X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fllite%2Fllite_mmap.c;h=e45f84db4e0061efdc2c32cbb092643430697a8d;hb=e4deafd378448a0ef27ec4c2d1c191cc55bf691c;hp=ff3eefba7edeff5755fa4d4e0c7e9d540a47e047;hpb=f63ae8e8a6f27d4454ceef553143fb67300fe6d8;p=fs%2Flustre-release.git diff --git a/lustre/llite/llite_mmap.c b/lustre/llite/llite_mmap.c index ff3eefb..e45f84d 100644 --- a/lustre/llite/llite_mmap.c +++ b/lustre/llite/llite_mmap.c @@ -19,7 +19,9 @@ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. */ +#ifndef AUTOCONF_INCLUDED #include +#endif #include #include #include @@ -44,11 +46,18 @@ #define DEBUG_SUBSYSTEM S_LLITE -#include -#include +//#include +#include #include "llite_internal.h" #include +#define VMA_DEBUG(vma, fmt, arg...) \ + CDEBUG(D_MMAP, "vma(%p) start(%ld) end(%ld) pgoff(%ld) inode(%p) " \ + "ino(%lu) iname(%s): " fmt, vma, vma->vm_start, vma->vm_end, \ + vma->vm_pgoff, vma->vm_file->f_dentry->d_inode, \ + vma->vm_file->f_dentry->d_inode->i_ino, \ + vma->vm_file->f_dentry->d_iname, ## arg); \ + struct ll_lock_tree_node { rb_node_t lt_node; @@ -57,15 +66,17 @@ struct ll_lock_tree_node { ldlm_policy_data_t lt_policy; struct lustre_handle lt_lockh; ldlm_mode_t lt_mode; + struct inode *lt_inode; }; -__u64 lov_merge_size(struct lov_stripe_md *lsm, int kms); -int lt_get_mmap_locks(struct ll_lock_tree *tree, struct inode *inode, +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 @@ -79,6 +90,7 @@ struct ll_lock_tree_node * ll_node_from_inode(struct inode *inode, __u64 start, if (node == NULL) RETURN(ERR_PTR(-ENOMEM)); + node->lt_inode = inode; node->lt_oid = ll_i2info(inode)->lli_smd->lsm_object_id; node->lt_policy.l_extent.start = start; node->lt_policy.l_extent.end = end; @@ -91,14 +103,20 @@ struct ll_lock_tree_node * ll_node_from_inode(struct inode *inode, __u64 start, int lt_compare(struct ll_lock_tree_node *one, struct ll_lock_tree_node *two) { - if ( one->lt_oid < two->lt_oid) + /* To avoid multiple fs deadlock */ + if (one->lt_inode->i_sb->s_dev < two->lt_inode->i_sb->s_dev) return -1; - if ( one->lt_oid > two->lt_oid) + if (one->lt_inode->i_sb->s_dev > two->lt_inode->i_sb->s_dev) return 1; - if ( one->lt_policy.l_extent.end < two->lt_policy.l_extent.start ) + if (one->lt_oid < two->lt_oid) return -1; - if ( one->lt_policy.l_extent.start > two->lt_policy.l_extent.end ) + if (one->lt_oid > two->lt_oid) + return 1; + + if (one->lt_policy.l_extent.end < two->lt_policy.l_extent.start) + return -1; + if (one->lt_policy.l_extent.start > two->lt_policy.l_extent.end) return 1; return 0; /* they are the same object and overlap */ @@ -169,10 +187,11 @@ static struct ll_lock_tree_node *lt_least_node(struct ll_lock_tree *tree) RETURN(node); } -int ll_tree_unlock(struct ll_lock_tree *tree, struct inode *inode) +int ll_tree_unlock(struct ll_lock_tree *tree) { struct ll_lock_tree_node *node; struct list_head *pos, *n; + struct inode *inode; int rc = 0; ENTRY; @@ -180,6 +199,7 @@ int ll_tree_unlock(struct ll_lock_tree *tree, struct inode *inode) node = list_entry(pos, struct ll_lock_tree_node, lt_locked_item); + inode = node->lt_inode; rc = ll_extent_unlock(tree->lt_fd, inode, ll_i2info(inode)->lli_smd, node->lt_mode, &node->lt_lockh); @@ -200,7 +220,7 @@ int ll_tree_unlock(struct ll_lock_tree *tree, struct inode *inode) } int ll_tree_lock(struct ll_lock_tree *tree, - struct ll_lock_tree_node *first_node, struct inode *inode, + struct ll_lock_tree_node *first_node, const char *buf, size_t count, int ast_flags) { struct ll_lock_tree_node *node; @@ -212,22 +232,19 @@ int ll_tree_lock(struct ll_lock_tree *tree, if (first_node != NULL) lt_insert(tree, first_node); - if (mapping_mapped(inode->i_mapping)) { - rc = lt_get_mmap_locks(tree, inode, (unsigned long)buf, count); - if (rc) - GOTO(out, rc); - } + /* To avoid such subtle deadlock case: client1 try to read file1 to + * mmapped file2, on the same time, client2 try to read file2 to + * mmapped file1.