From patchwork Thu Feb 13 17:04:01 2020 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: "H.J. Lu" X-Patchwork-Id: 38037 Received: (qmail 38594 invoked by alias); 13 Feb 2020 17:04:43 -0000 Mailing-List: contact libc-alpha-help@sourceware.org; run by ezmlm Precedence: bulk List-Id: List-Unsubscribe: List-Subscribe: List-Archive: List-Post: List-Help: , Sender: libc-alpha-owner@sourceware.org Delivered-To: mailing list libc-alpha@sourceware.org Received: (qmail 38570 invoked by uid 89); 13 Feb 2020 17:04:42 -0000 Authentication-Results: sourceware.org; auth=none X-Spam-SWARE-Status: No, score=-14.4 required=5.0 tests=AWL, BAYES_00, FREEMAIL_FROM, GIT_PATCH_0, GIT_PATCH_1, GIT_PATCH_2, GIT_PATCH_3, KAM_SHORT, RCVD_IN_DNSWL_NONE, SCC_5_SHORT_WORD_LINES, SPF_PASS autolearn=ham version=3.3.1 spammy=ESP, H*f:sk:2020010 X-HELO: mail-ot1-f68.google.com DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20161025; h=mime-version:references:in-reply-to:from:date:message-id:subject:to; bh=uyTMXy0eTMxfgDUr+4ogZNi8QsU9HOx3ihFJDbeNMBo=; b=m8AP/J/JdG8o8k5e25oo3Fw708m5XhE4hfip/uA21bsjeeH+9VC9hvoIP1JNN85rQe 4i1uRMnGUNcqQuvhGsT5+LZEERMbuDFvQPC3GnVOC0j9khjYv6nnXKrnlAY4wmVLNu43 EX8HTIt6EaRiE1oInff1quEDgfD7jkNKn0D9Q3f5owHSlBUgXESNtnRjgcFKRRpNzo7f Rt/r0B2RdHHX8Kn3iXyRZVXqNDw/2h81TC7KR8CMRa4bMYuSrQVoK/RxPbwJFgGWuD90 z32ScYDLDV4S3y2tQ3JOeMwBm6X7Vxx2aAhjO1Xu0/5osucddxnKHbeEUqHpJyQ0NSrv Ijog== MIME-Version: 1.0 References: <20200108161535.6141-1-hjl.tools@gmail.com> <20200108161535.6141-6-hjl.tools@gmail.com> In-Reply-To: From: "H.J. Lu" Date: Thu, 13 Feb 2020 09:04:01 -0800 Message-ID: Subject: V4 [PATCH] i386: Enable CET support in ucontext functions To: GNU C Library On Sat, Feb 1, 2020 at 6:09 AM H.J. Lu wrote: > > On Wed, Jan 8, 2020 at 8:15 AM H.J. Lu wrote: > > > > 1. getcontext and swapcontext are updated to save the caller's shadow > > stack pointer and return address. > > 2. setcontext and swapcontext are updated to restore shadow stack and > > jump to new context directly. > > 3. makecontext is updated to allocate a new shadow stack and set the > > caller's return address to the helper code, L(exitcode). > > > > Since makecontext allocates a new shadow stack when making a new > > context and kernel allocates a new shadow stack for clone/fork/vfork > > syscalls, we track the current shadow stack base. In setcontext and > > swapcontext, if the target shadow stack base is the same as the current > > shadow stack base, we unwind the shadow stack. Otherwise it is a stack > > switch and we look for a restore token. > > > > We enable shadow stack at run-time only if program and all used shared > > objects, including dlopened ones, are shadow stack enabled, which means > > that they must be compiled with GCC 8 or above and glibc 2.28 or above. > > We need to save and restore shadow stack only if shadow stack is enabled. > > When caller of getcontext, setcontext, swapcontext and makecontext is > > compiled with smaller ucontext_t, shadow stack won't be enabled at > > run-time. We check if shadow stack is enabled before accessing the > > extended field in ucontext_t. > > > > This is the updated patch. The only change is to use > > +oSCRATCH1 mreg (EAX) > +oSCRATCH2 mreg (ECX) > +oSCRATCH3 mreg (EDX) > > to replace oEAX, oECX and -oEDX. > > OK for master? > A small update. The difference between old and new are: --- sysdeps/unix/sysv/linux/i386/setcontext.S 2020-01-22 05:28:24.027273572 -0800 +++ sysdeps/unix/sysv/linux/i386/setcontext.S 2020-02-13 08:48:09.731359073 -0800 @@ -82,11 +82,11 @@ ENTRY(__setcontext) cmpl %gs:SSP_BASE_OFFSET, %ecx je L(unwind_shadow_stack) -L(find_restore_token_loop): /* Align the saved original shadow stack pointer to the next 8 byte aligned boundary. */ andl $-8, %esi +L(find_restore_token_loop): /* Look for a restore token. */ movl -8(%esi), %ebx andl $-8, %ebx --- sysdeps/unix/sysv/linux/i386/swapcontext.S 2020-01-22 05:28:24.028273573 -0800 +++ sysdeps/unix/sysv/linux/i386/swapcontext.S 2020-02-13 08:48:09.731359073 -0800 @@ -135,11 +135,11 @@ L(shadow_stack_bound_recorded): cmpl %gs:SSP_BASE_OFFSET, %ecx je L(unwind_shadow_stack) -L(find_restore_token_loop): /* Align the saved original shadow stack pointer to the next 8 byte aligned boundary. */ andl $-8, %esi +L(find_restore_token_loop): /* Look for a restore token. */ movl -8(%esi), %ebx andl $-8, %ebx We don't need to re-align shadow stack since shadow stack will always be decremented by 8 bytes after initial alignment to 8 bytes. Thanks. From 1b957edd0ed4ebee5ad401ac3f2de138d8495e79 Mon Sep 17 00:00:00 2001 From: "H.J. Lu" Date: Fri, 6 Dec 2019 16:45:17 -0800 Subject: [PATCH] i386: Enable CET support in ucontext functions 1. getcontext and swapcontext are updated to save the caller's shadow stack pointer and return address. 2. setcontext and swapcontext are updated to restore shadow stack and jump to new context directly. 3. makecontext is updated to allocate a new shadow stack and set the caller's return address to the helper code, L(exitcode). 4. Since we no longer save and restore EAX, ECX and EDX in getcontext, setcontext and swapcontext, we can use them as scratch register slots to enable CET in ucontext functions. Since makecontext allocates a new shadow stack when making a new context and kernel allocates a new shadow stack for clone/fork/vfork syscalls, we track the current shadow stack base. In setcontext and swapcontext, if the target shadow stack base is the same as the current shadow stack base, we unwind the shadow stack. Otherwise it is a stack switch and we look for a restore token. We enable shadow stack at run-time only if program and all used shared objects, including dlopened ones, are shadow stack enabled, which means that they must be compiled with GCC 8 or above and glibc 2.28 or above. We need to save and restore shadow stack only if shadow stack is enabled. When caller of getcontext, setcontext, swapcontext and makecontext is compiled with smaller ucontext_t, shadow stack won't be enabled at run-time. We check if shadow stack is enabled before accessing the extended field in ucontext_t. Tested on i386 CET/non-CET machines. --- sysdeps/unix/sysv/linux/i386/getcontext.S | 56 ++++++++ sysdeps/unix/sysv/linux/i386/makecontext.S | 123 +++++++++++++++++ sysdeps/unix/sysv/linux/i386/setcontext.S | 101 +++++++++++++- sysdeps/unix/sysv/linux/i386/swapcontext.S | 139 ++++++++++++++++++++ sysdeps/unix/sysv/linux/i386/sysdep.h | 5 + sysdeps/unix/sysv/linux/i386/ucontext_i.sym | 4 + 6 files changed, 425 insertions(+), 3 deletions(-) diff --git a/sysdeps/unix/sysv/linux/i386/getcontext.S b/sysdeps/unix/sysv/linux/i386/getcontext.S index 9c1df9a2aa..d91cfe4b1d 100644 --- a/sysdeps/unix/sysv/linux/i386/getcontext.S +++ b/sysdeps/unix/sysv/linux/i386/getcontext.S @@ -18,6 +18,7 @@ . */ #include +#include #include "ucontext_i.h" @@ -42,6 +43,61 @@ ENTRY(__getcontext) movw %fs, %dx movl %edx, oFS(%eax) +#if SHSTK_ENABLED + /* Check if shadow stack is enabled. */ + testl $X86_FEATURE_1_SHSTK, %gs:FEATURE_1_OFFSET + jz L(no_shstk) + + /* Save EAX in EDX. */ + movl %eax, %edx + + xorl %eax, %eax + cmpl %gs:SSP_BASE_OFFSET, %eax + jnz L(shadow_stack_bound_recorded) + + /* Save EBX in the first scratch register slot. */ + movl %ebx, oSCRATCH1(%edx) + + /* Get the base address and size of the default shadow stack + which must be the current shadow stack since nothing has + been recorded yet. */ + sub $24, %esp + mov %esp, %ecx + movl $ARCH_CET_STATUS, %ebx + movl $__NR_arch_prctl, %eax + ENTER_KERNEL + testl %eax, %eax + jz L(continue_no_err) + + /* This should never happen. */ + hlt + +L(continue_no_err): + /* Restore EBX from the first scratch register slot. */ + movl oSCRATCH1(%edx), %ebx + + /* Record the base of the current shadow stack. */ + movl 8(%esp), %eax + movl %eax, %gs:SSP_BASE_OFFSET + add $24, %esp + +L(shadow_stack_bound_recorded): + /* Load address of the context data structure. */ + movl 4(%esp), %eax + + /* Get the current shadow stack pointer. */ + rdsspd %edx + /* NB: Save the caller's shadow stack so that we can jump back + to the caller directly. */ + addl $4, %edx + movl %edx, oSSP(%eax) + + /* Save the current shadow stack base in ucontext. */ + movl %gs:SSP_BASE_OFFSET, %edx + movl %edx, (oSSP + 4)(%eax) + +L(no_shstk): +#endif /* We have separate floating-point register content memory on the stack. We use the __fpregs_mem block in the context. Set the links up correctly. */ diff --git a/sysdeps/unix/sysv/linux/i386/makecontext.S b/sysdeps/unix/sysv/linux/i386/makecontext.S index ad9ce5f977..91009675d1 100644 --- a/sysdeps/unix/sysv/linux/i386/makecontext.S +++ b/sysdeps/unix/sysv/linux/i386/makecontext.S @@ -18,6 +18,7 @@ . */ #include +#include #include "ucontext_i.h" @@ -68,6 +69,127 @@ ENTRY(__makecontext) jnz 1b 2: +#if SHSTK_ENABLED + /* Check if Shadow Stack is enabled. */ + testl $X86_FEATURE_1_SHSTK, %gs:FEATURE_1_OFFSET + jz L(skip_ssp) + + /* Reload the pointer to ucontext. */ + movl 4(%esp), %eax + + /* Shadow stack is enabled. We need to allocate a new shadow + stack. */ + subl oSS_SP(%eax), %edx + shrl $STACK_SIZE_TO_SHADOW_STACK_SIZE_SHIFT, %edx + + /* Align shadow stack size to 8 bytes. */ + addl $7, %edx + andl $-8, %edx + + /* Store shadow stack size in __ssp[2]. */ + movl %edx, (oSSP + 8)(%eax) + + /* Save ESI in the second scratch register slot. */ + movl %esi, oSCRATCH2(%eax) + /* Save EDI in the third scratch register slot. */ + movl %edi, oSCRATCH3(%eax) + + /* Save the pointer to ucontext. */ + movl %eax, %edi + + /* Get the original shadow stack pointer. */ + rdsspd %esi + + /* Align the saved original shadow stack pointer to the next + 8 byte aligned boundary. */ + andl $-8, %esi + + /* Load the top of the new stack into EDX. */ + movl oESP(%eax), %edx + + /* We need to terminate the FDE here because the unwinder looks + at ra-1 for unwind information. */ + cfi_endproc + + /* Swap the original stack pointer with the top of the new + stack. */ + xchgl %esp, %edx + + /* Add 4 bytes since CALL will push the 4-byte return address + onto stack. */ + addl $4, %esp + + /* Allocate the new shadow stack. Save EBX in the first scratch + register slot. */ + movl %ebx, oSCRATCH1(%eax) + + /* CET syscall takes 64-bit sizes. */ + subl $16, %esp + movl (oSSP + 8)(%eax), %ecx + movl %ecx, (%esp) + movl $0, 4(%esp) + movl %ecx, 8(%esp) + movl $0, 12(%esp) + movl %esp, %ecx + + movl $ARCH_CET_ALLOC_SHSTK, %ebx + movl $__NR_arch_prctl, %eax + ENTER_KERNEL + testl %eax, %eax + jne L(hlt) /* This should never happen. */ + + /* Copy the base address of the new shadow stack to __ssp[1]. */ + movl (%esp), %eax + movl %eax, (oSSP + 4)(%edi) + + addl $16, %esp + + /* Restore EBX from the first scratch register slot. */ + movl oSCRATCH1(%edi), %ebx + + /* Get the size of the new shadow stack. */ + movl (oSSP + 8)(%edi), %ecx + + /* Use the restore stoken to restore the new shadow stack. */ + rstorssp -8(%eax, %ecx) + + /* Save the restore token at the next 8 byte aligned boundary + on the original shadow stack. */ + saveprevssp + + /* Push the address of "jmp exitcode" onto the new stack as + well as the new shadow stack. */ + call 1f + jmp L(exitcode) +1: + + /* Get the new shadow stack pointer. */ + rdsspd %eax + + /* Use the restore stoken to restore the original shadow stack. */ + rstorssp -8(%esi) + + /* Save the restore token on the new shadow stack. */ + saveprevssp + + /* Store the new shadow stack pointer in __ssp[0]. */ + movl %eax, oSSP(%edi) + + /* Restore the original stack. */ + mov %edx, %esp + + cfi_startproc + + /* Restore ESI from the second scratch register slot. */ + movl oSCRATCH2(%edi), %esi + /* Restore EDI from the third scratch register slot. */ + movl oSCRATCH3(%edi), %edi + + ret + +L(skip_ssp): +#endif + /* If the function we call returns we must continue with the context which is given in the uc_link element. To do this set the return address for the function the user provides @@ -123,6 +245,7 @@ L(call_exit): call HIDDEN_JUMPTARGET(exit) /* The 'exit' call should never return. In case it does cause the process to terminate. */ +L(hlt): hlt cfi_startproc END(__makecontext) diff --git a/sysdeps/unix/sysv/linux/i386/setcontext.S b/sysdeps/unix/sysv/linux/i386/setcontext.S index f042d80bf4..332b5147bc 100644 --- a/sysdeps/unix/sysv/linux/i386/setcontext.S +++ b/sysdeps/unix/sysv/linux/i386/setcontext.S @@ -18,6 +18,7 @@ . */ #include +#include #include "ucontext_i.h" @@ -56,9 +57,6 @@ ENTRY(__setcontext) movl oFS(%eax), %ecx movw %cx, %fs - /* Fetch the address to return to. */ - movl oEIP(%eax), %ecx - /* Load the new stack pointer. */ cfi_def_cfa (eax, 0) cfi_offset (edi, oEDI) @@ -67,6 +65,103 @@ ENTRY(__setcontext) cfi_offset (ebx, oEBX) movl oESP(%eax), %esp +#if SHSTK_ENABLED + /* Check if Shadow Stack is enabled. */ + testl $X86_FEATURE_1_SHSTK, %gs:FEATURE_1_OFFSET + jz L(no_shstk) + + /* If the base of the target shadow stack is the same as the + base of the current shadow stack, we unwind the shadow + stack. Otherwise it is a stack switch and we look for a + restore token. */ + movl oSSP(%eax), %esi + movl %esi, %edi + + /* Get the base of the target shadow stack. */ + movl (oSSP + 4)(%eax), %ecx + cmpl %gs:SSP_BASE_OFFSET, %ecx + je L(unwind_shadow_stack) + + /* Align the saved original shadow stack pointer to the next + 8 byte aligned boundary. */ + andl $-8, %esi + +L(find_restore_token_loop): + /* Look for a restore token. */ + movl -8(%esi), %ebx + andl $-8, %ebx + cmpl %esi, %ebx + je L(restore_shadow_stack) + + /* Try the next slot. */ + subl $8, %esi + jmp L(find_restore_token_loop) + +L(restore_shadow_stack): + /* Pop return address from the shadow stack since setcontext + will not return. */ + movl $1, %ebx + incsspd %ebx + + /* Use the restore stoken to restore the target shadow stack. */ + rstorssp -8(%esi) + + /* Save the restore token on the old shadow stack. NB: This + restore token may be checked by setcontext or swapcontext + later. */ + saveprevssp + + /* Record the new shadow stack base that was switched to. */ + movl (oSSP + 4)(%eax), %ebx + movl %ebx, %gs:SSP_BASE_OFFSET + +L(unwind_shadow_stack): + rdsspd %ebx + subl %edi, %ebx + je L(skip_unwind_shadow_stack) + negl %ebx + shrl $2, %ebx + movl $255, %esi +L(loop): + cmpl %esi, %ebx + cmovb %ebx, %esi + incsspd %esi + subl %esi, %ebx + ja L(loop) + +L(skip_unwind_shadow_stack): + + /* Load the values of all the preserved registers (except ESP). */ + movl oEDI(%eax), %edi + movl oESI(%eax), %esi + movl oEBP(%eax), %ebp + movl oEBX(%eax), %ebx + + /* Get the return address set with getcontext. */ + movl oEIP(%eax), %ecx + + /* Check if return address is valid for the case when setcontext + is invoked from L(exitcode) with linked context. */ + rdsspd %eax + cmpl (%eax), %ecx + /* Clear EAX to indicate success. NB: Don't use xorl to keep + EFLAGS for jne. */ + movl $0, %eax + jne L(jmp) + /* Return to the new context if return address valid. */ + pushl %ecx + ret + +L(jmp): + /* Jump to the new context directly. */ + jmp *%ecx + +L(no_shstk): +#endif + + /* Fetch the address to return to. */ + movl oEIP(%eax), %ecx + /* Push the return address on the new stack so we can return there. */ pushl %ecx diff --git a/sysdeps/unix/sysv/linux/i386/swapcontext.S b/sysdeps/unix/sysv/linux/i386/swapcontext.S index 090c2d8c3e..203eafa2e7 100644 --- a/sysdeps/unix/sysv/linux/i386/swapcontext.S +++ b/sysdeps/unix/sysv/linux/i386/swapcontext.S @@ -18,6 +18,7 @@ . */ #include +#include #include "ucontext_i.h" @@ -76,6 +77,144 @@ ENTRY(__swapcontext) movl oFS(%eax), %edx movw %dx, %fs +#if SHSTK_ENABLED + /* Check if Shadow Stack is enabled. */ + testl $X86_FEATURE_1_SHSTK, %gs:FEATURE_1_OFFSET + jz L(no_shstk) + + xorl %eax, %eax + cmpl %gs:SSP_BASE_OFFSET, %eax + jnz L(shadow_stack_bound_recorded) + + /* Get the base address and size of the default shadow stack + which must be the current shadow stack since nothing has + been recorded yet. */ + sub $24, %esp + mov %esp, %ecx + movl $ARCH_CET_STATUS, %ebx + movl $__NR_arch_prctl, %eax + ENTER_KERNEL + testl %eax, %eax + jz L(continue_no_err) + + /* This should never happen. */ + hlt + +L(continue_no_err): + /* Record the base of the current shadow stack. */ + movl 8(%esp), %eax + movl %eax, %gs:SSP_BASE_OFFSET + add $24, %esp + +L(shadow_stack_bound_recorded): + /* Load address of the context data structure we save in. */ + movl 4(%esp), %eax + + /* Load address of the context data structure we swap in */ + movl 8(%esp), %edx + + /* If we unwind the stack, we can't undo stack unwinding. Just + save the target shadow stack pointer as the current shadow + stack pointer. */ + movl oSSP(%edx), %ecx + movl %ecx, oSSP(%eax) + + /* Save the current shadow stack base in ucontext. */ + movl %gs:SSP_BASE_OFFSET, %ecx + movl %ecx, (oSSP + 