shmem: make shmem_inode_info::lock irq-safe
authorKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Tue, 26 Jul 2016 22:26:29 +0000 (15:26 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 26 Jul 2016 23:19:19 +0000 (16:19 -0700)
We are going to need to call shmem_charge() under tree_lock to get
accoutning right on collapse of small tmpfs pages into a huge one.

The problem is that tree_lock is irq-safe and lockdep is not happy, that
we take irq-unsafe lock under irq-safe[1].

Let's convert the lock to irq-safe.

[1] https://gist.github.com/kiryl/80c0149e03ed35dfaf26628b8e03cdbc

Link: http://lkml.kernel.org/r/1466021202-61880-34-git-send-email-kirill.shutemov@linux.intel.com
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
ipc/shm.c
mm/shmem.c

index 7fa5cbebbf19f75fd33caadd12327b60c6276ae9..dbac8860c721ed64527cf87093dd9e717b7c703f 100644 (file)
--- a/ipc/shm.c
+++ b/ipc/shm.c
@@ -766,10 +766,10 @@ static void shm_add_rss_swap(struct shmid_kernel *shp,
        } else {
 #ifdef CONFIG_SHMEM
                struct shmem_inode_info *info = SHMEM_I(inode);
-               spin_lock(&info->lock);
+               spin_lock_irq(&info->lock);
                *rss_add += inode->i_mapping->nrpages;
                *swp_add += info->swapped;
-               spin_unlock(&info->lock);
+               spin_unlock_irq(&info->lock);
 #else
                *rss_add += inode->i_mapping->nrpages;
 #endif
index f19b6b44ae4661fe68c01afe3c2be78440352ffc..03eb915c82eb8472086a54fa771c2cc6388b1c6a 100644 (file)
@@ -258,14 +258,15 @@ bool shmem_charge(struct inode *inode, long pages)
 {
        struct shmem_inode_info *info = SHMEM_I(inode);
        struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
+       unsigned long flags;
 
        if (shmem_acct_block(info->flags, pages))
                return false;
-       spin_lock(&info->lock);
+       spin_lock_irqsave(&info->lock, flags);
        info->alloced += pages;
        inode->i_blocks += pages * BLOCKS_PER_PAGE;
        shmem_recalc_inode(inode);
-       spin_unlock(&info->lock);
+       spin_unlock_irqrestore(&info->lock, flags);
        inode->i_mapping->nrpages += pages;
 
        if (!sbinfo->max_blocks)
@@ -273,10 +274,10 @@ bool shmem_charge(struct inode *inode, long pages)
        if (percpu_counter_compare(&sbinfo->used_blocks,
                                sbinfo->max_blocks - pages) > 0) {
                inode->i_mapping->nrpages -= pages;
-               spin_lock(&info->lock);
+               spin_lock_irqsave(&info->lock, flags);
                info->alloced -= pages;
                shmem_recalc_inode(inode);
-               spin_unlock(&info->lock);
+               spin_unlock_irqrestore(&info->lock, flags);
 
                return false;
        }
@@ -288,12 +289,13 @@ void shmem_uncharge(struct inode *inode, long pages)
 {
        struct shmem_inode_info *info = SHMEM_I(inode);
        struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
+       unsigned long flags;
 
-       spin_lock(&info->lock);
+       spin_lock_irqsave(&info->lock, flags);
        info->alloced -= pages;
        inode->i_blocks -= pages * BLOCKS_PER_PAGE;
        shmem_recalc_inode(inode);
-       spin_unlock(&info->lock);
+       spin_unlock_irqrestore(&info->lock, flags);
 
        if (sbinfo->max_blocks)
                percpu_counter_sub(&sbinfo->used_blocks, pages);
@@ -818,10 +820,10 @@ static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
                index++;
        }
 
