Whamcloud - gitweb
3475469bf624b622ad5b551116b26c1a1380e8ad
[fs/lustre-release.git] / lustre / llite / rw26.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  * GPL HEADER START
5  *
6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
7  *
8  * This program is free software; you can redistribute it and/or modify
9  * it under the terms of the GNU General Public License version 2 only,
10  * as published by the Free Software Foundation.
11  *
12  * This program is distributed in the hope that it will be useful, but
13  * WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * General Public License version 2 for more details (a copy is included
16  * in the LICENSE file that accompanied this code).
17  *
18  * You should have received a copy of the GNU General Public License
19  * version 2 along with this program; If not, see
20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
21  *
22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23  * CA 95054 USA or visit www.sun.com if you need additional information or
24  * have any questions.
25  *
26  * GPL HEADER END
27  */
28 /*
29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30  * Use is subject to license terms.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * lustre/lustre/llite/rw26.c
37  *
38  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
39  */
40
41 #ifndef AUTOCONF_INCLUDED
42 #include <linux/config.h>
43 #endif
44 #include <linux/kernel.h>
45 #include <linux/mm.h>
46 #include <linux/string.h>
47 #include <linux/stat.h>
48 #include <linux/errno.h>
49 #include <linux/smp_lock.h>
50 #include <linux/unistd.h>
51 #include <linux/version.h>
52 #include <asm/system.h>
53 #include <asm/uaccess.h>
54
55 #include <linux/fs.h>
56 #include <linux/buffer_head.h>
57 #include <linux/writeback.h>
58 #include <linux/stat.h>
59 #include <asm/uaccess.h>
60 #include <linux/mm.h>
61 #include <linux/pagemap.h>
62 #include <linux/smp_lock.h>
63
64 #define DEBUG_SUBSYSTEM S_LLITE
65
66 //#include <lustre_mdc.h>
67 #include <lustre_lite.h>
68 #include "llite_internal.h"
69 #include <linux/lustre_compat25.h>
70
71 /**
72  * Implements Linux VM address_space::invalidatepage() method. This method is
73  * called when the page is truncate from a file, either as a result of
74  * explicit truncate, or when inode is removed from memory (as a result of
75  * final iput(), umount, or memory pressure induced icache shrinking).
76  *
77  * [0, offset] bytes of the page remain valid (this is for a case of not-page
78  * aligned truncate). Lustre leaves partially truncated page in the cache,
79  * relying on struct inode::i_size to limit further accesses.
80  */
81 static int cl_invalidatepage(struct page *vmpage, unsigned long offset)
82 {
83         struct inode     *inode;
84         struct lu_env    *env;
85         struct cl_page   *page;
86         struct cl_object *obj;
87
88         int result;
89         int refcheck;
90
91         LASSERT(PageLocked(vmpage));
92         LASSERT(!PageWriteback(vmpage));
93
94         /*
95          * It is safe to not check anything in invalidatepage/releasepage
96          * below because they are run with page locked and all our io is
97          * happening with locked page too
98          */
99         result = 0;
100         if (offset == 0) {
101                 env = cl_env_get(&refcheck);
102                 if (!IS_ERR(env)) {
103                         inode = vmpage->mapping->host;
104                         obj = ll_i2info(inode)->lli_clob;
105                         if (obj != NULL) {
106                                 page = cl_vmpage_page(vmpage, obj);
107                                 if (page != NULL) {
108                                         lu_ref_add(&page->cp_reference,
109                                                    "delete", vmpage);
110                                         cl_page_delete(env, page);
111                                         result = 1;
112                                         lu_ref_del(&page->cp_reference,
113                                                    "delete", vmpage);
114                                         cl_page_put(env, page);
115                                 }
116                         } else
117                                 LASSERT(vmpage->private == 0);
118                         cl_env_put(env, &refcheck);
119                 }
120         }
121         return result;
122 }
123
124 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
125 static int ll_invalidatepage(struct page *page, unsigned long offset)
126 {
127         return cl_invalidatepage(page, offset);
128 }
129 #else /* !HAVE_INVALIDATEPAGE_RETURN_INT */
130 static void ll_invalidatepage(struct page *page, unsigned long offset)
131 {
132         cl_invalidatepage(page, offset);
133 }
134 #endif
135
136 #ifdef HAVE_RELEASEPAGE_WITH_INT
137 #define RELEASEPAGE_ARG_TYPE int
138 #else
139 #define RELEASEPAGE_ARG_TYPE gfp_t
140 #endif
141 static int ll_releasepage(struct page *page, RELEASEPAGE_ARG_TYPE gfp_mask)
142 {
143         void *cookie;
144
145         cookie = cl_env_reenter();
146         ll_invalidatepage(page, 0);
147         cl_env_reexit(cookie);
148         return 1;
149 }
150
151 static int ll_set_page_dirty(struct page *vmpage)
152 {
153 #if 0
154         struct cl_page    *page = vvp_vmpage_page_transient(vmpage);
155         struct vvp_object *obj  = cl_inode2vvp(vmpage->mapping->host);
156         struct vvp_page   *cpg;
157
158         /*
159          * XXX should page method be called here?
