From 856fa79d4a581cb305c9eb2b281bb27146169cca Mon Sep 17 00:00:00 2001 From: "Alexander M. Merritt" Date: Thu, 28 May 2026 09:33:30 -0500 Subject: [PATCH] patches: xen: xlate_mmu: batch gfn mappings added to improve performance of PVH dom0 mapping in domU memory, esp. during zone launches Signed-off-by: Alexander M. Merritt --- config.yaml | 4 + ...n-xlate_mmu-relocate-remap_pfn-utils.patch | 70 +++++++ ...002-xen-xlate_mmu-batch-gfn-mappings.patch | 197 ++++++++++++++++++ 3 files changed, 271 insertions(+) create mode 100644 patches/0001-xen-xlate_mmu-relocate-remap_pfn-utils.patch create mode 100644 patches/0002-xen-xlate_mmu-batch-gfn-mappings.patch diff --git a/config.yaml b/config.yaml index 2e639ed..72a37ef 100644 --- a/config.yaml +++ b/config.yaml @@ -104,6 +104,10 @@ patches: - 0008-xen-netfront-place-per-queue-rings-on-per-queue-node.patch - 0009-xen-blkfront-place-per-ring-buffers-on-per-hctx-node.patch lower: '6.18' +- patches: + - 0001-xen-xlate_mmu-relocate-remap_pfn-utils.patch + - 0002-xen-xlate_mmu-batch-gfn-mappings.patch + lower: '6.10' images: - target: kernelsrc name: kernel-src diff --git a/patches/0001-xen-xlate_mmu-relocate-remap_pfn-utils.patch b/patches/0001-xen-xlate_mmu-relocate-remap_pfn-utils.patch new file mode 100644 index 0000000..afa4ed6 --- /dev/null +++ b/patches/0001-xen-xlate_mmu-relocate-remap_pfn-utils.patch @@ -0,0 +1,70 @@ +From 5bc298c60601161ec0f012d38e4cc021347b9831 Mon Sep 17 00:00:00 2001 +Message-ID: <5bc298c60601161ec0f012d38e4cc021347b9831.1779978695.git.alexander@edera.dev> +From: "Alexander M. Merritt" +Date: Wed, 27 May 2026 13:21:23 -0500 +Subject: [PATCH 1/2] xen: xlate_mmu: relocate remap_pfn utils + +Signed-off-by: Alexander M. Merritt +--- + drivers/xen/xlate_mmu.c | 38 +++++++++++++++++++------------------- + 1 file changed, 19 insertions(+), 19 deletions(-) + +diff --git a/drivers/xen/xlate_mmu.c b/drivers/xen/xlate_mmu.c +index f17c4c03db30..6a15396b08aa 100644 +--- a/drivers/xen/xlate_mmu.c ++++ b/drivers/xen/xlate_mmu.c +@@ -61,6 +61,25 @@ static void xen_for_each_gfn(struct page **pages, unsigned nr_gfn, + } + } + ++struct remap_pfn { ++ struct mm_struct *mm; ++ struct page **pages; ++ pgprot_t prot; ++ unsigned long i; ++}; ++ ++static int remap_pfn_fn(pte_t *ptep, unsigned long addr, void *data) ++{ ++ struct remap_pfn *r = data; ++ struct page *page = r->pages[r->i]; ++ pte_t pte = pte_mkspecial(pfn_pte(page_to_pfn(page), r->prot)); ++ ++ set_pte_at(r->mm, addr, ptep, pte); ++ r->i++; ++ ++ return 0; ++} ++ + struct remap_data { + xen_pfn_t *fgfn; /* foreign domain's gfn */ + int nr_fgfn; /* Number of foreign gfn left to map */ +@@ -262,25 +281,6 @@ int __init xen_xlate_map_ballooned_pages(xen_pfn_t **gfns, void **virt, + return 0; + } + +-struct remap_pfn { +- struct mm_struct *mm; +- struct page **pages; +- pgprot_t prot; +- unsigned long i; +-}; +- +-static int remap_pfn_fn(pte_t *ptep, unsigned long addr, void *data) +-{ +- struct remap_pfn *r = data; +- struct page *page = r->pages[r->i]; +- pte_t pte = pte_mkspecial(pfn_pte(page_to_pfn(page), r->prot)); +- +- set_pte_at(r->mm, addr, ptep, pte); +- r->i++; +- +- return 0; +-} +- + /* Used by the privcmd module, but has to be