linux-user: Fix siginfo_t contents when jumping to non-readable pages
v5 cover linux-user: Fix siginfo_t contents when jumping to non-readable pages
Hi,
I noticed that when we get a SEGV due to jumping to non-readable memory, sometimes si_addr and program counter in siginfo_t are slightly off. I tracked this down to the assumption that translators stop before the end of a page, while in reality they may stop right after it.
Patch 1 fixes an invalidation issue, which may prevent SEGV from happening altogether.
Patches 2-3 fix the main issue on x86_64 and s390x. Many other architectures have fixed-size instructions and are not affected. Patch 4 adds tests.
Note: this series depends on [1].
Best regards,
Ilya
v1: https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg00822.html
v1 -> v2: Fix individual translators instead of translator_loop
(Peter).
v2: https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg01079.html
v2 -> v3: Peek at the next instruction on s390x (Richard).
Undo more on i386 (Richard).
Check PAGE_EXEC, not PAGE_READ (Peter, Richard).
v3: https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg01306.html
v3 -> v4: Improve the commit message in patch 1 to better reflect what
exactly is being fixed there.
Factor out the is_same_page() patch (Richard).
Do not touch the common code in the i386 fix (Richard).
v4: https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg01747.html
v4 -> v5: Drop patch 2.
Use a different fix for the invalidation issue based on discussion with Richard [2].
[1] https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg02472.html
[2] https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg02556.html
Ilya Leoshkevich (4):
linux-user: Clear tb_jmp_cache on mprotect()
target/s390x: Make translator stop before the end of a page target/i386: Make translator stop before the end of a page tests/tcg: Test siginfo_t contents when jumping to non-readable pages
linux-user/mmap.c | 14 +++ target/i386/tcg/translate.c | 25 +++++- target/s390x/tcg/translate.c | 15 +++- tests/tcg/multiarch/noexec.h | 114 ++++++++++++++++++++++++ tests/tcg/s390x/Makefile.target | 1 + tests/tcg/s390x/noexec.c | 145 +++++++++++++++++++++++++++++++ tests/tcg/x86_64/Makefile.target | 3 +- tests/tcg/x86_64/noexec.c | 116 +++++++++++++++++++++++++ 8 files changed, 427 insertions(+), 6 deletions(-) create mode 100644 tests/tcg/multiarch/noexec.h create mode 100644 tests/tcg/s390x/noexec.c create mode 100644 tests/tcg/x86_64/noexec.c
Signature · 1 line
2.37.1
PATCH v5 1/4 linux-user: Clear tb_jmp_cache on mprotect()
Currently it's possible to execute pages that do not have PAGE_EXEC if there is an existing translation block. Fix by clearing tb_jmp_cache, which forces HELPER(lookup_tb_ptr)() to recheck permission bits the next time.
--- linux-user/mmap.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)
linux-user/mmap.c +14 -0
@@ -115,6 +115,7 @@ int target_mprotect(abi_ulong start, abi_ulong len, int target_prot) {abi_ulong end, host_start, host_end, addr;int prot1, ret, page_flags, host_prot; +CPUState *cpu;trace_target_mprotect (start, len, target_prot);
@@ -178,6 +179,19 @@ int target_mprotect(abi_ulong start, abi_ulong len, int target_prot) } }page_set_flags (start, start + len, page_flags); + +/* +* Unlike target_mmap(), target_munmap() and target_mremap(), we don't need +* to call tb_invalidate_phys_range() here, since pages still hold the same +* data. However, tb_jmp_cache needs to be cleared, otherwise after +* clearing PAGE_EXEC it would still be possible to jump to the existing +* translation blocks. We just clear the whole cache here: mprotect() is +* rare enough, so a more fine-grained approach is not necessary. +*/ +CPU_FOREACH (cpu) { +cpu_tb_jmp_cache_clear (cpu); + } +mmap_unlock ();return 0 ; error:
Signature · 1 line
2.37.1
PATCH v5 2/4 target/s390x: Make translator stop before the end of a page
In reply to Ilya LeoshkevichRight now translator stops right *after* the end of a page, which breaks reporting of fault locations when the last instruction of a multi-insn translation block crosses a page boundary.