160          */
161         LASSERT(&obj->co_cl == page->cp_obj);
162         cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
163         /*
164          * XXX cannot do much here, because page is possibly not locked:
165          * sys_munmap()->...
166          *     ->unmap_page_range()->zap_pte_range()->set_page_dirty().
167          */
168         vvp_write_pending(obj, cpg);
169 #endif
170         RETURN(__set_page_dirty_nobuffers(vmpage));
171 }
172
173 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
174
175 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
176                                     size_t size, struct page ***pages,
177                                     int *max_pages)
178 {
179         int result = -ENOMEM;
180
181         /* set an arbitrary limit to prevent arithmetic overflow */
182         if (size > MAX_DIRECTIO_SIZE) {
183                 *pages = NULL;
184                 return -EFBIG;
185         }
186
187         *max_pages = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
188         *max_pages -= user_addr >> CFS_PAGE_SHIFT;
189
190         OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
191         if (*pages) {
192                 down_read(&current->mm->mmap_sem);
193                 result = get_user_pages(current, current->mm, user_addr,
194                                         *max_pages, (rw == READ), 0, *pages,
195                                         NULL);
196                 up_read(&current->mm->mmap_sem);
197                 if (unlikely(result <= 0))
198                         OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
199         }
200
201         return result;
202 }
203
204 /*  ll_free_user_pages - tear down page struct array
205  *  @pages: array of page struct pointers underlying target buffer */
206 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
207 {
208         int i;
209
210         for (i = 0; i < npages; i++) {
211                 if (pages[i] == NULL)
212                         break;
213                 if (do_dirty)
214                         set_page_dirty_lock(pages[i]);
215                 page_cache_release(pages[i]);
216         }
217
218         OBD_FREE_LARGE(pages, npages * sizeof(*pages));
219 }
220
221 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
222                            int rw, struct inode *inode,
223                            struct ll_dio_pages *pv)
224 {
225         struct cl_page    *clp;
226         struct cl_2queue  *queue;
227         struct cl_object  *obj = io->ci_obj;
228         int i;
229         ssize_t rc = 0;
230         loff_t file_offset  = pv->ldp_start_offset;
231         long size           = pv->ldp_size;
232         int page_count      = pv->ldp_nr;
233         struct page **pages = pv->ldp_pages;
234         long page_size      = cl_page_size(obj);
235         ENTRY;
236
237         queue = &io->ci_queue;
238         cl_2queue_init(queue);
239         for (i = 0; i < page_count; i++) {
240                 if (pv->ldp_offsets)
241                     file_offset = pv->ldp_offsets[i];
242                 LASSERT(!(file_offset & (page_size - 1)));
243                 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
244                                    pv->ldp_pages[i], CPT_TRANSIENT);
245                 if (IS_ERR(clp)) {
246                         rc = PTR_ERR(clp);
247                         break;
248                 }
249
250                 /* check the page type: if the page is a host page, then do
251                  * write directly */
252                 /*
253                  * Very rare case that the host pages can be found for
254                  * directIO case, since linux kernel truncated all covered
255                  * pages before getting here. So, to make the OST happy(to
256                  * write a contiguous region), all pages are issued
257                  * here. -jay */
258                 if (clp->cp_type == CPT_CACHEABLE) {
259                         cfs_page_t *vmpage = cl_page_vmpage(env, clp);
260                         cfs_page_t *src_page;
261                         cfs_page_t *dst_page;
262                         void       *src;
263                         void       *dst;
264
265                         src_page = (rw == WRITE) ? pages[i] : vmpage;
266                         dst_page = (rw == WRITE) ? vmpage : pages[i];
267
268                         src = kmap_atomic(src_page, KM_USER0);
269                         dst = kmap_atomic(dst_page, KM_USER1);
270                         memcpy(dst, src, min(page_size, size));
271                         kunmap_atomic(dst, KM_USER1);
272                         kunmap_atomic(src, KM_USER0);
273
274                         /* make sure page will be added to the transfer by
275                          * cl_io_submit()->...->vvp_page_prep_write(). */
276                         if (rw == WRITE)
277                                 set_page_dirty(vmpage);
278                         /*
279                          * If direct-io read finds up-to-date page in the
280                          * cache, just copy it to the user space. Page will be
281                          * filtered out by vvp_page_prep_read(). This
282                          * preserves an invariant, that page is read at most
283                          * once, see cl_page_flags::CPF_READ_COMPLETED.
