[POWERPC] spufs: allow isolated mode apps by starting the SPE loader
authorarnd@arndb.de <arnd@arndb.de>
Tue, 24 Oct 2006 16:31:18 +0000 (18:31 +0200)
committerPaul Mackerras <paulus@samba.org>
Wed, 25 Oct 2006 04:20:21 +0000 (14:20 +1000)
This patch adds general support for isolated mode SPE apps.

Isolated apps are started indirectly, by a dedicated loader "kernel".
This patch starts the loader when spe_create is invoked with the
ISOLATE flag. We do this at spe_create time to allow libspe to pass the
isolated app in before calling spe_run.

The loader is read from the device tree, at the location
"/spu-isolation/loader". If the loader is not present, an attempt to
start an isolated SPE binary will fail with -ENODEV.

Update: loader needs to be correctly aligned - copy to a kmalloced buf.
Update: remove workaround for systemsim/spurom 'L-bit' bug, which has
        been fixed.
Update: don't write to runcntl on spu_run_init: SPU is already running.
Update: do spu_setup_isolated earlier

Tested on systemsim.

Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
arch/powerpc/platforms/cell/spu_base.c
arch/powerpc/platforms/cell/spufs/inode.c
arch/powerpc/platforms/cell/spufs/run.c

index f6c94087db4025384c280d498dace78860f5bcf1..d78b0af038e6ab0da8b25c5baedac271b20f697e 100644 (file)
@@ -89,7 +89,30 @@ static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
                printk("%s: invalid access during switch!\n", __func__);
                return 1;
        }
-       if (!mm || (REGION_ID(ea) != USER_REGION_ID)) {
+       esid = (ea & ESID_MASK) | SLB_ESID_V;
+
+       switch(REGION_ID(ea)) {
+       case USER_REGION_ID:
+#ifdef CONFIG_HUGETLB_PAGE
+               if (in_hugepage_area(mm->context, ea))
+                       llp = mmu_psize_defs[mmu_huge_psize].sllp;
+               else
+#endif
+                       llp = mmu_psize_defs[mmu_virtual_psize].sllp;
+               vsid = (get_vsid(mm->context.id, ea) << SLB_VSID_SHIFT) |
+                               SLB_VSID_USER | llp;
+               break;
+       case VMALLOC_REGION_ID:
+               llp = mmu_psize_defs[mmu_virtual_psize].sllp;
+               vsid = (get_kernel_vsid(ea) << SLB_VSID_SHIFT) |
+                       SLB_VSID_KERNEL | llp;
+               break;
+       case KERNEL_REGION_ID:
+               llp = mmu_psize_defs[mmu_linear_psize].sllp;
+               vsid = (get_kernel_vsid(ea) << SLB_VSID_SHIFT) |
+                       SLB_VSID_KERNEL | llp;
+               break;
+       default:
                /* Future: support kernel segments so that drivers
                 * can use SPUs.
                 */
@@ -97,16 +120,6 @@ static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
                return 1;
        }
 
-       esid = (ea & ESID_MASK) | SLB_ESID_V;
-#ifdef CONFIG_HUGETLB_PAGE
-       if (in_hugepage_area(mm->context, ea))
-               llp = mmu_psize_defs[mmu_huge_psize].sllp;
-       else
-#endif
-               llp = mmu_psize_defs[mmu_virtual_psize].sllp;
-       vsid = (get_vsid(mm->context.id, ea) << SLB_VSID_SHIFT) |
-                       SLB_VSID_USER | llp;
-
        out_be64(&priv2->slb_index_W, spu->slb_replace);
        out_be64(&priv2->slb_vsid_RW, vsid);
        out_be64(&priv2->slb_esid_RW, esid);
index 787ae71a6859c63cdacee241a9168fb5ed669bca..c8751936672aa5bfaa73a519b3e25344afaeba6f 100644 (file)
@@ -33,6 +33,8 @@
 #include <linux/slab.h>
 #include <linux/parser.h>
 
+#include <asm/prom.h>
+#include <asm/spu_priv1.h>
 #include <asm/io.h>
 #include <asm/semaphore.h>
 #include <asm/spu.h>
@@ -41,6 +43,7 @@
 #include "spufs.h"
 
 static kmem_cache_t *spufs_inode_cache;
+static char *isolated_loader;
 
 static struct inode *
 spufs_alloc_inode(struct super_block *sb)
@@ -232,6 +235,89 @@ struct file_operations spufs_context_fops = {
        .fsync          = simple_sync_file,
 };
 
