[PATCH] Xen-SWIOTBL v0.8.3 used for Xen PCI pass through for PV guests.

[PATCH] Xen-SWIOTBL v0.8.3 used for Xen PCI pass through for PV guests.

[Konrad Rzeszutek Wilk]

These nineteen patches lay the groundwork for Xen Paravirtualized (PV)
domains to access PCI pass-through devices. These patches utilize the
SWIOTLB library modifications (http://lkml.org/lkml/2010/6/4/272).

The end user of this is the Xen PCI frontend and Xen PCI [1] which
require a DMA API "backend" that understands Xen's MMU. This allows the
PV domains to use PCI devices.

This patch-set is split in two groups. The first alter the Xen components,
while the second introduces the SWIOTLB-Xen.

The Xen components patches consist of:

 [PATCH 01/19] xen: use _PAGE_IOMAP in ioremap to do machine mappings
 [PATCH 02/19] xen: Allow unprivileged Xen domains to create iomap pages
 [PATCH 03/19] xen: Rename the balloon lock
 [PATCH 04/19] xen: Add xen_create_contiguous_region

which alter the Xen MMU, which by default utilizes a layer of indirection
wherein the PFN is translated to the Machine Frame Number (MFN) and vice-versa.
This is required to "fool" the guest in thinking its memory starts at PFN 0 and
goes up to the available amount.  While in the background, PFN 0 might as well be
MFN 1048576 (4GB).

For PCI/DMA API calls (ioremap, pci_map_page, etc) having a PFN != MFN is
not too good, so two new mechanisms are introduced:
  a). For PTEs which manipulate the PCI/DMA pages the PFN == MFN. This is done
      via utilizing the _PAGE_IOMAP flag to distinguish the PTE as an IO type.
  b). Allow a mechanism to "swizzle" the MFN's under a PFN to be under a certain
      bus width (say 32). This allows us to provide a mechanism for the
      SWIOTLB Xen to allocate memory for its pool that will be guaranteed to be
      under the 4GB mark.

The SWIOTLB-Xen adds a library that is only used if the machine is detected
to be running under Xen. It utilizes the SWIOTLB library bookkeeping functions
(swiotlb_tbl_*) and only deals with the virtual to [physical, bus] (and vice-versa)
address translations.

The diffstat:

 arch/x86/include/asm/xen/page.h        |    8 +-
 arch/x86/include/asm/xen/swiotlb-xen.h |   14 +
 arch/x86/kernel/pci-dma.c              |    7 +-
 arch/x86/xen/Kconfig                   |    4 +
 arch/x86/xen/Makefile                  |    1 +
 arch/x86/xen/enlighten.c               |    4 +
 arch/x86/xen/mmu.c                     |  291 ++++++++++++++++++-
 arch/x86/xen/pci-swiotlb-xen.c         |   58 ++++
 drivers/xen/balloon.c                  |   15 +-
 include/linux/swiotlb-xen.h            |   65 ++++
 include/xen/interface/memory.h         |   50 +++
 include/xen/xen-ops.h                  |    6 +
 lib/Makefile                           |    1 +
 lib/swiotlb-xen.c                      |  515 ++++++++++++++++++++++++++++++++
 14 files changed, 1018 insertions(+), 21 deletions(-)


[1]: The Xen PCI and Xen PCI fronted patches, which have not yet been
posted are available at git://git.kernel.org/pub/scm/linux/kernel/git/konrad/xen.git
under the branch devel/xen-pcifront-0.2. A merge of all required branches
is under 'devel/merge.2.6.35-rc3'. Or 'devel/merge.2.6.34'. I will post them quite soon.

[PATCH 01/19] xen: use _PAGE_IOMAP in ioremap to do machine mappings

[Konrad Rzeszutek Wilk]

From: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>

In a Xen domain, ioremap operates on machine addresses, not
pseudo-physical addresses.  We use _PAGE_IOMAP to determine whether a
mapping is intended for machine addresses.

[ Impact: allow Xen domain to map real hardware ]

Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 arch/x86/include/asm/xen/page.h |    8 +---
 arch/x86/xen/mmu.c              |   71 +++++++++++++++++++++++++++++++++++++-
 2 files changed, 71 insertions(+), 8 deletions(-)

diff --git a/arch/x86/include/asm/xen/page.h b/arch/x86/include/asm/xen/page.h
index 018a0a4..bf5f7d3 100644
--- a/arch/x86/include/asm/xen/page.h
+++ b/arch/x86/include/asm/xen/page.h
@@ -112,13 +112,9 @@ static inline xpaddr_t machine_to_phys(xmaddr_t machine)
  */
 static inline unsigned long mfn_to_local_pfn(unsigned long mfn)
 {
-	extern unsigned long max_mapnr;
 	unsigned long pfn = mfn_to_pfn(mfn);
-	if ((pfn < max_mapnr)
-	    && !xen_feature(XENFEAT_auto_translated_physmap)
-	    && (get_phys_to_machine(pfn) != mfn))
-		return max_mapnr; /* force !pfn_valid() */
-	/* XXX fixme; not true with sparsemem */
+	if (get_phys_to_machine(pfn) != mfn)
+		return -1; /* force !pfn_valid() */
 	return pfn;
 }
 
diff --git a/arch/x86/xen/mmu.c b/arch/x86/xen/mmu.c
index 914f046..a4dea9d 100644
--- a/arch/x86/xen/mmu.c
+++ b/arch/x86/xen/mmu.c
@@ -56,9 +56,11 @@
 #include <asm/xen/hypercall.h>
 #include <asm/xen/hypervisor.h>
 
+#include <xen/xen.h>
 #include <xen/page.h>
 #include <xen/interface/xen.h>
 #include <xen/interface/version.h>
+#include <xen/interface/memory.h>
 #include <xen/hvc-console.h>
 
 #include "multicalls.h"
@@ -377,6 +379,28 @@ static bool xen_page_pinned(void *ptr)
 	return PagePinned(page);
 }
 
+static bool xen_iomap_pte(pte_t pte)
+{
+	return xen_initial_domain() && (pte_flags(pte) & _PAGE_IOMAP);
+}
+
+static void xen_set_iomap_pte(pte_t *ptep, pte_t pteval)
+{
+	struct multicall_space mcs;
+	struct mmu_update *u;
+
+	mcs = xen_mc_entry(sizeof(*u));
+	u = mcs.args;
+
+	/* ptep might be kmapped when using 32-bit HIGHPTE */
+	u->ptr = arbitrary_virt_to_machine(ptep).maddr;
+	u->val = pte_val_ma(pteval);
+
+	MULTI_mmu_update(mcs.mc, mcs.args, 1, NULL, DOMID_IO);
+
+	xen_mc_issue(PARAVIRT_LAZY_MMU);
+}
+
 static void xen_extend_mmu_update(const struct mmu_update *update)
 {
 	struct multicall_space mcs;
@@ -453,6 +477,11 @@ void set_pte_mfn(unsigned long vaddr, unsigned long mfn, pgprot_t flags)
 void xen_set_pte_at(struct mm_struct *mm, unsigned long addr,
 		    pte_t *ptep, pte_t pteval)
 {
+	if (xen_iomap_pte(pteval)) {
+		xen_set_iomap_pte(ptep, pteval);
+		goto out;
+	}
+
 	ADD_STATS(set_pte_at, 1);
 //	ADD_STATS(set_pte_at_pinned, xen_page_pinned(ptep));
 	ADD_STATS(set_pte_at_current, mm == current->mm);
@@ -523,8 +552,25 @@ static pteval_t pte_pfn_to_mfn(pteval_t val)
 	return val;
 }
 
+static pteval_t iomap_pte(pteval_t val)
+{
+	if (val & _PAGE_PRESENT) {
+		unsigned long pfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
+		pteval_t flags = val & PTE_FLAGS_MASK;
+
+		/* We assume the pte frame number is a MFN, so
+		   just use it as-is. */
+		val = ((pteval_t)pfn << PAGE_SHIFT) | flags;
+	}
+
+	return val;
+}
+
 pteval_t xen_pte_val(pte_t pte)
 {
+	if (xen_initial_domain() && (pte.pte & _PAGE_IOMAP))
+		return pte.pte;
+
 	return pte_mfn_to_pfn(pte.pte);
 }
 PV_CALLEE_SAVE_REGS_THUNK(xen_pte_val);
@@ -537,7 +583,11 @@ PV_CALLEE_SAVE_REGS_THUNK(xen_pgd_val);
 
 pte_t xen_make_pte(pteval_t pte)
 {
-	pte = pte_pfn_to_mfn(pte);
+	if (unlikely(xen_initial_domain() && (pte & _PAGE_IOMAP)))
+		pte = iomap_pte(pte);
+	else
+		pte = pte_pfn_to_mfn(pte);
+
 	return native_make_pte(pte);
 }
 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pte);
@@ -593,6 +643,11 @@ void xen_set_pud(pud_t *ptr, pud_t val)
 
 void xen_set_pte(pte_t *ptep, pte_t pte)
 {
+	if (xen_iomap_pte(pte)) {
+		xen_set_iomap_pte(ptep, pte);
+		return;
+	}
+
 	ADD_STATS(pte_update, 1);
 //	ADD_STATS(pte_update_pinned, xen_page_pinned(ptep));
 	ADD_STATS(pte_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
@@ -609,6 +664,11 @@ void xen_set_pte(pte_t *ptep, pte_t pte)
 #ifdef CONFIG_X86_PAE
 void xen_set_pte_atomic(pte_t *ptep, pte_t pte)
 {
+	if (xen_iomap_pte(pte)) {
+		xen_set_iomap_pte(ptep, pte);
+		return;
+	}
+
 	set_64bit((u64 *)ptep, native_pte_val(pte));
 }
 
@@ -1811,9 +1871,16 @@ static void xen_set_fixmap(unsigned idx, phys_addr_t phys, pgprot_t prot)
 		pte = pfn_pte(phys, prot);
 		break;
 
-	default:
+	case FIX_PARAVIRT_BOOTMAP:
+		/* This is an MFN, but it isn't an IO mapping from the
+		   IO domain */
 		pte = mfn_pte(phys, prot);
 		break;
+
+	default:
+		/* By default, set_fixmap is used for hardware mappings */
+		pte = mfn_pte(phys, __pgprot(pgprot_val(prot) | _PAGE_IOMAP));
+		break;
 	}
 
 	__native_set_fixmap(idx, pte);
-- 
1.7.0.1

[PATCH 02/19] xen: Allow unprivileged Xen domains to create iomap pages

[Konrad Rzeszutek Wilk]

From: Alex Nixon <alex.nixon@citrix.com>

PV DomU domains are allowed to map hardware MFNs for PCI passthrough,
but are not generally allowed to map raw machine pages.  In particular,
various pieces of code try to map DMI and ACPI tables in the ISA ROM
range.  We disallow _PAGE_IOMAP for those mappings, so that they are
redirected to a set of local zeroed pages we reserve for that purpose.

