Whamcloud - gitweb
LU-3487 llite: handle io init failure in ll_fault_io_init()
[fs/lustre-release.git] / lustre / llite / llite_mmap.c
index ec74d35..a42cb46 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, 2012, Whamcloud, Inc.
+ * Copyright (c) 2011, 2012, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -41,7 +41,6 @@
 #include <linux/errno.h>
 #include <linux/unistd.h>
 #include <linux/version.h>
-#include <asm/system.h>
 #include <asm/uaccess.h>
 
 #include <linux/fs.h>
@@ -66,7 +65,7 @@ void policy_from_vma(ldlm_policy_data_t *policy,
                             size_t count)
 {
         policy->l_extent.start = ((addr - vma->vm_start) & CFS_PAGE_MASK) +
-                                 (vma->vm_pgoff << CFS_PAGE_SHIFT);
+                                (vma->vm_pgoff << PAGE_CACHE_SHIFT);
         policy->l_extent.end = (policy->l_extent.start + count - 1) |
                                ~CFS_PAGE_MASK;
 }
@@ -104,16 +103,17 @@ struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
  * \return other error codes from cl_io_init.
  */
 struct cl_io *ll_fault_io_init(struct vm_area_struct *vma,
-                               struct lu_env **env_ret,
-                               struct cl_env_nest *nest,
-                               pgoff_t index, unsigned long *ra_flags)
+                              struct lu_env **env_ret,
+                              struct cl_env_nest *nest,
+                              pgoff_t index, unsigned long *ra_flags)
 {
-        struct file       *file  = vma->vm_file;
-        struct inode      *inode = file->f_dentry->d_inode;
-        struct cl_io      *io;
-        struct cl_fault_io *fio;
-        struct lu_env     *env;
-        ENTRY;
+       struct file            *file = vma->vm_file;
+       struct inode           *inode = file->f_dentry->d_inode;
+       struct cl_io           *io;
+       struct cl_fault_io     *fio;
+       struct lu_env          *env;
+       int                     rc;
+       ENTRY;
 
         *env_ret = NULL;
         if (ll_file_nolock(file))
@@ -153,65 +153,77 @@ struct cl_io *ll_fault_io_init(struct vm_area_struct *vma,
         CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
                fio->ft_index, fio->ft_executable);
 
-        if (cl_io_init(env, io, CIT_FAULT, io->ci_obj) == 0) {
-                struct ccc_io *cio = ccc_env_io(env);
-                struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
-
-                LASSERT(cio->cui_cl.cis_io == io);
-
-                /* mmap lock must be MANDATORY
-                 * it has to cache pages. */
-                io->ci_lockreq = CILR_MANDATORY;
-
-                cio->cui_fd  = fd;
-        }
+       rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
+       if (rc == 0) {
+               struct ccc_io *cio = ccc_env_io(env);
+               struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
+
+               LASSERT(cio->cui_cl.cis_io == io);
+
+               /* mmap lock must be MANDATORY it has to cache
+                * pages. */
+               io->ci_lockreq = CILR_MANDATORY;
+               cio->cui_fd = fd;
+       } else {
+               LASSERT(rc < 0);
+               cl_io_fini(env, io);
+               cl_env_nested_put(nest, env);
+               io = ERR_PTR(rc);
+       }
 
-        return io;
+       return io;
 }
 
 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
                             bool *retry)
 {
-        struct lu_env           *env;
-        struct cl_io            *io;
-        struct vvp_io           *vio;
-        struct cl_env_nest       nest;
-        int                      result;
+       struct lu_env           *env;
+       struct cl_io            *io;
+       struct vvp_io           *vio;
+       struct cl_env_nest       nest;
+       int                      result;
        cfs_sigset_t             set;
-        ENTRY;
+       struct inode             *inode;
+       struct ll_inode_info     *lli;
+       ENTRY;
 
-        LASSERT(vmpage != NULL);
+       LASSERT(vmpage != NULL);
 
-        io = ll_fault_io_init(vma, &env,  &nest, vmpage->index, NULL);
-        if (IS_ERR(io))
-                GOTO(out, result = PTR_ERR(io));
+       io = ll_fault_io_init(vma, &env,  &nest, vmpage->index, NULL);
+       if (IS_ERR(io))
+               GOTO(out, result = PTR_ERR(io));
 
-        result = io->ci_result;
-        if (result < 0)
-                GOTO(out, result);
+       result = io->ci_result;
+       if (result < 0)
+               GOTO(out_io, result);
 
-        /* Don't enqueue new locks for page_mkwrite().
-         * If the lock has been cancelled then page must have been
-         * truncated, in that case, kernel will handle it.
-         */
-        io->ci_lockreq = CILR_PEEK;
-        io->u.ci_fault.ft_mkwrite = 1;
-        io->u.ci_fault.ft_writable = 1;
+       io->u.ci_fault.ft_mkwrite = 1;
+       io->u.ci_fault.ft_writable = 1;
 
