As noted by vdavydov@ we've a race here if two tasks are trying to allocate vdso concurently. Fix it using lockless test and mutex for vdso allocation.
https://jira.sw.ru/browse/PSBM-30093 https://bugzilla.openvz.org/show_bug.cgi?id=2768 Signed-off-by: Cyrill Gorcunov <[email protected]> CC: Vladimir Davydov <[email protected]> CC: Konstantin Khorenko <[email protected]> --- arch/x86/vdso/vdso32-setup.c | 48 ++++++++++++++++++++++++++++++----------- arch/x86/vdso/vma.c | 50 ++++++++++++++++++++++++++++++------------- 2 files changed, 71 insertions(+), 27 deletions(-) Index: linux-pcs7.git/arch/x86/vdso/vdso32-setup.c =================================================================== --- linux-pcs7.git.orig/arch/x86/vdso/vdso32-setup.c +++ linux-pcs7.git/arch/x86/vdso/vdso32-setup.c @@ -307,6 +307,8 @@ int __init sysenter_setup(void) return 0; } +static DEFINE_MUTEX(vdso32_mutex); + static struct page **uts_prep_vdso_pages_locked(int map) { struct uts_namespace *uts_ns = current->nsproxy->uts_ns; @@ -314,6 +316,7 @@ static struct page **uts_prep_vdso_pages struct ve_struct *ve = get_exec_env(); struct page **pages = vdso32_pages; int n1, n2, n3, new_version; + struct page **new_pages, **p; void *addr; /* @@ -321,7 +324,16 @@ static struct page **uts_prep_vdso_pages */ if (uts_ns == &init_uts_ns) return vdso32_pages; - else if (uts_ns->vdso32.pages) + + /* + * Dirty lockless hack. Strictly speaking + * we need to return @p here if it's non-nil, + * but since there only one trasition possible + * { =0 ; !=0 } we simply return @uts_ns->vdso32.pages + */ + p = ACCESS_ONCE(uts_ns->vdso32.pages); + smp_read_barrier_depends(); + if (p) return uts_ns->vdso32.pages; up_write(&mm->mmap_sem); @@ -346,41 +358,51 @@ static struct page **uts_prep_vdso_pages goto out; } + mutex_lock(&vdso32_mutex); + if (uts_ns->vdso32.pages) { + pages = uts_ns->vdso32.pages; + goto out_unlock; + } + uts_ns->vdso32.nr_pages = 1; uts_ns->vdso32.size = PAGE_SIZE; uts_ns->vdso32.version_off = (unsigned long)VDSO32_SYMBOL(0, linux_version_code); - uts_ns->vdso32.pages = kmalloc(sizeof(struct page *), GFP_KERNEL); - if (!uts_ns->vdso32.pages) { + new_pages = kmalloc(sizeof(struct page *), GFP_KERNEL); + if (!new_pages) { pr_err("Can't allocate vDSO pages array for VE %d\n", ve->veid); pages = ERR_PTR(-ENOMEM); - goto out; + goto out_unlock; } - uts_ns->vdso32.pages[0] = alloc_page(GFP_KERNEL); - if (!uts_ns->vdso32.pages[0]) { + new_pages[0] = alloc_page(GFP_KERNEL); + if (!new_pages[0]) { pr_err("Can't allocate page for VE %d\n", ve->veid); - kfree(uts_ns->vdso32.pages); - uts_ns->vdso32.pages = NULL; + kfree(new_pages); pages = ERR_PTR(-ENOMEM); - goto out; + goto out_unlock; } - copy_page(page_address(uts_ns->vdso32.pages[0]), page_address(vdso32_pages[0])); - pages = uts_ns->vdso32.pages; + copy_page(page_address(new_pages[0]), page_address(vdso32_pages[0])); - addr = page_address(uts_ns->vdso32.pages[0]); + addr = page_address(new_pages[0]); *((int *)(addr + uts_ns->vdso32.version_off)) = new_version; + smp_wmb(); + + pages = uts_ns->vdso32.pages = new_pages; + pr_debug("vDSO version transition %d -> %d for VE %d\n", LINUX_VERSION_CODE, new_version, ve->veid); #ifdef CONFIG_X86_32 - __set_fixmap(FIX_VDSO, page_to_pfn(uts_ns->vdso32.pages[0]) << PAGE_SHIFT, + __set_fixmap(FIX_VDSO, page_to_pfn(new_pages[0]) << PAGE_SHIFT, map ? PAGE_READONLY_EXEC : PAGE_NONE); /* flush stray tlbs */ flush_tlb_all(); #endif +out_unlock: + mutex_unlock(&vdso32_mutex); out: down_write(&mm->mmap_sem); return pages; Index: linux-pcs7.git/arch/x86/vdso/vma.c =================================================================== --- linux-pcs7.git.orig/arch/x86/vdso/vma.c +++ linux-pcs7.git/arch/x86/vdso/vma.c @@ -194,11 +194,14 @@ up_fail: return ret; } +static DEFINE_MUTEX(vdso_mutex); + static int uts_arch_setup_additional_pages(struct linux_binprm *bprm, int uses_interp) { struct uts_namespace *uts_ns = current->nsproxy->uts_ns; struct ve_struct *ve = get_exec_env(); int i, n1, n2, n3, new_version; + struct page **new_pages, **p; /* * For node or in case we've not changed UTS simply @@ -209,7 +212,15 @@ static int uts_arch_setup_additional_pag */ if (uts_ns == &init_uts_ns) goto map_init_uts; - else if (uts_ns->vdso.pages) + /* + * Dirty lockless hack. Strictly speaking + * we need to return @p here if it's non-nil, + * but since there only one trasition possible + * { =0 ; !=0 } we simply return @uts_ns->vdso.pages + */ + p = ACCESS_ONCE(uts_ns->vdso.pages); + smp_read_barrier_depends(); + if (p) goto map_uts; if (sscanf(uts_ns->name.release, "%d.%d.%d", &n1, &n2, &n3) == 3) { @@ -232,13 +243,19 @@ static int uts_arch_setup_additional_pag goto map_init_uts; } + mutex_lock(&vdso_mutex); + if (uts_ns->vdso.pages) { + mutex_unlock(&vdso_mutex); + goto map_uts; + } + uts_ns->vdso.nr_pages = init_uts_ns.vdso.nr_pages; uts_ns->vdso.size = init_uts_ns.vdso.size; uts_ns->vdso.version_off= init_uts_ns.vdso.version_off; - uts_ns->vdso.pages = kmalloc(sizeof(struct page *) * init_uts_ns.vdso.nr_pages, GFP_KERNEL); - if (!uts_ns->vdso.pages) { + new_pages = kmalloc(sizeof(struct page *) * init_uts_ns.vdso.nr_pages, GFP_KERNEL); + if (!new_pages) { pr_err("Can't allocate vDSO pages array for VE %d\n", ve->veid); - return -ENOMEM; + goto out_unlock; } for (i = 0; i < uts_ns->vdso.nr_pages; i++) { @@ -246,26 +263,28 @@ static int uts_arch_setup_additional_pag if (!p) { pr_err("Can't allocate page for VE %d\n", ve->veid); for (; i > 0; i--) - put_page(uts_ns->vdso.pages[i - 1]); - kfree(uts_ns->vdso.pages); - uts_ns->vdso.pages = NULL; - return -ENOMEM; + put_page(new_pages[i - 1]); + kfree(new_pages); + goto out_unlock; } - uts_ns->vdso.pages[i] = p; + new_pages[i] = p; copy_page(page_address(p), page_address(init_uts_ns.vdso.pages[i])); } - uts_ns->vdso.addr = vmap(uts_ns->vdso.pages, uts_ns->vdso.nr_pages, 0, PAGE_KERNEL); + uts_ns->vdso.addr = vmap(new_pages, uts_ns->vdso.nr_pages, 0, PAGE_KERNEL); if (!uts_ns->vdso.addr) { pr_err("Can't map vDSO pages for VE %d\n", ve->veid); for (i = 0; i < uts_ns->vdso.nr_pages; i++) - put_page(uts_ns->vdso.pages[i]); - kfree(uts_ns->vdso.pages); - uts_ns->vdso.pages = NULL; - return -ENOMEM; + put_page(new_pages[i]); + kfree(new_pages); + goto out_unlock; } *((int *)(uts_ns->vdso.addr + uts_ns->vdso.version_off)) = new_version; + smp_wmb(); + uts_ns->vdso.pages = new_pages; + mutex_unlock(&vdso_mutex); + pr_debug("vDSO version transition %d -> %d for VE %d\n", LINUX_VERSION_CODE, new_version, ve->veid); @@ -273,6 +292,9 @@ map_uts: return setup_additional_pages(bprm, uses_interp, uts_ns->vdso.pages, uts_ns->vdso.size); map_init_uts: return setup_additional_pages(bprm, uses_interp, init_uts_ns.vdso.pages, init_uts_ns.vdso.size); +out_unlock: + mutex_unlock(&vdso_mutex); + return -ENOMEM; } int arch_setup_additional_pages(struct linux_binprm *bprm, int uses_interp) _______________________________________________ Devel mailing list [email protected] https://lists.openvz.org/mailman/listinfo/devel