[ Impact: prevent passthrough of ISA space, as we only allow PCI ]

Signed-off-by: Alex Nixon <alex.nixon@citrix.com>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 arch/x86/xen/enlighten.c |    4 ++++
 arch/x86/xen/mmu.c       |   18 +++++++++++++++---
 2 files changed, 19 insertions(+), 3 deletions(-)

diff --git a/arch/x86/xen/enlighten.c b/arch/x86/xen/enlighten.c
index 65d8d79..3254e8b 100644
--- a/arch/x86/xen/enlighten.c
+++ b/arch/x86/xen/enlighten.c
@@ -1145,6 +1145,10 @@ asmlinkage void __init xen_start_kernel(void)
 
 	pgd = (pgd_t *)xen_start_info->pt_base;
 
+	if (!xen_initial_domain())
+		__supported_pte_mask &= ~(_PAGE_PWT | _PAGE_PCD);
+
+	__supported_pte_mask |= _PAGE_IOMAP;
 	/* Don't do the full vcpu_info placement stuff until we have a
 	   possible map and a non-dummy shared_info. */
 	per_cpu(xen_vcpu, 0) = &HYPERVISOR_shared_info->vcpu_info[0];
diff --git a/arch/x86/xen/mmu.c b/arch/x86/xen/mmu.c
index a4dea9d..a5577f5 100644
--- a/arch/x86/xen/mmu.c
+++ b/arch/x86/xen/mmu.c
@@ -51,6 +51,7 @@
 #include <asm/mmu_context.h>
 #include <asm/setup.h>
 #include <asm/paravirt.h>
+#include <asm/e820.h>
 #include <asm/linkage.h>
 
 #include <asm/xen/hypercall.h>
@@ -381,7 +382,7 @@ static bool xen_page_pinned(void *ptr)
 
 static bool xen_iomap_pte(pte_t pte)
 {
-	return xen_initial_domain() && (pte_flags(pte) & _PAGE_IOMAP);
+	return pte_flags(pte) & _PAGE_IOMAP;
 }
 
 static void xen_set_iomap_pte(pte_t *ptep, pte_t pteval)
@@ -583,10 +584,21 @@ PV_CALLEE_SAVE_REGS_THUNK(xen_pgd_val);
 
 pte_t xen_make_pte(pteval_t pte)
 {
-	if (unlikely(xen_initial_domain() && (pte & _PAGE_IOMAP)))
+	phys_addr_t addr = (pte & PTE_PFN_MASK);
+
+	/*
+	 * Unprivileged domains are allowed to do IOMAPpings for
+	 * PCI passthrough, but not map ISA space.  The ISA
+	 * mappings are just dummy local mappings to keep other
+	 * parts of the kernel happy.
+	 */
+	if (unlikely(pte & _PAGE_IOMAP) &&
+	    (xen_initial_domain() || addr >= ISA_END_ADDRESS)) {
 		pte = iomap_pte(pte);
-	else
+	} else {
+		pte &= ~_PAGE_IOMAP;
 		pte = pte_pfn_to_mfn(pte);
+	}
 
 	return native_make_pte(pte);
 }
-- 
1.7.0.1

[PATCH 03/19] xen: Rename the balloon lock

[Konrad Rzeszutek Wilk]

From: Alex Nixon <alex.nixon@citrix.com>

* xen_create_contiguous_region needs access to the balloon lock to
  ensure memory doesn't change under its feet, so expose the balloon
  lock
* Change the name of the lock to xen_reservation_lock, to imply it's
  now less-specific usage.

[ Impact: cleanup ]

Signed-off-by: Alex Nixon <alex.nixon@citrix.com>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 arch/x86/xen/mmu.c             |    7 +++++++
 drivers/xen/balloon.c          |   15 ++++-----------
 include/xen/interface/memory.h |    8 ++++++++
 3 files changed, 19 insertions(+), 11 deletions(-)

diff --git a/arch/x86/xen/mmu.c b/arch/x86/xen/mmu.c
index a5577f5..9e0d82f 100644
--- a/arch/x86/xen/mmu.c
+++ b/arch/x86/xen/mmu.c
@@ -70,6 +70,13 @@
 
 #define MMU_UPDATE_HISTO	30
 
+/*
+ * Protects atomic reservation decrease/increase against concurrent increases.
+ * Also protects non-atomic updates of current_pages and driver_pages, and
+ * balloon lists.
+ */
+DEFINE_SPINLOCK(xen_reservation_lock);
+
 #ifdef CONFIG_XEN_DEBUG_FS
 
 static struct {
diff --git a/drivers/xen/balloon.c b/drivers/xen/balloon.c
index 1a0d8c2..500290b 100644
--- a/drivers/xen/balloon.c
+++ b/drivers/xen/balloon.c
@@ -85,13 +85,6 @@ static struct sys_device balloon_sysdev;
 
 static int register_balloon(struct sys_device *sysdev);
 
-/*
- * Protects atomic reservation decrease/increase against concurrent increases.
- * Also protects non-atomic updates of current_pages and driver_pages, and
- * balloon lists.
- */
-static DEFINE_SPINLOCK(balloon_lock);
-
 static struct balloon_stats balloon_stats;
 
 /* We increase/decrease in batches which fit in a page */
@@ -210,7 +203,7 @@ static int increase_reservation(unsigned long nr_pages)
 	if (nr_pages > ARRAY_SIZE(frame_list))
 		nr_pages = ARRAY_SIZE(frame_list);
 
-	spin_lock_irqsave(&balloon_lock, flags);
+	spin_lock_irqsave(&xen_reservation_lock, flags);
 
 	page = balloon_first_page();
 	for (i = 0; i < nr_pages; i++) {
@@ -254,7 +247,7 @@ static int increase_reservation(unsigned long nr_pages)
 	balloon_stats.current_pages += rc;
 
  out:
-	spin_unlock_irqrestore(&balloon_lock, flags);
+	spin_unlock_irqrestore(&xen_reservation_lock, flags);
 
 	return rc < 0 ? rc : rc != nr_pages;
 }
@@ -299,7 +292,7 @@ static int decrease_reservation(unsigned long nr_pages)
 	kmap_flush_unused();
 	flush_tlb_all();
 
-	spin_lock_irqsave(&balloon_lock, flags);
+	spin_lock_irqsave(&xen_reservation_lock, flags);
 
 	/* No more mappings: invalidate P2M and add to balloon. */
 	for (i = 0; i < nr_pages; i++) {
@@ -315,7 +308,7 @@ static int decrease_reservation(unsigned long nr_pages)
 
 	balloon_stats.current_pages -= nr_pages;
 
-	spin_unlock_irqrestore(&balloon_lock, flags);
+	spin_unlock_irqrestore(&xen_reservation_lock, flags);
 
 	return need_sleep;
 }
diff --git a/include/xen/interface/memory.h b/include/xen/interface/memory.h
index af36ead..e6adce6 100644
--- a/include/xen/interface/memory.h
+++ b/include/xen/interface/memory.h
@@ -9,6 +9,8 @@
 #ifndef __XEN_PUBLIC_MEMORY_H__
 #define __XEN_PUBLIC_MEMORY_H__
 
+#include <linux/spinlock.h>
+
 /*
  * Increase or decrease the specified domain's memory reservation. Returns a
  * -ve errcode on failure, or the # extents successfully allocated or freed.
@@ -142,4 +144,10 @@ struct xen_translate_gpfn_list {
 };
 DEFINE_GUEST_HANDLE_STRUCT(xen_translate_gpfn_list);
 
+
+/*
+ * Prevent the balloon driver from changing the memory reservation
+ * during a driver critical region.
+ */
+extern spinlock_t xen_reservation_lock;
 #endif /* __XEN_PUBLIC_MEMORY_H__ */
-- 
1.7.0.1

[PATCH 04/19] xen: Add xen_create_contiguous_region

[Konrad Rzeszutek Wilk]

From: Alex Nixon <alex.nixon@citrix.com>

A memory region must be physically contiguous in order to be accessed
through DMA.  This patch adds xen_create_contiguous_region, which
ensures a region of contiguous virtual memory is also physically
contiguous.

Based on Stephen Tweedie's port of the 2.6.18-xen version.

Remove contiguous_bitmap[] as it's no longer needed.