--- target/s390x/tcg/translate.c | 15 +++++++++++---- 1 file changed, 11 insertions(+), 4 deletions(-)
target/s390x/tcg/translate.c +11 -4
@@ -6609,6 +6609,14 @@ static void s390x_tr_insn_start(DisasContextBase *dcbase, CPUState *cs) dc->insn_start =tcg_last_op (); } +static target_ulong get_next_pc (CPUS390XState *env,DisasContext *s, +uint64_t pc) +{ +uint64_t insn =ld_code2 (env, s, pc); + +return pc +get_ilen ((insn >>8 ) &0xff ); +} +static void s390x_tr_translate_insn (DisasContextBase *dcbase,CPUState *cs) {CPUS390XState *env = cs->env_ptr;
@@ -6616,10 +6624,9 @@ static void s390x_tr_translate_insn(DisasContextBase *dcbase, CPUState *cs) dc->base.is_jmp =translate_one (env, dc);if (dc->base.is_jmp ==DISAS_NEXT ) { -uint64_t page_start; - - page_start = dc->base.pc_first &TARGET_PAGE_MASK ; -if (dc->base.pc_next - page_start >=TARGET_PAGE_SIZE || dc->ex_value) { +if (!is_same_page (dcbase, dc->base.pc_next) || + !is_same_page (dcbase,get_next_pc (env, dc, dc->base.pc_next)) || + dc->ex_value) { dc->base.is_jmp =DISAS_TOO_MANY ; } }
Signature · 1 line
2.37.1
PATCH v5 3/4 target/i386: Make translator stop before the end of a page
In reply to Ilya LeoshkevichRight now translator stops right *after* the end of a page, which breaks reporting of fault locations when the last instruction of a multi-insn translation block crosses a page boundary.
An implementation, like the one arm and s390x have, would require an i386 length disassembler, which is burdensome to maintain. Another alternative would be to single-step at the end of a guest page, but this may come with a performance impact.
Fix by snapshotting disassembly state and restoring it after we figure out we crossed a page boundary. This includes rolling back cc_op updates and emitted ops.
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/1143
--- target/i386/tcg/translate.c | 25 ++++++++++++++++++++++++- 1 file changed, 24 insertions(+), 1 deletion(-)
target/i386/tcg/translate.c +24 -1
@@ -130,6 +130,7 @@ typedef struct DisasContext {TCGv_i64 tmp1_i64;sigjmp_buf jmpbuf; +TCGOp *prev_insn_end; } DisasContext;/* The environment in which user-only runs is constrained. */
@@ -2008,6 +2009,12 @@ static uint64_t advance_pc(CPUX86State *env, DisasContext *s, int num_bytes) {uint64_t pc = s->pc; +/* This is a subsequent insn that crosses a page boundary. */ +if (s->base.num_insns >1 && + !is_same_page (&s->base, s->pc + num_bytes -1 )) { +siglongjmp (s->jmpbuf,2 ); + } + s->pc += num_bytes;if (unlikely (s->pc - s->pc_start >X86_MAX_INSN_LENGTH )) {/* If the instruction's 16th byte is on a different page than the 1st, a
@@ -4556,6 +4563,8 @@ static target_ulong disas_insn(DisasContext *s, CPUState *cpu)int modrm, reg, rm, mod, op, opreg, val;target_ulong next_eip, tval;target_ulong pc_start = s->base.pc_next; +bool orig_cc_op_dirty = s->cc_op_dirty; +CCOp orig_cc_op = s->cc_op; s->pc_start = s->pc = pc_start; s->override =-1 ;
@@ -4568,9 +4577,22 @@ static target_ulong disas_insn(DisasContext *s, CPUState *cpu) s->rip_offset =0 ;/* for relative ip address */ s->vex_l =0 ; s->vex_v =0 ; -if (sigsetjmp (s->jmpbuf,0 ) !=0 ) { +switch (sigsetjmp (s->jmpbuf,0 )) { +case 0 : +break ; +case 1 :gen_exception_gpf (s);return s->pc; +case 2 : +/* Restore state that may affect the next instruction. */ + s->cc_op_dirty = orig_cc_op_dirty; + s->cc_op = orig_cc_op; + s->base.num_insns--; +tcg_remove_ops_after (s->prev_insn_end); + s->base.is_jmp =DISAS_TOO_MANY ; +return pc_start; +default : +g_assert_not_reached (); } prefixes =0 ;
@@ -8632,6 +8654,7 @@ static void i386_tr_insn_start(DisasContextBase *dcbase, CPUState *cpu) {DisasContext *dc =container_of (dcbase, DisasContext, base); + dc->prev_insn_end =tcg_last_op ();tcg_gen_insn_start (dc->base.pc_next, dc->cc_op); }
Signature · 1 line
2.37.1
PATCH v5 4/4 tests/tcg: Test siginfo_t contents when jumping to non-readable pages
In reply to Ilya LeoshkevichAdd x86_64 and s390x tests to prevent regressions.
