From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from foss.arm.com (foss.arm.com [217.140.110.172]) by smtp.subspace.kernel.org (Postfix) with ESMTP id 973C82DF13A for ; Fri, 2 Jan 2026 11:03:58 +0000 (UTC) Authentication-Results: smtp.subspace.kernel.org; arc=none smtp.client-ip=217.140.110.172 ARC-Seal:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1767351841; cv=none; b=OZBsNWwFNNdiR4Quho2LIQqflraXQrE0S9E3G10R8EAJpFFgvmQ1umEhaNn2kvU8dDyf+o3oT+XFFj9uuit1vzAhrKPqWVwc9/A0OM0n/fpT++LW4nShU1Cu3nMW3WT2flhWIgL12BhSP0j+ubVtFxWP33rbxqUPZ8y1jIb1fHg= ARC-Message-Signature:i=1; a=rsa-sha256; d=subspace.kernel.org; s=arc-20240116; t=1767351841; c=relaxed/simple; bh=XeNfI7L6c84wQYdnmJuitQE+ESE4IJ9CmZa2yw2DnaE=; h=Message-ID:Date:MIME-Version:Subject:To:Cc:References:From: In-Reply-To:Content-Type; b=sivY71X0OPtUJMNPZvvQKKlsDR2vEsuNfyPWQzXZPcSXlx/MZ522VTiRpKg/p0jkaN6Mi0NhFB9JUZO/JIKRx8ft0gB3MSiDwi3mHAqdZkF+Ms0nUIUtCbFaSMBIJfjdxzhwznjPOTg0/5KD5DMxHu5/o7YTSQSWrsYhSrmxkKo= ARC-Authentication-Results:i=1; smtp.subspace.kernel.org; dmarc=pass (p=none dis=none) header.from=arm.com; spf=pass smtp.mailfrom=arm.com; arc=none smtp.client-ip=217.140.110.172 Authentication-Results: smtp.subspace.kernel.org; dmarc=pass (p=none dis=none) header.from=arm.com Authentication-Results: smtp.subspace.kernel.org; spf=pass smtp.mailfrom=arm.com Received: from usa-sjc-imap-foss1.foss.arm.com (unknown [10.121.207.14]) by usa-sjc-mx-foss1.foss.arm.com (Postfix) with ESMTP id C8E90497; Fri, 2 Jan 2026 03:03:50 -0800 (PST) Received: from [10.57.94.221] (unknown [10.57.94.221]) by usa-sjc-imap-foss1.foss.arm.com (Postfix) with ESMTPSA id 391AD3F63F; Fri, 2 Jan 2026 03:03:54 -0800 (PST) Message-ID: Date: Fri, 2 Jan 2026 11:03:52 +0000 Precedence: bulk X-Mailing-List: linux-kernel@vger.kernel.org List-Id: List-Subscribe: List-Unsubscribe: MIME-Version: 1.0 User-Agent: Mozilla Thunderbird Subject: Re: [PATCH v3 1/2] arm64: mmu: avoid allocating pages while splitting the linear mapping Content-Language: en-GB To: Yeoreum Yun , catalin.marinas@arm.com, will@kernel.org, akpm@linux-foundation.org, david@kernel.org, kevin.brodsky@arm.com, quic_zhenhuah@quicinc.com, dev.jain@arm.com, yang@os.amperecomputing.com, chaitanyas.prakash@arm.com, bigeasy@linutronix.de, clrkwllms@kernel.org, rostedt@goodmis.org, lorenzo.stoakes@oracle.com, ardb@kernel.org, jackmanb@google.com, vbabka@suse.cz, mhocko@suse.com Cc: linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org, linux-rt-devel@lists.linux.dev References: <20251218194750.395301-1-yeoreum.yun@arm.com> <20251218194750.395301-2-yeoreum.yun@arm.com> From: Ryan Roberts In-Reply-To: <20251218194750.395301-2-yeoreum.yun@arm.com> Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 7bit On 18/12/2025 19:47, Yeoreum Yun wrote: > linear_map_split_to_ptes() currently allocates page tables while > splitting the linear mapping into PTEs under stop_machine() using GFP_ATOMIC. > > This is fine for non-PREEMPT_RT configurations. > However, it becomes problematic on PREEMPT_RT, because > generic memory allocation/free APIs (e.g. pgtable_alloc(), __get_free_pages(), etc.) > cannot be called from a non-preemptible context, except for the _nolock() variants. > This is because generic memory allocation/free paths are sleepable, > as they rely on spin_lock(), which becomes sleepable on PREEMPT_RT. > > In other words, even calling pgtable_alloc() with GFP_ATOMIC is not permitted > in __linear_map_split_to_pte() when