Ported from linux-2.6.18-xen.hg 707:e410857fd83c

[ Impact: add Xen-internal API to make pages phys-contig ]

Signed-off-by: Alex Nixon <alex.nixon@citrix.com>
Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 arch/x86/xen/mmu.c             |  201 ++++++++++++++++++++++++++++++++++++++++
 include/xen/interface/memory.h |   42 ++++++++
 include/xen/xen-ops.h          |    6 +
 3 files changed, 249 insertions(+), 0 deletions(-)

diff --git a/arch/x86/xen/mmu.c b/arch/x86/xen/mmu.c
index 9e0d82f..eb51402 100644
--- a/arch/x86/xen/mmu.c
+++ b/arch/x86/xen/mmu.c
@@ -53,6 +53,7 @@
 #include <asm/paravirt.h>
 #include <asm/e820.h>
 #include <asm/linkage.h>
+#include <asm/page.h>
 
 #include <asm/xen/hypercall.h>
 #include <asm/xen/hypervisor.h>
@@ -2027,6 +2028,206 @@ void __init xen_init_mmu_ops(void)
 	pv_mmu_ops = xen_mmu_ops;
 }
 
+/* Protected by xen_reservation_lock. */
+#define MAX_CONTIG_ORDER 9 /* 2MB */
+static unsigned long discontig_frames[1<<MAX_CONTIG_ORDER];
+
+#define VOID_PTE (mfn_pte(0, __pgprot(0)))
+static void xen_zap_pfn_range(unsigned long vaddr, unsigned int order,
+				unsigned long *in_frames,
+				unsigned long *out_frames)
+{
+	int i;
+	struct multicall_space mcs;
+
+	xen_mc_batch();
+	for (i = 0; i < (1UL<<order); i++, vaddr += PAGE_SIZE) {
+		mcs = __xen_mc_entry(0);
+
+		if (in_frames)
+			in_frames[i] = virt_to_mfn(vaddr);
+
+		MULTI_update_va_mapping(mcs.mc, vaddr, VOID_PTE, 0);
+		set_phys_to_machine(virt_to_pfn(vaddr), INVALID_P2M_ENTRY);
+
+		if (out_frames)
+			out_frames[i] = virt_to_pfn(vaddr);
+	}
+	xen_mc_issue(0);
+}
+
+/*
+ * Update the pfn-to-mfn mappings for a virtual address range, either to
+ * point to an array of mfns, or contiguously from a single starting
+ * mfn.
+ */
+static void xen_remap_exchanged_ptes(unsigned long vaddr, int order,
+				     unsigned long *mfns,
+				     unsigned long first_mfn)
+{
+	unsigned i, limit;
+	unsigned long mfn;
+
+	xen_mc_batch();
+
+	limit = 1u << order;
+	for (i = 0; i < limit; i++, vaddr += PAGE_SIZE) {
+		struct multicall_space mcs;
+		unsigned flags;
+
+		mcs = __xen_mc_entry(0);
+		if (mfns)
+			mfn = mfns[i];
+		else
+			mfn = first_mfn + i;
+
+		if (i < (limit - 1))
+			flags = 0;
+		else {
+			if (order == 0)
+				flags = UVMF_INVLPG | UVMF_ALL;
+			else
+				flags = UVMF_TLB_FLUSH | UVMF_ALL;
+		}
+
+		MULTI_update_va_mapping(mcs.mc, vaddr,
+				mfn_pte(mfn, PAGE_KERNEL), flags);
+
+		set_phys_to_machine(virt_to_pfn(vaddr), mfn);
+	}
+
+	xen_mc_issue(0);
+}
+
+/*
+ * Perform the hypercall to exchange a region of our pfns to point to
+ * memory with the required contiguous alignment.  Takes the pfns as
+ * input, and populates mfns as output.
+ *
+ * Returns a success code indicating whether the hypervisor was able to
+ * satisfy the request or not.
+ */
+static int xen_exchange_memory(unsigned long extents_in, unsigned int order_in,
+			       unsigned long *pfns_in,
+			       unsigned long extents_out,
+			       unsigned int order_out,
+			       unsigned long *mfns_out,
+			       unsigned int address_bits)
+{
+	long rc;
+	int success;
+
+	struct xen_memory_exchange exchange = {
+		.in = {
+			.nr_extents   = extents_in,
+			.extent_order = order_in,
+			.extent_start = pfns_in,
+			.domid        = DOMID_SELF
+		},
+		.out = {
+			.nr_extents   = extents_out,
+			.extent_order = order_out,
+			.extent_start = mfns_out,
+			.address_bits = address_bits,
+			.domid        = DOMID_SELF
+		}
+	};
+
+	BUG_ON(extents_in << order_in != extents_out << order_out);
+
+	rc = HYPERVISOR_memory_op(XENMEM_exchange, &exchange);
+	success = (exchange.nr_exchanged == extents_in);
+
+	BUG_ON(!success && ((exchange.nr_exchanged != 0) || (rc == 0)));
+	BUG_ON(success && (rc != 0));
+
+	return success;
+}
+
+int xen_create_contiguous_region(unsigned long vstart, unsigned int order,
+				 unsigned int address_bits)
+{
+	unsigned long *in_frames = discontig_frames, out_frame;
+	unsigned long  flags;
+	int            success;
+
+	/*
+	 * Currently an auto-translated guest will not perform I/O, nor will
+	 * it require PAE page directories below 4GB. Therefore any calls to
+	 * this function are redundant and can be ignored.
+	 */
+
+	if (xen_feature(XENFEAT_auto_translated_physmap))
+		return 0;
+
+	if (unlikely(order > MAX_CONTIG_ORDER))
+		return -ENOMEM;
+
+	memset((void *) vstart, 0, PAGE_SIZE << order);
+
+	vm_unmap_aliases();
+
+	spin_lock_irqsave(&xen_reservation_lock, flags);
+
+	/* 1. Zap current PTEs, remembering MFNs. */
+	xen_zap_pfn_range(vstart, order, in_frames, NULL);
+
+	/* 2. Get a new contiguous memory extent. */
+	out_frame = virt_to_pfn(vstart);
+	success = xen_exchange_memory(1UL << order, 0, in_frames,
+				      1, order, &out_frame,
+				      address_bits);
+
+	/* 3. Map the new extent in place of old pages. */
+	if (success)
+		xen_remap_exchanged_ptes(vstart, order, NULL, out_frame);
+	else
+		xen_remap_exchanged_ptes(vstart, order, in_frames, 0);
+
+	spin_unlock_irqrestore(&xen_reservation_lock, flags);
+
+	return success ? 0 : -ENOMEM;
+}
+EXPORT_SYMBOL_GPL(xen_create_contiguous_region);
+
+void xen_destroy_contiguous_region(unsigned long vstart, unsigned int order)
+{
+	unsigned long *out_frames = discontig_frames, in_frame;
+	unsigned long  flags;
+	int success;
+
+	if (xen_feature(XENFEAT_auto_translated_physmap))
+		return;
+
+	if (unlikely(order > MAX_CONTIG_ORDER))
+		return;
+
+	memset((void *) vstart, 0, PAGE_SIZE << order);
+
+	vm_unmap_aliases();
+
+	spin_lock_irqsave(&xen_reservation_lock, flags);
+
+	/* 1. Find start MFN of contiguous extent. */
+	in_frame = virt_to_mfn(vstart);
+
+	/* 2. Zap current PTEs. */
+	xen_zap_pfn_range(vstart, order, NULL, out_frames);
+
+	/* 3. Do the exchange for non-contiguous MFNs. */
+	success = xen_exchange_memory(1, order, &in_frame, 1UL << order,
+					0, out_frames, 0);
+
+	/* 4. Map new pages in place of old pages. */
+	if (success)
+		xen_remap_exchanged_ptes(vstart, order, out_frames, 0);
+	else
+		xen_remap_exchanged_ptes(vstart, order, NULL, in_frame);
+
+	spin_unlock_irqrestore(&xen_reservation_lock, flags);
+}
+EXPORT_SYMBOL_GPL(xen_destroy_contiguous_region);
+
 #ifdef CONFIG_XEN_DEBUG_FS
 
 static struct dentry *d_mmu_debug;
diff --git a/include/xen/interface/memory.h b/include/xen/interface/memory.h
index e6adce6..d3938d3 100644
--- a/include/xen/interface/memory.h
+++ b/include/xen/interface/memory.h
@@ -55,6 +55,48 @@ struct xen_memory_reservation {
 DEFINE_GUEST_HANDLE_STRUCT(xen_memory_reservation);
 
 /*
+ * An atomic exchange of memory pages. If return code is zero then
+ * @out.extent_list provides GMFNs of the newly-allocated memory.
+ * Returns zero on complete success, otherwise a negative error code.
+ * On complete success then always @nr_exchanged == @in.nr_extents.
+ * On partial success @nr_exchanged indicates how much work was done.
+ */
+#define XENMEM_exchange             11
+struct xen_memory_exchange {
+    /*
+     * [IN] Details of memory extents to be exchanged (GMFN bases).
+     * Note that @in.address_bits is ignored and unused.
+     */
+    struct xen_memory_reservation in;
+
+    /*
+     * [IN/OUT] Details of new memory extents.
+     * We require that:
+     *  1. @in.domid == @out.domid
+     *  2. @in.nr_extents  << @in.extent_order ==
+     *     @out.nr_extents << @out.extent_order
+     *  3. @in.extent_start and @out.extent_start lists must not overlap
+     *  4. @out.extent_start lists GPFN bases to be populated
+     *  5. @out.extent_start is overwritten with allocated GMFN bases
+     */
+    struct xen_memory_reservation out;
+
+    /*
+     * [OUT] Number of input extents that were successfully exchanged:
+     *  1. The first @nr_exchanged input extents were successfully
+     *     deallocated.
+     *  2. The corresponding first entries in the output extent list correctly
+     *     indicate the GMFNs that were successfully exchanged.
+     *  3. All other input and output extents are untouched.
+     *  4. If not all input exents are exchanged then the return code of this
+     *     command will be non-zero.
+     *  5. THIS FIELD MUST BE INITIALISED TO ZERO BY THE CALLER!
+     */
+    unsigned long nr_exchanged;
+};
+
+DEFINE_GUEST_HANDLE_STRUCT(xen_memory_exchange);
+/*
  * Returns the maximum machine frame number of mapped RAM in this system.
  * This command always succeeds (it never returns an error code).
  * arg == NULL.
diff --git a/include/xen/xen-ops.h b/include/xen/xen-ops.h
index 883a21b..d789c93 100644
--- a/include/xen/xen-ops.h
+++ b/include/xen/xen-ops.h
@@ -14,4 +14,10 @@ void xen_mm_unpin_all(void);
 void xen_timer_resume(void);
 void xen_arch_resume(void);
 
+extern unsigned long *xen_contiguous_bitmap;
+int xen_create_contiguous_region(unsigned long vstart, unsigned int order,
+				unsigned int address_bits);
+
+void xen_destroy_contiguous_region(unsigned long vstart, unsigned int order);
+
 #endif /* INCLUDE_XEN_OPS_H */
-- 
1.7.0.1

[PATCH 05/19] swiotlb-xen: Early skeleton code and explanation.

