tsx_async_abort=off [X86]
                                               kvm.nx_huge_pages=off [X86]
                                               no_entry_flush [PPC]
+                                              no_uaccess_flush [PPC]
 
                                Exceptions:
                                               This does not have any effect on
        nospec_store_bypass_disable
                        [HW] Disable all mitigations for the Speculative Store Bypass vulnerability
 
+       no_uaccess_flush
+                       [PPC] Don't flush the L1-D cache after accessing user data.
+
        noxsave         [BUGS=X86] Disables x86 extended register state save
                        and restore using xsave. The kernel will fallback to
                        enabling legacy floating-point and sse state.
 
--- /dev/null
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_POWERPC_BOOK3S_64_KUP_RADIX_H
+#define _ASM_POWERPC_BOOK3S_64_KUP_RADIX_H
+
+DECLARE_STATIC_KEY_FALSE(uaccess_flush_key);
+
+/* Prototype for function defined in exceptions-64s.S */
+void do_uaccess_flush(void);
+
+static __always_inline void allow_user_access(void __user *to, const void __user *from,
+                                             unsigned long size)
+{
+}
+
+static inline void prevent_user_access(void __user *to, const void __user *from,
+                                      unsigned long size)
+{
+       if (static_branch_unlikely(&uaccess_flush_key))
+               do_uaccess_flush();
+}
+
+#endif /* _ASM_POWERPC_BOOK3S_64_KUP_RADIX_H */
 
        FTR_ENTRY_OFFSET 955b-956b;                     \
        .popsection;
 
+#define UACCESS_FLUSH_FIXUP_SECTION                    \
+959:                                                   \
+       .pushsection __uaccess_flush_fixup,"a";         \
+       .align 2;                                       \
+960:                                                   \
+       FTR_ENTRY_OFFSET 959b-960b;                     \
+       .popsection;
+
 #define ENTRY_FLUSH_FIXUP_SECTION                      \
 957:                                                   \
        .pushsection __entry_flush_fixup,"a";           \
 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___uaccess_flush_fixup, __stop___uaccess_flush_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;
 
 
 #include <asm/pgtable.h>
 
+#ifdef CONFIG_PPC_BOOK3S_64
+#include <asm/book3s/64/kup-radix.h>
+#else
 static inline void allow_user_access(void __user *to, const void __user *from,
                                     unsigned long size) { }
 static inline void prevent_user_access(void __user *to, const void __user *from,
                                       unsigned long size) { }
+#endif /* CONFIG_PPC_BOOK3S_64 */
 
 static inline void allow_read_from_user(const void __user *from, unsigned long size)
 {
 
 // The L1-D cache should be flushed when entering the kernel
 #define SEC_FTR_L1D_FLUSH_ENTRY                0x0000000000004000ull
 
+// The L1-D cache should be flushed after user accesses from the kernel
+#define SEC_FTR_L1D_FLUSH_UACCESS      0x0000000000008000ull
 
 // Features enabled by default
 #define SEC_FTR_DEFAULT \
         SEC_FTR_L1D_FLUSH_PR | \
         SEC_FTR_BNDS_CHK_SPEC_BAR | \
         SEC_FTR_L1D_FLUSH_ENTRY | \
+        SEC_FTR_L1D_FLUSH_UACCESS | \
         SEC_FTR_FAVOUR_SECURITY)
 
 #endif /* _ASM_POWERPC_SECURITY_FEATURES_H */
 
 #else
 static inline void setup_barrier_nospec(void) { };
 #endif
+void do_uaccess_flush_fixups(enum l1d_flush_type types);
 void do_entry_flush_fixups(enum l1d_flush_type types);
 void do_barrier_nospec_fixups(bool enable);
 extern bool barrier_nospec_enabled;
 
        .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
+/* Clobbers r10, r11, ctr */
+.macro L1D_DISPLACEMENT_FLUSH
        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 */
        ld      r11,(0x80 + 8)*7(r10)
        addi    r10,r10,0x80*8
        bdnz    1b
+.endm
 
+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
+       L1D_DISPLACEMENT_FLUSH
        mtctr   r9
        ld      r9,PACA_EXRFI+EX_R9(r13)
        ld      r10,PACA_EXRFI+EX_R10(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 adresses 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
-
+       L1D_DISPLACEMENT_FLUSH
        mtctr   r9
        ld      r9,PACA_EXRFI+EX_R9(r13)
        ld      r10,PACA_EXRFI+EX_R10(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 adresses 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
-
+       L1D_DISPLACEMENT_FLUSH
        mtctr   r9
        ld      r9,PACA_EXRFI+EX_R9(r13)
        ld      r10,PACA_EXRFI+EX_R10(r13)
        GET_SCRATCH0(r13);
        hrfid
 
+USE_TEXT_SECTION()
+
+_GLOBAL(do_uaccess_flush)
+       UACCESS_FLUSH_FIXUP_SECTION
+       nop
+       nop
+       nop
+       blr
+       L1D_DISPLACEMENT_FLUSH
+       blr
+_ASM_NOKPROBE_SYMBOL(do_uaccess_flush)
+EXPORT_SYMBOL(do_uaccess_flush)
+
 /*
  * Real mode exceptions actually use this too, but alternate
  * instruction code patches (which end up in the common .text area)
 
 static void *l1d_flush_fallback_area;
 static bool no_rfi_flush;
 static bool no_entry_flush;
+static bool no_uaccess_flush;
 bool rfi_flush;
 bool entry_flush;
+bool uaccess_flush;
+DEFINE_STATIC_KEY_FALSE(uaccess_flush_key);
+EXPORT_SYMBOL(uaccess_flush_key);
 
 static int __init handle_no_rfi_flush(char *p)
 {
 }
 early_param("no_entry_flush", handle_no_entry_flush);
 
