Re: [PATCH RFC v2 12/27] arm64: mte: Add tag storage pages to the MIGRATE_CMA migratetype

From: David Hildenbrand
Date: Fri Nov 24 2023 - 14:41:06 EST


On 19.11.23 17:57, Alexandru Elisei wrote:
Add the MTE tag storage pages to the MIGRATE_CMA migratetype, which allows
the page allocator to manage them like regular pages.

Ths migratype lends the pages some very desirable properties:

* They cannot be longterm pinned, meaning they will always be migratable.

* The pages can be allocated explicitely by using their PFN (with
alloc_contig_range()) when they are needed to store tags.

Signed-off-by: Alexandru Elisei <alexandru.elisei@xxxxxxx>
---
arch/arm64/Kconfig | 1 +
arch/arm64/kernel/mte_tag_storage.c | 68 +++++++++++++++++++++++++++++
include/linux/mmzone.h | 5 +++
mm/internal.h | 3 --
4 files changed, 74 insertions(+), 3 deletions(-)

diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index fe8276fdc7a8..047487046e8f 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -2065,6 +2065,7 @@ config ARM64_MTE
if ARM64_MTE
config ARM64_MTE_TAG_STORAGE
bool "Dynamic MTE tag storage management"
+ select CONFIG_CMA
help
Adds support for dynamic management of the memory used by the hardware
for storing MTE tags. This memory, unlike normal memory, cannot be
diff --git a/arch/arm64/kernel/mte_tag_storage.c b/arch/arm64/kernel/mte_tag_storage.c
index fa6267ef8392..427f4f1909f3 100644
--- a/arch/arm64/kernel/mte_tag_storage.c
+++ b/arch/arm64/kernel/mte_tag_storage.c
@@ -5,10 +5,12 @@
* Copyright (C) 2023 ARM Ltd.
*/
+#include <linux/cma.h>
#include <linux/memblock.h>
#include <linux/mm.h>
#include <linux/of_device.h>
#include <linux/of_fdt.h>
+#include <linux/pageblock-flags.h>
#include <linux/range.h>
#include <linux/string.h>
#include <linux/xarray.h>
@@ -189,6 +191,14 @@ static int __init fdt_init_tag_storage(unsigned long node, const char *uname,
return ret;
}
+ /* Pages are managed in pageblock_nr_pages chunks */
+ if (!IS_ALIGNED(tag_range->start | range_len(tag_range), pageblock_nr_pages)) {
+ pr_err("Tag storage region 0x%llx-0x%llx not aligned to pageblock size 0x%llx",
+ PFN_PHYS(tag_range->start), PFN_PHYS(tag_range->end),
+ PFN_PHYS(pageblock_nr_pages));
+ return -EINVAL;
+ }
+
ret = tag_storage_get_memory_node(node, &mem_node);
if (ret)
return ret;
@@ -254,3 +264,61 @@ void __init mte_tag_storage_init(void)
pr_info("MTE tag storage region management disabled");
}
}
+
+static int __init mte_tag_storage_activate_regions(void)
+{
+ phys_addr_t dram_start, dram_end;
+ struct range *tag_range;
+ unsigned long pfn;
+ int i, ret;
+
+ if (num_tag_regions == 0)
+ return 0;
+
+ dram_start = memblock_start_of_DRAM();
+ dram_end = memblock_end_of_DRAM();
+
+ for (i = 0; i < num_tag_regions; i++) {
+ tag_range = &tag_regions[i].tag_range;
+ /*
+ * Tag storage region was clipped by arm64_bootmem_init()
+ * enforcing addressing limits.
+ */
+ if (PFN_PHYS(tag_range->start) < dram_start ||
+ PFN_PHYS(tag_range->end) >= dram_end) {
+ pr_err("Tag storage region 0x%llx-0x%llx outside addressable memory",
+ PFN_PHYS(tag_range->start), PFN_PHYS(tag_range->end));
+ ret = -EINVAL;
+ goto out_disabled;
+ }
+ }
+
+ /*
+ * MTE disabled, tag storage pages can be used like any other pages. The
+ * only restriction is that the pages cannot be used by kexec because
+ * the memory remains marked as reserved in the memblock allocator.
+ */
+ if (!system_supports_mte()) {
+ for (i = 0; i< num_tag_regions; i++) {
+ tag_range = &tag_regions[i].tag_range;
+ for (pfn = tag_range->start; pfn <= tag_range->end; pfn++)
+ free_reserved_page(pfn_to_page(pfn));
+ }
+ ret = 0;
+ goto out_disabled;
+ }
+
+ for (i = 0; i < num_tag_regions; i++) {
+ tag_range = &tag_regions[i].tag_range;
+ for (pfn = tag_range->start; pfn <= tag_range->end; pfn += pageblock_nr_pages)
+ init_cma_reserved_pageblock(pfn_to_page(pfn));
+ totalcma_pages += range_len(tag_range);
+ }

You shouldn't be doing that manually in arm code. Likely you want some cma.c helper for something like that.

But, can you elaborate on why you took this hacky (sorry) approach as documented in the cover letter:

"The arm64 code manages this memory directly instead of using
cma_declare_contiguous/cma_alloc for performance reasons."

What is the exact problem?

--
Cheers,

David / dhildenb