[Konrad Rzeszutek Wilk]

Adding in base code for address translation routines.

This patchset:

PV guests under Xen are running in an non-contiguous memory architecture.

When PCI pass-through is utilized, this necessitates an IOMMU for
translating bus (DMA) to virtual and vice-versa and also providing a
mechanism to have contiguous pages for device drivers operations (say DMA
operations).

Specifically, under Xen the Linux idea of pages is an illusion. It
assumes that pages start at zero and go up to the available memory. To
help with that, the Linux Xen MMU provides a lookup mechanism to
translate the page frame numbers (PFN) to machine frame numbers (MFN)
and vice-versa. The MFN are the "real" frame numbers. Furthermore
memory is not contiguous. Xen hypervisor stitches memory for guests
from different pools, which means there is no guarantee that PFN==MFN
and PFN+1==MFN+1. Lastly with Xen 4.0, pages (in debug mode) are
allocated in descending order (high to low), meaning the guest might
never get any MFN's under the 4GB mark.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 arch/x86/xen/Kconfig |    4 ++
 lib/Makefile         |    1 +
 lib/swiotlb-xen.c    |  118 ++++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 123 insertions(+), 0 deletions(-)
 create mode 100644 lib/swiotlb-xen.c

diff --git a/arch/x86/xen/Kconfig b/arch/x86/xen/Kconfig
index b83e119..faed31d 100644
--- a/arch/x86/xen/Kconfig
+++ b/arch/x86/xen/Kconfig
@@ -36,3 +36,7 @@ config XEN_DEBUG_FS
 	help
 	  Enable statistics output and various tuning options in debugfs.
 	  Enabling this option may incur a significant performance overhead.
+
+config SWIOTLB_XEN
+	def_bool y
+	depends on XEN && SWIOTLB
diff --git a/lib/Makefile b/lib/Makefile
index 0d40152..c25f6bf 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -80,6 +80,7 @@ obj-$(CONFIG_SMP) += percpu_counter.o
 obj-$(CONFIG_AUDIT_GENERIC) += audit.o
 
 obj-$(CONFIG_SWIOTLB) += swiotlb.o
+obj-$(CONFIG_SWIOTLB_XEN) += swiotlb-xen.o
 obj-$(CONFIG_IOMMU_HELPER) += iommu-helper.o
 obj-$(CONFIG_FAULT_INJECTION) += fault-inject.o
 
diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
new file mode 100644
index 0000000..11207e0
--- /dev/null
+++ b/lib/swiotlb-xen.c
@@ -0,0 +1,118 @@
+/*
+ *  Copyright 2010
+ *  by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
+ *
+ * This code provides a IOMMU for Xen PV guests with PCI passthrough.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License v2.0 as published by
+ * the Free Software Foundation
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * PV guests under Xen are running in an non-contiguous memory architecture.
+ *
+ * When PCI pass-through is utilized, this necessitates an IOMMU for
+ * translating bus (DMA) to virtual and vice-versa and also providing a
+ * mechanism to have contiguous pages for device drivers operations (say DMA
+ * operations).
+ *
+ * Specifically, under Xen the Linux idea of pages is an illusion. It
+ * assumes that pages start at zero and go up to the available memory. To
+ * help with that, the Linux Xen MMU provides a lookup mechanism to
+ * translate the page frame numbers (PFN) to machine frame numbers (MFN)
+ * and vice-versa. The MFN are the "real" frame numbers. Furthermore
+ * memory is not contiguous. Xen hypervisor stitches memory for guests
+ * from different pools, which means there is no guarantee that PFN==MFN
+ * and PFN+1==MFN+1. Lastly with Xen 4.0, pages (in debug mode) are
+ * allocated in descending order (high to low), meaning the guest might
+ * never get any MFN's under the 4GB mark.
+ *
+ */
+
+#include <linux/dma-mapping.h>
+#include <xen/page.h>
+
+/*
+ * Used to do a quick range check in swiotlb_tbl_unmap_single and
+ * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
+ * API.
+ */
+
+static char *xen_io_tlb_start, *xen_io_tlb_end;
+
+static dma_addr_t xen_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
+{
+	return phys_to_machine(XPADDR(paddr)).maddr;;
+}
+
+static phys_addr_t xen_bus_to_phys(struct device *hwdev, dma_addr_t baddr)
+{
+	return machine_to_phys(XMADDR(baddr)).paddr;
+}
+
+static dma_addr_t xen_virt_to_bus(struct device *hwdev,
+				  void *address)
+{
+	return xen_phys_to_bus(hwdev, virt_to_phys(address));
+}
+
+static int check_pages_physically_contiguous(unsigned long pfn,
+					     unsigned int offset,
+					     size_t length)
+{
+	unsigned long next_mfn;
+	int i;
+	int nr_pages;
+
+	next_mfn = pfn_to_mfn(pfn);
+	nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
+
+	for (i = 1; i < nr_pages; i++) {
+		if (pfn_to_mfn(++pfn) != ++next_mfn)
+			return 0;
+	}
+	return 1;
+}
+
+static int range_straddles_page_boundary(phys_addr_t p, size_t size)
+{
+	unsigned long pfn = PFN_DOWN(p);
+	unsigned int offset = p & ~PAGE_MASK;
+
+	if (offset + size <= PAGE_SIZE)
+		return 0;
+	if (check_pages_physically_contiguous(pfn, offset, size))
+		return 0;
+	return 1;
+}
+
+static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
+{
+	unsigned long mfn = PFN_DOWN(dma_addr);
+	unsigned long pfn = mfn_to_local_pfn(mfn);
+	phys_addr_t paddr;
+
+	/* If the address is outside our domain, it CAN
+	 * have the same virtual address as another address
+	 * in our domain. Therefore _only_ check address within our domain.
+	 */
+	if (pfn_valid(pfn)) {
+		paddr = PFN_PHYS(pfn);
+		return paddr >= virt_to_phys(xen_io_tlb_start) &&
+		       paddr < virt_to_phys(xen_io_tlb_end);
+	}
+	return 0;
+}
+
+static void *
+xen_map_single(struct device *hwdev, phys_addr_t phys, size_t size,
+	       enum dma_data_direction dir)
+{
+	u64 start_dma_addr = xen_virt_to_bus(hwdev, xen_io_tlb_start);
+
+	return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
+}
-- 
1.7.0.1

[PATCH 06/19] swiotlb-xen: Copied swiotlb.c in, added xen_ prefix.

[Konrad Rzeszutek Wilk]

Furthemore converted all of the EXPORT_SYMBOL to EXPORT_SYMBOL_GPL,
ran checkpatch and fixed the outstanding issues. Lastly added temporary
function decleration and macro defines which will be removed in
later commits.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |  349 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 files changed, 349 insertions(+), 0 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index 11207e0..c177f32 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -44,6 +44,14 @@
 
 static char *xen_io_tlb_start, *xen_io_tlb_end;
 
+/* Temporary scaffolding. Will be removed later. */
+void
+xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
+			   int nelems, enum dma_data_direction dir,
+			   struct dma_attrs *attrs);
+#define swiotlb_full(dev, size, dir, panic)
+#define io_tlb_overflow_buffer DMA_ERROR_CODE
+
 static dma_addr_t xen_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
 {
 	return phys_to_machine(XPADDR(paddr)).maddr;;
@@ -116,3 +124,344 @@ xen_map_single(struct device *hwdev, phys_addr_t phys, size_t size,
 
