[PATCH 0/5] xen: fix dom0 PV boot on some AMD machines

[PATCH 0/5] xen: fix dom0 PV boot on some AMD machines

[Juergen Gross]

There have been reports of failed boots with Xen due to an overlap of
the kernel's memory with ACPI NVS reported in the E820 map of the host.

This series fixes this issue by moving the NVS area in dom0 to some
higher address.

Juergen Gross (5):
  xen: use correct end address of kernel for conflict checking
  xen: introduce generic helper checking for memory map conflicts
  xen: move checks for e820 conflicts further up
  xen: move max_pfn in xen_memory_setup() out of function scope
  xen: tolerate ACPI NVS memory overlapping with Xen allocated memory

 arch/x86/xen/mmu_pv.c  |   5 +-
 arch/x86/xen/setup.c   | 242 +++++++++++++++++++++++++++++++++--------
 arch/x86/xen/xen-ops.h |   3 +-
 3 files changed, 201 insertions(+), 49 deletions(-)

-- 
2.43.0

[PATCH 1/5] xen: use correct end address of kernel for conflict checking

[Juergen Gross]

When running as a Xen PV dom0 the kernel is loaded by the hypervisor
using a different memory map than that of the host. In order to
minimize the required changes in the kernel, the kernel adapts its
memory map to that of the host. In order to do that it is checking
for conflicts of its load address with the host memory map.

Unfortunately the tested memory range does not include the .brk
area, which might result in crashes or memory corruption when this
area does conflict withe the memory map of the host.

Fix the test by using the _end label instead of __bss_stop.

Fixes: 808fdb71936c ("xen: check for kernel memory conflicting with memory layout")

Signed-off-by: Juergen Gross <jgross@suse.com>
Tested-by: Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
---
 arch/x86/xen/setup.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/arch/x86/xen/setup.c b/arch/x86/xen/setup.c
index 806ddb2391d9..4bcc70a71b7d 100644
--- a/arch/x86/xen/setup.c
+++ b/arch/x86/xen/setup.c
@@ -825,7 +825,7 @@ char * __init xen_memory_setup(void)
 	 * to relocating (and even reusing) pages with kernel text or data.
 	 */
 	if (xen_is_e820_reserved(__pa_symbol(_text),
-			__pa_symbol(__bss_stop) - __pa_symbol(_text))) {
+				 __pa_symbol(_end) - __pa_symbol(_text))) {
 		xen_raw_console_write("Xen hypervisor allocated kernel memory conflicts with E820 map\n");
 		BUG();
 	}
-- 
2.43.0

[PATCH 2/5] xen: introduce generic helper checking for memory map conflicts

[Juergen Gross]

When booting as a Xen PV dom0 the memory layout of the dom0 is
modified to match that of the host, as this requires less changes in
the kernel for supporting Xen.

There are some cases, though, which are problematic, as it is the Xen
hypervisor selecting the kernel's load address plus some other data,
which might conflict with the host's memory map.

These conflicts are detected at boot time and result in a boot error.
In order to support handling at least some of these conflicts in
future, introduce a generic helper function which will later gain the
ability to adapt the memory layout when possible.

Add the missing check for the xen_start_info area.

Signed-off-by: Juergen Gross <jgross@suse.com>
Tested-by: Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
---
 arch/x86/xen/mmu_pv.c  |  5 +----
 arch/x86/xen/setup.c   | 34 ++++++++++++++++++++++++++++------
 arch/x86/xen/xen-ops.h |  3 ++-
 3 files changed, 31 insertions(+), 11 deletions(-)

diff --git a/arch/x86/xen/mmu_pv.c b/arch/x86/xen/mmu_pv.c
index f1ce39d6d32c..839e6613753d 100644
--- a/arch/x86/xen/mmu_pv.c
+++ b/arch/x86/xen/mmu_pv.c
@@ -2018,10 +2018,7 @@ void __init xen_reserve_special_pages(void)
 
 void __init xen_pt_check_e820(void)
 {
-	if (xen_is_e820_reserved(xen_pt_base, xen_pt_size)) {
-		xen_raw_console_write("Xen hypervisor allocated page table memory conflicts with E820 map\n");
-		BUG();
-	}
+	xen_chk_is_e820_usable(xen_pt_base, xen_pt_size, "page table");
 }
 
 static unsigned char dummy_mapping[PAGE_SIZE] __page_aligned_bss;
diff --git a/arch/x86/xen/setup.c b/arch/x86/xen/setup.c
index 4bcc70a71b7d..96765180514b 100644
--- a/arch/x86/xen/setup.c
+++ b/arch/x86/xen/setup.c
@@ -567,7 +567,7 @@ static void __init xen_ignore_unusable(void)
 	}
 }
 
-bool __init xen_is_e820_reserved(phys_addr_t start, phys_addr_t size)
+static bool __init xen_is_e820_reserved(phys_addr_t start, phys_addr_t size)
 {
 	struct e820_entry *entry;
 	unsigned mapcnt;
@@ -624,6 +624,23 @@ phys_addr_t __init xen_find_free_area(phys_addr_t size)
 	return 0;
 }
 
