2 files changed, 114 insertions, 76 deletions

diff --git a/arch/arm/vfp/vfphw.S b/arch/arm/vfp/vfphw.S
index 9897dcfc16d6..2d30c7f6edd3 100644
--- a/arch/arm/vfp/vfphw.S
+++ b/arch/arm/vfp/vfphw.S
@@ -77,27 +77,27 @@ ENTRY(vfp_support_entry)
         bne     look_for_VFP_exceptions @ VFP is already enabled
 
         DBGSTR1 "enable %x", r10
-        ldr     r3, last_VFP_context_address
+        ldr     r3, vfp_current_hw_state_address
         orr     r1, r1, #FPEXC_EN       @ user FPEXC has the enable bit set
-        ldr     r4, [r3, r11, lsl #2]   @ last_VFP_context pointer
+        ldr     r4, [r3, r11, lsl #2]   @ vfp_current_hw_state pointer
         bic     r5, r1, #FPEXC_EX       @ make sure exceptions are disabled
-        cmp     r4, r10
-        beq     check_for_exception     @ we are returning to the same
-                                        @ process, so the registers are
-                                        @ still there.  In this case, we do
-                                        @ not want to drop a pending exception.
+        cmp     r4, r10                 @ this thread owns the hw context?
+#ifndef CONFIG_SMP
+        @ For UP, checking that this thread owns the hw context is
+        @ sufficient to determine that the hardware state is valid.
+        beq     vfp_hw_state_valid
+
+        @ On UP, we lazily save the VFP context.  As a different
+        @ thread wants ownership of the VFP hardware, save the old
+        @ state if there was a previous (valid) owner.
 
         VFPFMXR FPEXC, r5               @ enable VFP, disable any pending
                                         @ exceptions, so we can get at the
                                         @ rest of it
 
-#ifndef CONFIG_SMP
-        @ Save out the current registers to the old thread state
-        @ No need for SMP since this is not done lazily
-
         DBGSTR1 "save old state %p", r4
-        cmp     r4, #0
-        beq     no_old_VFP_process
+        cmp     r4, #0                  @ if the vfp_current_hw_state is NULL
+        beq     vfp_reload_hw           @ then the hw state needs reloading
         VFPFSTMIA r4, r5                @ save the working registers
         VFPFMRX r5, FPSCR               @ current status
 #ifndef CONFIG_CPU_FEROCEON
@@ -110,13 +110,35 @@ ENTRY(vfp_support_entry)
 1:
 #endif
         stmia   r4, {r1, r5, r6, r8}    @ save FPEXC, FPSCR, FPINST, FPINST2
-                                        @ and point r4 at the word at the
-                                        @ start of the register dump
+vfp_reload_hw:
+
+#else
+        @ For SMP, if this thread does not own the hw context, then we
+        @ need to reload it.  No need to save the old state as on SMP,
+        @ we always save the state when we switch away from a thread.
+        bne     vfp_reload_hw
+
+        @ This thread has ownership of the current hardware context.
+        @ However, it may have been migrated to another CPU, in which
+        @ case the saved state is newer than the hardware context.
+        @ Check this by looking at the CPU number which the state was
+        @ last loaded onto.
+        ldr     ip, [r10, #VFP_CPU]
+        teq     ip, r11
+        beq     vfp_hw_state_valid
+
+vfp_reload_hw:
+        @ We're loading this threads state into the VFP hardware. Update
+        @ the CPU number which contains the most up to date VFP context.
+        str     r11, [r10, #VFP_CPU]
+
+        VFPFMXR FPEXC, r5               @ enable VFP, disable any pending
+                                        @ exceptions, so we can get at the
+                                        @ rest of it
 #endif
 
-no_old_VFP_process:
         DBGSTR1 "load state %p", r10
-        str     r10, [r3, r11, lsl #2]  @ update the last_VFP_context pointer
+        str     r10, [r3, r11, lsl #2]  @ update the vfp_current_hw_state pointer
                                         @ Load the saved state back into the VFP
         VFPFLDMIA r10, r5               @ reload the working registers while
                                         @ FPEXC is in a safe state
@@ -132,7 +154,8 @@ no_old_VFP_process:
 #endif
         VFPFMXR FPSCR, r5               @ restore status
 