 	return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
 }
+
+void *
+xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
+			   dma_addr_t *dma_handle, gfp_t flags)
+{
+	dma_addr_t dev_addr;
+	void *ret;
+	int order = get_order(size);
+	u64 dma_mask = DMA_BIT_MASK(32);
+
+	if (hwdev && hwdev->coherent_dma_mask)
+		dma_mask = hwdev->coherent_dma_mask;
+
+	ret = (void *)__get_free_pages(flags, order);
+	if (ret && xen_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
+		/*
+		 * The allocated memory isn't reachable by the device.
+		 */
+		free_pages((unsigned long) ret, order);
+		ret = NULL;
+	}
+	if (!ret) {
+		/*
+		 * We are either out of memory or the device can't DMA to
+		 * GFP_DMA memory; fall back on map_single(), which
+		 * will grab memory from the lowest available address range.
+		 */
+		ret = xen_map_single(hwdev, 0, size, DMA_FROM_DEVICE);
+		if (!ret)
+			return NULL;
+	}
+
+	memset(ret, 0, size);
+	dev_addr = xen_virt_to_bus(hwdev, ret);
+
+	/* Confirm address can be DMA'd by device */
+	if (dev_addr + size - 1 > dma_mask) {
+		printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
+		       (unsigned long long)dma_mask,
+		       (unsigned long long)dev_addr);
+
+		/* DMA_TO_DEVICE to avoid memcpy in unmap_single */
+		swiotlb_tbl_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
+		return NULL;
+	}
+	*dma_handle = dev_addr;
+	return ret;
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
+
+void
+xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
+			  dma_addr_t dev_addr)
+{
+	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
+
+	WARN_ON(irqs_disabled());
+	if (!is_xen_swiotlb_buffer(paddr))
+		free_pages((unsigned long)vaddr, get_order(size));
+	else
+		/* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
+		swiotlb_tbl_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
+
+/*
+ * Map a single buffer of the indicated size for DMA in streaming mode.  The
+ * physical address to use is returned.
+ *
+ * Once the device is given the dma address, the device owns this memory until
+ * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
+ */
+dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
+				unsigned long offset, size_t size,
+				enum dma_data_direction dir,
+				struct dma_attrs *attrs)
+{
+	phys_addr_t phys = page_to_phys(page) + offset;
+	dma_addr_t dev_addr = phys_to_dma(dev, phys);
+	void *map;
+
+	BUG_ON(dir == DMA_NONE);
+	/*
+	 * If the address happens to be in the device's DMA window,
+	 * we can safely return the device addr and not worry about bounce
+	 * buffering it.
+	 */
+	if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
+		return dev_addr;
+
+	/*
+	 * Oh well, have to allocate and map a bounce buffer.
+	 */
+	map = xen_map_single(dev, phys, size, dir);
+	if (!map) {
+		swiotlb_full(dev, size, dir, 1);
+		map = io_tlb_overflow_buffer;
+	}
+
+	dev_addr = xen_virt_to_bus(dev, map);
+
+	/*
+	 * Ensure that the address returned is DMA'ble
+	 */
+	if (!dma_capable(dev, dev_addr, size))
+		panic("map_single: bounce buffer is not DMA'ble");
+
+	return dev_addr;
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
+
+/*
+ * Unmap a single streaming mode DMA translation.  The dma_addr and size must
+ * match what was provided for in a previous xen_swiotlb_map_page call.  All
+ * other usages are undefined.
+ *
+ * After this call, reads by the cpu to the buffer are guaranteed to see
+ * whatever the device wrote there.
+ */
+static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
+			     size_t size, enum dma_data_direction dir)
+{
+	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
+
+	BUG_ON(dir == DMA_NONE);
+
+	if (is_xen_swiotlb_buffer(paddr)) {
+		swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
+		return;
+	}
+
+	if (dir != DMA_FROM_DEVICE)
+		return;
+
+	/*
+	 * phys_to_virt doesn't work with hihgmem page but we could
+	 * call dma_mark_clean() with hihgmem page here. However, we
+	 * are fine since dma_mark_clean() is null on POWERPC. We can
+	 * make dma_mark_clean() take a physical address if necessary.
+	 */
+	dma_mark_clean(phys_to_virt(paddr), size);
+}
+
+void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
+			    size_t size, enum dma_data_direction dir,
+			    struct dma_attrs *attrs)
+{
+	xen_unmap_single(hwdev, dev_addr, size, dir);
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
+
+/*
+ * Make physical memory consistent for a single streaming mode DMA translation
+ * after a transfer.
+ *
+ * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
+ * using the cpu, yet do not wish to teardown the dma mapping, you must
+ * call this function before doing so.  At the next point you give the dma
+ * address back to the card, you must first perform a
+ * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
+ */
+static void
+xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
+			size_t size, enum dma_data_direction dir,
+			enum dma_sync_target target)
+{
+	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
+
+	BUG_ON(dir == DMA_NONE);
+
+	if (is_xen_swiotlb_buffer(paddr)) {
+		swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
+				       target);
+		return;
+	}
+
+	if (dir != DMA_FROM_DEVICE)
+		return;
+
+	dma_mark_clean(phys_to_virt(paddr), size);
+}
+
+void
+xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
+				size_t size, enum dma_data_direction dir)
+{
+	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
+
+void
+xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
+				   size_t size, enum dma_data_direction dir)
+{
+	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
+
+/*
+ * Map a set of buffers described by scatterlist in streaming mode for DMA.
+ * This is the scatter-gather version of the above xen_swiotlb_map_page
+ * interface.  Here the scatter gather list elements are each tagged with the
+ * appropriate dma address and length.  They are obtained via
+ * sg_dma_{address,length}(SG).
+ *
+ * NOTE: An implementation may be able to use a smaller number of
+ *       DMA address/length pairs than there are SG table elements.
+ *       (for example via virtual mapping capabilities)
+ *       The routine returns the number of addr/length pairs actually
+ *       used, at most nents.
+ *
+ * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
+ * same here.
+ */
+int
+xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
+			 int nelems, enum dma_data_direction dir,
+			 struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(dir == DMA_NONE);
+
+	for_each_sg(sgl, sg, nelems, i) {
+		phys_addr_t paddr = sg_phys(sg);
+		dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
+
+		if (swiotlb_force ||
+		    !dma_capable(hwdev, dev_addr, sg->length)) {
+			void *map = xen_map_single(hwdev, sg_phys(sg),
+						   sg->length, dir);
+			if (!map) {
+				/* Don't panic here, we expect map_sg users
+				   to do proper error handling. */
+				swiotlb_full(hwdev, sg->length, dir, 0);
+				xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
+							   attrs);
+				sgl[0].dma_length = 0;
+				return 0;
+			}
+			sg->dma_address = xen_virt_to_bus(hwdev, map);
+		} else
+			sg->dma_address = dev_addr;
+		sg->dma_length = sg->length;
+	}
+	return nelems;
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
+
+int
+xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
+		   enum dma_data_direction dir)
+{
+	return xen_swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg);
+
+/*
+ * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
+ * concerning calls here are the same as for swiotlb_unmap_page() above.
+ */
+void
+xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
+			   int nelems, enum dma_data_direction dir,
+			   struct dma_attrs *attrs)
+{
+	struct scatterlist *sg;
+	int i;
+
+	BUG_ON(dir == DMA_NONE);
+
+	for_each_sg(sgl, sg, nelems, i)
+		xen_unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
+
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
+
+void
+xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
+		     enum dma_data_direction dir)
+{
+	return xen_swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg);
+
+/*
+ * Make physical memory consistent for a set of streaming mode DMA translations
+ * after a transfer.
+ *
+ * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
+ * and usage.
+ */
+static void
+xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
+		    int nelems, enum dma_data_direction dir,
+		    enum dma_sync_target target)
+{
+	struct scatterlist *sg;
+	int i;
+
+	for_each_sg(sgl, sg, nelems, i)
+		xen_swiotlb_sync_single(hwdev, sg->dma_address,
+					sg->dma_length, dir, target);
+}
+
+void
+xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
+			    int nelems, enum dma_data_direction dir)
+{
+	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
+
+void
+xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
+			       int nelems, enum dma_data_direction dir)
+{
+	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
+
+int
+xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
+{
+	return (dma_addr == xen_virt_to_bus(hwdev, io_tlb_overflow_buffer));
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
+
+/*
+ * Return whether the given device DMA address mask can be supported
+ * properly.  For example, if your device can only drive the low 24-bits
+ * during bus mastering, then you would pass 0x00ffffff as the mask to
+ * this function.
+ */
+int
+xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
+{
+	return xen_virt_to_bus(hwdev, xen_io_tlb_end - 1) <= mask;
+}
+EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);
-- 
1.7.0.1

[PATCH 07/19] swiotlb-xen: Make 'xen_swiotlb_alloc_coherent' work.

[Konrad Rzeszutek Wilk]

We allocate the requested page from wherever (we ignore the DMA32 flags)
and after allocation make a call for the kernel & hypervisor to replace
the memory at the virtual addresses with physical memory that is under
the 4GB mark.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |   55 ++++++++++++++++++++++------------------------------
 1 files changed, 23 insertions(+), 32 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index c177f32..89443e4 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -35,6 +35,7 @@
 
 #include <linux/dma-mapping.h>
 #include <xen/page.h>
+#include <xen/xen-ops.h>
 
 /*
  * Used to do a quick range check in swiotlb_tbl_unmap_single and
@@ -129,47 +130,37 @@ void *
 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 			   dma_addr_t *dma_handle, gfp_t flags)
 {
-	dma_addr_t dev_addr;
 	void *ret;
 	int order = get_order(size);
 	u64 dma_mask = DMA_BIT_MASK(32);
+	unsigned long vstart;
 
-	if (hwdev && hwdev->coherent_dma_mask)
-		dma_mask = hwdev->coherent_dma_mask;
+	/*
+	* Ignore region specifiers - the kernel's ideas of
+	* pseudo-phys memory layout has nothing to do with the
+	* machine physical layout.  We can't allocate highmem
+	* because we can't return a pointer to it.
+	*/
+	flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
 
-	ret = (void *)__get_free_pages(flags, order);
-	if (ret && xen_virt_to_bus(hwdev, ret) + size - 1 > dma_mask) {
-		/*
-		 * The allocated memory isn't reachable by the device.
-		 */
-		free_pages((unsigned long) ret, order);
-		ret = NULL;
-	}
-	if (!ret) {
-		/*
-		 * We are either out of memory or the device can't DMA to
-		 * GFP_DMA memory; fall back on map_single(), which
-		 * will grab memory from the lowest available address range.
-		 */
-		ret = xen_map_single(hwdev, 0, size, DMA_FROM_DEVICE);
-		if (!ret)
-			return NULL;
-	}
+	if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
+		return ret;
 
-	memset(ret, 0, size);
-	dev_addr = xen_virt_to_bus(hwdev, ret);
+	vstart = __get_free_pages(flags, order);
+	ret = (void *)vstart;
 
-	/* Confirm address can be DMA'd by device */
-	if (dev_addr + size - 1 > dma_mask) {
-		printk("hwdev DMA mask = 0x%016Lx, dev_addr = 0x%016Lx\n",
-		       (unsigned long long)dma_mask,
-		       (unsigned long long)dev_addr);
+	if (hwdev && hwdev->coherent_dma_mask)
+		dma_mask = hwdev->coherent_dma_mask;
 
-		/* DMA_TO_DEVICE to avoid memcpy in unmap_single */
-		swiotlb_tbl_unmap_single(hwdev, ret, size, DMA_TO_DEVICE);
-		return NULL;
+	if (ret) {
+		if (xen_create_contiguous_region(vstart, order,
+						 fls64(dma_mask)) != 0) {
+			free_pages(vstart, order);
+			return NULL;
+		}
+		memset(ret, 0, size);
+		*dma_handle = virt_to_machine(ret).maddr;
 	}
-	*dma_handle = dev_addr;
 	return ret;
 }
 EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
-- 
1.7.0.1

[PATCH 08/19] swiotlb-xen: Don't allocate DMA-memory beyond 4GB in 32-bit mode.

[Konrad Rzeszutek Wilk]

From: Ronny.Hegewald@online.de <Ronny.Hegewald@online.de>

When running in 32-bit PV environment various drivers try to allocate
a coherent DMA memory and the Xen SWIOTLB would return memory beyond 4GB.

