]> www.infradead.org Git - users/dwmw2/linux.git/commitdiff
powerpc/64s: flush L1D on kernel entry
authorNicholas Piggin <npiggin@gmail.com>
Thu, 19 Nov 2020 23:52:39 +0000 (10:52 +1100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 22 Nov 2020 09:00:23 +0000 (10:00 +0100)
commit f79643787e0a0762d2409b7b8334e83f22d85695 upstream.

IBM Power9 processors can speculatively operate on data in the L1 cache before
it has been completely validated, via a way-prediction mechanism. It is not possible
for an attacker to determine the contents of impermissible memory using this method,
since these systems implement a combination of hardware and software security measures
to prevent scenarios where protected data could be leaked.

However these measures don't address the scenario where an attacker induces
the operating system to speculatively execute instructions using data that the
attacker controls. This can be used for example to speculatively bypass "kernel
user access prevention" techniques, as discovered by Anthony Steinhauser of
Google's Safeside Project. This is not an attack by itself, but there is a possibility
it could be used in conjunction with side-channels or other weaknesses in the
privileged code to construct an attack.

This issue can be mitigated by flushing the L1 cache between privilege boundaries
of concern. This patch flushes the L1 cache on kernel entry.

This is part of the fix for CVE-2020-4788.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Documentation/admin-guide/kernel-parameters.txt
arch/powerpc/include/asm/exception-64s.h
arch/powerpc/include/asm/feature-fixups.h
arch/powerpc/include/asm/security_features.h
arch/powerpc/include/asm/setup.h
arch/powerpc/kernel/exceptions-64s.S
arch/powerpc/kernel/setup_64.c
arch/powerpc/kernel/vmlinux.lds.S
arch/powerpc/lib/feature-fixups.c
arch/powerpc/platforms/powernv/setup.c
arch/powerpc/platforms/pseries/setup.c

index e0ce14f028d82679df158c29537f4aa15a3ffed2..d61d15de22cb8404906ee942be18339816dde7dc 100644 (file)
                                               mds=off [X86]
                                               tsx_async_abort=off [X86]
                                               kvm.nx_huge_pages=off [X86]
+                                              no_entry_flush [PPC]
 
                                Exceptions:
                                               This does not have any effect on
 
        noefi           Disable EFI runtime services support.
 
+       no_entry_flush  [PPC] Don't flush the L1-D cache when entering the kernel.
+
        noexec          [IA-64]
 
        noexec          [X86]
index a99a455bc267036cf3674909fed403cfd36bd78f..88254597865143580d9c79200b10887d453bebd3 100644 (file)
        nop;                                                            \
        nop
 
+#define ENTRY_FLUSH_SLOT                                               \
+       ENTRY_FLUSH_FIXUP_SECTION;                                      \
+       nop;                                                            \
+       nop;                                                            \
+       nop;
+
 /*
  * r10 must be free to use, r13 must be paca
  */
 #define INTERRUPT_TO_KERNEL                                            \
-       STF_ENTRY_BARRIER_SLOT
+       STF_ENTRY_BARRIER_SLOT;                                         \
+       ENTRY_FLUSH_SLOT
 
 /*
  * Macros for annotating the expected destination of (h)rfid
index b1d478acbaecf7fb095f8bbf04a42b36ddcf6f40..e282e8146c9df616464234682e15ffcbd3f9e185 100644 (file)
@@ -203,6 +203,14 @@ label##3:                                          \
        FTR_ENTRY_OFFSET 955b-956b;                     \
        .popsection;
 
