1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
22 #include <linux/config.h>
23 #include <linux/kernel.h>
25 #include <linux/string.h>
26 #include <linux/stat.h>
27 #include <linux/errno.h>
28 #include <linux/smp_lock.h>
29 #include <linux/unistd.h>
30 #include <linux/version.h>
31 #include <asm/system.h>
32 #include <asm/uaccess.h>
35 #include <linux/stat.h>
36 #include <asm/uaccess.h>
37 #include <asm/segment.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp_lock.h>
41 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
42 #include <linux/iobuf.h>
45 #define DEBUG_SUBSYSTEM S_LLITE
47 #include <linux/lustre_mds.h>
48 #include <linux/lustre_lite.h>
49 #include "llite_internal.h"
50 #include <linux/lustre_compat25.h>
53 struct ll_lock_tree_node {
55 struct list_head lt_locked_item;
57 ldlm_policy_data_t lt_policy;
58 struct lustre_handle lt_lockh;
62 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
63 int lt_get_mmap_locks(struct ll_lock_tree *tree, struct inode *inode,
64 unsigned long addr, size_t count);
65 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
66 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
69 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
73 struct ll_lock_tree_node * ll_node_from_inode(struct inode *inode, __u64 start,
74 __u64 end, ldlm_mode_t mode)
76 struct ll_lock_tree_node *node;
78 OBD_ALLOC(node, sizeof(*node));
80 RETURN(ERR_PTR(-ENOMEM));
82 node->lt_oid = ll_i2info(inode)->lli_smd->lsm_object_id;
83 node->lt_policy.l_extent.start = start;
84 node->lt_policy.l_extent.end = end;
85 memset(&node->lt_lockh, 0, sizeof(node->lt_lockh));
86 INIT_LIST_HEAD(&node->lt_locked_item);
92 int lt_compare(struct ll_lock_tree_node *one, struct ll_lock_tree_node *two)
94 /* XXX remove this assert when we really want to use this function
95 * to compare different file's region */
96 LASSERT(one->lt_oid == two->lt_oid);
98 if ( one->lt_oid < two->lt_oid)
100 if ( one->lt_oid > two->lt_oid)
103 if ( one->lt_policy.l_extent.end < two->lt_policy.l_extent.start )
105 if ( one->lt_policy.l_extent.start > two->lt_policy.l_extent.end )
108 return 0; /* they are the same object and overlap */
111 static void lt_merge(struct ll_lock_tree_node *dst,
112 struct ll_lock_tree_node *src)
114 dst->lt_policy.l_extent.start = min(dst->lt_policy.l_extent.start,
115 src->lt_policy.l_extent.start);
116 dst->lt_policy.l_extent.end = max(dst->lt_policy.l_extent.end,
117 src->lt_policy.l_extent.end);
119 /* XXX could be a real call to the dlm to find superset modes */
120 if (src->lt_mode == LCK_PW && dst->lt_mode != LCK_PW)
121 dst->lt_mode = LCK_PW;
124 static void lt_insert(struct ll_lock_tree *tree,
125 struct ll_lock_tree_node *node)
127 struct ll_lock_tree_node *walk;
128 rb_node_t **p, *parent;
132 p = &tree->lt_root.rb_node;
136 walk = rb_entry(parent, struct ll_lock_tree_node, lt_node);
137 switch (lt_compare(node, walk)) {
145 lt_merge(node, walk);
146 rb_erase(&walk->lt_node, &tree->lt_root);
147 OBD_FREE(walk, sizeof(*walk));
155 rb_link_node(&node->lt_node, parent, p);
156 rb_insert_color(&node->lt_node, &tree->lt_root);
160 static struct ll_lock_tree_node *lt_least_node(struct ll_lock_tree *tree)
163 struct ll_lock_tree_node *node = NULL;
165 for ( rbnode = tree->lt_root.rb_node; rbnode != NULL;
166 rbnode = rbnode->rb_left) {
167 if (rbnode->rb_left == NULL) {
168 node = rb_entry(rbnode, struct ll_lock_tree_node,
176 int ll_tree_unlock(struct ll_lock_tree *tree, struct inode *inode)
178 struct ll_lock_tree_node *node;
179 struct list_head *pos, *n;
183 list_for_each_safe(pos, n, &tree->lt_locked_list) {
184 node = list_entry(pos, struct ll_lock_tree_node,
187 rc = ll_extent_unlock(tree->lt_fd, inode,
188 ll_i2info(inode)->lli_smd, node->lt_mode,
191 /* XXX better message */
192 CERROR("couldn't unlock %d\n", rc);
194 list_del(&node->lt_locked_item);
195 OBD_FREE(node, sizeof(*node));
198 while ((node = lt_least_node(tree))) {
199 rb_erase(&node->lt_node, &tree->lt_root);
200 OBD_FREE(node, sizeof(*node));
206 int ll_tree_lock(struct ll_lock_tree *tree,
207 struct ll_lock_tree_node *first_node, struct inode *inode,
208 const char *buf, size_t count, int ast_flags)
210 struct ll_lock_tree_node *node;
214 tree->lt_root.rb_node = NULL;
215 INIT_LIST_HEAD(&tree->lt_locked_list);
216 if (first_node != NULL)
217 lt_insert(tree, first_node);
219 /* order locking. what we have to concern about is ONLY double lock:
220 * the buffer is mapped to exactly this file. */
221 if (mapping_mapped(inode->i_mapping)) {
222 rc = lt_get_mmap_locks(tree, inode, (unsigned long)buf, count);
227 while ((node = lt_least_node(tree))) {
228 struct obd_service_time *stime;
229 stime = (node->lt_mode & LCK_PW) ?