On bare-metal the coherent DMA-memory using the native SWIOTLB would
always allocate memory inside the 32-bit address-range by calling
dma_alloc_coherent_mask.

The attached patch adds the same functionality to Xen SWIOTLB.

Signed-off-by: Ronny Hegewald <Ronny.Hegewald@online.de>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index 89443e4..2ebfcbd 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -150,7 +150,7 @@ xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 	ret = (void *)vstart;
 
 	if (hwdev && hwdev->coherent_dma_mask)
-		dma_mask = hwdev->coherent_dma_mask;
+		dma_mask = dma_alloc_coherent_mask(hwdev, flags);
 
 	if (ret) {
 		if (xen_create_contiguous_region(vstart, order,
-- 
1.7.0.1

[PATCH 09/19] swiotlb-xen: Make 'xen_swiotlb_free_coherent' work.

[Konrad Rzeszutek Wilk]

See "swiotlb-xen: Make 'xen_swiotlb_free_coherent' work." for
detailed description.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |   15 +++++++--------
 1 files changed, 7 insertions(+), 8 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index 2ebfcbd..2640052 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -169,14 +169,13 @@ void
 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
 			  dma_addr_t dev_addr)
 {
-	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
+	int order = get_order(size);
+
+	if (dma_release_from_coherent(hwdev, order, vaddr))
+		return;
 
-	WARN_ON(irqs_disabled());
-	if (!is_xen_swiotlb_buffer(paddr))
-		free_pages((unsigned long)vaddr, get_order(size));
-	else
-		/* DMA_TO_DEVICE to avoid memcpy in swiotlb_tbl_unmap_single */
-		swiotlb_tbl_unmap_single(hwdev, vaddr, size, DMA_TO_DEVICE);
+	xen_destroy_contiguous_region((unsigned long)vaddr, order);
+	free_pages((unsigned long)vaddr, order);
 }
 EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
 
-- 
1.7.0.1

[PATCH 10/19] swiotlb-xen: Make 'xen_swiotlb_[map|unmap]_page' work.

[Konrad Rzeszutek Wilk]

We utilize the SWIOTLB proper book-keeping functions and compared
to bare-metal SWIOTLB, we use a different virt->bus address
translation. This is necessary as under Xen, the PFN is a pseudo-number
that is not necessarily the true MFN. Because of that, successive pages
might not physically contiguous, meaning mfn++ != pfn_to_mfn(pfn++), so
we have to check for that too.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |   10 ++++++----
 1 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index 2640052..efafcce 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -192,7 +192,7 @@ dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
 				struct dma_attrs *attrs)
 {
 	phys_addr_t phys = page_to_phys(page) + offset;
-	dma_addr_t dev_addr = phys_to_dma(dev, phys);
+	dma_addr_t dev_addr = xen_phys_to_bus(dev, phys);
 	void *map;
 
 	BUG_ON(dir == DMA_NONE);
@@ -201,7 +201,8 @@ dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
 	 * we can safely return the device addr and not worry about bounce
 	 * buffering it.
 	 */
-	if (dma_capable(dev, dev_addr, size) && !swiotlb_force)
+	if (dma_capable(dev, dev_addr, size) &&
+	    !range_straddles_page_boundary(phys, size) && !swiotlb_force)
 		return dev_addr;
 
 	/*
@@ -236,11 +237,12 @@ EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
 			     size_t size, enum dma_data_direction dir)
 {
-	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
+	phys_addr_t paddr = xen_bus_to_phys(hwdev, dev_addr);
 
 	BUG_ON(dir == DMA_NONE);
 
-	if (is_xen_swiotlb_buffer(paddr)) {
+	/* NOTE: We use dev_addr here, not paddr! */
+	if (is_xen_swiotlb_buffer(dev_addr)) {
 		swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
 		return;
 	}
-- 
1.7.0.1

[PATCH 11/19] swiotlb-xen: Make 'xen_swiotlb_sync_single' work.

[Konrad Rzeszutek Wilk]

Refer to "swiotlb-xen: Make 'xen_swiotlb_[map|unmap]_page' work." for
details why we want to use our own address translation mechanism.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |    5 +++--
 1 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index efafcce..ca6c537 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -282,11 +282,12 @@ xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
 			size_t size, enum dma_data_direction dir,
 			enum dma_sync_target target)
 {
-	phys_addr_t paddr = dma_to_phys(hwdev, dev_addr);
+	phys_addr_t paddr = xen_bus_to_phys(hwdev, dev_addr);
 
 	BUG_ON(dir == DMA_NONE);
 
-	if (is_xen_swiotlb_buffer(paddr)) {
+	/* NOTE: We use dev_addr here, not paddr! */
+	if (is_xen_swiotlb_buffer(dev_addr)) {
 		swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
 				       target);
 		return;
-- 
1.7.0.1

[PATCH 12/19] swiotlb-xen: Make 'xen_swiotlb_map_sg_attrs' work.

[Konrad Rzeszutek Wilk]

Refer to "swiotlb-xen: Make 'xen_swiotlb_[map|unmap]_page' work." for
details why we want to use our own address translation mechanism.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |    5 +++--
 1 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index ca6c537..d457f7e 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -343,10 +343,11 @@ xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
 
 	for_each_sg(sgl, sg, nelems, i) {
 		phys_addr_t paddr = sg_phys(sg);
-		dma_addr_t dev_addr = phys_to_dma(hwdev, paddr);
+		dma_addr_t dev_addr = xen_phys_to_bus(hwdev, paddr);
 
 		if (swiotlb_force ||
-		    !dma_capable(hwdev, dev_addr, sg->length)) {
+		    !dma_capable(hwdev, dev_addr, sg->length) ||
+		    range_straddles_page_boundary(paddr, sg->length)) {
 			void *map = xen_map_single(hwdev, sg_phys(sg),
 						   sg->length, dir);
 			if (!map) {
-- 
1.7.0.1

[PATCH 13/19] swiotlb-xen: Remove io_tlb_overflow usage.

[Konrad Rzeszutek Wilk]

We are sheding the usage of the overflow buffer and returning
the value of zero (DMA_ERROR_CODE), following the lead of the
Intel and AMD IOMMU code.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |   13 ++++---------
 1 files changed, 4 insertions(+), 9 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index d457f7e..e0f944e 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -50,8 +50,6 @@ void
 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
 			   int nelems, enum dma_data_direction dir,
 			   struct dma_attrs *attrs);
-#define swiotlb_full(dev, size, dir, panic)
-#define io_tlb_overflow_buffer DMA_ERROR_CODE
 
 static dma_addr_t xen_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
 {
@@ -209,10 +207,8 @@ dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
 	 * Oh well, have to allocate and map a bounce buffer.
 	 */
 	map = xen_map_single(dev, phys, size, dir);
-	if (!map) {
-		swiotlb_full(dev, size, dir, 1);
-		map = io_tlb_overflow_buffer;
-	}
+	if (!map)
+		return DMA_ERROR_CODE;
 
 	dev_addr = xen_virt_to_bus(dev, map);
 
@@ -353,11 +349,10 @@ xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
 			if (!map) {
 				/* Don't panic here, we expect map_sg users
 				   to do proper error handling. */
-				swiotlb_full(hwdev, sg->length, dir, 0);
 				xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
 							   attrs);
 				sgl[0].dma_length = 0;
-				return 0;
+				return DMA_ERROR_CODE;
 			}
 			sg->dma_address = xen_virt_to_bus(hwdev, map);
 		} else
@@ -443,7 +438,7 @@ EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
 int
 xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
 {
-	return (dma_addr == xen_virt_to_bus(hwdev, io_tlb_overflow_buffer));
+	return !dma_addr;
 }
 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
 
-- 
1.7.0.1

[PATCH 14/19] swiotlb-xen: Add 'xen_swiotlb_init' function.

[Konrad Rzeszutek Wilk]

We utilize the alloc_bootmem to allocate any memory (even
past the 4GB) and then follow it with a xen kernel
to replace the memory pointed by the virtual address with memory
that is under the 4GB mark.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |   68 ++++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 files changed, 67 insertions(+), 1 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index e0f944e..3c5bfde 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -33,10 +33,10 @@
  *
  */
 
+#include <linux/bootmem.h>
 #include <linux/dma-mapping.h>
 #include <xen/page.h>
 #include <xen/xen-ops.h>
-
 /*
  * Used to do a quick range check in swiotlb_tbl_unmap_single and
  * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
@@ -44,6 +44,7 @@
  */
 
 static char *xen_io_tlb_start, *xen_io_tlb_end;
+static unsigned long xen_io_tlb_nslabs;
 
 /* Temporary scaffolding. Will be removed later. */
 void
@@ -124,6 +125,70 @@ xen_map_single(struct device *hwdev, phys_addr_t phys, size_t size,
 	return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
 }
 
+static int max_dma_bits = 32;
+
+static int
+xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
+{
+	int i, rc;
+	int dma_bits;
+
+	dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
+
+	i = 0;
+	do {
+		int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
+
+		do {
+			rc = xen_create_contiguous_region(
+				(unsigned long)buf + (i << IO_TLB_SHIFT),
+				get_order(slabs << IO_TLB_SHIFT),
+				dma_bits);
+		} while (rc && dma_bits++ < max_dma_bits);
+		if (rc)
+			return rc;
+
+		i += slabs;
+	} while (i < nslabs);
+	return 0;
+}
+
+void __init xen_swiotlb_init(int verbose)
+{
+	unsigned long bytes;
+	int rc;
+
+	xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
+	xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
+
+	bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
+
+	/*
+	 * Get IO TLB memory from any location.
+	 */
+	xen_io_tlb_start = alloc_bootmem(bytes);
+	if (!xen_io_tlb_start)
+		panic("Cannot allocate SWIOTLB buffer");
+
+	xen_io_tlb_end = xen_io_tlb_start + bytes;
+	/*
+	 * And replace that memory with pages under 4GB.
+	 */
+	rc = xen_swiotlb_fixup(xen_io_tlb_start,
+			       bytes,
+			       xen_io_tlb_nslabs);
+	if (rc)
+		goto error;
+
+	swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, verbose);
+
+	return;
+error:
+	panic("DMA(%d): Failed to exchange pages allocated for DMA with Xen! "\
+	      "We either don't have the permission or you do not have enough"\
+	      "free memory under 4GB!\n", rc);
+}
+
 void *
 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
 			   dma_addr_t *dma_handle, gfp_t flags)
@@ -177,6 +242,7 @@ xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
 }
 EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
 
+
 /*
  * Map a single buffer of the indicated size for DMA in streaming mode.  The
  * physical address to use is returned.
-- 
1.7.0.1

[PATCH 15/19] swiotlb-xen: Put 'swiotlb-xen.c' function declarations in the header.