--- tests/tcg/multiarch/noexec.h | 114 ++++++++++++++++++++++++ tests/tcg/s390x/Makefile.target | 1 + tests/tcg/s390x/noexec.c | 145 +++++++++++++++++++++++++++++++ tests/tcg/x86_64/Makefile.target | 3 +- tests/tcg/x86_64/noexec.c | 116 +++++++++++++++++++++++++ 5 files changed, 378 insertions(+), 1 deletion(-) create mode 100644 tests/tcg/multiarch/noexec.h create mode 100644 tests/tcg/s390x/noexec.c create mode 100644 tests/tcg/x86_64/noexec.c
tests/tcg/multiarch/noexec.h new file +114 -0
@@ -0,0 +1,114 @@ +/* +* Common code for arch-specific MMU_INST_FETCH fault testing. +* +* Declare struct arch_noexec_test before including this file and define +* arch_check_mcontext() after that. +*/ + +#include <assert.h> +#include <signal.h> +#include <stdlib.h> +#include <string.h> +#include <sys/mman.h> +#include <sys/ucontext.h> +#include <unistd.h> + +/* Forward declarations. */ + +static void arch_check_mcontext (const struct arch_noexec_test *test, +const mcontext_t *ctx); + +/* Utility functions. */ + +static void safe_print (const char *s) +{ +write (0 , s,strlen (s)); +} + +static void safe_puts (const char *s) +{ +safe_print (s); +safe_print ("\n" ); +} + +#define PAGE_ALIGN (p) (void *)((unsigned long)(p) & ~0xfffUL) + +/* Testing infrastructure. */ + +struct noexec_test { +const char *name; +void (*func)(int ); +void *page; +void *expected_si_addr; +struct arch_noexec_test arch; +}; + +static const struct noexec_test *current_noexec_test; + +static void handle_segv (int sig,siginfo_t *info,void *ucontext) +{ +int err; + +if (current_noexec_test ==NULL ) { +safe_puts ("[ FAILED ] unexpected SEGV" ); +_exit (1 ); + } + +if (info->si_addr != current_noexec_test->expected_si_addr) { +safe_puts ("[ FAILED ] wrong si_addr" ); +_exit (1 ); + } + +arch_check_mcontext (¤t_noexec_test->arch, + &((ucontext_t *)ucontext)->uc_mcontext); + + err =mprotect (current_noexec_test->page,0x1000 ,PROT_READ |PROT_EXEC ); +if (err !=0 ) { +safe_puts ("[ FAILED ] mprotect() failed" ); +_exit (1 ); + } + + current_noexec_test =NULL ; +} + +static void test_noexec_1 (const struct noexec_test *test) +{ +int ret; + +/* Trigger TB creation in order to test invalidation. */ + test->func (0 ); + + ret =mprotect (test->page,0x1000 ,PROT_NONE ); +assert (ret ==0 ); + +/* Trigger SEGV and check that handle_segv() ran. */ + current_noexec_test = test; + test->func (0 ); +assert (current_noexec_test ==NULL ); +} + +static int test_noexec (struct noexec_test *tests,size_t n_tests) +{ +struct sigaction act; +size_t i; +int err; + +memset (&act,0 ,sizeof (act)); + act.sa_sigaction = handle_segv; + act.sa_flags =SA_SIGINFO ; + err =sigaction (SIGSEGV , &act,NULL ); +assert (err ==0 ); + +for (i =0 ; i < n_tests; i++) { +struct noexec_test *test = &tests[i]; + +safe_print ("[ RUN ] " ); +safe_puts (test->name); +test_noexec_1 (test); +safe_puts ("[ OK ]" ); + } + +safe_puts ("[ PASSED ]" ); + +return EXIT_SUCCESS ; +}
tests/tcg/s390x/Makefile.target +1 -0
@@ -16,6 +16,7 @@ TESTS+=shift TESTS+=trap TESTS+=signals-s390x TESTS+=branch-relative-long +TESTS+=noexec Z14_TESTS=vfminmax vfminmax: LDFLAGS+=-lm
tests/tcg/s390x/noexec.c new file +145 -0
145 lines not on this page · raw message
tests/tcg/x86_64/Makefile.target +2 -1