*/ + rc = lt_get_mmap_locks(tree, (unsigned long)buf, count); + if (rc) + GOTO(out, rc); while ((node = lt_least_node(tree))) { - struct obd_service_time *stime; - stime = (node->lt_mode & LCK_PW) ? - &ll_i2sbi(inode)->ll_write_stime : - &ll_i2sbi(inode)->ll_read_stime; - + struct inode *inode = node->lt_inode; rc = ll_extent_lock(tree->lt_fd, inode, ll_i2info(inode)->lli_smd, node->lt_mode, &node->lt_policy, &node->lt_lockh, - ast_flags, stime); + ast_flags); if (rc != 0) GOTO(out, rc); @@ -236,8 +253,8 @@ int ll_tree_lock(struct ll_lock_tree *tree, } RETURN(rc); out: - ll_tree_unlock(tree, inode); - return rc; + ll_tree_unlock(tree); + RETURN(rc); } static ldlm_mode_t mode_from_vma(struct vm_area_struct *vma) @@ -254,21 +271,27 @@ static void policy_from_vma(ldlm_policy_data_t *policy, struct vm_area_struct *vma, unsigned long addr, size_t count) { - policy->l_extent.start = ((addr - vma->vm_start) & PAGE_CACHE_MASK) + - (vma->vm_pgoff << PAGE_CACHE_SHIFT); + policy->l_extent.start = ((addr - vma->vm_start) & CFS_PAGE_MASK) + + (vma->vm_pgoff << CFS_PAGE_SHIFT); policy->l_extent.end = (policy->l_extent.start + count - 1) | - (PAGE_CACHE_SIZE - 1); + ~CFS_PAGE_MASK; } + static struct vm_area_struct * our_vma(unsigned long addr, size_t count) { struct mm_struct *mm = current->mm; struct vm_area_struct *vma, *ret = NULL; ENTRY; + /* No MM (e.g. NFS)? No vmas too. */ + if (!mm) + RETURN(NULL); + spin_lock(&mm->page_table_lock); for(vma = find_vma(mm, addr); vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) { - if (vma->vm_ops && vma->vm_ops->nopage == ll_nopage) { + if (vma->vm_ops && vma->vm_ops->nopage == ll_nopage && + vma->vm_flags & VM_SHARED) { ret = vma; break; } @@ -277,23 +300,26 @@ static struct vm_area_struct * our_vma(unsigned long addr, size_t count) RETURN(ret); } -int lt_get_mmap_locks(struct ll_lock_tree *tree, struct inode *inode, +int lt_get_mmap_locks(struct ll_lock_tree *tree, unsigned long addr, size_t count) { struct vm_area_struct *vma; struct ll_lock_tree_node *node; ldlm_policy_data_t policy; + struct inode *inode; ENTRY; if (count == 0) RETURN(0); /* we need to look up vmas on page aligned addresses */ - count += addr & (PAGE_SIZE - 1); - addr -= addr & (PAGE_SIZE - 1); + count += addr & (~CFS_PAGE_MASK); + addr &= CFS_PAGE_MASK; while ((vma = our_vma(addr, count)) != NULL) { + LASSERT(vma->vm_file); + inode = vma->vm_file->f_dentry->d_inode; policy_from_vma(&policy, vma, addr, count); node = ll_node_from_inode(inode, policy.l_extent.start, policy.l_extent.end, @@ -311,7 +337,8 @@ int lt_get_mmap_locks(struct ll_lock_tree *tree, struct inode *inode, } RETURN(0); } -/* FIXME: there is a pagefault race goes as follow: + +/* FIXME: there is a pagefault race goes as follow (only 2.4): * 1. A user process on node A accesses a portion of a mapped file, * resulting in a page fault. The pagefault handler invokes the * ll_nopage function, which reads the page into memory. @@ -323,59 +350,94 @@ int lt_get_mmap_locks(struct ll_lock_tree *tree, struct inode *inode, * * Make the whole do_no_page as a hook to cover both the page cache * and page mapping installing with dlm lock would eliminate this race. + * + * 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 unused) + int type /* unused */) #endif { struct file *filp = vma->vm_file; - struct ll_file_data *fd = filp->private_data; + struct ll_file_data *fd = LUSTRE_FPRIVATE(filp); struct inode *inode = filp->f_dentry->d_inode; struct lustre_handle lockh = { 0 }; ldlm_policy_data_t policy; ldlm_mode_t mode; - struct page *page; - struct obd_service_time *stime; - __u64 kms; + struct page *page = NULL; + struct ll_inode_info *lli = ll_i2info(inode); + struct lov_stripe_md *lsm; + struct ost_lvb lvb; + __u64 kms, old_mtime; unsigned long pgoff, size, rand_read, seq_read; int rc = 0; ENTRY; - if (ll_i2info(inode)->lli_smd == NULL) { + if (lli->lli_smd == NULL) { CERROR("No lsm on fault?