+/*
+ * Check for an area in physical memory to be usable for non-movable purposes.
+ * An area is considered to usable if the used E820 map lists it to be RAM.
+ * In case the area is not usable, crash the system with an error message.
+ */
+void __init xen_chk_is_e820_usable(phys_addr_t start, phys_addr_t size,
+				   const char *component)
+{
+	if (!xen_is_e820_reserved(start, size))
+		return;
+
+	xen_raw_console_write("Xen hypervisor allocated ");
+	xen_raw_console_write(component);
+	xen_raw_console_write(" memory conflicts with E820 map\n");
+	BUG();
+}
+
 /*
  * Like memcpy, but with physical addresses for dest and src.
  */
@@ -824,11 +841,16 @@ char * __init xen_memory_setup(void)
 	 * Failing now is better than running into weird problems later due
 	 * to relocating (and even reusing) pages with kernel text or data.
 	 */
-	if (xen_is_e820_reserved(__pa_symbol(_text),
-				 __pa_symbol(_end) - __pa_symbol(_text))) {
-		xen_raw_console_write("Xen hypervisor allocated kernel memory conflicts with E820 map\n");
-		BUG();
-	}
+	xen_chk_is_e820_usable(__pa_symbol(_text),
+			       __pa_symbol(_end) - __pa_symbol(_text),
+			       "kernel");
+
+	/*
+	 * Check for a conflict of the xen_start_info memory with the target
+	 * E820 map.
+	 */
+	xen_chk_is_e820_usable(__pa(xen_start_info), sizeof(*xen_start_info),
+			       "xen_start_info");
 
 	/*
 	 * Check for a conflict of the hypervisor supplied page tables with
diff --git a/arch/x86/xen/xen-ops.h b/arch/x86/xen/xen-ops.h
index 0cf16fc79e0b..9a27d1d653d3 100644
--- a/arch/x86/xen/xen-ops.h
+++ b/arch/x86/xen/xen-ops.h
@@ -48,7 +48,8 @@ void xen_mm_unpin_all(void);
 void __init xen_relocate_p2m(void);
 #endif
 
-bool __init xen_is_e820_reserved(phys_addr_t start, phys_addr_t size);
+void __init xen_chk_is_e820_usable(phys_addr_t start, phys_addr_t size,
+				   const char *component);
 unsigned long __ref xen_chk_extra_mem(unsigned long pfn);
 void __init xen_inv_extra_mem(void);
 void __init xen_remap_memory(void);
-- 
2.43.0

[PATCH 3/5] xen: move checks for e820 conflicts further up

[Juergen Gross]

Move the checks for e820 memory map conflicts using the
xen_chk_is_e820_usable() helper further up in order to prepare
resolving some of the possible conflicts by doing some e820 map
modifications, which must happen before evaluating the RAM layout.

Signed-off-by: Juergen Gross <jgross@suse.com>
Tested-by: Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
---
 arch/x86/xen/setup.c | 44 ++++++++++++++++++++++----------------------
 1 file changed, 22 insertions(+), 22 deletions(-)

diff --git a/arch/x86/xen/setup.c b/arch/x86/xen/setup.c
index 96765180514b..dba68951ed6b 100644
--- a/arch/x86/xen/setup.c
+++ b/arch/x86/xen/setup.c
@@ -764,6 +764,28 @@ char * __init xen_memory_setup(void)
 	/* Make sure the Xen-supplied memory map is well-ordered. */
 	e820__update_table(&xen_e820_table);
 
+	/*
+	 * Check whether the kernel itself conflicts with the target E820 map.
+	 * Failing now is better than running into weird problems later due
+	 * to relocating (and even reusing) pages with kernel text or data.
+	 */
+	xen_chk_is_e820_usable(__pa_symbol(_text),
+			       __pa_symbol(_end) - __pa_symbol(_text),
+			       "kernel");
+
+	/*
+	 * Check for a conflict of the xen_start_info memory with the target
+	 * E820 map.
+	 */
+	xen_chk_is_e820_usable(__pa(xen_start_info), sizeof(*xen_start_info),
+			       "xen_start_info");
+
+	/*
+	 * Check for a conflict of the hypervisor supplied page tables with
+	 * the target E820 map.
+	 */
+	xen_pt_check_e820();
+
 	max_pages = xen_get_max_pages();
 
 	/* How many extra pages do we need due to remapping? */
@@ -836,28 +858,6 @@ char * __init xen_memory_setup(void)
 
 	e820__update_table(e820_table);
 
-	/*
-	 * Check whether the kernel itself conflicts with the target E820 map.
-	 * Failing now is better than running into weird problems later due
-	 * to relocating (and even reusing) pages with kernel text or data.
-	 */
-	xen_chk_is_e820_usable(__pa_symbol(_text),
-			       __pa_symbol(_end) - __pa_symbol(_text),
-			       "kernel");
-
-	/*
-	 * Check for a conflict of the xen_start_info memory with the target
-	 * E820 map.
-	 */
-	xen_chk_is_e820_usable(__pa(xen_start_info), sizeof(*xen_start_info),
-			       "xen_start_info");
-
-	/*
-	 * Check for a conflict of the hypervisor supplied page tables with
-	 * the target E820 map.
-	 */
-	xen_pt_check_e820();
-
 	xen_reserve_xen_mfnlist();
 
 	/* Check for a conflict of the initrd with the target E820 map. */
-- 
2.43.0

[PATCH 4/5] xen: move max_pfn in xen_memory_setup() out of function scope

[Juergen Gross]

Instead of having max_pfn as a local variable of xen_memory_setup(),
make it a static variable in setup.c instead. This avoids having to
pass it to subfunctions, which will be needed in more cases in future.