+#define ENTRY_FLUSH_FIXUP_SECTION                      \
+957:                                                   \
+       .pushsection __entry_flush_fixup,"a";           \
+       .align 2;                                       \
+958:                                                   \
+       FTR_ENTRY_OFFSET 957b-958b;                     \
+       .popsection;
+
 #define RFI_FLUSH_FIXUP_SECTION                                \
 951:                                                   \
        .pushsection __rfi_flush_fixup,"a";             \
@@ -235,8 +243,10 @@ label##3:                                          \
 #include <linux/types.h>
 
 extern long stf_barrier_fallback;
+extern long entry_flush_fallback;
 extern long __start___stf_entry_barrier_fixup, __stop___stf_entry_barrier_fixup;
 extern long __start___stf_exit_barrier_fixup, __stop___stf_exit_barrier_fixup;
+extern long __start___entry_flush_fixup, __stop___entry_flush_fixup;
 extern long __start___rfi_flush_fixup, __stop___rfi_flush_fixup;
 extern long __start___barrier_nospec_fixup, __stop___barrier_nospec_fixup;
 extern long __start__btb_flush_fixup, __stop__btb_flush_fixup;
index ccf44c135389a111023a244498f9fab60a8583f5..082b56bf678dd33b2fa7ee49e8491a103f99bf94 100644 (file)
@@ -84,12 +84,16 @@ static inline bool security_ftr_enabled(unsigned long feature)
 // Software required to flush link stack on context switch
 #define SEC_FTR_FLUSH_LINK_STACK       0x0000000000001000ull
 
+// The L1-D cache should be flushed when entering the kernel
+#define SEC_FTR_L1D_FLUSH_ENTRY                0x0000000000004000ull
+
 
 // Features enabled by default
 #define SEC_FTR_DEFAULT \
        (SEC_FTR_L1D_FLUSH_HV | \
         SEC_FTR_L1D_FLUSH_PR | \
         SEC_FTR_BNDS_CHK_SPEC_BAR | \
+        SEC_FTR_L1D_FLUSH_ENTRY | \
         SEC_FTR_FAVOUR_SECURITY)
 
 #endif /* _ASM_POWERPC_SECURITY_FEATURES_H */
index 5ceab440ecb9ba6cb7e6dff55d45cc0f6a6fb739..02dfa8b1147e64b8cdd95c24247ba69208f898e6 100644 (file)
@@ -51,12 +51,15 @@ enum l1d_flush_type {
 };
 
 void setup_rfi_flush(enum l1d_flush_type, bool enable);
+void setup_entry_flush(bool enable);
+void setup_uaccess_flush(bool enable);
 void do_rfi_flush_fixups(enum l1d_flush_type types);
 #ifdef CONFIG_PPC_BARRIER_NOSPEC
 void setup_barrier_nospec(void);
 #else
 static inline void setup_barrier_nospec(void) { };
 #endif
+void do_entry_flush_fixups(enum l1d_flush_type types);
 void do_barrier_nospec_fixups(bool enable);
 extern bool barrier_nospec_enabled;
 
index f619e0deb23b7af61dd9105ece164744459b524e..2712331190af4bf16fe566a6ff3c84c3b23472f9 100644 (file)
@@ -484,7 +484,7 @@ EXC_COMMON_BEGIN(unrecover_mce)
        b       1b
 
 
-EXC_REAL(data_access, 0x300, 0x80)
+EXC_REAL_OOL(data_access, 0x300, 0x80)
 EXC_VIRT(data_access, 0x4300, 0x80, 0x300)
 TRAMP_KVM_SKIP(PACA_EXGEN, 0x300)
 
@@ -540,7 +540,7 @@ EXC_VIRT_END(data_access_slb, 0x4380, 0x80)
 TRAMP_KVM_SKIP(PACA_EXSLB, 0x380)
 
 
-EXC_REAL(instruction_access, 0x400, 0x80)
+EXC_REAL_OOL(instruction_access, 0x400, 0x80)
 EXC_VIRT(instruction_access, 0x4400, 0x80, 0x400)
 TRAMP_KVM(PACA_EXGEN, 0x400)
 
@@ -836,13 +836,13 @@ END_FTR_SECTION_IFSET(CPU_FTR_TM)
 