-check_for_exception:
+@ The context stored in the VFP hardware is up to date with this thread
+vfp_hw_state_valid:
         tst     r1, #FPEXC_EX
         bne     process_exception       @ might as well handle the pending
                                         @ exception before retrying branch
@@ -207,8 +230,8 @@ ENTRY(vfp_save_state)
 ENDPROC(vfp_save_state)
 
         .align
-last_VFP_context_address:
-        .word   last_VFP_context
+vfp_current_hw_state_address:
+        .word   vfp_current_hw_state
 
         .macro  tbl_branch, base, tmp, shift
 #ifdef CONFIG_THUMB2_KERNEL
diff --git a/arch/arm/vfp/vfpmodule.c b/arch/arm/vfp/vfpmodule.c
index f25e7ec89416..0a96f71f0abd 100644
--- a/arch/arm/vfp/vfpmodule.c
+++ b/arch/arm/vfp/vfpmodule.c
@@ -33,7 +33,6 @@ void vfp_support_entry(void);
 void vfp_null_entry(void);
 
 void (*vfp_vector)(void) = vfp_null_entry;
-union vfp_state *last_VFP_context[NR_CPUS];
 
 /*
  * Dual-use variable.
@@ -43,6 +42,46 @@ union vfp_state *last_VFP_context[NR_CPUS];
 unsigned int VFP_arch;
 
 /*
+ * The pointer to the vfpstate structure of the thread which currently
+ * owns the context held in the VFP hardware, or NULL if the hardware
+ * context is invalid.
+ *
+ * For UP, this is sufficient to tell which thread owns the VFP context.
+ * However, for SMP, we also need to check the CPU number stored in the
+ * saved state too to catch migrations.
+ */
+union vfp_state *vfp_current_hw_state[NR_CPUS];
+
+/*
+ * Is 'thread's most up to date state stored in this CPUs hardware?
+ * Must be called from non-preemptible context.
+ */
+static bool vfp_state_in_hw(unsigned int cpu, struct thread_info *thread)
+{
+#ifdef CONFIG_SMP
+        if (thread->vfpstate.hard.cpu != cpu)
+                return false;
+#endif
+        return vfp_current_hw_state[cpu] == &thread->vfpstate;
+}
+
+/*
+ * Force a reload of the VFP context from the thread structure.  We do
+ * this by ensuring that access to the VFP hardware is disabled, and
+ * clear last_VFP_context.  Must be called from non-preemptible context.
+ */
+static void vfp_force_reload(unsigned int cpu, struct thread_info *thread)
+{
+        if (vfp_state_in_hw(cpu, thread)) {
+                fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
+                vfp_current_hw_state[cpu] = NULL;
+        }
+#ifdef CONFIG_SMP
+        thread->vfpstate.hard.cpu = NR_CPUS;
+#endif
+}
+
+/*
  * Per-thread VFP initialization.
  */
 static void vfp_thread_flush(struct thread_info *thread)
@@ -50,21 +89,27 @@ static void vfp_thread_flush(struct thread_info *thread)
         union vfp_state *vfp = &thread->vfpstate;
         unsigned int cpu;
 
-        memset(vfp, 0, sizeof(union vfp_state));
-
-        vfp->hard.fpexc = FPEXC_EN;
-        vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
-
         /*
          * Disable VFP to ensure we initialize it first.  We must ensure
-         * that the modification of last_VFP_context[] and hardware disable
-         * are done for the same CPU and without preemption.
+         * that the modification of vfp_current_hw_state[] and hardware
+         * disable are done for the same CPU and without preemption.
+         *
+         * Do this first to ensure that preemption won't overwrite our
+         * state saving should access to the VFP be enabled at this point.
          */
         cpu = get_cpu();
-        if (last_VFP_context[cpu] == vfp)
-                last_VFP_context[cpu] = NULL;
+        if (vfp_current_hw_state[cpu] == vfp)
+                vfp_current_hw_state[cpu] = NULL;
         fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
         put_cpu();
+
+        memset(vfp, 0, sizeof(union vfp_state));
+
+        vfp->hard.fpexc = FPEXC_EN;
+        vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
+#ifdef CONFIG_SMP
+        vfp->hard.cpu = NR_CPUS;
+#endif
 }
 
 static void vfp_thread_exit(struct thread_info *thread)
@@ -73,8 +118,8 @@ static void vfp_thread_exit(struct thread_info *thread)
         union vfp_state *vfp = &thread->vfpstate;
         unsigned int cpu = get_cpu();
 