-       spin_lock(&info->lock);
+       spin_lock_irq(&info->lock);
        info->swapped -= nr_swaps_freed;
        shmem_recalc_inode(inode);
-       spin_unlock(&info->lock);
+       spin_unlock_irq(&info->lock);
 }
 
 void shmem_truncate_range(struct inode *inode, loff_t lstart, loff_t lend)
@@ -838,9 +840,9 @@ static int shmem_getattr(struct vfsmount *mnt, struct dentry *dentry,
        struct shmem_inode_info *info = SHMEM_I(inode);
 
        if (info->alloced - info->swapped != inode->i_mapping->nrpages) {
-               spin_lock(&info->lock);
+               spin_lock_irq(&info->lock);
                shmem_recalc_inode(inode);
-               spin_unlock(&info->lock);
+               spin_unlock_irq(&info->lock);
        }
        generic_fillattr(inode, stat);
        return 0;
@@ -984,9 +986,9 @@ static int shmem_unuse_inode(struct shmem_inode_info *info,
                delete_from_swap_cache(*pagep);
                set_page_dirty(*pagep);
                if (!error) {
-                       spin_lock(&info->lock);
+                       spin_lock_irq(&info->lock);
                        info->swapped--;
-                       spin_unlock(&info->lock);
+                       spin_unlock_irq(&info->lock);
                        swap_free(swap);
                }
        }
@@ -1134,10 +1136,10 @@ static int shmem_writepage(struct page *page, struct writeback_control *wbc)
                list_add_tail(&info->swaplist, &shmem_swaplist);
 
        if (add_to_swap_cache(page, swap, GFP_ATOMIC) == 0) {
-               spin_lock(&info->lock);
+               spin_lock_irq(&info->lock);
                shmem_recalc_inode(inode);
                info->swapped++;
-               spin_unlock(&info->lock);
+               spin_unlock_irq(&info->lock);
 
                swap_shmem_alloc(swap);
                shmem_delete_from_page_cache(page, swp_to_radix_entry(swap));
@@ -1523,10 +1525,10 @@ repeat:
 
                mem_cgroup_commit_charge(page, memcg, true, false);
 
-               spin_lock(&info->lock);
+               spin_lock_irq(&info->lock);
                info->swapped--;
                shmem_recalc_inode(inode);
-               spin_unlock(&info->lock);
+               spin_unlock_irq(&info->lock);
 
                if (sgp == SGP_WRITE)
                        mark_page_accessed(page);
@@ -1603,11 +1605,11 @@ alloc_nohuge:           page = shmem_alloc_and_acct_page(gfp, info, sbinfo,
                                PageTransHuge(page));
                lru_cache_add_anon(page);
 
-               spin_lock(&info->lock);
+               spin_lock_irq(&info->lock);
                info->alloced += 1 << compound_order(page);
                inode->i_blocks += BLOCKS_PER_PAGE << compound_order(page);
                shmem_recalc_inode(inode);
-               spin_unlock(&info->lock);
+               spin_unlock_irq(&info->lock);
                alloced = true;
 
                /*
@@ -1639,9 +1641,9 @@ clear:
                if (alloced) {
                        ClearPageDirty(page);
                        delete_from_page_cache(page);
-                       spin_lock(&info->lock);
+                       spin_lock_irq(&info->lock);
                        shmem_recalc_inode(inode);
-                       spin_unlock(&info->lock);
+                       spin_unlock_irq(&info->lock);
                }
                error = -EINVAL;
                goto unlock;
@@ -1673,9 +1675,9 @@ unlock:
        }
        if (error == -ENOSPC && !once++) {
                info = SHMEM_I(inode);
-               spin_lock(&info->lock);
+               spin_lock_irq(&info->lock);
                shmem_recalc_inode(inode);
-               spin_unlock(&info->lock);
+               spin_unlock_irq(&info->lock);
                goto repeat;
        }
        if (error == -EEXIST)   /* from above or from radix_tree_insert */
@@ -1874,7 +1876,7 @@ int shmem_lock(struct file *file, int lock, struct user_struct *user)
        struct shmem_inode_info *info = SHMEM_I(inode);
        int retval = -ENOMEM;
 
-       spin_lock(&info->lock);
+       spin_lock_irq(&info->lock);
        if (lock && !(info->flags & VM_LOCKED)) {
                if (!user_shm_lock(inode->i_size, user))
                        goto out_nomem;
@@ -1889,7 +1891,7 @@ int shmem_lock(struct file *file, int lock, struct user_struct *user)
        retval = 0;
 
 out_nomem:
-       spin_unlock(&info->lock);
+       spin_unlock_irq(&info->lock);
        return retval;
 }