+static int spu_setup_isolated(struct spu_context *ctx)
+{
+       int ret;
+       u64 __iomem *mfc_cntl;
+       u64 sr1;
+       u32 status;
+       unsigned long timeout;
+       const u32 status_loading = SPU_STATUS_RUNNING
+               | SPU_STATUS_ISOLATED_STATE | SPU_STATUS_ISOLATED_LOAD_STATUS;
+
+       if (!isolated_loader)
+               return -ENODEV;
+
+       if ((ret = spu_acquire_runnable(ctx)) != 0)
+               return ret;
+
+       mfc_cntl = &ctx->spu->priv2->mfc_control_RW;
+
+       /* purge the MFC DMA queue to ensure no spurious accesses before we
+        * enter kernel mode */
+       timeout = jiffies + HZ;
+       out_be64(mfc_cntl, MFC_CNTL_PURGE_DMA_REQUEST);
+       while ((in_be64(mfc_cntl) & MFC_CNTL_PURGE_DMA_STATUS_MASK)
+                       != MFC_CNTL_PURGE_DMA_COMPLETE) {
+               if (time_after(jiffies, timeout)) {
+                       printk(KERN_ERR "%s: timeout flushing MFC DMA queue\n",
+                                       __FUNCTION__);
+                       ret = -EIO;
+                       goto out_unlock;
+               }
+               cond_resched();
+       }
+
+       /* put the SPE in kernel mode to allow access to the loader */
+       sr1 = spu_mfc_sr1_get(ctx->spu);
+       sr1 &= ~MFC_STATE1_PROBLEM_STATE_MASK;
+       spu_mfc_sr1_set(ctx->spu, sr1);
+
+       /* start the loader */
+       ctx->ops->signal1_write(ctx, (unsigned long)isolated_loader >> 32);
+       ctx->ops->signal2_write(ctx,
+                       (unsigned long)isolated_loader & 0xffffffff);
+
+       ctx->ops->runcntl_write(ctx,
+                       SPU_RUNCNTL_RUNNABLE | SPU_RUNCNTL_ISOLATE);
+
+       ret = 0;
+       timeout = jiffies + HZ;
+       while (((status = ctx->ops->status_read(ctx)) & status_loading) ==
+                               status_loading) {
+               if (time_after(jiffies, timeout)) {
+                       printk(KERN_ERR "%s: timeout waiting for loader\n",
+                                       __FUNCTION__);
+                       ret = -EIO;
+                       goto out_drop_priv;
+               }
+               cond_resched();
+       }
+
+       if (!(status & SPU_STATUS_RUNNING)) {
+               /* If isolated LOAD has failed: run SPU, we will get a stop-and
+                * signal later. */
+               pr_debug("%s: isolated LOAD failed\n", __FUNCTION__);
+               ctx->ops->runcntl_write(ctx, SPU_RUNCNTL_RUNNABLE);
+               ret = -EACCES;
+
+       } else if (!(status & SPU_STATUS_ISOLATED_STATE)) {
+               /* This isn't allowed by the CBEA, but check anyway */
+               pr_debug("%s: SPU fell out of isolated mode?\n", __FUNCTION__);
+               ctx->ops->runcntl_write(ctx, SPU_RUNCNTL_STOP);
+               ret = -EINVAL;
+       }
+
+out_drop_priv:
+       /* Finished accessing the loader. Drop kernel mode */
+       sr1 |= MFC_STATE1_PROBLEM_STATE_MASK;
+       spu_mfc_sr1_set(ctx->spu, sr1);
+
+out_unlock:
+       up_write(&ctx->state_sema);
+       return ret;
+}
+
 static int
 spufs_mkdir(struct inode *dir, struct dentry *dentry, unsigned int flags,
                int mode)
@@ -255,6 +341,11 @@ spufs_mkdir(struct inode *dir, struct dentry *dentry, unsigned int flags,
                goto out_iput;
 
        ctx->flags = flags;
+       if (flags & SPU_CREATE_ISOLATE) {
+               ret = spu_setup_isolated(ctx);
+               if (ret)
+                       goto out_iput;
+       }
 
        inode->i_op = &spufs_dir_inode_operations;
        inode->i_fop = &simple_dir_operations;
@@ -555,6 +646,30 @@ spufs_parse_options(char *options, struct inode *root)
        return 1;
 }
 
+static void
+spufs_init_isolated_loader(void)
+{
+       struct device_node *dn;
+       const char *loader;
+       int size;
+
+       dn = of_find_node_by_path("/spu-isolation");
+       if (!dn)
+               return;
+
+       loader = get_property(dn, "loader", &size);
+       if (!loader)
+               return;
+
+       /* kmalloc should align on a 16 byte boundary..* */
+       isolated_loader = kmalloc(size, GFP_KERNEL);
+       if (!isolated_loader)
+               return;
+
+       memcpy(isolated_loader, loader, size);
+       printk(KERN_INFO "spufs: SPU isolation mode enabled\n");
+}
+
 static int
 spufs_create_root(struct super_block *sb, void *data)
 {
@@ -640,6 +755,8 @@ static int __init spufs_init(void)
        ret = register_spu_syscalls(&spufs_calls);
        if (ret)
                goto out_fs;
+
+       spufs_init_isolated_loader();
        return 0;
 out_fs:
        unregister_filesystem(&spufs_type);
index 0c03a04b6a3045df9a5322c252b5ae68d6f37029..a4a0080c22335da1fdbea413963007db996c6a95 100644 (file)
@@ -1,3 +1,5 @@
+#define DEBUG
+
 #include <linux/wait.h>
 #include <linux/ptrace.h>
 
@@ -56,12 +58,12 @@ static inline int spu_run_init(struct spu_context *ctx, u32 * npc)
        if ((ret = spu_acquire_runnable(ctx)) != 0)
                return ret;
 
-       if (ctx->flags & SPU_CREATE_ISOLATE)
-               runcntl |= SPU_RUNCNTL_ISOLATE;
-       else
+       /* if we're in isolated mode, we would have started the SPU
+        * earlier, so don't do it again now. */
+       if (!(ctx->flags & SPU_CREATE_ISOLATE)) {
                ctx->ops->npc_write(ctx, *npc);
-
-       ctx->ops->runcntl_write(ctx, runcntl);
+               ctx->ops->runcntl_write(ctx, runcntl);
+       }
        return 0;
 }