Rename it to ini_nr_pages, as the value denotes the currently usable
number of memory pages as passed from the hypervisor at boot time.

Signed-off-by: Juergen Gross <jgross@suse.com>
Tested-by: Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
---
 arch/x86/xen/setup.c | 53 ++++++++++++++++++++++----------------------
 1 file changed, 27 insertions(+), 26 deletions(-)

diff --git a/arch/x86/xen/setup.c b/arch/x86/xen/setup.c
index dba68951ed6b..d678c0330971 100644
--- a/arch/x86/xen/setup.c
+++ b/arch/x86/xen/setup.c
@@ -46,6 +46,9 @@ bool xen_pv_pci_possible;
 /* E820 map used during setting up memory. */
 static struct e820_table xen_e820_table __initdata;
 
+/* Number of initially usable memory pages. */
+static unsigned long ini_nr_pages __initdata;
+
 /*
  * Buffer used to remap identity mapped pages. We only need the virtual space.
  * The physical page behind this address is remapped as needed to different
@@ -212,7 +215,7 @@ static int __init xen_free_mfn(unsigned long mfn)
  * as a fallback if the remapping fails.
  */
 static void __init xen_set_identity_and_release_chunk(unsigned long start_pfn,
-			unsigned long end_pfn, unsigned long nr_pages)
+						      unsigned long end_pfn)
 {
 	unsigned long pfn, end;
 	int ret;
@@ -220,7 +223,7 @@ static void __init xen_set_identity_and_release_chunk(unsigned long start_pfn,
 	WARN_ON(start_pfn > end_pfn);
 
 	/* Release pages first. */
-	end = min(end_pfn, nr_pages);
+	end = min(end_pfn, ini_nr_pages);
 	for (pfn = start_pfn; pfn < end; pfn++) {
 		unsigned long mfn = pfn_to_mfn(pfn);
 
@@ -341,15 +344,14 @@ static void __init xen_do_set_identity_and_remap_chunk(
  * to Xen and not remapped.
  */
 static unsigned long __init xen_set_identity_and_remap_chunk(
-	unsigned long start_pfn, unsigned long end_pfn, unsigned long nr_pages,
-	unsigned long remap_pfn)
+	unsigned long start_pfn, unsigned long end_pfn, unsigned long remap_pfn)
 {
 	unsigned long pfn;
 	unsigned long i = 0;
 	unsigned long n = end_pfn - start_pfn;
 
 	if (remap_pfn == 0)
-		remap_pfn = nr_pages;
+		remap_pfn = ini_nr_pages;
 
 	while (i < n) {
 		unsigned long cur_pfn = start_pfn + i;
@@ -358,19 +360,19 @@ static unsigned long __init xen_set_identity_and_remap_chunk(
 		unsigned long remap_range_size;
 
 		/* Do not remap pages beyond the current allocation */
-		if (cur_pfn >= nr_pages) {
+		if (cur_pfn >= ini_nr_pages) {
 			/* Identity map remaining pages */
 			set_phys_range_identity(cur_pfn, cur_pfn + size);
 			break;
 		}
-		if (cur_pfn + size > nr_pages)
-			size = nr_pages - cur_pfn;
+		if (cur_pfn + size > ini_nr_pages)
+			size = ini_nr_pages - cur_pfn;
 
 		remap_range_size = xen_find_pfn_range(&remap_pfn);
 		if (!remap_range_size) {
 			pr_warn("Unable to find available pfn range, not remapping identity pages\n");
 			xen_set_identity_and_release_chunk(cur_pfn,
-						cur_pfn + left, nr_pages);
+							   cur_pfn + left);
 			break;
 		}
 		/* Adjust size to fit in current e820 RAM region */
@@ -397,18 +399,18 @@ static unsigned long __init xen_set_identity_and_remap_chunk(
 }
 
 static unsigned long __init xen_count_remap_pages(
-	unsigned long start_pfn, unsigned long end_pfn, unsigned long nr_pages,
+	unsigned long start_pfn, unsigned long end_pfn,
 	unsigned long remap_pages)
 {
-	if (start_pfn >= nr_pages)
+	if (start_pfn >= ini_nr_pages)
 		return remap_pages;
 
-	return remap_pages + min(end_pfn, nr_pages) - start_pfn;
+	return remap_pages + min(end_pfn, ini_nr_pages) - start_pfn;
 }
 