284                          */
285                 }
286
287                 rc = cl_page_own(env, io, clp);
288                 if (rc) {
289                         LASSERT(clp->cp_state == CPS_FREEING);
290                         cl_page_put(env, clp);
291                         break;
292                 }
293
294                 cl_2queue_add(queue, clp);
295
296                 /* drop the reference count for cl_page_find, so that the page
297                  * will be freed in cl_2queue_fini. */
298                 cl_page_put(env, clp);
299                 /*
300                  * Set page clip to tell transfer formation engine that page
301                  * has to be sent even if it is beyond KMS.
302                  */
303                 cl_page_clip(env, clp, 0, min(size, page_size));
304                 size -= page_size;
305                 file_offset += page_size;
306         }
307
308         if (rc == 0) {
309                 rc = cl_io_submit_sync(env, io,
310                                        rw == READ ? CRT_READ : CRT_WRITE,
311                                        queue, CRP_NORMAL, 0);
312                 if (rc == 0)
313                         rc = pv->ldp_size;
314         }
315
316         cl_2queue_discard(env, io, queue);
317         cl_2queue_disown(env, io, queue);
318         cl_2queue_fini(env, queue);
319         RETURN(rc);
320 }
321 EXPORT_SYMBOL(ll_direct_rw_pages);
322
323 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
324                                    int rw, struct inode *inode,
325                                    struct address_space *mapping,
326                                    size_t size, loff_t file_offset,
327                                    struct page **pages, int page_count)
328 {
329     struct ll_dio_pages pvec = { .ldp_pages        = pages,
330                                  .ldp_nr           = page_count,
331                                  .ldp_size         = size,
332                                  .ldp_offsets      = NULL,
333                                  .ldp_start_offset = file_offset
334                                };
335
336     return ll_direct_rw_pages(env, io, rw, inode, &pvec);
337 }
338
339 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
340  * kmalloc limit.  We need to fit all of the brw_page structs, each one
341  * representing PAGE_SIZE worth of user data, into a single buffer, and
342  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
343 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
344                       ~(PTLRPC_MAX_BRW_SIZE - 1))
345 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
346                                const struct iovec *iov, loff_t file_offset,
347                                unsigned long nr_segs)
348 {
349         struct lu_env *env;
350         struct cl_io *io;
351         struct file *file = iocb->ki_filp;
352         struct inode *inode = file->f_mapping->host;
353         struct ccc_object *obj = cl_inode2ccc(inode);
354         long count = iov_length(iov, nr_segs);
355         long tot_bytes = 0, result = 0;
356         struct ll_inode_info *lli = ll_i2info(inode);
357         struct lov_stripe_md *lsm = lli->lli_smd;
358         unsigned long seg = 0;
359         long size = MAX_DIO_SIZE;
360         int refcheck;
361         ENTRY;
362
363         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
364                 RETURN(-EBADF);
365
366         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
367         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
368                 RETURN(-EINVAL);
369
370         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), "
371                "offset=%lld=%llx, pages %lu (max %lu)\n",
372                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
373                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
374                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
375
376         /* Check that all user buffers are aligned as well */
377         for (seg = 0; seg < nr_segs; seg++) {
378                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
379                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
380                         RETURN(-EINVAL);
381         }
382
383         env = cl_env_get(&refcheck);
384         LASSERT(!IS_ERR(env));
385         io = ccc_env_io(env)->cui_cl.cis_io;
386         LASSERT(io != NULL);
387
388         /* 0. Need locking between buffered and direct access. and race with
389          *size changing by concurrent truncates and writes.