 
 EXC_REAL_OOL_MASKABLE(decrementer, 0x900, 0x80)
-EXC_VIRT_MASKABLE(decrementer, 0x4900, 0x80, 0x900)
+EXC_VIRT_OOL_MASKABLE(decrementer, 0x4900, 0x80, 0x900)
 TRAMP_KVM(PACA_EXGEN, 0x900)
 EXC_COMMON_ASYNC(decrementer_common, 0x900, timer_interrupt)
 
 
-EXC_REAL_HV(hdecrementer, 0x980, 0x80)
-EXC_VIRT_HV(hdecrementer, 0x4980, 0x80, 0x980)
+EXC_REAL_OOL_HV(hdecrementer, 0x980, 0x80)
+EXC_VIRT_OOL_HV(hdecrementer, 0x4980, 0x80, 0x980)
 TRAMP_KVM_HV(PACA_EXGEN, 0x980)
 EXC_COMMON(hdecrementer_common, 0x980, hdec_interrupt)
 
@@ -1459,6 +1459,43 @@ TRAMP_REAL_BEGIN(stf_barrier_fallback)
        .endr
        blr
 
+TRAMP_REAL_BEGIN(entry_flush_fallback)
+       std     r9,PACA_EXRFI+EX_R9(r13)
+       std     r10,PACA_EXRFI+EX_R10(r13)
+       std     r11,PACA_EXRFI+EX_R11(r13)
+       mfctr   r9
+       ld      r10,PACA_RFI_FLUSH_FALLBACK_AREA(r13)
+       ld      r11,PACA_L1D_FLUSH_SIZE(r13)
+       srdi    r11,r11,(7 + 3) /* 128 byte lines, unrolled 8x */
+       mtctr   r11
+       DCBT_STOP_ALL_STREAM_IDS(r11) /* Stop prefetch streams */
+
+       /* order ld/st prior to dcbt stop all streams with flushing */
+       sync
+
+       /*
+        * The load addresses are at staggered offsets within cachelines,
+        * which suits some pipelines better (on others it should not
+        * hurt).
+        */
+1:
+       ld      r11,(0x80 + 8)*0(r10)
+       ld      r11,(0x80 + 8)*1(r10)
+       ld      r11,(0x80 + 8)*2(r10)
+       ld      r11,(0x80 + 8)*3(r10)
+       ld      r11,(0x80 + 8)*4(r10)
+       ld      r11,(0x80 + 8)*5(r10)
+       ld      r11,(0x80 + 8)*6(r10)
+       ld      r11,(0x80 + 8)*7(r10)
+       addi    r10,r10,0x80*8
+       bdnz    1b
+
+       mtctr   r9
+       ld      r9,PACA_EXRFI+EX_R9(r13)
+       ld      r10,PACA_EXRFI+EX_R10(r13)
+       ld      r11,PACA_EXRFI+EX_R11(r13)
+       blr
+
 TRAMP_REAL_BEGIN(rfi_flush_fallback)
        SET_SCRATCH0(r13);
        GET_PACA(r13);
index a1e336901cc83acd516065ca40dd4ba7c22d28f6..62a3c2e5d918c50f94bebd75f997f369f5e5b943 100644 (file)
@@ -792,7 +792,9 @@ early_initcall(disable_hardlockup_detector);
 static enum l1d_flush_type enabled_flush_types;
 static void *l1d_flush_fallback_area;
 static bool no_rfi_flush;
+static bool no_entry_flush;
 bool rfi_flush;
+bool entry_flush;
 
 static int __init handle_no_rfi_flush(char *p)
 {
@@ -802,6 +804,14 @@ static int __init handle_no_rfi_flush(char *p)
 }
 early_param("no_rfi_flush", handle_no_rfi_flush);
 
+static int __init handle_no_entry_flush(char *p)
+{
+       pr_info("entry-flush: disabled on command line.");
+       no_entry_flush = true;
+       return 0;
+}
+early_param("no_entry_flush", handle_no_entry_flush);
+
 /*
  * The RFI flush is not KPTI, but because users will see doco that says to use
  * nopti we hijack that option here to also disable the RFI flush.
@@ -833,6 +843,18 @@ void rfi_flush_enable(bool enable)
        rfi_flush = enable;
 }
 