@@ -10,6 +10,7 @@ include $(SRC_PATH)/tests/tcg/i386/Makefile.target ifeq ($(filter %-linux-user, $(TARGET)),$(TARGET)) X86_64_TESTS += vsyscall +X86_64_TESTS += noexec TESTS=$(MULTIARCH_TESTS) $(X86_64_TESTS) test-x86_64 else TESTS=$(MULTIARCH_TESTS)
@@ -20,5 +21,5 @@ test-x86_64: LDFLAGS+=-lm -lc test-x86_64: test-i386.c test-i386.h test-i386-shift.h test-i386-muldiv.h $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS) -vsyscall: $(SRC_PATH)/tests/tcg/x86_64/vsyscall.c +%: $(SRC_PATH)/tests/tcg/x86_64/%.c $(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
tests/tcg/x86_64/noexec.c new file +116 -0
116 lines not on this page · raw message
Signature · 1 line
2.37.1
Quoted 7 lines · Ilya Leoshkevich
Hi,
I noticed that when we get a SEGV due to jumping to non-readable memory, sometimes
si_addrand program counter insiginfo_tare slightly off. I tracked this down to the assumption that translators stop before the end of a page, while in reality they may stop right after it.
Quoted 7 lines
Patch 1 fixes an invalidation issue, which may prevent SEGV from happening altogether.
Patches 2-3 fix the main issue onx86_64and s390x. Many other architectures have fixed-size instructions and are not affected. Patch 4 adds tests.Note: this series depends on [1].
Hah. I was just thinking that I should queue your patch set to tcg-next-7.2, and then rebase my stuff off of that. It would ensure that I have your test cases in tree so that I don't keep regressing them on you. :-)
I'll cherry pick the one patch you're depending on.
r~
Quoted 24 lines · Richard Henderson
Quoted 9 lines · Ilya Leoshkevich
Hi,
I noticed that when we get a SEGV due to jumping to non-readable memory, sometimes
si_addrand program counter insiginfo_tare slightly
off. I tracked this down to the assumption that translators stop before
the end of a page, while in reality they may stop right after it.Quoted 7 lines
Patch 1 fixes an invalidation issue, which may prevent SEGV from happening altogether.
Patches 2-3 fix the main issue onx86_64and s390x. Many other architectures have fixed-size instructions and are not affected. Patch 4 adds tests.Note: this series depends on [1].
Hah. I was just thinking that I should queue your patch set to tcg- next-7.2, and then
rebase my stuff off of that. It would ensure that I have your test cases in tree so that
I don't keep regressing them on you. :-)
Quoted 4 lines
I'll cherry pick the one patch you're depending on.
r~
I just checked and cherry-picking [1] and [2] before this series should be enough.
[1] https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg02462.html
[2] https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg02461.html
Quoted 6 lines · Ilya Leoshkevich
Hi,
I noticed that when we get a SEGV due to jumping to non-readable memory, sometimes
si_addrand program counter insiginfo_tare slightly off. I tracked this down to the assumption that translators stop before the end of a page, while in reality they may stop right after it.
Hi,
Could this be related to issue 1155 [1]? On RISC-V, I'm getting incorrect [m|s]tval/[m|s]epc combinations for page faults in system emulation and incorrect si_addr and program counter on SIGSEGV in user emulation. Since it seems to only affect instructions that cross page boundaries, and RISC-V also has variable length instructions, it seems that I've run into the same problem as what is fixed here.
