 1.21 2005/04/16 07:53:35 yamt Exp */
 
 /*-
@@ -39,7 +39,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: xen_bus_dma.c,v 1.6 2006/01/15 22:09:52 bouyer Exp $");
+__KERNEL_RCSID(0, "$NetBSD: xen_bus_dma.c,v 1.7 2006/08/28 19:58:57 bouyer Exp $");
 
 #include <sys/param.h>
 #include <sys/systm.h>
@@ -75,14 +75,11 @@ _xen_alloc_contig(bus_size_t size, bus_size_t alignment, bus_size_t boundary,
 	unsigned long npagesreq, npages, mfn;
 	bus_addr_t pa;
 	struct vm_page *pg, *pgnext;
-	struct pglist freelist;
 	int s, error;
 #ifdef XEN3
 	struct xen_memory_reservation res;
 #endif
 
-	TAILQ_INIT(&freelist);
-
 	/*
 	 * When requesting a contigous memory region, the hypervisor will
 	 * return a memory range aligned on size. This will automagically
@@ -114,14 +111,21 @@ _xen_alloc_contig(bus_size_t size, bus_size_t alignment, bus_size_t boundary,
 		    < 0) {
 			printf("xen_alloc_contig: XENMEM_decrease_reservation "
 			    "failed!\n");
-			return ENOMEM;
+			xpmap_phys_to_machine_mapping[
+			    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
+
+			error = ENOMEM;
+			goto failed;
 		}
 #else
 		if (HYPERVISOR_dom_mem_op(MEMOP_decrease_reservation,
 		    &mfn, 1, 0) != 1) {
 			printf("xen_alloc_contig: MEMOP_decrease_reservation "
 			    "failed!\n");
-			return ENOMEM;
+			xpmap_phys_to_machine_mapping[
+			    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
+			error = ENOMEM;
+			goto failed;
 		}
 #endif
 	}
@@ -135,14 +139,18 @@ _xen_alloc_contig(bus_size_t size, bus_size_t alignment, bus_size_t boundary,
 	if (HYPERVISOR_memory_op(XENMEM_increase_reservation, &res) < 0) {
 		printf("xen_alloc_contig: XENMEM_increase_reservation "
 		    "failed!\n");
-		return ENOMEM;
+		error = ENOMEM;
+		pg = NULL;
+		goto failed;
 	}
 #else
 	if (HYPERVISOR_dom_mem_op(MEMOP_increase_reservation,
 	    &mfn, 1, order) != 1) {
 		printf("xen_alloc_contig: MEMOP_increase_reservation "
 		    "failed!\n");
-		return ENOMEM;
+		error = ENOMEM;
+		pg = NULL;
+		goto failed;
 	}
 #endif
 	s = splvm();
@@ -165,6 +173,58 @@ _xen_alloc_contig(bus_size_t size, bus_size_t alignment, bus_size_t boundary,
 	xpq_flush_queue();
 	splx(s);
 	return 0;
+
+failed:
+	/*
+	 * Attempt to recover from a failed decrease or increase reservation:
+	 * if decrease_reservation failed, we don't have given all pages
+	 * back to Xen; give them back to UVM, and get the missing pages
+	 * from Xen.
+	 * if increase_reservation failed, we expect pg to be NULL and we just
+	 * get back the missing pages from Xen one by one.
+	 */
+	/* give back remaining pages to UVM */
+	for (; pg != NULL; pg = pgnext) {
+		pgnext = pg->pageq.tqe_next;
+		TAILQ_REMOVE(mlistp, pg, pageq);
+		uvm_pagefree(pg);
+	}
+	/* remplace the pages that we already gave to Xen */
+	s = splvm();
+	for (pg = mlistp->tqh_first; pg != NULL; pg = pgnext) {
+		pgnext = pg->pageq.tqe_next;
+#ifdef XEN3
+		res.extent_start = &mfn;
+		res.nr_extents = 1;
+		res.extent_order = 0;
+		res.address_bits = 31;
+		res.domid = DOMID_SELF;
+		if (HYPERVISOR_memory_op(XENMEM_increase_reservation, &res)
+		    < 0) {
+			printf("xen_alloc_contig: recovery "
+			    "XENMEM_increase_reservation failed!\n");
+			break;
+		}
+#else
+		if (HYPERVISOR_dom_mem_op(MEMOP_increase_reservation,
+		    &mfn, 1, 0) != 1) {
+			printf("xen_alloc_contig: recovery "
+			    "MEMOP_increase_reservation failed!\n");
+			break;
+		}
+#endif
+		pa = VM_PAGE_TO_PHYS(pg);
+		xpmap_phys_to_machine_mapping[
+		    (pa - XPMAP_OFFSET) >> PAGE_SHIFT] = mfn;
+		xpq_queue_machphys_update((mfn) << PAGE_SHIFT, pa);
+		TAILQ_REMOVE(mlistp, pg, pageq);
+		uvm_pagefree(pg);
+	}
+	/* Flush updates through and flush the TLB */
+	xpq_queue_tlb_flush();
+	xpq_flush_queue();
+	splx(s);
+	return error;
 }
 