-static unsigned long __init xen_foreach_remap_area(unsigned long nr_pages,
+static unsigned long __init xen_foreach_remap_area(
 	unsigned long (*func)(unsigned long start_pfn, unsigned long end_pfn,
-			      unsigned long nr_pages, unsigned long last_val))
+			      unsigned long last_val))
 {
 	phys_addr_t start = 0;
 	unsigned long ret_val = 0;
@@ -436,8 +438,7 @@ static unsigned long __init xen_foreach_remap_area(unsigned long nr_pages,
 				end_pfn = PFN_UP(entry->addr);
 
 			if (start_pfn < end_pfn)
-				ret_val = func(start_pfn, end_pfn, nr_pages,
-					       ret_val);
+				ret_val = func(start_pfn, end_pfn, ret_val);
 			start = end;
 		}
 	}
@@ -700,7 +701,7 @@ static void __init xen_reserve_xen_mfnlist(void)
  **/
 char * __init xen_memory_setup(void)
 {
-	unsigned long max_pfn, pfn_s, n_pfns;
+	unsigned long pfn_s, n_pfns;
 	phys_addr_t mem_end, addr, size, chunk_size;
 	u32 type;
 	int rc;
@@ -712,9 +713,9 @@ char * __init xen_memory_setup(void)
 	int op;
 
 	xen_parse_512gb();
-	max_pfn = xen_get_pages_limit();
-	max_pfn = min(max_pfn, xen_start_info->nr_pages);
-	mem_end = PFN_PHYS(max_pfn);
+	ini_nr_pages = xen_get_pages_limit();
+	ini_nr_pages = min(ini_nr_pages, xen_start_info->nr_pages);
+	mem_end = PFN_PHYS(ini_nr_pages);
 
 	memmap.nr_entries = ARRAY_SIZE(xen_e820_table.entries);
 	set_xen_guest_handle(memmap.buffer, xen_e820_table.entries);
@@ -789,10 +790,10 @@ char * __init xen_memory_setup(void)
 	max_pages = xen_get_max_pages();
 
 	/* How many extra pages do we need due to remapping? */
-	max_pages += xen_foreach_remap_area(max_pfn, xen_count_remap_pages);
+	max_pages += xen_foreach_remap_area(xen_count_remap_pages);
 
-	if (max_pages > max_pfn)
-		extra_pages += max_pages - max_pfn;
+	if (max_pages > ini_nr_pages)
+		extra_pages += max_pages - ini_nr_pages;
 
 	/*
 	 * Clamp the amount of extra memory to a EXTRA_MEM_RATIO
@@ -801,8 +802,8 @@ char * __init xen_memory_setup(void)
 	 * Make sure we have no memory above max_pages, as this area
 	 * isn't handled by the p2m management.
 	 */
-	maxmem_pages = EXTRA_MEM_RATIO * min(max_pfn, PFN_DOWN(MAXMEM));
-	extra_pages = min3(maxmem_pages, extra_pages, max_pages - max_pfn);
+	maxmem_pages = EXTRA_MEM_RATIO * min(ini_nr_pages, PFN_DOWN(MAXMEM));
+	extra_pages = min3(maxmem_pages, extra_pages, max_pages - ini_nr_pages);
 	i = 0;
 	addr = xen_e820_table.entries[0].addr;
 	size = xen_e820_table.entries[0].size;
@@ -885,7 +886,7 @@ char * __init xen_memory_setup(void)
 	 * Set identity map on non-RAM pages and prepare remapping the
 	 * underlying RAM.
 	 */
-	xen_foreach_remap_area(max_pfn, xen_set_identity_and_remap_chunk);
+	xen_foreach_remap_area(xen_set_identity_and_remap_chunk);
 
 	pr_info("Released %ld page(s)\n", xen_released_pages);
 
-- 
2.43.0

[PATCH 5/5] xen: tolerate ACPI NVS memory overlapping with Xen allocated memory

[Juergen Gross]

In order to minimize required special handling for running as Xen PV
dom0, the memory layout is modified to match that of the host. This
requires to have only RAM at the locations where Xen allocated memory
is living. Unfortunately there seem to be some machines, where ACPI
NVS is located at 64 MB, resulting in a conflict with the loaded
kernel or the initial page tables built by Xen.

As ACPI NVS needs to be accessed by the kernel only for saving and
restoring it across suspend operations, it can be relocated in the
dom0's memory map by swapping it with unused RAM (this is possible
via modification of the dom0 P2M map).

While the E820 map can (and should) be modified right away, the P2M
map can be updated only after memory allocation is working, as the P2M
map might need to be extended.

Fixes: 808fdb71936c ("xen: check for kernel memory conflicting with memory layout")
Signed-off-by: Juergen Gross <jgross@suse.com>
Tested-by: Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
---
 arch/x86/xen/setup.c | 133 ++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 132 insertions(+), 1 deletion(-)

diff --git a/arch/x86/xen/setup.c b/arch/x86/xen/setup.c
index d678c0330971..dbb5d13ca61a 100644
--- a/arch/x86/xen/setup.c
+++ b/arch/x86/xen/setup.c
@@ -49,6 +49,15 @@ static struct e820_table xen_e820_table __initdata;
 /* Number of initially usable memory pages. */
 static unsigned long ini_nr_pages __initdata;
 