built-in on ARM */ + int xen_remap_vma_range(struct vm_area_struct *vma, unsigned long addr, unsigned long len) + { +-- +2.48.1 + diff --git a/patches/0002-xen-xlate_mmu-batch-gfn-mappings.patch b/patches/0002-xen-xlate_mmu-batch-gfn-mappings.patch new file mode 100644 index 0000000..991a27f --- /dev/null +++ b/patches/0002-xen-xlate_mmu-batch-gfn-mappings.patch @@ -0,0 +1,197 @@ +From 0b3410862d1c9a6be6973314f6979cf309b13e13 Mon Sep 17 00:00:00 2001 +Message-ID: <0b3410862d1c9a6be6973314f6979cf309b13e13.1779978695.git.alexander@edera.dev> +In-Reply-To: <5bc298c60601161ec0f012d38e4cc021347b9831.1779978695.git.alexander@edera.dev> +References: <5bc298c60601161ec0f012d38e4cc021347b9831.1779978695.git.alexander@edera.dev> +From: "Alexander M. Merritt" +Date: Wed, 27 May 2026 13:47:13 -0500 +Subject: [PATCH 2/2] xen: xlate_mmu: batch gfn mappings + +PVH dom0 wanting to map in guest memory would iterate over each page +individually invoking hypercall add_to_physmap_range, leading to high +mapping latencies. + +Refactor xen_xlate_remap_gfn_array to separately batch GFN mappings to +the hypervisor to amortize hypercall overhead. + +Signed-off-by: Alexander M. Merritt +--- + drivers/xen/xlate_mmu.c | 133 +++++++++++++++++++++++----------------- + 1 file changed, 76 insertions(+), 57 deletions(-) + +diff --git a/drivers/xen/xlate_mmu.c b/drivers/xen/xlate_mmu.c +index 6a15396b08aa..f8dd332b7f46 100644 +--- a/drivers/xen/xlate_mmu.c ++++ b/drivers/xen/xlate_mmu.c +@@ -42,6 +42,9 @@ + #include + #include + ++/* Issue up to this many GFN mapping requests in a batch at a time. */ ++#define XLATE_BATCH_SIZE 16 ++ + typedef void (*xen_gfn_fn_t)(unsigned long gfn, void *data); + + /* Break down the pages in 4KB chunk and call fn for each gfn */ +@@ -92,12 +95,19 @@ struct remap_data { + int *err_ptr; + int mapped; + +- /* Hypercall parameters */ +- int h_errs[XEN_PFN_PER_PAGE]; +- xen_ulong_t h_idxs[XEN_PFN_PER_PAGE]; +- xen_pfn_t h_gpfns[XEN_PFN_PER_PAGE]; ++ /* Hypercall parameters. ++ * ++ * idxs: source GFNs in foreign domain P2M (combined with XENMAPSPACE_gmfn_foreign) ++ * gpfns: destination GFNs in dom0's P2M. Extracted from pages[] ++ * ++ * h_idxs and h_gpfns exist as pairs: we map into dom0's GFN ++ * the same MFN for the GFN in the foreign domain. ++ */ ++ int h_errs[XLATE_BATCH_SIZE]; ++ xen_ulong_t h_idxs[XLATE_BATCH_SIZE]; ++ xen_pfn_t h_gpfns[XLATE_BATCH_SIZE]; + +- int h_iter; /* Iterator */ ++ int h_iter; + }; + + static void setup_hparams(unsigned long gfn, void *data) +@@ -112,53 +122,6 @@ static void setup_hparams(unsigned long gfn, void *data) + info->fgfn++; + } + +-static int remap_pte_fn(pte_t *ptep, unsigned long addr, void *data) +-{ +- struct remap_data *info = data; +- struct page *page = info->pages[info->index++]; +- pte_t pte = pte_mkspecial(pfn_pte(page_to_pfn(page), info->prot)); +- int rc, nr_gfn; +- uint32_t i; +- struct xen_add_to_physmap_range xatp = { +- .domid = DOMID_SELF, +- .foreign_domid = info->domid, +- .space = XENMAPSPACE_gmfn_foreign, +- }; +- +- nr_gfn = min_t(typeof(info->nr_fgfn), XEN_PFN_PER_PAGE, info->nr_fgfn); +- info->nr_fgfn -= nr_gfn; +- +- info->h_iter = 0; +- xen_for_each_gfn(&page, nr_gfn, setup_hparams, info); +- BUG_ON(info->h_iter != nr_gfn); +- +- set_xen_guest_handle(xatp.idxs, info->h_idxs); +- set_xen_guest_handle(xatp.gpfns, info->h_gpfns); +- set_xen_guest_handle(xatp.errs, info->h_errs); +- xatp.size = nr_gfn; +- +- rc = HYPERVISOR_memory_op(XENMEM_add_to_physmap_range, &xatp); +- +- /* info->err_ptr expect to have one error status per Xen PFN */ +- for (i = 0; i < nr_gfn; i++) { +- int err = (rc < 0) ? rc : info->h_errs[i]; +- +- *(info->err_ptr++) = err; +- if (!err) +- info->mapped++; +- } +- +- /* +- * Note: The hypercall will return 0 in most of the case if even if +- * all the fgmfn are not mapped. We still have to update the pte +- * as the userspace may decide to continue. +- */ +- if (!rc) +- set_pte_at(info->vma->vm_mm, addr, ptep, pte); +- +- return 0; +-} +- + int xen_xlate_remap_gfn_array(struct vm_area_struct *vma, + unsigned long addr, + xen_pfn_t *gfn, int nr, +@@ -166,9 +129,10 @@ int xen_xlate_remap_gfn_array(struct vm_area_struct *vma, + unsigned domid, + struct page **pages) + { +- int err; + struct remap_data data; +- unsigned long range = DIV_ROUND_UP(nr, XEN_PFN_PER_PAGE) << PAGE_SHIFT; ++ int page_off = 0; ++ ++ BUILD_BUG_ON(XLATE_BATCH_SIZE % XEN_PFN_PER_PAGE != 0); + + /* Kept here for the purpose of making sure code doesn't break + x86 PVOPS */ +@@ -184,9 +148,64 @@ int xen_xlate_remap_gfn_array(struct vm_area_struct *vma, + data.err_ptr = err_ptr; + data.mapped = 0; + +- err = apply_to_page_range(vma->vm_mm, addr, range, +- remap_pte_fn, &data); +- return err < 0 ? err : data.mapped; ++ while (data.nr_fgfn > 0) { ++ int batch = min_t(int, XLATE_BATCH_SIZE, data.nr_fgfn); /* xen GFN units */ ++ /* NOTE: on ARM, page size may be larger than in Xen */ ++ int batch_pages = DIV_ROUND_UP(batch, XEN_PFN_PER_PAGE); ++ unsigned long batch_range = (unsigned long)batch_pages << PAGE_SHIFT; ++ ++ struct xen_add_to_physmap_range xatp = { ++ .domid = DOMID_SELF, ++ .foreign_domid = domid, ++ .space = XENMAPSPACE_gmfn_foreign, ++ .size = batch, ++ }; ++ ++ int rc, i; ++ ++ data.h_iter = 0; ++ xen_for_each_gfn(&pages[page_off], batch, setup_hparams, &data); ++ data.nr_fgfn -= batch; ++ ++ set_xen_guest_handle(xatp.idxs, data.h_idxs); ++ set_xen_guest_handle(xatp.gpfns, data.h_gpfns); ++ set_xen_guest_handle(xatp.errs, data.h_errs); ++ ++ rc = HYPERVISOR_memory_op(XENMEM_add_to_physmap_range, &xatp); ++ ++ /* ++ * Note: The hypercall will return 0 in most of the case if even if ++ * all the fgmfn are not mapped. We still have to update the pte ++ * as the userspace may decide to continue. ++ */ ++ if (rc == 0) { ++ struct remap_pfn r = { ++ .mm = vma->vm_mm, ++ .pages = &pages[page_off], ++ .prot = prot, ++ .i = 0, ++ }; ++ int walk_err = apply_to_page_range(vma->vm_mm, addr, ++ batch_range, ++ remap_pfn_fn, &r); ++ if (walk_err < 0) ++ return walk_err; ++ } ++ ++ /* err_ptr expected to have one error status per Xen PFN */ ++ for (i = 0; i < batch; i++) { ++ int err = (rc < 0) ? rc : data.h_errs[i]; ++ *(data.err_ptr++) = err; ++ if (!err) ++ data.mapped++; ++ } ++ ++ addr += batch_range; ++ page_off += batch_pages; ++ cond_resched(); ++ } ++ ++ return data.mapped; + } + EXPORT_SYMBOL_GPL(xen_xlate_remap_gfn_array); + +-- +2.48.1 +