\n"); RETURN(NULL); } /* start and end the lock on the first and last bytes in the page */ - policy_from_vma(&policy, vma, address, PAGE_CACHE_SIZE); + policy_from_vma(&policy, vma, address, CFS_PAGE_SIZE); CDEBUG(D_MMAP, "nopage vma %p inode %lu, locking ["LPU64", "LPU64"]\n", - vma, inode->i_ino, policy.l_extent.start, - policy.l_extent.end); + vma, inode->i_ino, policy.l_extent.start, policy.l_extent.end); mode = mode_from_vma(vma); - stime = (mode & LCK_PW) ? &ll_i2sbi(inode)->ll_write_stime : - &ll_i2sbi(inode)->ll_read_stime; + old_mtime = LTIME_S(inode->i_mtime); - rc = ll_extent_lock(fd, inode, ll_i2info(inode)->lli_smd, mode, &policy, - &lockh, LDLM_FL_CBPENDING, stime); + lsm = lli->lli_smd; + rc = ll_extent_lock(fd, inode, lsm, mode, &policy, + &lockh, LDLM_FL_CBPENDING | LDLM_FL_NO_LRU); if (rc != 0) RETURN(NULL); - /* XXX change inode size without i_sem hold! there is a race condition - * with truncate path. (see ll_extent_lock) */ - kms = lov_merge_size(ll_i2info(inode)->lli_smd, 1); - pgoff = ((address - vma->vm_start) >> PAGE_CACHE_SHIFT) + vma->vm_pgoff; - size = (kms + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT; + if (vma->vm_flags & VM_EXEC && LTIME_S(inode->i_mtime) != old_mtime) + CWARN("binary changed. inode %lu\n", inode->i_ino); + + lov_stripe_lock(lsm); + inode_init_lvb(inode, &lvb); + obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1); + kms = lvb.lvb_size; + + pgoff = ((address - vma->vm_start) >> CFS_PAGE_SHIFT) + vma->vm_pgoff; + size = (kms + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT; + + if (pgoff >= size) { + lov_stripe_unlock(lsm); + ll_glimpse_size(inode, LDLM_FL_BLOCK_GRANTED); + } else { + /* 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). + * 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 (i_size_read(inode) < kms) { + i_size_write(inode, kms); + CDEBUG(D_INODE, "ino=%lu, updating i_size %llu\n", + inode->i_ino, i_size_read(inode)); + } + lov_stripe_unlock(lsm); + } - if (pgoff >= size) - ll_glimpse_size(inode); - else - inode->i_size = kms; + /* If mapping is writeable, adjust kms to cover this page, + * but do not extend kms beyond actual file size. + * policy.l_extent.end is set to the end of the page by policy_from_vma + * bug 10919 */ + lov_stripe_lock(lsm); + if (mode == LCK_PW) + obd_adjust_kms(ll_i2dtexp(inode), lsm, + min_t(loff_t, policy.l_extent.end, + i_size_read(inode)), 0); + lov_stripe_unlock(lsm); /* disable VM_SEQ_READ and use VM_RAND_READ to make sure that * the kernel will not read other pages not covered by ldlm in @@ -386,11 +448,9 @@ struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address, vma->vm_flags &= ~ VM_SEQ_READ; vma->vm_flags |= VM_RAND_READ; -#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)) page = filemap_nopage(vma, address, type); -#else - page = filemap_nopage(vma, address, unused); -#endif + LL_CDEBUG_PAGE(D_PAGE, page, "got addr %lu type %lx\n", address, + (long)type); vma->vm_flags &= ~VM_RAND_READ; vma->vm_flags |= (rand_read | seq_read); @@ -398,19 +458,93 @@ struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address, RETURN(page); } -/* 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) +/* To avoid cancel the locks covering mmapped region for lock cache pressure, + * we track the mapped vma count by lli_mmap_cnt. + * ll_vm_open(): when first vma is linked, split locks from lru. + * ll_vm_close(): when last vma is unlinked, join all this file's locks to lru. + * + * XXX we don't check the if the region of vma/lock for performance. + */ +static void ll_vm_open(struct vm_area_struct * vma) { - return vma->vm_start + - (byte - ((__u64)vma->vm_pgoff << PAGE_CACHE_SHIFT)); + struct inode *inode = vma->vm_file->f_dentry->d_inode; + struct ll_inode_info *lli = ll_i2info(inode); + ENTRY; + + LASSERT(vma->vm_file); + + spin_lock(&lli->lli_lock); + LASSERT(atomic_read(&lli->lli_mmap_cnt) >= 0); + + atomic_inc(&lli->lli_mmap_cnt); + if (atomic_read(&lli->lli_mmap_cnt) == 1) { + struct lov_stripe_md *lsm = lli->lli_smd; + struct ll_sb_info *sbi = ll_i2sbi(inode); + int count; + + spin_unlock(&lli->lli_lock); + + if (!lsm) + return; + 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); + } } -#define VMA_DEBUG(vma, fmt, arg...) \ - CDEBUG(D_MMAP, "vma(%p) start(%ld) end(%ld) pgoff(%ld) inode(%p): " \ - fmt, vma, vma->vm_start, vma->vm_end, vma->vm_pgoff, \ - vma->vm_file->f_dentry->d_inode, ## arg); +static void ll_vm_close(struct vm_area_struct *vma) +{ + struct inode *inode = vma->vm_file->f_dentry->d_inode; + struct ll_inode_info *lli = ll_i2info(inode); + ENTRY; + + LASSERT(vma->vm_file); + + spin_lock(&lli->lli_lock); + LASSERT(atomic_read(&lli->lli_mmap_cnt) > 0); + + atomic_dec(&lli->lli_mmap_cnt); + if (atomic_read(&lli->lli_mmap_cnt) == 0) { + struct lov_stripe_md *lsm = lli->lli_smd; + struct ll_sb_info *sbi = ll_i2sbi(inode); + int count; + + spin_unlock(&lli->lli_lock); + + if (!lsm) + return; + 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 +static int ll_populate(struct vm_area_struct *area, unsigned long address, + unsigned long len, pgprot_t prot, unsigned long pgoff, + int nonblock) +{ + int rc = 0; + ENTRY; + + /* always set nonblock as true to avoid page read ahead */ + 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) +{ + return vma->vm_start + (byte - ((__u64)vma->vm_pgoff << CFS_PAGE_SHIFT)); + +} #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0)) /* [first, last] are the byte offsets affected. @@ -422,10 +556,22 @@ static void teardown_vmas(struct vm_area_struct *vma, __u64 first, { unsigned long address, len; for (; vma ; vma = vma->vm_next_share) { - if (last >> PAGE_CACHE_SHIFT < vma->vm_pgoff) + if (last >> CFS_PAGE_SHIFT < vma->vm_pgoff) continue; - if (first >> PAGE_CACHE_SHIFT > (vma->vm_pgoff + - ((vma->vm_end - vma->vm_start) >> PAGE_CACHE_SHIFT))) + 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, @@ -433,9 +579,9 @@ static void teardown_vmas(struct vm_area_struct *vma, __u64 first, len = min((unsigned long)vma->vm_end, file_to_user(vma, last) + 1) - address; - VMA_DEBUG(vma, "zapping vma [address=%ld len=%ld]\n", - address, len); - LASSERT(vma->vm_mm); + 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); } } @@ -443,17 +589,17 @@ static void teardown_vmas(struct vm_area_struct *vma, __u64 first, /* 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) +int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last) { int rc = -ENOENT; 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 + PAGE_SIZE - 1, - last - first + 1, 1); + unmap_mapping_range(mapping, first + CFS_PAGE_SIZE - 1, + last - first + 1, 0); } #else spin_lock(&mapping->i_shared_lock); @@ -467,12 +613,16 @@ int ll_teardown_mmaps(struct address_space *mapping, __u64 first, } spin_unlock(&mapping->i_shared_lock); #endif - RETURN(rc); } 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) @@ -480,10 +630,19 @@ int ll_file_mmap(struct file * file, struct vm_area_struct * vma) int rc; ENTRY; + 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 (rc == 0) { +#if !defined(HAVE_FILEMAP_POPULATE) && \ + (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0)) + if (!filemap_populate) + filemap_populate = vma->vm_ops->populate; +#endif vma->vm_ops = &ll_file_vm_ops; + vma->vm_ops->open(vma); + /* update the inode's size and mtime */ + rc = ll_glimpse_size(file->f_dentry->d_inode, 0); + } RETURN(rc); } -