+/* Remapped non-RAM areas */
+#define NR_NONRAM_REMAP 4
+static struct nonram_remap {
+	unsigned long maddr;
+	unsigned long size;
+	unsigned long paddr;
+} xen_nonram_remap[NR_NONRAM_REMAP] __initdata;
+static unsigned int nr_nonram_remap __initdata;
+
 /*
  * Buffer used to remap identity mapped pages. We only need the virtual space.
  * The physical page behind this address is remapped as needed to different
@@ -452,6 +461,8 @@ static unsigned long __init xen_foreach_remap_area(
  * to be remapped memory itself in a linked list anchored at xen_remap_mfn.
  * This scheme allows to remap the different chunks in arbitrary order while
  * the resulting mapping will be independent from the order.
+ * In case xen_e820_resolve_conflicts() did relocate some non-RAM E820
+ * entries, set the correct P2M information for the affected pages.
  */
 void __init xen_remap_memory(void)
 {
@@ -495,6 +506,29 @@ void __init xen_remap_memory(void)
 	set_pte_mfn(buf, mfn_save, PAGE_KERNEL);
 
 	pr_info("Remapped %ld page(s)\n", remapped);
+
+	if (nr_nonram_remap == 0)
+		return;
+
+	remapped = 0;
+	for (i = 0; i < nr_nonram_remap; i++) {
+		struct nonram_remap *remap = xen_nonram_remap + i;
+
+		pfn = PFN_DOWN(remap->paddr);
+		mfn_save = PFN_DOWN(remap->maddr);
+		for (len = 0; len < remap->size; len += PAGE_SIZE) {
+			if (!set_phys_to_machine(pfn, mfn_save)) {
+				WARN(1, "Failed to set p2m mapping for pfn=%ld mfn=%ld\n",
+				     pfn, mfn_save);
+				BUG();
+			}
+			pfn++;
+			mfn_save++;
+			remapped++;
+		}
+	}
+
+	pr_info("Remapped %ld non-RAM page(s)\n", remapped);
 }
 
 static unsigned long __init xen_get_pages_limit(void)
@@ -625,14 +659,111 @@ phys_addr_t __init xen_find_free_area(phys_addr_t size)
 	return 0;
 }
 
+/*
+ * Swap a non-RAM E820 map entry with RAM above ini_nr_pages.
+ * Note that the E820 map is modified accordingly, but the P2M map isn't yet.
+ * The adaption of the P2M must be deferred until page allocation is possible.
+ */
+static void __init xen_e820_swap_entry_with_ram(struct e820_entry *swap_entry)
+{
+	struct e820_entry *entry;
+	unsigned int mapcnt;
+	phys_addr_t mem_end = PFN_PHYS(ini_nr_pages);
+	struct nonram_remap *remap;
+	phys_addr_t swap_addr, swap_size, entry_end;
+
+	if (nr_nonram_remap == NR_NONRAM_REMAP) {
+		xen_raw_console_write("Number of required E820 entry remapping actions exceed maximum value\n");
+		BUG();
+	}
+
+	swap_addr = PAGE_ALIGN_DOWN(swap_entry->addr);
+	swap_size = PAGE_ALIGN(swap_entry->addr - swap_addr + swap_entry->size);
+	remap = xen_nonram_remap + nr_nonram_remap;
+	entry = xen_e820_table.entries;
+
+	for (mapcnt = 0; mapcnt < xen_e820_table.nr_entries; mapcnt++) {
+		entry_end = entry->addr + entry->size;
+		if (entry->type == E820_TYPE_RAM && entry->size >= swap_size &&
+		    entry_end - swap_size >= mem_end) {
+			/* Reduce RAM entry by needed space (whole pages). */
+			entry->size -= swap_size;
+
+			/* Add new entry at the end of E820 map. */
+			entry = xen_e820_table.entries +
+				xen_e820_table.nr_entries;
+			xen_e820_table.nr_entries++;
+
+			/* Fill new entry (keep size and page offset). */
+			entry->type = swap_entry->type;
+			entry->addr = entry_end - swap_size +
+				      swap_addr - swap_entry->addr;
+			entry->size = swap_entry->size;
+
+			/* Convert old entry to RAM, align to pages. */
+			swap_entry->type = E820_TYPE_RAM;
+			swap_entry->addr = swap_addr;
+			swap_entry->size = swap_size;
+
+			/* Remember PFN<->MFN relation for P2M update. */
+			remap->maddr = swap_addr;
+			remap->size = swap_size;
+			remap->paddr = entry_end - swap_size;
+			nr_nonram_remap++;
+
+			/* Order E820 table and merge entries. */
+			e820__update_table(&xen_e820_table);
+
+			return;
+		}
+
+		entry++;
+	}
+
+	xen_raw_console_write("No suitable area found for required E820 entry remapping action\n");
+	BUG();
+}
+
+/*
+ * Look for non-RAM memory types in a specific guest physical area and move
+ * those away if possible (ACPI NVS only for now).
+ */
+static void __init xen_e820_resolve_conflicts(phys_addr_t start,
+					      phys_addr_t size)
+{
+	struct e820_entry *entry;
+	unsigned int mapcnt;
+	phys_addr_t end;
+
+	if (!size)
+		return;
+
+	end = start + size;
+	entry = xen_e820_table.entries;
+
+	for (mapcnt = 0; mapcnt < xen_e820_table.nr_entries; mapcnt++) {
+		if (entry->addr >= end)
+			return;
+
+		if (entry->addr + entry->size > start &&
+		    entry->type == E820_TYPE_NVS)
+			xen_e820_swap_entry_with_ram(entry);
+
+		entry++;
+	}
+}
+
 /*
  * Check for an area in physical memory to be usable for non-movable purposes.
- * An area is considered to usable if the used E820 map lists it to be RAM.
+ * An area is considered to usable if the used E820 map lists it to be RAM or
+ * some other type which can be moved to higher PFNs while keeping the MFNs.
  * In case the area is not usable, crash the system with an error message.
  */
 void __init xen_chk_is_e820_usable(phys_addr_t start, phys_addr_t size,
 				   const char *component)
 {
+	xen_e820_resolve_conflicts(start, size);
+
 	if (!xen_is_e820_reserved(start, size))
 		return;
 