-        if (last_VFP_context[cpu] == vfp)
-                last_VFP_context[cpu] = NULL;
+        if (vfp_current_hw_state[cpu] == vfp)
+                vfp_current_hw_state[cpu] = NULL;
         put_cpu();
 }
 
@@ -84,6 +129,9 @@ static void vfp_thread_copy(struct thread_info *thread)
 
         vfp_sync_hwstate(parent);
         thread->vfpstate = parent->vfpstate;
+#ifdef CONFIG_SMP
+        thread->vfpstate.hard.cpu = NR_CPUS;
+#endif
 }
 
 /*
@@ -129,17 +177,8 @@ static int vfp_notifier(struct notifier_block *self, unsigned long cmd, void *v)
                  * case the thread migrates to a different CPU. The
                  * restoring is done lazily.
                  */
-                if ((fpexc & FPEXC_EN) && last_VFP_context[cpu]) {
-                        vfp_save_state(last_VFP_context[cpu], fpexc);
-                        last_VFP_context[cpu]->hard.cpu = cpu;
-                }
-                /*
-                 * Thread migration, just force the reloading of the
-                 * state on the new CPU in case the VFP registers
-                 * contain stale data.
-                 */
-                if (thread->vfpstate.hard.cpu != cpu)
-                        last_VFP_context[cpu] = NULL;
+                if ((fpexc & FPEXC_EN) && vfp_current_hw_state[cpu])
+                        vfp_save_state(vfp_current_hw_state[cpu], fpexc);
 #endif
 
                 /*
@@ -415,7 +454,7 @@ static int vfp_pm_suspend(void)
         }
 
         /* clear any information we had about last context state */
-        memset(last_VFP_context, 0, sizeof(last_VFP_context));
+        memset(vfp_current_hw_state, 0, sizeof(vfp_current_hw_state));
 
         return 0;
 }
@@ -443,15 +482,15 @@ static void vfp_pm_init(void)
 static inline void vfp_pm_init(void) { }
 #endif /* CONFIG_PM */
 
+/*
+ * Ensure that the VFP state stored in 'thread->vfpstate' is up to date
+ * with the hardware state.
+ */
 void vfp_sync_hwstate(struct thread_info *thread)
 {
         unsigned int cpu = get_cpu();
 
-        /*
-         * If the thread we're interested in is the current owner of the
-         * hardware VFP state, then we need to save its state.
-         */
-        if (last_VFP_context[cpu] == &thread->vfpstate) {
+        if (vfp_state_in_hw(cpu, thread)) {
                 u32 fpexc = fmrx(FPEXC);
 
                 /*
@@ -465,36 +504,13 @@ void vfp_sync_hwstate(struct thread_info *thread)
         put_cpu();
 }
 
+/* Ensure that the thread reloads the hardware VFP state on the next use. */
 void vfp_flush_hwstate(struct thread_info *thread)
 {
         unsigned int cpu = get_cpu();
 
-        /*
-         * If the thread we're interested in is the current owner of the
-         * hardware VFP state, then we need to save its state.
-         */
-        if (last_VFP_context[cpu] == &thread->vfpstate) {
-                u32 fpexc = fmrx(FPEXC);
-
-                fmxr(FPEXC, fpexc & ~FPEXC_EN);
-
-                /*
-                 * Set the context to NULL to force a reload the next time
-                 * the thread uses the VFP.
-                 */
-                last_VFP_context[cpu] = NULL;
-        }
+        vfp_force_reload(cpu, thread);
 
-#ifdef CONFIG_SMP
-        /*
-         * For SMP we still have to take care of the case where the thread
-         * migrates to another CPU and then back to the original CPU on which
-         * the last VFP user is still the same thread. Mark the thread VFP
-         * state as belonging to a non-existent CPU so that the saved one will
-         * be reloaded in the above case.
-         */
-        thread->vfpstate.hard.cpu = NR_CPUS;
-#endif
         put_cpu();
 }
 
@@ -513,8 +529,7 @@ static int vfp_hotplug(struct notifier_block *b, unsigned long action,
         void *hcpu)
 {
         if (action == CPU_DYING || action == CPU_DYING_FROZEN) {
-                unsigned int cpu = (long)hcpu;
-                last_VFP_context[cpu] = NULL;
+                vfp_force_reload((long)hcpu, current_thread_info());
         } else if (action == CPU_STARTING || action == CPU_STARTING_FROZEN)
                 vfp_enable(NULL);
         return NOTIFY_OK;