it is executed by the stopper thread, > where preemption is disabled on PREEMPT_RT. > > To address this, the required number of page tables is first collected > and preallocated, and the preallocated page tables are then used > when splitting the linear mapping in __linear_map_split_to_pte(). > > Fixes: 3df6979d222b ("arm64: mm: split linear mapping if BBML2 unsupported on secondary CPUs") > Signed-off-by: Yeoreum Yun > --- > arch/arm64/mm/mmu.c | 232 +++++++++++++++++++++++++++++++++++--------- > 1 file changed, 184 insertions(+), 48 deletions(-) > > diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c > index 9ae7ce00a7ef..96a9fa505e71 100644 > --- a/arch/arm64/mm/mmu.c > +++ b/arch/arm64/mm/mmu.c > @@ -527,18 +527,15 @@ static void early_create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys, > panic("Failed to create page tables\n"); > } > > -static phys_addr_t __pgd_pgtable_alloc(struct mm_struct *mm, gfp_t gfp, > - enum pgtable_type pgtable_type) > -{ > - /* Page is zeroed by init_clear_pgtable() so don't duplicate effort. */ > - struct ptdesc *ptdesc = pagetable_alloc(gfp & ~__GFP_ZERO, 0); > - phys_addr_t pa; > - > - if (!ptdesc) > - return INVALID_PHYS_ADDR; > - > - pa = page_to_phys(ptdesc_page(ptdesc)); > +static struct ptdesc **split_pgtables; > +static int split_pgtables_order; > +static unsigned long split_pgtables_count; > +static unsigned long split_pgtables_idx; > > +static __always_inline void __pgd_pgtable_init(struct mm_struct *mm, > + struct ptdesc *ptdesc, > + enum pgtable_type pgtable_type) > +{ > switch (pgtable_type) { > case TABLE_PTE: > BUG_ON(!pagetable_pte_ctor(mm, ptdesc)); > @@ -554,19 +551,28 @@ static phys_addr_t __pgd_pgtable_alloc(struct mm_struct *mm, gfp_t gfp, > break; > } > nit: no need for this empty line > - return pa; > } > > -static phys_addr_t > -pgd_pgtable_alloc_init_mm_gfp(enum pgtable_type pgtable_type, gfp_t gfp) > +static phys_addr_t __pgd_pgtable_alloc(struct mm_struct *mm, gfp_t gfp, nit: all remaining callers pass gfp=GFP_PGTABLE_KERNEL so you could remove the param? > + enum pgtable_type pgtable_type) > { > - return __pgd_pgtable_alloc(&init_mm, gfp, pgtable_type); > + /* Page is zeroed by init_clear_pgtable() so don't duplicate effort. */ > + struct ptdesc *ptdesc = pagetable_alloc(gfp & ~__GFP_ZERO, 0); > + > + if (!ptdesc) > + return INVALID_PHYS_ADDR; > + > + __pgd_pgtable_init(mm, ptdesc, pgtable_type); > + > + return page_to_phys(ptdesc_page(ptdesc)); > } > > +typedef phys_addr_t (split_pgtable_alloc_fn)(enum pgtable_type); This type is used more generally than just for splitting. Perhaps simply call it "pgtable_alloc_fn"? We also pass this type around in __create_pgd_mapping() and friends; perhaps we should have a preparatory patch to define this type and consistently use it everywhere instead of passing around "phys_addr_t (*pgtable_alloc)(enum pgtable_type)"? > + > static phys_addr_t __maybe_unused This is no longer __maybe_unused; you can drop the decorator. > pgd_pgtable_alloc_init_mm(enum pgtable_type pgtable_type) > { > - return pgd_pgtable_alloc_init_mm_gfp(pgtable_type, GFP_PGTABLE_KERNEL); > + return __pgd_pgtable_alloc(&init_mm, GFP_PGTABLE_KERNEL, pgtable_type); > } > > static phys_addr_t > @@ -575,6 +581,23 @@ pgd_pgtable_alloc_special_mm(enum pgtable_type pgtable_type) > return __pgd_pgtable_alloc(NULL, GFP_PGTABLE_KERNEL, pgtable_type); > } > > +static phys_addr_t > +pgd_pgtable_get_preallocated(enum