390          * 1. Need inode sem to operate transient pages. */
391         if (rw == READ)
392                 LOCK_INODE_MUTEX(inode);
393
394         LASSERT(obj->cob_transient_pages == 0);
395         for (seg = 0; seg < nr_segs; seg++) {
396                 long iov_left = iov[seg].iov_len;
397                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
398
399                 if (rw == READ) {
400                         if (file_offset >= inode->i_size)
401                                 break;
402                         if (file_offset + iov_left > inode->i_size)
403                                 iov_left = inode->i_size - file_offset;
404                 }
405
406                 while (iov_left > 0) {
407                         struct page **pages;
408                         int page_count, max_pages = 0;
409                         long bytes;
410
411                         bytes = min(size,iov_left);
412                         page_count = ll_get_user_pages(rw, user_addr, bytes,
413                                                        &pages, &max_pages);
414                         if (likely(page_count > 0)) {
415                                 if (unlikely(page_count <  max_pages))
416                                         bytes = page_count << CFS_PAGE_SHIFT;
417                                 result = ll_direct_IO_26_seg(env, io, rw, inode,
418                                                              file->f_mapping,
419                                                              bytes,
420                                                              file_offset, pages,
421                                                              page_count);
422                                 ll_free_user_pages(pages, max_pages, rw==READ);
423                         } else if (page_count == 0) {
424                                 GOTO(out, result = -EFAULT);
425                         } else {
426                                 result = page_count;
427                         }
428                         if (unlikely(result <= 0)) {
429                                 /* If we can't allocate a large enough buffer
430                                  * for the request, shrink it to a smaller
431                                  * PAGE_SIZE multiple and try again.
432                                  * We should always be able to kmalloc for a
433                                  * page worth of page pointers = 4MB on i386. */
434                                 if (result == -ENOMEM &&
435                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
436                                            CFS_PAGE_SIZE) {
437                                         size = ((((size / 2) - 1) |
438                                                  ~CFS_PAGE_MASK) + 1) &
439                                                 CFS_PAGE_MASK;
440                                         CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
441                                                size);
442                                         continue;
443                                 }
444
445                                 GOTO(out, result);
446                         }
447
448                         tot_bytes += result;
449                         file_offset += result;
450                         iov_left -= result;
451                         user_addr += result;
452                 }
453         }
454 out:
455         LASSERT(obj->cob_transient_pages == 0);
456         if (rw == READ)
457                 UNLOCK_INODE_MUTEX(inode);
458
459         if (tot_bytes > 0) {
460                 if (rw == WRITE) {
461                         lov_stripe_lock(lsm);
462                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
463                         lov_stripe_unlock(lsm);
464                 }
465         }
466
467         cl_env_put(env, &refcheck);
468         RETURN(tot_bytes ? : result);
469 }
470
471 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
472 static int ll_write_begin(struct file *file, struct address_space *mapping,
473                          loff_t pos, unsigned len, unsigned flags,
474                          struct page **pagep, void **fsdata)
475 {
476         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
477         struct page *page;
478         int rc;
479         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
480         ENTRY;
481
482         page = grab_cache_page_write_begin(mapping, index, flags);
483         if (!page)
484                 RETURN(-ENOMEM);
485
486         *pagep = page;
487
488         rc = ll_prepare_write(file, page, from, from + len);
489         if (rc) {
490                 unlock_page(page);
491                 page_cache_release(page);
492         }
493         RETURN(rc);
494 }
495
496 static int ll_write_end(struct file *file, struct address_space *mapping,
497                         loff_t pos, unsigned len, unsigned copied,
498                         struct page *page, void *fsdata)
499 {
500         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
501         int rc;
502         rc = ll_commit_write(file, page, from, from + copied);
503
504         unlock_page(page);
505         page_cache_release(page);
506         return rc?rc:copied;
507 }
508 #endif
509
510 struct address_space_operations ll_aops = {
511         .readpage       = ll_readpage,
512 //        .readpages      = ll_readpages,
513         .direct_IO      = ll_direct_IO_26,
514         .writepage      = ll_writepage,
515         .writepages     = generic_writepages,
516         .set_page_dirty = ll_set_page_dirty,
517         .sync_page      = NULL,
518 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
519         .write_begin    = ll_write_begin,
520         .write_end      = ll_write_end,
521 #else
522         .prepare_write  = ll_prepare_write,
523         .commit_write   = ll_commit_write,
524 #endif
525         .invalidatepage = ll_invalidatepage,
526         .releasepage    = (void *)ll_releasepage,
527         .bmap           = NULL
528 };