[POWERPC] Fix spu SLB invalidations

The SPU code doesn't properly invalidate SPUs SLBs when necessary,
for example when changing a segment size from the hugetlbfs code. In
addition, it saves and restores the SLB content on context switches
which makes it harder to properly handle those invalidations.

This patch removes the saving & restoring for now, something more
efficient might be found later on. It also adds a spu_flush_all_slbs(mm)
that can be used by the core mm code to flush the SLBs of all SPEs that
are running a given mm at the time of the flush.

In order to do that, it adds a spinlock to the list of all SPEs and move
some bits & pieces from spufs to spu_base.c

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>

1 files changed, 69 insertions, 12 deletions

diff --git a/arch/powerpc/platforms/cell/spu_base.c b/arch/powerpc/platforms/cell/spu_base.c
index c43999a10deb..eba7a2641dce 100644
--- a/arch/powerpc/platforms/cell/spu_base.c
+++ b/arch/powerpc/platforms/cell/spu_base.c
@@ -38,8 +38,61 @@
 const struct spu_management_ops *spu_management_ops;
 const struct spu_priv1_ops *spu_priv1_ops;
 
+static struct list_head spu_list[MAX_NUMNODES];
+static LIST_HEAD(spu_full_list);
+static DEFINE_MUTEX(spu_mutex);
+static spinlock_t spu_list_lock = SPIN_LOCK_UNLOCKED;
+
 EXPORT_SYMBOL_GPL(spu_priv1_ops);
 
+void spu_invalidate_slbs(struct spu *spu)
+{
+        struct spu_priv2 __iomem *priv2 = spu->priv2;
+
+        if (spu_mfc_sr1_get(spu) & MFC_STATE1_RELOCATE_MASK)
+                out_be64(&priv2->slb_invalidate_all_W, 0UL);
+}
+EXPORT_SYMBOL_GPL(spu_invalidate_slbs);
+
+/* This is called by the MM core when a segment size is changed, to
+ * request a flush of all the SPEs using a given mm
+ */
+void spu_flush_all_slbs(struct mm_struct *mm)
+{
+        struct spu *spu;
+        unsigned long flags;
+
+        spin_lock_irqsave(&spu_list_lock, flags);
+        list_for_each_entry(spu, &spu_full_list, full_list) {
+                if (spu->mm == mm)
+                        spu_invalidate_slbs(spu);
+        }
+        spin_unlock_irqrestore(&spu_list_lock, flags);
+}
+
+/* The hack below stinks... try to do something better one of
+ * these days... Does it even work properly with NR_CPUS == 1 ?
+ */
+static inline void mm_needs_global_tlbie(struct mm_struct *mm)
+{
+        int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
+
+        /* Global TLBIE broadcast required with SPEs. */
+        __cpus_setall(&mm->cpu_vm_mask, nr);
+}
+
+void spu_associate_mm(struct spu *spu, struct mm_struct *mm)
+{
+        unsigned long flags;
+
+        spin_lock_irqsave(&spu_list_lock, flags);
+        spu->mm = mm;
+        spin_unlock_irqrestore(&spu_list_lock, flags);
+        if (mm)
+                mm_needs_global_tlbie(mm);
+}
+EXPORT_SYMBOL_GPL(spu_associate_mm);
+
 static int __spu_trap_invalid_dma(struct spu *spu)
 {
         pr_debug("%s\n", __FUNCTION__);
@@ -74,6 +127,7 @@ static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
         struct spu_priv2 __iomem *priv2 = spu->priv2;
         struct mm_struct *mm = spu->mm;
         u64 esid, vsid, llp;
+        int psize;
 
         pr_debug("%s\n", __FUNCTION__);
 
@@ -90,22 +144,25 @@ static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
         case USER_REGION_ID:
 #ifdef CONFIG_HUGETLB_PAGE
                 if (in_hugepage_area(mm->context, ea))
-                        llp = mmu_psize_defs[mmu_huge_psize].sllp;
+                        psize = mmu_huge_psize;
                 else
 #endif
-                        llp = mmu_psize_defs[mmu_virtual_psize].sllp;
+                        psize = mm->context.user_psize;
                 vsid = (get_vsid(mm->context.id, ea) << SLB_VSID_SHIFT) |
-                                SLB_VSID_USER | llp;
+                                SLB_VSID_USER;
                 break;
         case VMALLOC_REGION_ID:
-                llp = mmu_psize_defs[mmu_virtual_psize].sllp;
+                if (ea < VMALLOC_END)
+                        psize = mmu_vmalloc_psize;
+                else
+                        psize = mmu_io_psize;
                 vsid = (get_kernel_vsid(ea) << SLB_VSID_SHIFT) |
-                        SLB_VSID_KERNEL | llp;
+                        SLB_VSID_KERNEL;
                 break;
         case KERNEL_REGION_ID:
-                llp = mmu_psize_defs[mmu_linear_psize].sllp;
+                psize = mmu_linear_psize;
                 vsid = (get_kernel_vsid(ea) << SLB_VSID_SHIFT) |
-                        SLB_VSID_KERNEL | llp;
+                        SLB_VSID_KERNEL;
                 break;
         default:
                 /* Future: support kernel segments so that drivers
@@ -114,9 +171,10 @@ static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
                 pr_debug("invalid region access at %016lx\n", ea);
                 return 1;
         }
+        llp = mmu_psize_defs[psize].sllp;
 
         out_be64(&priv2->slb_index_W, spu->slb_replace);
-        out_be64(&priv2->slb_vsid_RW, vsid);
+        out_be64(&priv2->slb_vsid_RW, vsid | llp);
         out_be64(&priv2->slb_esid_RW, esid);
 
         spu->slb_replace++;
@@ -330,10 +388,6 @@ static void spu_free_irqs(struct spu *spu)
                 free_irq(spu->irqs[2], spu);
 }
 
-static struct list_head spu_list[MAX_NUMNODES];
-static LIST_HEAD(spu_full_list);
-static DEFINE_MUTEX(spu_mutex);
-
 static void spu_init_channels(struct spu *spu)
 {
         static const struct {
@@ -593,6 +647,7 @@ static int __init create_spu(void *data)
         struct spu *spu;
         int ret;
         static int number;
+        unsigned long flags;
 
         ret = -ENOMEM;
         spu = kzalloc(sizeof (*spu), GFP_KERNEL);
@@ -620,8 +675,10 @@ static int __init create_spu(void *data)
                 goto out_free_irqs;
 
         mutex_lock(&spu_mutex);
+        spin_lock_irqsave(&spu_list_lock, flags);
         list_add(&spu->list, &spu_list[spu->node]);
         list_add(&spu->full_list, &spu_full_list);
+        spin_unlock_irqrestore(&spu_list_lock, flags);
         mutex_unlock(&spu_mutex);
 
         goto out;