-- 
2.43.0

Re: [PATCH 5/5] xen: tolerate ACPI NVS memory overlapping with Xen allocated memory

[Jan Beulich]

On 07.08.2024 12:41, Juergen Gross wrote:
> In order to minimize required special handling for running as Xen PV
> dom0, the memory layout is modified to match that of the host. This
> requires to have only RAM at the locations where Xen allocated memory
> is living. Unfortunately there seem to be some machines, where ACPI
> NVS is located at 64 MB, resulting in a conflict with the loaded
> kernel or the initial page tables built by Xen.
> 
> As ACPI NVS needs to be accessed by the kernel only for saving and
> restoring it across suspend operations, it can be relocated in the
> dom0's memory map by swapping it with unused RAM (this is possible
> via modification of the dom0 P2M map).

While the kernel may not (directly) need to access it for other purposes,
what about AML accessing it? As you can't advertise the movement to ACPI,
and as non-RAM mappings are carried out by drivers/acpi/osl.c:acpi_map()
using acpi_os_ioremap(), phys-to-machine translations won't cover for
that (unless I'm overlooking something, which unfortunately seems like I
might be).

Even without that I wonder in how far tools (kexec?) may be misguided by
relocating non-RAM memory ranges. Even more so when stuff traditionally
living below the 4G boundary suddenly moves far beyond that.

Jan

Re: [PATCH 5/5] xen: tolerate ACPI NVS memory overlapping with Xen allocated memory

[Jürgen Groß]

On 07.08.24 14:05, Jan Beulich wrote:
> On 07.08.2024 12:41, Juergen Gross wrote:
>> In order to minimize required special handling for running as Xen PV
>> dom0, the memory layout is modified to match that of the host. This
>> requires to have only RAM at the locations where Xen allocated memory
>> is living. Unfortunately there seem to be some machines, where ACPI
>> NVS is located at 64 MB, resulting in a conflict with the loaded
>> kernel or the initial page tables built by Xen.
>>
>> As ACPI NVS needs to be accessed by the kernel only for saving and
>> restoring it across suspend operations, it can be relocated in the
>> dom0's memory map by swapping it with unused RAM (this is possible
>> via modification of the dom0 P2M map).
> 
> While the kernel may not (directly) need to access it for other purposes,
> what about AML accessing it? As you can't advertise the movement to ACPI,
> and as non-RAM mappings are carried out by drivers/acpi/osl.c:acpi_map()
> using acpi_os_ioremap(), phys-to-machine translations won't cover for
> that (unless I'm overlooking something, which unfortunately seems like I
> might be).

Seems I need to hook into acpi_os_ioremap() in order to handle that case.

> Even without that I wonder in how far tools (kexec?) may be misguided by
> relocating non-RAM memory ranges. Even more so when stuff traditionally
> living below the 4G boundary suddenly moves far beyond that.

I don't think kexec is working in PV mode.

Other tools: time will tell, I guess. I prefer a generally working system
with maybe some tools failing over one not booting at all.

The only other general solution I could think of would be something similar
to vmlinuz on bare metal: a "small" boot code looking for an appropriate
location for the kernel and then relocating the "real" kernel accordingly.
This would require quite some effort, though, as the boot code would need
to create the related page tables and p2m table, too. I'd like to avoid
that.


Juergen

Re: [PATCH 5/5] xen: tolerate ACPI NVS memory overlapping with Xen allocated memory

[Jan Beulich]

On 08.08.2024 16:42, Jürgen Groß wrote:
> On 07.08.24 14:05, Jan Beulich wrote:
>> Even without that I wonder in how far tools (kexec?) may be misguided by
>> relocating non-RAM memory ranges. Even more so when stuff traditionally
>> living below the 4G boundary suddenly moves far beyond that.
> 
> I don't think kexec is working in PV mode.

I wasn't thinking of in-Dom0 kexec, but about host one. Iirc the kexec tool
will collect information it needs from various data sources, not uniformly
everything from the hypervisor. But I may be wrong with this.

Jan

Re: [PATCH 5/5] xen: tolerate ACPI NVS memory overlapping with Xen allocated memory

[Jan Beulich]

On 07.08.2024 14:05, Jan Beulich wrote:
> On 07.08.2024 12:41, Juergen Gross wrote:
>> In order to minimize required special handling for running as Xen PV
>> dom0, the memory layout is modified to match that of the host. This
>> requires to have only RAM at the locations where Xen allocated memory
>> is living. Unfortunately there seem to be some machines, where ACPI
>> NVS is located at 64 MB, resulting in a conflict with the loaded
>> kernel or the initial page tables built by Xen.
>>
>> As ACPI NVS needs to be accessed by the kernel only for saving and
>> restoring it across suspend operations, it can be relocated in the
>> dom0's memory map by swapping it with unused RAM (this is possible
>> via modification of the dom0 P2M map).
> 
> While the kernel may not (directly) need to access it for other purposes,
> what about AML accessing it? As you can't advertise the movement to ACPI,
> and as non-RAM mappings are carried out by drivers/acpi/osl.c:acpi_map()
> using acpi_os_ioremap(), phys-to-machine translations won't cover for
> that (unless I'm overlooking something, which unfortunately seems like I
> might be).