+void entry_flush_enable(bool enable)
+{
+       if (enable) {
+               do_entry_flush_fixups(enabled_flush_types);
+               on_each_cpu(do_nothing, NULL, 1);
+       } else {
+               do_entry_flush_fixups(L1D_FLUSH_NONE);
+       }
+
+       entry_flush = enable;
+}
+
 static void __ref init_fallback_flush(void)
 {
        u64 l1d_size, limit;
@@ -874,10 +896,19 @@ void setup_rfi_flush(enum l1d_flush_type types, bool enable)
 
        enabled_flush_types = types;
 
-       if (!no_rfi_flush && !cpu_mitigations_off())
+       if (!cpu_mitigations_off() && !no_rfi_flush)
                rfi_flush_enable(enable);
 }
 
+void setup_entry_flush(bool enable)
+{
+       if (cpu_mitigations_off())
+               return;
+
+       if (!no_entry_flush)
+               entry_flush_enable(enable);
+}
+
 #ifdef CONFIG_DEBUG_FS
 static int rfi_flush_set(void *data, u64 val)
 {
@@ -905,9 +936,36 @@ static int rfi_flush_get(void *data, u64 *val)
 
 DEFINE_SIMPLE_ATTRIBUTE(fops_rfi_flush, rfi_flush_get, rfi_flush_set, "%llu\n");
 
+static int entry_flush_set(void *data, u64 val)
+{
+       bool enable;
+
+       if (val == 1)
+               enable = true;
+       else if (val == 0)
+               enable = false;
+       else
+               return -EINVAL;
+
+       /* Only do anything if we're changing state */
+       if (enable != entry_flush)
+               entry_flush_enable(enable);
+
+       return 0;
+}
+
+static int entry_flush_get(void *data, u64 *val)
+{
+       *val = entry_flush ? 1 : 0;
+       return 0;
+}
+
+DEFINE_SIMPLE_ATTRIBUTE(fops_entry_flush, entry_flush_get, entry_flush_set, "%llu\n");
+
 static __init int rfi_flush_debugfs_init(void)
 {
        debugfs_create_file("rfi_flush", 0600, powerpc_debugfs_root, NULL, &fops_rfi_flush);
+       debugfs_create_file("entry_flush", 0600, powerpc_debugfs_root, NULL, &fops_entry_flush);
        return 0;
 }
 device_initcall(rfi_flush_debugfs_init);
index e4da937d6cf919918277166f80e44c24b9575a34..6c3cd47bdb8196d82f5e3e004472de3901b57932 100644 (file)
@@ -140,6 +140,13 @@ SECTIONS
                __stop___stf_entry_barrier_fixup = .;
        }
 
+       . = ALIGN(8);
+       __entry_flush_fixup : AT(ADDR(__entry_flush_fixup) - LOAD_OFFSET) {
+               __start___entry_flush_fixup = .;
+               *(__entry_flush_fixup)
+               __stop___entry_flush_fixup = .;
+       }
+
        . = ALIGN(8);
        __stf_exit_barrier_fixup : AT(ADDR(__stf_exit_barrier_fixup) - LOAD_OFFSET) {
                __start___stf_exit_barrier_fixup = .;
index de7861e09b41ced8a4756fe5a58c9ddfffc87799..d21c05bc94a0699200910e112355c977f23cf943 100644 (file)
@@ -232,6 +232,60 @@ void do_stf_barrier_fixups(enum stf_barrier_type types)
        do_stf_exit_barrier_fixups(types);
 }
 