pgtable_type pgtable_type) > +{ > + struct ptdesc *ptdesc; > + > + if (WARN_ON(split_pgtables_idx >= split_pgtables_count)) > + return INVALID_PHYS_ADDR; > + > + ptdesc = split_pgtables[split_pgtables_idx++]; > + if (!ptdesc) > + return INVALID_PHYS_ADDR; > + > + __pgd_pgtable_init(&init_mm, ptdesc, pgtable_type); > + > + return page_to_phys(ptdesc_page(ptdesc)); > +} > + > static void split_contpte(pte_t *ptep) > { > int i; > @@ -584,7 +607,9 @@ static void split_contpte(pte_t *ptep) > __set_pte(ptep, pte_mknoncont(__ptep_get(ptep))); > } > > -static int split_pmd(pmd_t *pmdp, pmd_t pmd, gfp_t gfp, bool to_cont) > +static int split_pmd(pmd_t *pmdp, pmd_t pmd, > + split_pgtable_alloc_fn *pgtable_alloc_fn, nit: I believe the * has no effect when passing function pointers and the usual convention in Linux is to not use the *. Existing functions are also calling it "pgtable_alloc" so perhaps this is a bit more consistent: pgtable_alloc_fn pgtable_alloc (same nitty comment for all uses below :) ) > + bool to_cont) > { > pmdval_t tableprot = PMD_TYPE_TABLE | PMD_TABLE_UXN | PMD_TABLE_AF; > unsigned long pfn = pmd_pfn(pmd); > @@ -593,7 +618,7 @@ static int split_pmd(pmd_t *pmdp, pmd_t pmd, gfp_t gfp, bool to_cont) > pte_t *ptep; > int i; > > - pte_phys = pgd_pgtable_alloc_init_mm_gfp(TABLE_PTE, gfp); > + pte_phys = pgtable_alloc_fn(TABLE_PTE); > if (pte_phys == INVALID_PHYS_ADDR) > return -ENOMEM; > ptep = (pte_t *)phys_to_virt(pte_phys); > @@ -628,7 +653,9 @@ static void split_contpmd(pmd_t *pmdp) > set_pmd(pmdp, pmd_mknoncont(pmdp_get(pmdp))); > } > > -static int split_pud(pud_t *pudp, pud_t pud, gfp_t gfp, bool to_cont) > +static int split_pud(pud_t *pudp, pud_t pud, > + split_pgtable_alloc_fn *pgtable_alloc_fn, > + bool to_cont) > { > pudval_t tableprot = PUD_TYPE_TABLE | PUD_TABLE_UXN | PUD_TABLE_AF; > unsigned int step = PMD_SIZE >> PAGE_SHIFT; > @@ -638,7 +665,7 @@ static int split_pud(pud_t *pudp, pud_t pud, gfp_t gfp, bool to_cont) > pmd_t *pmdp; > int i; > > - pmd_phys = pgd_pgtable_alloc_init_mm_gfp(TABLE_PMD, gfp); > + pmd_phys = pgtable_alloc_fn(TABLE_PMD); > if (pmd_phys == INVALID_PHYS_ADDR) > return -ENOMEM; > pmdp = (pmd_t *)phys_to_virt(pmd_phys); > @@ -707,7 +734,7 @@ static int split_kernel_leaf_mapping_locked(unsigned long addr) > if (!pud_present(pud)) > goto out; > if (pud_leaf(pud)) { > - ret = split_pud(pudp, pud, GFP_PGTABLE_KERNEL, true); > + ret = split_pud(pudp, pud, pgd_pgtable_alloc_init_mm, true); > if (ret) > goto out; > } > @@ -732,7 +759,7 @@ static int split_kernel_leaf_mapping_locked(unsigned long addr) > */ > if (ALIGN_DOWN(addr, PMD_SIZE) == addr) > goto out; > - ret = split_pmd(pmdp, pmd, GFP_PGTABLE_KERNEL, true); > + ret = split_pmd(pmdp, pmd, pgd_pgtable_alloc_init_mm, true); > if (ret) > goto out; > } > @@ -831,34 +858,35 @@ int split_kernel_leaf_mapping(unsigned long start, unsigned long end) > static int split_to_ptes_pud_entry(pud_t *pudp, unsigned long addr, > unsigned long next, struct mm_walk *walk) > { > - gfp_t gfp = *(gfp_t *)walk->private; > + split_pgtable_alloc_fn *pgtable_alloc_fn = walk->private; > pud_t pud = pudp_get(pudp); > - int ret = 0; > > - if (pud_leaf(pud)) > - ret = split_pud(pudp, pud, gfp, false); > + if (!pud_leaf(pud)) > + return 0; > > - return ret; > + return split_pud(pudp, pud, pgtable_alloc_fn, false); why are you changing the layout of this function? Seems like unneccessary churn. Just pass pgtable_alloc to split_pud() instead of