Thinking some more about this, the above may be pointing in the wrong
direction. If from acpi_os_ioremap() downwards no address translation
(PFN -> MFN) occurred, then what's coming from AML would still be
handled correctly as far as page table entries go. The problem then
might instead be that the mapping would appear to be covering RAM, not
the ACPI NVS region (and there may be checks for that).

Jan

Re: [PATCH 5/5] xen: tolerate ACPI NVS memory overlapping with Xen allocated memory

[Jürgen Groß]

On 09.08.24 11:45, Jan Beulich wrote:
> On 07.08.2024 14:05, Jan Beulich wrote:
>> On 07.08.2024 12:41, Juergen Gross wrote:
>>> In order to minimize required special handling for running as Xen PV
>>> dom0, the memory layout is modified to match that of the host. This
>>> requires to have only RAM at the locations where Xen allocated memory
>>> is living. Unfortunately there seem to be some machines, where ACPI
>>> NVS is located at 64 MB, resulting in a conflict with the loaded
>>> kernel or the initial page tables built by Xen.
>>>
>>> As ACPI NVS needs to be accessed by the kernel only for saving and
>>> restoring it across suspend operations, it can be relocated in the
>>> dom0's memory map by swapping it with unused RAM (this is possible
>>> via modification of the dom0 P2M map).
>>
>> While the kernel may not (directly) need to access it for other purposes,
>> what about AML accessing it? As you can't advertise the movement to ACPI,
>> and as non-RAM mappings are carried out by drivers/acpi/osl.c:acpi_map()
>> using acpi_os_ioremap(), phys-to-machine translations won't cover for
>> that (unless I'm overlooking something, which unfortunately seems like I
>> might be).
> 
> Thinking some more about this, the above may be pointing in the wrong
> direction. If from acpi_os_ioremap() downwards no address translation
> (PFN -> MFN) occurred, then what's coming from AML would still be
> handled correctly as far as page table entries go. The problem then
> might instead be that the mapping would appear to be covering RAM, not
> the ACPI NVS region (and there may be checks for that).

All PTE entries written go through the P->M translation.


Juergen

[PATCH 5/5] xen: tolerate ACPI NVS memory overlapping with Xen allocated memory

[Juergen Gross]

In order to minimize required special handling for running as Xen PV
dom0, the memory layout is modified to match that of the host. This
requires to have only RAM at the locations where Xen allocated memory
is living. Unfortunately there seem to be some machines, where ACPI
NVS is located at 64 MB, resulting in a conflict with the loaded
kernel or the initial page tables built by Xen.

As ACPI NVS needs to be accessed by the kernel only for saving and
restoring it across suspend operations, it can be relocated in the
dom0's memory map by swapping it with unused RAM (this is possible
via modification of the dom0 P2M map).

While the E820 map can (and should) be modified right away, the P2M
map can be updated only after memory allocation is working, as the P2M
map might need to be extended.

Fixes: 808fdb71936c ("xen: check for kernel memory conflicting with memory layout")
Signed-off-by: Juergen Gross <jgross@suse.com>
Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
---
 arch/x86/xen/setup.c | 133 ++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 132 insertions(+), 1 deletion(-)

diff --git a/arch/x86/xen/setup.c b/arch/x86/xen/setup.c
index d678c0330971..dbb5d13ca61a 100644
--- a/arch/x86/xen/setup.c
+++ b/arch/x86/xen/setup.c
@@ -49,6 +49,15 @@ static struct e820_table xen_e820_table __initdata;
 /* Number of initially usable memory pages. */
 static unsigned long ini_nr_pages __initdata;
 
+/* Remapped non-RAM areas */
+#define NR_NONRAM_REMAP 4
+static struct nonram_remap {
+	unsigned long maddr;
+	unsigned long size;
+	unsigned long paddr;
+} xen_nonram_remap[NR_NONRAM_REMAP] __initdata;
+static unsigned int nr_nonram_remap __initdata;
+
 /*
  * Buffer used to remap identity mapped pages. We only need the virtual space.
  * The physical page behind this address is remapped as needed to different
@@ -452,6 +461,8 @@ static unsigned long __init xen_foreach_remap_area(
  * to be remapped memory itself in a linked list anchored at xen_remap_mfn.
  * This scheme allows to remap the different chunks in arbitrary order while
  * the resulting mapping will be independent from the order.
+ * In case xen_e820_resolve_conflicts() did relocate some non-RAM E820
+ * entries, set the correct P2M information for the affected pages.
  */
 void __init xen_remap_memory(void)
 {
@@ -495,6 +506,29 @@ void __init xen_remap_memory(void)
 	set_pte_mfn(buf, mfn_save, PAGE_KERNEL);
 