230 &ll_i2sbi(inode)->ll_write_stime :
231 &ll_i2sbi(inode)->ll_read_stime;
233 rc = ll_extent_lock(tree->lt_fd, inode,
234 ll_i2info(inode)->lli_smd, node->lt_mode,
235 &node->lt_policy, &node->lt_lockh,
240 rb_erase(&node->lt_node, &tree->lt_root);
241 list_add_tail(&node->lt_locked_item, &tree->lt_locked_list);
245 ll_tree_unlock(tree, inode);
249 static ldlm_mode_t mode_from_vma(struct vm_area_struct *vma)
251 /* we only want to hold PW locks if the mmap() can generate
252 * writes back to the file and that only happens in shared
254 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
259 static void policy_from_vma(ldlm_policy_data_t *policy,
260 struct vm_area_struct *vma, unsigned long addr,
263 policy->l_extent.start = ((addr - vma->vm_start) & PAGE_CACHE_MASK) +
264 (vma->vm_pgoff << PAGE_CACHE_SHIFT);
265 policy->l_extent.end = (policy->l_extent.start + count - 1) |
266 (PAGE_CACHE_SIZE - 1);
269 static struct vm_area_struct *our_vma(unsigned long addr, size_t count,
272 struct mm_struct *mm = current->mm;
273 struct vm_area_struct *vma, *ret = NULL;
276 spin_lock(&mm->page_table_lock);
277 for(vma = find_vma(mm, addr);
278 vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
279 if (vma->vm_ops && vma->vm_ops->nopage == ll_nopage &&
280 vma->vm_file && vma->vm_file->f_dentry->d_inode == inode) {
285 spin_unlock(&mm->page_table_lock);
289 int lt_get_mmap_locks(struct ll_lock_tree *tree, struct inode *inode,
290 unsigned long addr, size_t count)
292 struct vm_area_struct *vma;
293 struct ll_lock_tree_node *node;
294 ldlm_policy_data_t policy;
300 /* we need to look up vmas on page aligned addresses */
301 count += addr & (PAGE_SIZE - 1);
302 addr -= addr & (PAGE_SIZE - 1);
304 while ((vma = our_vma(addr, count, inode)) != NULL) {
306 policy_from_vma(&policy, vma, addr, count);
307 node = ll_node_from_inode(inode, policy.l_extent.start,
311 CERROR("not enough mem for lock_tree_node!\n");
314 lt_insert(tree, node);
316 if (vma->vm_end - addr >= count)
318 count -= vma->vm_end - addr;
323 /* FIXME: there is a pagefault race goes as follow:
324 * 1. A user process on node A accesses a portion of a mapped file,
325 * resulting in a page fault. The pagefault handler invokes the
326 * ll_nopage function, which reads the page into memory.
327 * 2. A user process on node B writes to the same portion of the file
328 * (either via mmap or write()), that cause node A to cancel the
329 * lock and truncate the page.
330 * 3. Node A then executes the rest of do_no_page(), entering the
331 * now-invalid page into the PTEs.
333 * Make the whole do_no_page as a hook to cover both the page cache
334 * and page mapping installing with dlm lock would eliminate this race.