[Konrad Rzeszutek Wilk]

We also remove the temporary scaffolding required to compile
swiotlb-xen.c file.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 include/linux/swiotlb-xen.h |   65 +++++++++++++++++++++++++++++++++++++++++++
 lib/swiotlb-xen.c           |    7 +----
 2 files changed, 66 insertions(+), 6 deletions(-)
 create mode 100644 include/linux/swiotlb-xen.h

diff --git a/include/linux/swiotlb-xen.h b/include/linux/swiotlb-xen.h
new file mode 100644
index 0000000..21083f2
--- /dev/null
+++ b/include/linux/swiotlb-xen.h
@@ -0,0 +1,65 @@
+#ifndef __LINUX_SWIOTLB_XEN_H
+#define __LINUX_SWIOTLB_XEN_H
+
+#include <linux/swiotlb.h>
+
+extern void xen_swiotlb_init(int verbose);
+
+extern void
+*xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
+			    dma_addr_t *dma_handle, gfp_t flags);
+
+extern void
+xen_swiotlb_free_coherent(struct device *hwdev, size_t size,
+			  void *vaddr, dma_addr_t dma_handle);
+
+extern dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
+				       unsigned long offset, size_t size,
+				       enum dma_data_direction dir,
+				       struct dma_attrs *attrs);
+
+extern void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
+				   size_t size, enum dma_data_direction dir,
+				   struct dma_attrs *attrs);
+
+extern int
+xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sg, int nents,
+		   enum dma_data_direction dir);
+
+extern void
+xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sg, int nents,
+		     enum dma_data_direction dir);
+
+extern int
+xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
+			 int nelems, enum dma_data_direction dir,
+			 struct dma_attrs *attrs);
+
+extern void
+xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
+			   int nelems, enum dma_data_direction dir,
+			   struct dma_attrs *attrs);
+
+extern void
+xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
+				size_t size, enum dma_data_direction dir);
+
+extern void
+xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
+			    int nelems, enum dma_data_direction dir);
+
+extern void
+xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
+				   size_t size, enum dma_data_direction dir);
+
+extern void
+xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
+			       int nelems, enum dma_data_direction dir);
+
+extern int
+xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr);
+
+extern int
+xen_swiotlb_dma_supported(struct device *hwdev, u64 mask);
+
+#endif /* __LINUX_SWIOTLB_XEN_H */
diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index 3c5bfde..6bba1a5 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -35,6 +35,7 @@
 
 #include <linux/bootmem.h>
 #include <linux/dma-mapping.h>
+#include <linux/swiotlb-xen.h>
 #include <xen/page.h>
 #include <xen/xen-ops.h>
 /*
@@ -46,12 +47,6 @@
 static char *xen_io_tlb_start, *xen_io_tlb_end;
 static unsigned long xen_io_tlb_nslabs;
 
-/* Temporary scaffolding. Will be removed later. */
-void
-xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
-			   int nelems, enum dma_data_direction dir,
-			   struct dma_attrs *attrs);
-
 static dma_addr_t xen_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
 {
 	return phys_to_machine(XPADDR(paddr)).maddr;;
-- 
1.7.0.1

[PATCH 16/19] swiotlb-xen: Removing the 'struct device' in the address translation routines.

[Konrad Rzeszutek Wilk]

We don't use it at all.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |   25 ++++++++++++-------------
 1 files changed, 12 insertions(+), 13 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index 6bba1a5..12e9dcd 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -47,20 +47,19 @@
 static char *xen_io_tlb_start, *xen_io_tlb_end;
 static unsigned long xen_io_tlb_nslabs;
 
-static dma_addr_t xen_phys_to_bus(struct device *hwdev, phys_addr_t paddr)
+static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
 {
 	return phys_to_machine(XPADDR(paddr)).maddr;;
 }
 
-static phys_addr_t xen_bus_to_phys(struct device *hwdev, dma_addr_t baddr)
+static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
 {
 	return machine_to_phys(XMADDR(baddr)).paddr;
 }
 
-static dma_addr_t xen_virt_to_bus(struct device *hwdev,
-				  void *address)
+static dma_addr_t xen_virt_to_bus(void *address)
 {
-	return xen_phys_to_bus(hwdev, virt_to_phys(address));
+	return xen_phys_to_bus(virt_to_phys(address));
 }
 
 static int check_pages_physically_contiguous(unsigned long pfn,
@@ -115,7 +114,7 @@ static void *
 xen_map_single(struct device *hwdev, phys_addr_t phys, size_t size,
 	       enum dma_data_direction dir)
 {
-	u64 start_dma_addr = xen_virt_to_bus(hwdev, xen_io_tlb_start);
+	u64 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
 
 	return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
 }
@@ -251,7 +250,7 @@ dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
 				struct dma_attrs *attrs)
 {
 	phys_addr_t phys = page_to_phys(page) + offset;
-	dma_addr_t dev_addr = xen_phys_to_bus(dev, phys);
+	dma_addr_t dev_addr = xen_phys_to_bus(phys);
 	void *map;
 
 	BUG_ON(dir == DMA_NONE);
@@ -271,7 +270,7 @@ dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
 	if (!map)
 		return DMA_ERROR_CODE;
 
-	dev_addr = xen_virt_to_bus(dev, map);
+	dev_addr = xen_virt_to_bus(map);
 
 	/*
 	 * Ensure that the address returned is DMA'ble
@@ -294,7 +293,7 @@ EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
 			     size_t size, enum dma_data_direction dir)
 {
-	phys_addr_t paddr = xen_bus_to_phys(hwdev, dev_addr);
+	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
 
 	BUG_ON(dir == DMA_NONE);
 
@@ -339,7 +338,7 @@ xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
 			size_t size, enum dma_data_direction dir,
 			enum dma_sync_target target)
 {
-	phys_addr_t paddr = xen_bus_to_phys(hwdev, dev_addr);
+	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
 
 	BUG_ON(dir == DMA_NONE);
 
@@ -400,7 +399,7 @@ xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
 
 	for_each_sg(sgl, sg, nelems, i) {
 		phys_addr_t paddr = sg_phys(sg);
-		dma_addr_t dev_addr = xen_phys_to_bus(hwdev, paddr);
+		dma_addr_t dev_addr = xen_phys_to_bus(paddr);
 
 		if (swiotlb_force ||
 		    !dma_capable(hwdev, dev_addr, sg->length) ||
@@ -415,7 +414,7 @@ xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
 				sgl[0].dma_length = 0;
 				return DMA_ERROR_CODE;
 			}
-			sg->dma_address = xen_virt_to_bus(hwdev, map);
+			sg->dma_address = xen_virt_to_bus(map);
 		} else
 			sg->dma_address = dev_addr;
 		sg->dma_length = sg->length;
@@ -512,6 +511,6 @@ EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
 int
 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
 {
-	return xen_virt_to_bus(hwdev, xen_io_tlb_end - 1) <= mask;
+	return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
 }
 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);
-- 
1.7.0.1

[PATCH 17/19] swiotlb-xen: Coalesce usage of xen_swiotlb_map.

[Konrad Rzeszutek Wilk]

We don't use the 'struct device' anymore, and the xen_swiotlb_map ended
up just passing on the values and recomputing the same value
(start_dma_addr) every time - so now we do it only once.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 lib/swiotlb-xen.c |   23 +++++++++++------------
 1 files changed, 11 insertions(+), 12 deletions(-)

diff --git a/lib/swiotlb-xen.c b/lib/swiotlb-xen.c
index 12e9dcd..b15f85c 100644
--- a/lib/swiotlb-xen.c
+++ b/lib/swiotlb-xen.c
@@ -46,6 +46,11 @@
 
 static char *xen_io_tlb_start, *xen_io_tlb_end;
 static unsigned long xen_io_tlb_nslabs;
+/*
+ * Quick lookup value of the bus address of the IOTLB.
+ */
+
+u64 start_dma_addr;
 
 static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
 {
@@ -110,15 +115,6 @@ static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
 	return 0;
 }
 
-static void *
-xen_map_single(struct device *hwdev, phys_addr_t phys, size_t size,
-	       enum dma_data_direction dir)
-{
-	u64 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
-
-	return swiotlb_tbl_map_single(hwdev, start_dma_addr, phys, size, dir);
-}
-
 static int max_dma_bits = 32;
 
 static int
@@ -174,6 +170,7 @@ void __init xen_swiotlb_init(int verbose)
 	if (rc)
 		goto error;
 
+	start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
 	swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, verbose);
 
 	return;
@@ -266,7 +263,7 @@ dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
 	/*
 	 * Oh well, have to allocate and map a bounce buffer.
 	 */
-	map = xen_map_single(dev, phys, size, dir);
+	map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
 	if (!map)
 		return DMA_ERROR_CODE;
 
@@ -404,8 +401,10 @@ xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
 		if (swiotlb_force ||
 		    !dma_capable(hwdev, dev_addr, sg->length) ||
 		    range_straddles_page_boundary(paddr, sg->length)) {
-			void *map = xen_map_single(hwdev, sg_phys(sg),
-						   sg->length, dir);
+			void *map = swiotlb_tbl_map_single(hwdev,
+							   start_dma_addr,
+							   sg_phys(sg),
+							   sg->length, dir);
 			if (!map) {
 				/* Don't panic here, we expect map_sg users
 				   to do proper error handling. */
-- 
1.7.0.1

[PATCH 18/19] pci-swiotlb-xen: Add glue code to setup dma_ops utilizing xen_swiotlb_* functions.