Could this fix be extended be extended to targets/riscv?
dram
[1]: https://gitlab.com/qemu-project/qemu/-/issues/1155
Quoted 53 lines
Patch 1 fixes an invalidation issue, which may prevent SEGV from happening altogether.
Patches 2-3 fix the main issue onx86_64and s390x. Many other architectures have fixed-size instructions and are not affected. Patch 4 adds tests.Note: this series depends on [1].
Best regards,
Ilyav1: https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg00822.html
v1 -> v2: Fix individual translators instead oftranslator_loop(Peter).v2: https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg01079.html
v2 -> v3: Peek at the next instruction on s390x (Richard).
Undo more on i386 (Richard).
CheckPAGE_EXEC, notPAGE_READ(Peter, Richard).v3: https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg01306.html
v3 -> v4: Improve the commit message in patch 1 to better reflect what
exactly is being fixed there.
Factor out theis_same_page()patch (Richard).
Do not touch the common code in the i386 fix (Richard).v4: https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg01747.html
v4 -> v5: Drop patch 2.
Use a different fix for the invalidation issue based on discussion with Richard [2].[1] https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg02472.html
[2] https://lists.gnu.org/archive/html/qemu-devel/2022-08/msg02556.htmlIlya Leoshkevich (4):
linux-user: Cleartb_jmp_cacheonmprotect()
target/s390x: Make translator stop before the end of a page target/i386: Make translator stop before the end of a page tests/tcg: Testsiginfo_tcontents when jumping to non-readable pageslinux-user/mmap.c | 14 +++ target/i386/tcg/translate.c | 25 +++++- target/s390x/tcg/translate.c | 15 +++- tests/tcg/multiarch/noexec.h | 114 ++++++++++++++++++++++++ tests/tcg/s390x/Makefile.target | 1 + tests/tcg/s390x/noexec.c | 145 +++++++++++++++++++++++++++++++ tests/tcg/x86_64/Makefile.target | 3 +- tests/tcg/x86_64/noexec.c | 116 +++++++++++++++++++++++++ 8 files changed, 427 insertions(+), 6 deletions(-) create mode 100644 tests/tcg/multiarch/noexec.h create mode 100644 tests/tcg/s390x/noexec.c create mode 100644 tests/tcg/x86_64/noexec.c
Quoted 10 lines · Vivian Wang
Hi,
Could this be related to issue 1155 [1]? On RISC-V, I'm getting incorrect [m|s]tval/[m|s]epc combinations for page faults in system emulation and incorrectsi_addrand program counter on SIGSEGV in user emulation. Since it seems to only affect instructions that cross page boundaries, and RISC-V also has variable length instructions, it seems that I've run into the same problem as what is fixed here. Could this fix be extended be extended to targets/riscv? dram
[1]: https://gitlab.com/qemu-project/qemu/-/issues/1155
Yes, this looks quite similar.
I'm not too familiar with riscv, but I just googled [1]. If the following is correct:
---
However, the instruction set reserves enough opcode space to make it possible to differentiate between 16-bit, 32-bit, 48-bit, and 64-bit instructions. Instructions that start with binary 11 (in the lowest bit position of the instruction) are 32-bit sized instructions (but one pattern is reserved: so they cannot start with 11111). The compact instructions use 00, 01, and 10 in that same position. 48-bit instructions use starting sequence 011111, and 64-bit instructions start with 0111111.
---
then we can fix this the same way s390x is being fixed here.
Quoted 11 lines · Vivian Wang
Quoted 6 lines · Ilya Leoshkevich
Hi,
I noticed that when we get a SEGV due to jumping to non-readable memory, sometimes
si_addrand program counter insiginfo_tare slightly off. I tracked this down to the assumption that translators stop before the end of a page, while in reality they may stop right after it.Hi,
Could this be related to issue 1155 [1]? On RISC-V, I'm getting incorrect [m|s]tval/[m|s]epc combinations for page faults in system emulation and incorrect
si_addrand program counter on SIGSEGV in user emulation. Since it seems to only affect instructions that cross page boundaries, and RISC-V also has variable length instructions, it seems that I've run into the same problem as what is fixed here.
It seems likely, and the code at the end of riscv_tr_translate_insn is wrong.
Quoted 1 line
Could this fix be extended be extended to targets/riscv?
I'll write up something.
r~