-        vio = vvp_env_io(env);
-        vio->u.fault.ft_vma    = vma;
-        vio->u.fault.ft_vmpage = vmpage;
+       vio = vvp_env_io(env);
+       vio->u.fault.ft_vma    = vma;
+       vio->u.fault.ft_vmpage = vmpage;
 
        set = cfs_block_sigsinv(sigmask(SIGKILL) | sigmask(SIGTERM));
+
+       /* we grab lli_trunc_sem to exclude truncate case.
+        * Otherwise, we could add dirty pages into osc cache
+        * while truncate is on-going. */
+       inode = ccc_object_inode(io->ci_obj);
+       lli = ll_i2info(inode);
+       down_read(&lli->lli_trunc_sem);
+
        result = cl_io_loop(env, io);
+
+       up_read(&lli->lli_trunc_sem);
+
        cfs_restore_sigs(set);
 
-        if (result == -ENODATA) /* peek failed, no lock caching. */
-                CDEBUG(D_MMAP, "race on page_mkwrite: %lx (%lu %p)\n",
-                       vma->vm_flags, io->u.ci_fault.ft_index, vmpage);
+        if (result == 0) {
+               struct inode *inode = vma->vm_file->f_dentry->d_inode;
+               struct ll_inode_info *lli = ll_i2info(inode);
 
-        if (result == 0 || result == -ENODATA) {
                 lock_page(vmpage);
                 if (vmpage->mapping == NULL) {
                         unlock_page(vmpage);
@@ -221,18 +233,6 @@ static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
                          * to handle_mm_fault(). */
                         if (result == 0)
                                 result = -ENODATA;
-                } else if (result == -ENODATA) {
-                        /* Invalidate it if the cl_lock is being revoked.
-                         * This piece of code is definitely needed for RHEL5,
-                         * otherwise, SIGBUS will be wrongly returned to
-                         * applications. */
-                        write_one_page(vmpage, 1);
-                        lock_page(vmpage);
-                        if (vmpage->mapping != NULL) {
-                                ll_invalidate_page(vmpage);
-                                LASSERT(vmpage->mapping == NULL);
-                        }
-                        unlock_page(vmpage);
                 } else if (!PageDirty(vmpage)) {
                         /* race, the page has been cleaned by ptlrpcd after
                          * it was unlocked, it has to be added into dirty
@@ -249,17 +249,23 @@ static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
                         *retry = true;
                         result = -EAGAIN;
                 }
+
+               if (result == 0) {
+                       spin_lock(&lli->lli_lock);
+                       lli->lli_flags |= LLIF_DATA_MODIFIED;
+                       spin_unlock(&lli->lli_lock);
+               }
         }
         EXIT;
 
+out_io:
+       cl_io_fini(env, io);
+       cl_env_nested_put(&nest, env);
 out:
-        cl_io_fini(env, io);
-        cl_env_nested_put(&nest, env);
-
-        CDEBUG(D_MMAP, "%s mkwrite with %d\n", cfs_current()->comm, result);
+       CDEBUG(D_MMAP, "%s mkwrite with %d\n", cfs_current()->comm, result);
+       LASSERT(ergo(result == 0, PageLocked(vmpage)));
 
-        LASSERT(ergo(result == 0, PageLocked(vmpage)));
-        return(result);
+       return result;
 }
 
 
@@ -576,7 +582,8 @@ static int ll_populate(struct vm_area_struct *area, unsigned long address,
 /* return the user space pointer that maps to a file offset via a vma */
 static inline unsigned long file_to_user(struct vm_area_struct *vma, __u64 byte)
 {
-        return vma->vm_start + (byte - ((__u64)vma->vm_pgoff << CFS_PAGE_SHIFT));
+       return vma->vm_start +
+              (byte - ((__u64)vma->vm_pgoff << PAGE_CACHE_SHIFT));
 
 }
 
@@ -590,7 +597,7 @@ int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
         LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
         if (mapping_mapped(mapping)) {
                 rc = 0;
-                unmap_mapping_range(mapping, first + CFS_PAGE_SIZE - 1,
+               unmap_mapping_range(mapping, first + PAGE_CACHE_SIZE - 1,
                                     last - first + 1, 0);
         }
 
@@ -599,15 +606,18 @@ int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
 
 static struct vm_operations_struct ll_file_vm_ops = {
 #ifndef HAVE_VM_OP_FAULT
-        .nopage         = ll_nopage,
-        .populate       = ll_populate,
-
+       .nopage                 = ll_nopage,
+       .populate               = ll_populate,
+#else
+       .fault                  = ll_fault,
+#endif
+#ifndef HAVE_PGMKWRITE_COMPACT
+       .page_mkwrite           = ll_page_mkwrite,
 #else
-        .fault          = ll_fault,
+       ._pmkw.page_mkwrite     = ll_page_mkwrite,
 #endif
-        .page_mkwrite   = ll_page_mkwrite,
-        .open           = ll_vm_open,
-        .close          = ll_vm_close,
+       .open                   = ll_vm_open,
+       .close                  = ll_vm_close,
 };
 
 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)