4)(%eax) + + /* If the base of the target shadow stack is the same as the + base of the current shadow stack, we unwind the shadow + stack. Otherwise it is a stack switch and we look for a + restore token. */ + movl oSSP(%edx), %esi + movl %esi, %edi + + /* Get the base of the target shadow stack. */ + movl (oSSP + 4)(%edx), %ecx + cmpl %gs:SSP_BASE_OFFSET, %ecx + je L(unwind_shadow_stack) + + /* Align the saved original shadow stack pointer to the next + 8 byte aligned boundary. */ + andl $-8, %esi + +L(find_restore_token_loop): + /* Look for a restore token. */ + movl -8(%esi), %ebx + andl $-8, %ebx + cmpl %esi, %ebx + je L(restore_shadow_stack) + + /* Try the next slot. */ + subl $8, %esi + jmp L(find_restore_token_loop) + +L(restore_shadow_stack): + /* The target shadow stack will be restored. Save the current + shadow stack pointer. */ + rdsspd %ecx + movl %ecx, oSSP(%eax) + + /* Use the restore stoken to restore the target shadow stack. */ + rstorssp -8(%esi) + + /* Save the restore token on the old shadow stack. NB: This + restore token may be checked by setcontext or swapcontext + later. */ + saveprevssp + + /* Record the new shadow stack base that was switched to. */ + movl (oSSP + 4)(%edx), %ebx + movl %ebx, %gs:SSP_BASE_OFFSET + +L(unwind_shadow_stack): + rdsspd %ebx + subl %edi, %ebx + je L(skip_unwind_shadow_stack) + negl %ebx + shrl $2, %ebx + movl $255, %esi +L(loop): + cmpl %esi, %ebx + cmovb %ebx, %esi + incsspd %esi + subl %esi, %ebx + ja L(loop) + +L(skip_unwind_shadow_stack): + + /* Load the new stack pointer. */ + movl oESP(%edx), %esp + + /* Load the values of all the preserved registers (except ESP). */ + movl oEDI(%edx), %edi + movl oESI(%edx), %esi + movl oEBP(%edx), %ebp + movl oEBX(%edx), %ebx + + /* Get the return address set with getcontext. */ + movl oEIP(%edx), %ecx + + /* Check if return address is valid for the case when setcontext + is invoked from L(exitcode) with linked context. */ + rdsspd %eax + cmpl (%eax), %ecx + /* Clear EAX to indicate success. NB: Don't use xorl to keep + EFLAGS for jne. */ + movl $0, %eax + jne L(jmp) + /* Return to the new context if return address valid. */ + pushl %ecx + ret + +L(jmp): + /* Jump to the new context directly. */ + jmp *%ecx + +L(no_shstk): +#endif + /* Fetch the address to return to. */ movl oEIP(%eax), %ecx diff --git a/sysdeps/unix/sysv/linux/i386/sysdep.h b/sysdeps/unix/sysv/linux/i386/sysdep.h index 4aa7bb496a..420b6a7912 100644 --- a/sysdeps/unix/sysv/linux/i386/sysdep.h +++ b/sysdeps/unix/sysv/linux/i386/sysdep.h @@ -662,4 +662,9 @@ struct libc_do_syscall_args # endif #endif +/* Each shadow stack slot takes 4 bytes. Assuming that each stack + frame takes 128 bytes, this is used to compute shadow stack size + from stack size. */ +#define STACK_SIZE_TO_SHADOW_STACK_SIZE_SHIFT 5 + #endif /* linux/i386/sysdep.h */ diff --git a/sysdeps/unix/sysv/linux/i386/ucontext_i.sym b/sysdeps/unix/sysv/linux/i386/ucontext_i.sym index 1dfe03d2cc..1d8608eafc 100644 --- a/sysdeps/unix/sysv/linux/i386/ucontext_i.sym +++ b/sysdeps/unix/sysv/linux/i386/ucontext_i.sym @@ -22,6 +22,10 @@ oEBP mreg (EBP) oESP mreg (ESP) oEBX mreg (EBX) oEIP mreg (EIP) +oSCRATCH1 mreg (EAX) +oSCRATCH2 mreg (ECX) +oSCRATCH3 mreg (EDX) oFPREGS mcontext (fpregs) oSIGMASK ucontext (uc_sigmask) oFPREGSMEM ucontext (__fpregs_mem) +oSSP ucontext (__ssp) -- 2.24.1