+static int __init handle_no_uaccess_flush(char *p)
+{
+       pr_info("uaccess-flush: disabled on command line.");
+       no_uaccess_flush = true;
+       return 0;
+}
+early_param("no_uaccess_flush", handle_no_uaccess_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.
        entry_flush = enable;
 }
 
+void uaccess_flush_enable(bool enable)
+{
+       if (enable) {
+               do_uaccess_flush_fixups(enabled_flush_types);
+               static_branch_enable(&uaccess_flush_key);
+               on_each_cpu(do_nothing, NULL, 1);
+       } else {
+               static_branch_disable(&uaccess_flush_key);
+               do_uaccess_flush_fixups(L1D_FLUSH_NONE);
+       }
+
+       uaccess_flush = enable;
+}
+
 static void __ref init_fallback_flush(void)
 {
        u64 l1d_size, limit;
                entry_flush_enable(enable);
 }
 
+void setup_uaccess_flush(bool enable)
+{
+       if (cpu_mitigations_off())
+               return;
+
+       if (!no_uaccess_flush)
+               uaccess_flush_enable(enable);
+}
+
 #ifdef CONFIG_DEBUG_FS
 static int rfi_flush_set(void *data, u64 val)
 {
 
 DEFINE_SIMPLE_ATTRIBUTE(fops_entry_flush, entry_flush_get, entry_flush_set, "%llu\n");
 
+static int uaccess_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 != uaccess_flush)
+               uaccess_flush_enable(enable);
+
+       return 0;
+}
+
+static int uaccess_flush_get(void *data, u64 *val)
+{
+       *val = uaccess_flush ? 1 : 0;
+       return 0;
+}
+
+DEFINE_SIMPLE_ATTRIBUTE(fops_uaccess_flush, uaccess_flush_get, uaccess_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);
+       debugfs_create_file("uaccess_flush", 0600, powerpc_debugfs_root, NULL, &fops_uaccess_flush);
        return 0;
 }
 device_initcall(rfi_flush_debugfs_init);
 
                __stop___stf_entry_barrier_fixup = .;
        }
 
+       . = ALIGN(8);
+       __uaccess_flush_fixup : AT(ADDR(__uaccess_flush_fixup) - LOAD_OFFSET) {
+               __start___uaccess_flush_fixup = .;
+               *(__uaccess_flush_fixup)
+               __stop___uaccess_flush_fixup = .;
+       }
+
        . = ALIGN(8);
        __entry_flush_fixup : AT(ADDR(__entry_flush_fixup) - LOAD_OFFSET) {
                __start___entry_flush_fixup = .;
 
        do_stf_exit_barrier_fixups(types);
 }
 
+void do_uaccess_flush_fixups(enum l1d_flush_type types)
+{
+       unsigned int instrs[4], *dest;
+       long *start, *end;
+       int i;
+
+       start = PTRRELOC(&__start___uaccess_flush_fixup);
+       end = PTRRELOC(&__stop___uaccess_flush_fixup);
+
+       instrs[0] = 0x60000000; /* nop */
+       instrs[1] = 0x60000000; /* nop */
+       instrs[2] = 0x60000000; /* nop */
+       instrs[3] = 0x4e800020; /* blr */
+
+       i = 0;
+       if (types == L1D_FLUSH_FALLBACK) {
+               instrs[3] = 0x60000000; /* nop */
+               /* fallthrough to fallback flush */
+       }
+
+       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]);
+
+               patch_instruction((dest + 1), instrs[1]);
+               patch_instruction((dest + 2), instrs[2]);
+               patch_instruction((dest + 3), instrs[3]);
+       }
+
+       printk(KERN_DEBUG "uaccess-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_entry_flush_fixups(enum l1d_flush_type types)
 {
        unsigned int instrs[3], *dest;
 
 
        /*
         * If we are non-Power9 bare metal, we don't need to flush on kernel
-        * entry: it fixes a P9 specific vulnerability.
+        * entry or after user access: they fix a P9 specific vulnerability.
         */
-       if (!pvr_version_is(PVR_POWER9))
+       if (!pvr_version_is(PVR_POWER9)) {
                security_ftr_clear(SEC_FTR_L1D_FLUSH_ENTRY);
+               security_ftr_clear(SEC_FTR_L1D_FLUSH_UACCESS);
+       }
 
        enable = security_ftr_enabled(SEC_FTR_FAVOUR_SECURITY) && \
                 (security_ftr_enabled(SEC_FTR_L1D_FLUSH_PR)   || \
        enable = security_ftr_enabled(SEC_FTR_FAVOUR_SECURITY) &&
                 security_ftr_enabled(SEC_FTR_L1D_FLUSH_ENTRY);
        setup_entry_flush(enable);
+
+       enable = security_ftr_enabled(SEC_FTR_FAVOUR_SECURITY) &&
+                security_ftr_enabled(SEC_FTR_L1D_FLUSH_UACCESS);
+       setup_uaccess_flush(enable);
 }
 
 static void __init pnv_setup_arch(void)
 
        enable = security_ftr_enabled(SEC_FTR_FAVOUR_SECURITY) &&
                 security_ftr_enabled(SEC_FTR_L1D_FLUSH_ENTRY);
        setup_entry_flush(enable);
+
+       enable = security_ftr_enabled(SEC_FTR_FAVOUR_SECURITY) &&
+                security_ftr_enabled(SEC_FTR_L1D_FLUSH_UACCESS);
+       setup_uaccess_flush(enable);
 }
 
 static void __init pSeries_setup_arch(void)