gfp. > } > > static int split_to_ptes_pmd_entry(pmd_t *pmdp, unsigned long addr, > unsigned long next, struct mm_walk *walk) > { > - gfp_t gfp = *(gfp_t *)walk->private; > + split_pgtable_alloc_fn *pgtable_alloc_fn = walk->private; > pmd_t pmd = pmdp_get(pmdp); > - int ret = 0; > + int ret; > > - if (pmd_leaf(pmd)) { > - if (pmd_cont(pmd)) > - split_contpmd(pmdp); > - ret = split_pmd(pmdp, pmd, gfp, false); > + if (!pmd_leaf(pmd)) > + return 0; > > - /* > - * We have split the pmd directly to ptes so there is no need to > - * visit each pte to check if they are contpte. > - */ > - walk->action = ACTION_CONTINUE; > - } > + if (pmd_cont(pmd)) > + split_contpmd(pmdp); > + > + ret = split_pmd(pmdp, pmd, pgtable_alloc_fn, false); > + > + /* > + * We have split the pmd directly to ptes so there is no need to > + * visit each pte to check if they are contpte. > + */ > + walk->action = ACTION_CONTINUE; Same comment; no need to change the layout of the function. > > return ret; > } > @@ -880,13 +908,15 @@ static const struct mm_walk_ops split_to_ptes_ops = { > .pte_entry = split_to_ptes_pte_entry, > }; > > -static int range_split_to_ptes(unsigned long start, unsigned long end, gfp_t gfp) > +static int range_split_to_ptes(unsigned long start, unsigned long end, > + split_pgtable_alloc_fn *pgtable_alloc_fn) > { > int ret; > > arch_enter_lazy_mmu_mode(); > ret = walk_kernel_page_table_range_lockless(start, end, > - &split_to_ptes_ops, NULL, &gfp); > + &split_to_ptes_ops, NULL, > + pgtable_alloc_fn); > arch_leave_lazy_mmu_mode(); > > return ret; > @@ -903,6 +933,105 @@ static void __init init_idmap_kpti_bbml2_flag(void) > smp_mb(); > } > > +static int __init > +collect_to_split_pud_entry(pud_t *pudp, unsigned long addr, > + unsigned long next, struct mm_walk *walk) > +{ > + pud_t pud = pudp_get(pudp); > + > + if (pud_leaf(pud)) > + split_pgtables_count += 1 + PTRS_PER_PMD; I think you probably want: walk->action = ACTION_CONTINUE; Likely you will end up with the same behaviour regardless. But you should at least we consistent with collect_to_split_pmd_entry(). > + > + return 0; > +} > + > +static int __init > +collect_to_split_pmd_entry(pmd_t *pmdp, unsigned long addr, > + unsigned long next, struct mm_walk *walk) > +{ > + pmd_t pmd = pmdp_get(pmdp); > + > + if (pmd_leaf(pmd)) > + split_pgtables_count++; > + > + walk->action = ACTION_CONTINUE; > + > + return 0; > +} > + > +static void __init linear_map_free_split_pgtables(void) > +{ > + int i; > + > + if (!split_pgtables_count || !split_pgtables) > + goto skip_free; > + > + for (i = split_pgtables_idx; i < split_pgtables_count; i++) { > + if (split_pgtables[i]) > + pagetable_free(split_pgtables[i]); > + } > + > + free_pages((unsigned long)split_pgtables, split_pgtables_order); > + > +skip_free: > + split_pgtables = NULL; > + split_pgtables_count = 0; > + split_pgtables_idx = 0; > + split_pgtables_order = 0; > +} > + > +static int __init linear_map_prealloc_split_pgtables(void) > +{ > + int ret, i; > + unsigned long lstart = _PAGE_OFFSET(vabits_actual); > + unsigned long lend = PAGE_END; > + unsigned long kstart = (unsigned long)lm_alias(_stext); > + unsigned long kend = (unsigned long)lm_alias(__init_begin); > + > + const struct mm_walk_ops collect_to_split_ops = { > + .pud_entry = collect_to_split_pud_entry, > + .pmd_entry = collect_to_split_pmd_entry > + }; > + > + split_pgtables_idx = 0; > + split_pgtables_count = 0; > + > + ret = walk_kernel_page_table_range_lockless(lstart, kstart, > + &collect_to_split_ops, > + NULL, NULL); > + if (!ret) > + ret = walk_kernel_page_table_range_lockless(kend, lend, > + &collect_to_split_ops, > + NULL, NULL); > + if (ret || !split_pgtables_count) > + goto error; > + > + ret = -ENOMEM; > + > + split_pgtables_order = > + order_base_2(PAGE_ALIGN(split_pgtables_count * > + sizeof(struct ptdesc *)) >> PAGE_SHIFT); Wouldn't this be simpler? split_pgtables_order = get_order(split_pgtables_count * sizeof(struct ptdesc *)); > + > + split_pgtables = (struct ptdesc **) __get_free_pages(GFP_KERNEL | __GFP_ZERO, > + split_pgtables_order); Do you really need the cast? (I'm not sure). > + if (!split_pgtables) > + goto error; > + > + for (i = 0; i < split_pgtables_count; i++) { > + split_pgtables[i] = pagetable_alloc(GFP_KERNEL, 0); For consistency with other code, perhaps spell it out?: /* Page is zeroed by init_clear_pgtable() so don't duplicate effort. */ gfp_t gfp = GFP_PGTABLE_KERNEL & ~__GFP_ZERO; for (i = 0; i < split_pgtables_count; i++) { split_pgtables[i] = pagetable_alloc(gfp, 0); > + if (!split_pgtables[i]) > + goto error; > + } > + > + ret = 0; > + > +error: > + if (ret) > + linear_map_free_split_pgtables(); > + > + return ret; > +} I wonder if there is value in generalizing this a bit to separate out the determination of the number of pages from the actual pre-allocation and free functions? Then you have a reusable pre-allocation function that you could also use for the KPTI case instead of having it have yet another private pre-allocator? > + > static int __init linear_map_split_to_ptes(void *__unused) > { > /* > @@ -928,9 +1057,9 @@ static int __init linear_map_split_to_ptes(void *__unused) > * PTE. The kernel alias remains static throughout runtime so > * can continue to be safely mapped with large mappings. > */ > - ret = range_split_to_ptes(lstart, kstart, GFP_ATOMIC); > + ret = range_split_to_ptes(lstart, kstart, pgd_pgtable_get_preallocated); > if (!ret) > - ret = range_split_to_ptes(kend, lend, GFP_ATOMIC); > + ret = range_split_to_ptes(kend, lend, pgd_pgtable_get_preallocated); > if (ret) > panic("Failed to split linear map\n"); > flush_tlb_kernel_range(lstart, lend); > @@ -963,10 +1092,16 @@ static int __init linear_map_split_to_ptes(void *__unused) > > void __init linear_map_maybe_split_to_ptes(void) > { > - if (linear_map_requires_bbml2 && !system_supports_bbml2_noabort()) { > - init_idmap_kpti_bbml2_flag(); > - stop_machine(linear_map_split_to_ptes, NULL, cpu_online_mask); > - } > + if (!linear_map_requires_bbml2 || system_supports_bbml2_noabort()) > + return; > + > + if (linear_map_prealloc_split_pgtables()) > + panic("Failed to split linear map\n"); > + > + init_idmap_kpti_bbml2_flag(); > + stop_machine(linear_map_split_to_ptes, NULL, cpu_online_mask); > + > + linear_map_free_split_pgtables(); > } > > /* > @@ -1088,6 +1223,7 @@ bool arch_kfence_init_pool(void) > unsigned long end = start + KFENCE_POOL_SIZE; > int ret; > > + nit: Remove extra empty line. This is looking much cleaner now; nearly there! Thanks, Ryan > /* Exit early if we know the linear map is already pte-mapped. */ > if (!split_leaf_mapping_possible()) > return true; > @@ -1097,7 +1233,7 @@ bool arch_kfence_init_pool(void) > return true; > > mutex_lock(&pgtable_split_lock); > - ret = range_split_to_ptes(start, end, GFP_PGTABLE_KERNEL); > + ret = range_split_to_ptes(start, end, pgd_pgtable_alloc_init_mm); > mutex_unlock(&pgtable_split_lock); > > /*