336 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
337 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
340 struct page *ll_nopage(struct vm_area_struct *vma, unsigned long address,
344 struct file *filp = vma->vm_file;
345 struct ll_file_data *fd = filp->private_data;
346 struct inode *inode = filp->f_dentry->d_inode;
347 struct lustre_handle lockh = { 0 };
348 ldlm_policy_data_t policy;
351 struct obd_service_time *stime;
353 unsigned long pgoff, size, rand_read, seq_read;
357 if (ll_i2info(inode)->lli_smd == NULL) {
358 CERROR("No lsm on fault?\n");
362 /* start and end the lock on the first and last bytes in the page */
363 policy_from_vma(&policy, vma, address, PAGE_CACHE_SIZE);
365 CDEBUG(D_MMAP, "nopage vma %p inode %lu, locking ["LPU64", "LPU64"]\n",
366 vma, inode->i_ino, policy.l_extent.start,
367 policy.l_extent.end);
369 mode = mode_from_vma(vma);
370 stime = (mode & LCK_PW) ? &ll_i2sbi(inode)->ll_write_stime :
371 &ll_i2sbi(inode)->ll_read_stime;
373 rc = ll_extent_lock(fd, inode, ll_i2info(inode)->lli_smd, mode, &policy,
374 &lockh, LDLM_FL_CBPENDING, stime);
378 /* XXX change inode size without i_sem hold! there is a race condition
379 * with truncate path. (see ll_extent_lock) */
380 kms = lov_merge_size(ll_i2info(inode)->lli_smd, 1);
381 pgoff = ((address - vma->vm_start) >> PAGE_CACHE_SHIFT) + vma->vm_pgoff;
382 size = (kms + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
385 ll_glimpse_size(inode);
389 /* disable VM_SEQ_READ and use VM_RAND_READ to make sure that
390 * the kernel will not read other pages not covered by ldlm in
391 * filemap_nopage. we do our readahead in ll_readpage.
393 rand_read = vma->vm_flags & VM_RAND_READ;
394 seq_read = vma->vm_flags & VM_SEQ_READ;
395 vma->vm_flags &= ~ VM_SEQ_READ;
396 vma->vm_flags |= VM_RAND_READ;
398 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
399 page = filemap_nopage(vma, address, type);
401 page = filemap_nopage(vma, address, unused);
403 vma->vm_flags &= ~VM_RAND_READ;
404 vma->vm_flags |= (rand_read | seq_read);
406 ll_extent_unlock(fd, inode, ll_i2info(inode)->lli_smd, mode, &lockh);
410 /* return the user space pointer that maps to a file offset via a vma */
411 static inline unsigned long file_to_user(struct vm_area_struct *vma,
414 return vma->vm_start +
415 (byte - ((__u64)vma->vm_pgoff << PAGE_CACHE_SHIFT));
418 #define VMA_DEBUG(vma, fmt, arg...) \
419 CDEBUG(D_MMAP, "vma(%p) start(%ld) end(%ld) pgoff(%ld) inode(%p) " \
420 "ino(%lu) iname(%s): " fmt, vma, vma->vm_start, vma->vm_end, \
421 vma->vm_pgoff, vma->vm_file->f_dentry->d_inode, \
422 vma->vm_file->f_dentry->d_inode->i_ino, \
423 vma->vm_file->f_dentry->d_iname, ## arg); \
425 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
426 /* [first, last] are the byte offsets affected.
427 * vm_{start, end} are user addresses of the first byte of the mapping and
428 * the next byte beyond it
429 * vm_pgoff is the page index of the first byte in the mapping */
430 static void teardown_vmas(struct vm_area_struct *vma, __u64 first,
433 unsigned long address, len;
434 for (; vma ; vma = vma->vm_next_share) {
435 if (last >> PAGE_SHIFT < vma->vm_pgoff)
437 if (first >> PAGE_CACHE_SHIFT > (vma->vm_pgoff +
438 ((vma->vm_end - vma->vm_start) >> PAGE_CACHE_SHIFT)))
441 /* XXX in case of unmap the cow pages of a running file,
442 * don't unmap these private writeable mapping here!
443 * though that will break private mappping a little.
445 * the clean way is to check the mapping of every page
446 * and just unmap the non-cow pages, just like
447 * unmap_mapping_range() with even_cow=0 in kernel 2.6.
449 if (!(vma->vm_flags & VM_SHARED) &&
450 (vma->vm_flags & VM_WRITE))
453 address = max((unsigned long)vma->vm_start,
454 file_to_user(vma, first));
455 len = min((unsigned long)vma->vm_end,
456 file_to_user(vma, last) + 1) - address;
458 VMA_DEBUG(vma, "zapping vma [first="LPU64" last="LPU64" "
459 "address=%ld len=%ld]\n", first, last, address, len);
461 ll_zap_page_range(vma, address, len);
466 /* XXX put nice comment here. talk about __free_pte -> dirty pages and
467 * nopage's reference passing to the pte */
468 int ll_teardown_mmaps(struct address_space *mapping, __u64 first,
474 LASSERT(last > first);
475 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
476 if (mapping_mapped(mapping)) {
478 unmap_mapping_range(mapping, first + PAGE_SIZE - 1,
479 last - first + 1, 0);
482 spin_lock(&mapping->i_shared_lock);
483 if (mapping->i_mmap != NULL) {
485 teardown_vmas(mapping->i_mmap, first, last);
487 if (mapping->i_mmap_shared != NULL) {
489 teardown_vmas(mapping->i_mmap_shared, first, last);
491 spin_unlock(&mapping->i_shared_lock);
497 static struct vm_operations_struct ll_file_vm_ops = {
501 int ll_file_mmap(struct file * file, struct vm_area_struct * vma)
506 rc = generic_file_mmap(file, vma);
508 vma->vm_ops = &ll_file_vm_ops;