[Konrad Rzeszutek Wilk]

We add the glue code that sets up a dma_ops structure with the
xen_swiotlb_* functions. The code turns on xen_swiotlb flag
when it detects it is running under Xen and it is either
in privileged mode or the iommu=soft flag was passed in.

It also disables the bare-metal SWIOTLB if the Xen-SWIOTLB has
been enabled.

Note: The Xen-SWIOTLB is only built when CONFIG_XEN is enabled.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 arch/x86/include/asm/xen/swiotlb-xen.h |   14 ++++++++
 arch/x86/xen/Makefile                  |    1 +
 arch/x86/xen/pci-swiotlb-xen.c         |   58 ++++++++++++++++++++++++++++++++
 3 files changed, 73 insertions(+), 0 deletions(-)
 create mode 100644 arch/x86/include/asm/xen/swiotlb-xen.h
 create mode 100644 arch/x86/xen/pci-swiotlb-xen.c

diff --git a/arch/x86/include/asm/xen/swiotlb-xen.h b/arch/x86/include/asm/xen/swiotlb-xen.h
new file mode 100644
index 0000000..c2bc047
--- /dev/null
+++ b/arch/x86/include/asm/xen/swiotlb-xen.h
@@ -0,0 +1,14 @@
+#ifndef _ASM_X86_SWIOTLB_XEN_H
+#define _ASM_X86_SWIOTLB_XEN_H
+
+#ifdef CONFIG_SWIOTLB_XEN
+extern int xen_swiotlb;
+extern int __init pci_xen_swiotlb_detect(void);
+extern void __init pci_xen_swiotlb_init(void);
+#else
+#define xen_swiotlb 0
+static inline int __init pci_xen_swiotlb_detect(void) { return 0; }
+static inline void __init pci_xen_swiotlb_init(void) { }
+#endif
+
+#endif /* _ASM_X86_SWIOTLB_XEN_H */
diff --git a/arch/x86/xen/Makefile b/arch/x86/xen/Makefile
index 3bb4fc2..32af238 100644
--- a/arch/x86/xen/Makefile
+++ b/arch/x86/xen/Makefile
@@ -18,3 +18,4 @@ obj-$(CONFIG_SMP)		+= smp.o
 obj-$(CONFIG_PARAVIRT_SPINLOCKS)+= spinlock.o
 obj-$(CONFIG_XEN_DEBUG_FS)	+= debugfs.o
 
+obj-$(CONFIG_SWIOTLB_XEN)	+= pci-swiotlb-xen.o
diff --git a/arch/x86/xen/pci-swiotlb-xen.c b/arch/x86/xen/pci-swiotlb-xen.c
new file mode 100644
index 0000000..fc3169d
--- /dev/null
+++ b/arch/x86/xen/pci-swiotlb-xen.c
@@ -0,0 +1,58 @@
+/* Glue code to lib/swiotlb-xen.c */
+
+#include <linux/dma-mapping.h>
+#include <linux/swiotlb-xen.h>
+
+#include <asm/xen/hypervisor.h>
+#include <xen/xen.h>
+
+int xen_swiotlb __read_mostly;
+
+static const struct dma_map_ops xen_swiotlb_dma_ops = {
+	.mapping_error = xen_swiotlb_dma_mapping_error,
+	.alloc_coherent = xen_swiotlb_alloc_coherent,
+	.free_coherent = xen_swiotlb_free_coherent,
+	.sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
+	.sync_single_for_device = xen_swiotlb_sync_single_for_device,
+	.sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
+	.sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
+	.map_sg = xen_swiotlb_map_sg_attrs,
+	.unmap_sg = xen_swiotlb_unmap_sg_attrs,
+	.map_page = xen_swiotlb_map_page,
+	.unmap_page = xen_swiotlb_unmap_page,
+	.dma_supported = xen_swiotlb_dma_supported,
+};
+
+/*
+ * pci_xen_swiotlb_detect - set xen_swiotlb to 1 if necessary
+ *
+ * This returns non-zero if we are forced to use xen_swiotlb (by the boot
+ * option).
+ */
+int __init pci_xen_swiotlb_detect(void)
+{
+
+	/* If running as PV guest, either iommu=soft, or swiotlb=force will
+	 * activate this IOMMU. If running as PV privileged, activate it
+	 * irregardlesss.
+	 */
+	if ((xen_initial_domain() || swiotlb || swiotlb_force) &&
+	    (xen_pv_domain()))
+		xen_swiotlb = 1;
+
+	/* If we are running under Xen, we MUST disable the native SWIOTLB.
+	 * Don't worry about swiotlb_force flag activating the native, as
+	 * the 'swiotlb' flag is the only one turning it on. */
+	if (xen_pv_domain())
+		swiotlb = 0;
+
+	return xen_swiotlb;
+}
+
+void __init pci_xen_swiotlb_init(void)
+{
+	if (xen_swiotlb) {
+		xen_swiotlb_init(1);
+		dma_ops = &xen_swiotlb_dma_ops;
+	}
+}
-- 
1.7.0.1

[PATCH 19/19] x86: Detect whether we should use Xen SWIOTLB.

[Konrad Rzeszutek Wilk]

It is paramount that we call pci_xen_swiotlb_detect before
pci_swiotlb_detect as both implementations use the 'swiotlb'
and 'swiotlb_force' flags. The pci-xen_swiotlb_detect inhibits
the swiotlb_force and swiotlb flag so that the native SWIOTLB
implementation is not enabled when running under Xen.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
---
 arch/x86/kernel/pci-dma.c |    7 +++++--
 1 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kernel/pci-dma.c b/arch/x86/kernel/pci-dma.c
index 4b7e3d8..9f07cfc 100644
--- a/arch/x86/kernel/pci-dma.c
+++ b/arch/x86/kernel/pci-dma.c
@@ -13,6 +13,7 @@
 #include <asm/calgary.h>
 #include <asm/amd_iommu.h>
 #include <asm/x86_init.h>
+#include <asm/xen/swiotlb-xen.h>
 
 static int forbid_dac __read_mostly;
 
@@ -132,7 +133,7 @@ void __init pci_iommu_alloc(void)
 	/* free the range so iommu could get some range less than 4G */
 	dma32_free_bootmem();
 
-	if (pci_swiotlb_detect())
+	if (pci_xen_swiotlb_detect() || pci_swiotlb_detect())
 		goto out;
 
 	gart_iommu_hole_init();
@@ -144,6 +145,8 @@ void __init pci_iommu_alloc(void)
 	/* needs to be called after gart_iommu_hole_init */
 	amd_iommu_detect();
 out:
+	pci_xen_swiotlb_init();
+
 	pci_swiotlb_init();
 }
 
@@ -296,7 +299,7 @@ static int __init pci_iommu_init(void)
 #endif
 	x86_init.iommu.iommu_init();
 
-	if (swiotlb) {
+	if (swiotlb || xen_swiotlb) {
 		printk(KERN_INFO "PCI-DMA: "
 		       "Using software bounce buffering for IO (SWIOTLB)\n");
 		swiotlb_print_info();
-- 
1.7.0.1

Re: [PATCH] Xen-SWIOTBL v0.8.3 used for Xen PCI pass through for PV guests.

[Alex Williamson]

On Tue, 2010-06-22 at 15:42 -0400, Konrad Rzeszutek Wilk wrote:
> These nineteen patches lay the groundwork for Xen Paravirtualized (PV)
> domains to access PCI pass-through devices. These patches utilize the
> SWIOTLB library modifications (http://lkml.org/lkml/2010/6/4/272).
> 
> The end user of this is the Xen PCI frontend and Xen PCI [1] which
> require a DMA API "backend" that understands Xen's MMU. This allows the
> PV domains to use PCI devices.

Hi Konrad,

Sorry if I missed it, but I didn't see any mention or apparent
requirement of a hardware iommu in xen for this code.  Is that true?  If
so, is there anything to stop a PV guest with ownership of a DMA capable
PCI device from reading all sorts of memory that the domain wouldn't
otherwise have access to?  I was under the impression that the old PCI
front/back for PV guests was mainly an interesting hack with limited
applications due to security.  Thanks,

Alex

Re: [PATCH] Xen-SWIOTBL v0.8.3 used for Xen PCI pass through for PV guests.

[Konrad Rzeszutek Wilk]

> > The end user of this is the Xen PCI frontend and Xen PCI [1] which
> > require a DMA API "backend" that understands Xen's MMU. This allows the
> > PV domains to use PCI devices.
> 
> Hi Konrad,

Hey Alex,

Congratulations on your new job at Red Hat!
> 
> Sorry if I missed it, but I didn't see any mention or apparent
> requirement of a hardware iommu in xen for this code.  Is that true?  If

Ah, I completely missed to put that in the writeup (and as well some
other things that I thought off overnight). The answer is: both.

You can run this without an IOMMU in which case the security threat you
mentioned is feasible. Or you can run with an hardware IOMMU, if you
pass in the iommu=pv argument to the Xen hypervisor.

> so, is there anything to stop a PV guest with ownership of a DMA capable
> PCI device from reading all sorts of memory that the domain wouldn't
> otherwise have access to?  I was under the impression that the old PCI
> front/back for PV guests was mainly an interesting hack with limited
> applications due to security.  Thanks,

I thought as well but it looks as many folks are using it. The other
thing that I forgot to mention in the writeup are these two other use cases:

 1) One of them is the privileged PV (PPV?) domain drivers. This
    idea came about a year ago and was suggested on LKML by Eric
    Biederman. I can't find the link to it thought. Eric,
    would you by any chance remember it?

    The idea is to have multiple PPV domains which serve specific
    device drivers. An implementation using the hardware IOMMU is the
    Qubes OS (http://quebes-os.org), while this would be the same but
    using PV guests.

 2) Xen Dom0 support. Without the SWIOTLB-Xen patchset Dom0 is incapable
    of working with PCI devices.