 	pr_info("Remapped %ld page(s)\n", remapped);
+
+	if (nr_nonram_remap == 0)
+		return;
+
+	remapped = 0;
+	for (i = 0; i < nr_nonram_remap; i++) {
+		struct nonram_remap *remap = xen_nonram_remap + i;
+
+		pfn = PFN_DOWN(remap->paddr);
+		mfn_save = PFN_DOWN(remap->maddr);
+		for (len = 0; len < remap->size; len += PAGE_SIZE) {
+			if (!set_phys_to_machine(pfn, mfn_save)) {
+				WARN(1, "Failed to set p2m mapping for pfn=%ld mfn=%ld\n",
+				     pfn, mfn_save);
+				BUG();
+			}
+			pfn++;
+			mfn_save++;
+			remapped++;
+		}
+	}
+
+	pr_info("Remapped %ld non-RAM page(s)\n", remapped);
 }
 
 static unsigned long __init xen_get_pages_limit(void)
@@ -625,14 +659,111 @@ phys_addr_t __init xen_find_free_area(phys_addr_t size)
 	return 0;
 }
 
+/*
+ * Swap a non-RAM E820 map entry with RAM above ini_nr_pages.
+ * Note that the E820 map is modified accordingly, but the P2M map isn't yet.
+ * The adaption of the P2M must be deferred until page allocation is possible.
+ */
+static void __init xen_e820_swap_entry_with_ram(struct e820_entry *swap_entry)
+{
+	struct e820_entry *entry;
+	unsigned int mapcnt;
+	phys_addr_t mem_end = PFN_PHYS(ini_nr_pages);
+	struct nonram_remap *remap;
+	phys_addr_t swap_addr, swap_size, entry_end;
+
+	if (nr_nonram_remap == NR_NONRAM_REMAP) {
+		xen_raw_console_write("Number of required E820 entry remapping actions exceed maximum value\n");
+		BUG();
+	}
+
+	swap_addr = PAGE_ALIGN_DOWN(swap_entry->addr);
+	swap_size = PAGE_ALIGN(swap_entry->addr - swap_addr + swap_entry->size);
+	remap = xen_nonram_remap + nr_nonram_remap;
+	entry = xen_e820_table.entries;
+
+	for (mapcnt = 0; mapcnt < xen_e820_table.nr_entries; mapcnt++) {
+		entry_end = entry->addr + entry->size;
+		if (entry->type == E820_TYPE_RAM && entry->size >= swap_size &&
+		    entry_end - swap_size >= mem_end) {
+			/* Reduce RAM entry by needed space (whole pages). */
+			entry->size -= swap_size;
+
+			/* Add new entry at the end of E820 map. */
+			entry = xen_e820_table.entries +
+				xen_e820_table.nr_entries;
+			xen_e820_table.nr_entries++;
+
+			/* Fill new entry (keep size and page offset). */
+			entry->type = swap_entry->type;
+			entry->addr = entry_end - swap_size +
+				      swap_addr - swap_entry->addr;
+			entry->size = swap_entry->size;
+
+			/* Convert old entry to RAM, align to pages. */
+			swap_entry->type = E820_TYPE_RAM;
+			swap_entry->addr = swap_addr;
+			swap_entry->size = swap_size;
+
+			/* Remember PFN<->MFN relation for P2M update. */
+			remap->maddr = swap_addr;
+			remap->size = swap_size;
+			remap->paddr = entry_end - swap_size;
+			nr_nonram_remap++;
+
+			/* Order E820 table and merge entries. */
+			e820__update_table(&xen_e820_table);
+
+			return;
+		}
+
+		entry++;
+	}
+
+	xen_raw_console_write("No suitable area found for required E820 entry remapping action\n");
+	BUG();
+}
+
+/*
+ * Look for non-RAM memory types in a specific guest physical area and move
+ * those away if possible (ACPI NVS only for now).
+ */
+static void __init xen_e820_resolve_conflicts(phys_addr_t start,
+					      phys_addr_t size)
+{
+	struct e820_entry *entry;
+	unsigned int mapcnt;
+	phys_addr_t end;
+
+	if (!size)
+		return;
+
+	end = start + size;
+	entry = xen_e820_table.entries;
+
+	for (mapcnt = 0; mapcnt < xen_e820_table.nr_entries; mapcnt++) {
+		if (entry->addr >= end)
+			return;
+
+		if (entry->addr + entry->size > start &&
+		    entry->type == E820_TYPE_NVS)
+			xen_e820_swap_entry_with_ram(entry);
+
+		entry++;
+	}
+}
+
 /*
  * Check for an area in physical memory to be usable for non-movable purposes.
- * An area is considered to usable if the used E820 map lists it to be RAM.
+ * An area is considered to usable if the used E820 map lists it to be RAM or
+ * some other type which can be moved to higher PFNs while keeping the MFNs.
  * In case the area is not usable, crash the system with an error message.
  */
 void __init xen_chk_is_e820_usable(phys_addr_t start, phys_addr_t size,
 				   const char *component)
 {
+	xen_e820_resolve_conflicts(start, size);
+
 	if (!xen_is_e820_reserved(start, size))
 		return;
 
-- 
2.43.0