Whamcloud - gitweb
Branch HEAD
[fs/lustre-release.git] / lustre / llite / llite_mmap.c
index f1ff91a..147a4d7 100644 (file)
@@ -18,7 +18,8 @@
  *   along with Lustre; if not, write to the Free Software
  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
-#ifdef HAVE_KERNEL_CONFIG_H
+
+#ifndef AUTOCONF_INCLUDED
 #include <linux/config.h>
 #endif
 #include <linux/kernel.h>
 #include <linux/fs.h>
 #include <linux/stat.h>
 #include <asm/uaccess.h>
-#include <asm/segment.h>
 #include <linux/mm.h>
 #include <linux/pagemap.h>
 #include <linux/smp_lock.h>
-#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
-#include <linux/iobuf.h>
-#endif
 
 #define DEBUG_SUBSYSTEM S_LLITE
 
+//#include <lustre_mdc.h>
 #include <lustre_lite.h>
 #include "llite_internal.h"
 #include <linux/lustre_compat25.h>
@@ -70,14 +68,8 @@ struct ll_lock_tree_node {
 int lt_get_mmap_locks(struct ll_lock_tree *tree,
                       unsigned long addr, size_t count);
 
-#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
                        int *type);
-#else
-
-struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
-                       int unused);
-#endif
 
 struct ll_lock_tree_node * ll_node_from_inode(struct inode *inode, __u64 start,
                                               __u64 end, ldlm_mode_t mode)
@@ -270,7 +262,7 @@ static void policy_from_vma(ldlm_policy_data_t *policy,
                             size_t count)
 {
         policy->l_extent.start = ((addr - vma->vm_start) & CFS_PAGE_MASK) +
-                                 ((__u64)vma->vm_pgoff << CFS_PAGE_SHIFT);
+                                 (vma->vm_pgoff << CFS_PAGE_SHIFT);
         policy->l_extent.end = (policy->l_extent.start + count - 1) |
                                ~CFS_PAGE_MASK;
 }
@@ -351,13 +343,8 @@ int lt_get_mmap_locks(struct ll_lock_tree *tree,
  *
  * In 2.6, the truncate_count of address_space can cover this race.
  */
-#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
                        int *type)
-#else
-struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
-                       int type /* unused */)
-#endif
 {
         struct file *filp = vma->vm_file;
         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
@@ -379,6 +366,8 @@ struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
                 RETURN(NULL);
         }
 
+        ll_clear_file_contended(inode);
+
         /* start and end the lock on the first and last bytes in the page */
         policy_from_vma(&policy, vma, address, CFS_PAGE_SIZE);
 
@@ -399,7 +388,7 @@ struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
 
         lov_stripe_lock(lsm);
         inode_init_lvb(inode, &lvb);
-        obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
+        obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
         kms = lvb.lvb_size;
 
         pgoff = ((address - vma->vm_start) >> CFS_PAGE_SHIFT) + vma->vm_pgoff;
@@ -412,16 +401,22 @@ struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
                 /* XXX change inode size without ll_inode_size_lock() held!
                  *     there is a race condition with truncate path. (see
                  *     ll_extent_lock) */
-               /* region is within kms and, hence, within real file size (A).
+                /* XXX i_size_write() is not used because it is not safe to
+                 *     take the ll_inode_size_lock() due to a potential lock
+                 *     inversion (bug 6077).  And since it's not safe to use
+                 *     i_size_write() without a covering mutex we do the
+                 *     assignment directly.  It is not critical that the
+                 *     size be correct. */
+                /* region is within kms and, hence, within real file size (A).
                  * We need to increase i_size to cover the read region so that
                  * generic_file_read() will do its job, but that doesn't mean
                  * the kms size is _correct_, it is only the _minimum_ size.
                  * If someone does a stat they will get the correct size which
                  * will always be >= the kms value here.  b=11081 */
-                if (inode->i_size < kms) {
+                if (i_size_read(inode) < kms) {
                         inode->i_size = kms;
-                        CDEBUG(D_INODE, "ino=%lu, updating i_size %llu\n",
-                               inode->i_ino, inode->i_size);
+                        CDEBUG(D_INODE, "ino=%lu, updating i_size %llu\n",
+                               inode->i_ino, i_size_read(inode));
                 }
                 lov_stripe_unlock(lsm);
         }
@@ -432,9 +427,9 @@ struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
          * bug 10919 */
         lov_stripe_lock(lsm);
         if (mode == LCK_PW)
-                obd_adjust_kms(ll_i2obdexp(inode), lsm,
-                               min_t(loff_t, policy.l_extent.end,inode->i_size),
-                               0);
+                obd_adjust_kms(ll_i2dtexp(inode), lsm,
+                               min_t(loff_t, policy.l_extent.end + 1,
+                               i_size_read(inode)), 0);
         lov_stripe_unlock(lsm);
 