+void do_entry_flush_fixups(enum l1d_flush_type types)
+{
+       unsigned int instrs[3], *dest;
+       long *start, *end;
+       int i;
+
+       start = PTRRELOC(&__start___entry_flush_fixup);
+       end = PTRRELOC(&__stop___entry_flush_fixup);
+
+       instrs[0] = 0x60000000; /* nop */
+       instrs[1] = 0x60000000; /* nop */
+       instrs[2] = 0x60000000; /* nop */
+
+       i = 0;
+       if (types == L1D_FLUSH_FALLBACK) {
+               instrs[i++] = 0x7d4802a6; /* mflr r10           */
+               instrs[i++] = 0x60000000; /* branch patched below */
+               instrs[i++] = 0x7d4803a6; /* mtlr r10           */
+       }
+
+       if (types & L1D_FLUSH_ORI) {
+               instrs[i++] = 0x63ff0000; /* ori 31,31,0 speculation barrier */
+               instrs[i++] = 0x63de0000; /* ori 30,30,0 L1d flush*/
+       }
+
+       if (types & L1D_FLUSH_MTTRIG)
+               instrs[i++] = 0x7c12dba6; /* mtspr TRIG2,r0 (SPR #882) */
+
+       for (i = 0; start < end; start++, i++) {
+               dest = (void *)start + *start;
+
+               pr_devel("patching dest %lx\n", (unsigned long)dest);
+
+               patch_instruction(dest, instrs[0]);
+
+               if (types == L1D_FLUSH_FALLBACK)
+                       patch_branch((dest + 1), (unsigned long)&entry_flush_fallback,
+                                    BRANCH_SET_LINK);
+               else
+                       patch_instruction((dest + 1), instrs[1]);
+
+               patch_instruction((dest + 2), instrs[2]);
+       }
+
+       printk(KERN_DEBUG "entry-flush: patched %d locations (%s flush)\n", i,
+               (types == L1D_FLUSH_NONE)       ? "no" :
+               (types == L1D_FLUSH_FALLBACK)   ? "fallback displacement" :
+               (types &  L1D_FLUSH_ORI)        ? (types & L1D_FLUSH_MTTRIG)
+                                                       ? "ori+mttrig type"
+                                                       : "ori type" :
+               (types &  L1D_FLUSH_MTTRIG)     ? "mttrig type"
+                                               : "unknown");
+}
+
 void do_rfi_flush_fixups(enum l1d_flush_type types)
 {
        unsigned int instrs[3], *dest;
index 888aa9584e94f0a5789294061611f3a1874e8e44..28cb14a26503dee978c849c684fc70151c7f3717 100644 (file)
@@ -124,12 +124,23 @@ static void pnv_setup_rfi_flush(void)
                        type = L1D_FLUSH_ORI;
        }
 
+       /*
+        * If we are non-Power9 bare metal, we don't need to flush on kernel
+        * entry: it fixes a P9 specific vulnerability.
+        */
+       if (!pvr_version_is(PVR_POWER9))
+               security_ftr_clear(SEC_FTR_L1D_FLUSH_ENTRY);
+
        enable = security_ftr_enabled(SEC_FTR_FAVOUR_SECURITY) && \
                 (security_ftr_enabled(SEC_FTR_L1D_FLUSH_PR)   || \
                  security_ftr_enabled(SEC_FTR_L1D_FLUSH_HV));
 
        setup_rfi_flush(type, enable);
        setup_count_cache_flush();
+
+       enable = security_ftr_enabled(SEC_FTR_FAVOUR_SECURITY) &&
+                security_ftr_enabled(SEC_FTR_L1D_FLUSH_ENTRY);
+       setup_entry_flush(enable);
 }
 
 static void __init pnv_setup_arch(void)
index 7a9945b350536420d3bf9a8729766f23a0f57e2d..5e58b0cf92df8e22ca50508cbfb48dac3f4d60f3 100644 (file)
@@ -544,6 +544,10 @@ void pseries_setup_rfi_flush(void)
 
        setup_rfi_flush(types, enable);
        setup_count_cache_flush();
+
+       enable = security_ftr_enabled(SEC_FTR_FAVOUR_SECURITY) &&
+                security_ftr_enabled(SEC_FTR_L1D_FLUSH_ENTRY);
+       setup_entry_flush(enable);
 }
 
 static void __init pSeries_setup_arch(void)