 
@@ -178,8 +238,7 @@ _xen_bus_dmamem_alloc_range(bus_dma_tag_t t, bus_size_t size,
     bus_size_t alignment, bus_size_t boundary, bus_dma_segment_t *segs,
     int nsegs, int *rsegs, int flags, bus_addr_t low, bus_addr_t high)
 {
-	paddr_t curaddr, lastaddr;
+	bus_addr_t curaddr, lastaddr;
-	bus_addr_t bus_curaddr, bus_lastaddr;
 	struct vm_page *m;
 	struct pglist mlist;
 	int curseg, error;
@@ -210,18 +269,16 @@ again:
 	 */
 	m = mlist.tqh_first;
 	curseg = 0;
-	lastaddr = segs[curseg].ds_addr = VM_PAGE_TO_PHYS(m);
+	lastaddr = segs[curseg].ds_addr = _BUS_VM_PAGE_TO_BUS(m);
 	segs[curseg].ds_len = PAGE_SIZE;
 	m = m->pageq.tqe_next;
-	if ((_BUS_PHYS_TO_BUS(segs[curseg].ds_addr) & (alignment - 1)) != 0)
+	if ((segs[curseg].ds_addr & (alignment - 1)) != 0)
 		goto dorealloc;
 
 	for (; m != NULL; m = m->pageq.tqe_next) {
-		curaddr = VM_PAGE_TO_PHYS(m);
+		curaddr = _BUS_VM_PAGE_TO_BUS(m);
-		bus_curaddr = _BUS_PHYS_TO_BUS(curaddr);
+		if ((lastaddr < low || lastaddr >= high) ||
-		bus_lastaddr = _BUS_PHYS_TO_BUS(lastaddr);
+		    (curaddr < low || curaddr >= high)) {
-		if ((bus_lastaddr < low || bus_lastaddr >= high) ||
-		    (bus_curaddr < low || bus_curaddr >= high)) {
 			/*
 			 * If machine addresses are outside the allowed
 			 * range we have to bail. Xen2 doesn't offer an
@@ -233,16 +290,15 @@ again:
 			uvm_pglistfree(&mlist);
 			return EINVAL;
 		}
-		if (bus_curaddr == (bus_lastaddr + PAGE_SIZE)) {
+		if (curaddr == (lastaddr + PAGE_SIZE)) {
 			segs[curseg].ds_len += PAGE_SIZE;
-			if ((bus_lastaddr & boundary) !=
+			if ((lastaddr & boundary) !=
-			    (bus_curaddr & boundary))
+			    (curaddr & boundary))
 				goto dorealloc;
-		}
+		} else {
-		else {
 			curseg++;
 			if (curseg >= nsegs ||
-			    (bus_curaddr & (alignment - 1)) != 0) {
+			    (curaddr & (alignment - 1)) != 0) {
 dorealloc:
 				if (doingrealloc == 1)
 					panic("_xen_bus_dmamem_alloc_range: "