         /* disable VM_SEQ_READ and use VM_RAND_READ to make sure that
@@ -484,7 +479,7 @@ static void ll_vm_open(struct vm_area_struct * vma)
 
                 if (!lsm)
                         return;
-                count = obd_join_lru(sbi->ll_osc_exp, lsm, 0);
+                count = obd_join_lru(sbi->ll_dt_exp, lsm, 0);
                 VMA_DEBUG(vma, "split %d unused locks from lru\n", count);
         } else {
                 spin_unlock(&lli->lli_lock);
@@ -513,14 +508,13 @@ static void ll_vm_close(struct vm_area_struct *vma)
 
                 if (!lsm)
                         return;
-                count = obd_join_lru(sbi->ll_osc_exp, lsm, 1);
+                count = obd_join_lru(sbi->ll_dt_exp, lsm, 1);
                 VMA_DEBUG(vma, "join %d unused locks to lru\n", count);
         } else {
                 spin_unlock(&lli->lli_lock);
         }
 }
 
-#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
 #ifndef HAVE_FILEMAP_POPULATE
 static int (*filemap_populate)(struct vm_area_struct * area, unsigned long address, unsigned long len, pgprot_t prot, unsigned long pgoff, int nonblock);
 #endif
@@ -535,7 +529,6 @@ static int ll_populate(struct vm_area_struct *area, unsigned long address,
         rc = filemap_populate(area, address, len, prot, pgoff, 1);
         RETURN(rc);
 }
-#endif
 
 /* 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)
@@ -544,47 +537,6 @@ static inline unsigned long file_to_user(struct vm_area_struct *vma, __u64 byte)
 
 }
 
-#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
-/* [first, last] are the byte offsets affected.
- * vm_{start, end} are user addresses of the first byte of the mapping and
- *      the next byte beyond it
- * vm_pgoff is the page index of the first byte in the mapping */
-static void teardown_vmas(struct vm_area_struct *vma, __u64 first,
-                          __u64 last)
-{
-        unsigned long address, len;
-        for (; vma ; vma = vma->vm_next_share) {
-                if (last >> CFS_PAGE_SHIFT < vma->vm_pgoff)
-                        continue;
-                if (first >> CFS_PAGE_SHIFT >= (vma->vm_pgoff +
-                    ((vma->vm_end - vma->vm_start) >> CFS_PAGE_SHIFT)))
-                        continue;
-
-                /* XXX in case of unmap the cow pages of a running file,
-                 * don't unmap these private writeable mapping here!
-                 * though that will break private mappping a little.
-                 *
-                 * the clean way is to check the mapping of every page
-                 * and just unmap the non-cow pages, just like
-                 * unmap_mapping_range() with even_cow=0 in kernel 2.6.
-                 */
-                if (!(vma->vm_flags & VM_SHARED) &&
-                    (vma->vm_flags & VM_WRITE))
-                        continue;
-
-                address = max((unsigned long)vma->vm_start,
-                              file_to_user(vma, first));
-                len = min((unsigned long)vma->vm_end,
-                          file_to_user(vma, last) + 1) - address;
-
-                VMA_DEBUG(vma, "zapping vma [first="LPU64" last="LPU64" "
-                          "address=%ld len=%ld]\n", first, last, address, len);
-                LASSERT(len > 0);
-                ll_zap_page_range(vma, address, len);
-        }
-}
-#endif
-
 /* XXX put nice comment here.  talk about __free_pte -> dirty pages and
  * nopage's reference passing to the pte */
 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
@@ -593,24 +545,12 @@ int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
         ENTRY;
 
         LASSERTF(last > first, "last "LPU64" first "LPU64"\n", last, first);
-#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
         if (mapping_mapped(mapping)) {
                 rc = 0;
                 unmap_mapping_range(mapping, first + CFS_PAGE_SIZE - 1,
                                     last - first + 1, 0);
         }
-#else
-        spin_lock(&mapping->i_shared_lock);
-        if (mapping->i_mmap != NULL) {
-                rc = 0;
-                teardown_vmas(mapping->i_mmap, first, last);
-        }
-        if (mapping->i_mmap_shared != NULL) {
-                rc = 0;
-                teardown_vmas(mapping->i_mmap_shared, first, last);
-        }
-        spin_unlock(&mapping->i_shared_lock);
-#endif
+
         RETURN(rc);
 }
 
@@ -618,9 +558,7 @@ static struct vm_operations_struct ll_file_vm_ops = {
         .nopage         = ll_nopage,
         .open           = ll_vm_open,
         .close          = ll_vm_close,
-#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
         .populate       = ll_populate,
-#endif
 };
 
 int ll_file_mmap(struct file * file, struct vm_area_struct * vma)
@@ -628,11 +566,10 @@ int ll_file_mmap(struct file * file, struct vm_area_struct * vma)
         int rc;
         ENTRY;
 
-        ll_vfs_ops_tally(ll_i2sbi(file->f_dentry->d_inode), VFS_OPS_MMAP);
+        ll_stats_ops_tally(ll_i2sbi(file->f_dentry->d_inode), LPROC_LL_MAP, 1);
         rc = generic_file_mmap(file, vma);
         if (rc == 0) {
-#if !defined(HAVE_FILEMAP_POPULATE) && \
-    (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
+#if !defined(HAVE_FILEMAP_POPULATE)
                 if (!filemap_populate)
                         filemap_populate = vma->vm_ops->populate;
 #endif