/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
* vim:expandtab:shiftwidth=8:tabstop=8:
*
- * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
+ * GPL HEADER START
+ *
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
*
- * Copyright (c) 2001-2003 Cluster File Systems, Inc.
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
*
- * This file is part of Lustre, http://www.lustre.org.
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
*
- * Lustre is free software; you can redistribute it and/or
- * modify it under the terms of version 2 of the GNU General Public
- * License as published by the Free Software Foundation.
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Use is subject to license terms.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
*
- * Lustre is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
- * GNU General Public License for more details.
+ * lustre/lustre/llite/rw26.c
*
- * You should have received a copy of the GNU General Public License
- * along with Lustre; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
*/
+#ifndef AUTOCONF_INCLUDED
#include <linux/config.h>
+#endif
#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/string.h>
#include <linux/fs.h>
#include <linux/buffer_head.h>
-#include <linux/mpage.h>
#include <linux/writeback.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>
#define DEBUG_SUBSYSTEM S_LLITE
-#include <linux/lustre_mds.h>
-#include <linux/lustre_lite.h>
+//#include <lustre_mdc.h>
+#include <lustre_lite.h>
#include "llite_internal.h"
#include <linux/lustre_compat25.h>
-/* called for each page in a completed rpc.*/
-void ll_ap_completion_26(void *data, int cmd, int rc)
+/**
+ * Implements Linux VM address_space::invalidatepage() method. This method is
+ * called when the page is truncate from a file, either as a result of
+ * explicit truncate, or when inode is removed from memory (as a result of
+ * final iput(), umount, or memory pressure induced icache shrinking).
+ *
+ * [0, offset] bytes of the page remain valid (this is for a case of not-page
+ * aligned truncate). Lustre leaves partially truncated page in the cache,
+ * relying on struct inode::i_size to limit further accesses.
+ */
+static int cl_invalidatepage(struct page *vmpage, unsigned long offset)
{
- struct ll_async_page *llap;
- struct page *page;
+ struct inode *inode;
+ struct lu_env *env;
+ struct cl_page *page;
+ struct cl_object *obj;
- llap = llap_from_cookie(data);
- if (IS_ERR(llap)) {
- EXIT;
- return;
- }
+ int result;
+ int refcheck;
- page = llap->llap_page;
- LASSERT(PageLocked(page));
+ LASSERT(PageLocked(vmpage));
+ LASSERT(!PageWriteback(vmpage));
- if (rc == 0) {
- if (cmd == OBD_BRW_READ) {
- if (!llap->llap_defer_uptodate)
- SetPageUptodate(page);
- } else {
- llap->llap_write_queued = 0;
+ /*
+ * It is safe to not check anything in invalidatepage/releasepage
+ * below because they are run with page locked and all our io is
+ * happening with locked page too
+ */
+ result = 0;
+ if (offset == 0) {
+ env = cl_env_get(&refcheck);
+ if (!IS_ERR(env)) {
+ inode = vmpage->mapping->host;
+ obj = ll_i2info(inode)->lli_clob;
+ if (obj != NULL) {
+ page = cl_vmpage_page(vmpage, obj);
+ if (page != NULL) {
+ lu_ref_add(&page->cp_reference,
+ "delete", vmpage);
+ cl_page_delete(env, page);
+ result = 1;
+ lu_ref_del(&page->cp_reference,
+ "delete", vmpage);
+ cl_page_put(env, page);
+ }
+ } else
+ LASSERT(vmpage->private == 0);
+ cl_env_put(env, &refcheck);
}
- } else {
- SetPageError(page);
+ }
+ return result;
+}
+
+#ifdef HAVE_INVALIDATEPAGE_RETURN_INT
+static int ll_invalidatepage(struct page *page, unsigned long offset)
+{
+ return cl_invalidatepage(page, offset);
+}
+#else /* !HAVE_INVALIDATEPAGE_RETURN_INT */
+static void ll_invalidatepage(struct page *page, unsigned long offset)
+{
+ cl_invalidatepage(page, offset);
+}
+#endif
+
+#ifdef HAVE_RELEASEPAGE_WITH_INT
+#define RELEASEPAGE_ARG_TYPE int
+#else
+#define RELEASEPAGE_ARG_TYPE gfp_t
+#endif
+static int ll_releasepage(struct page *page, RELEASEPAGE_ARG_TYPE gfp_mask)
+{
+ void *cookie;
+
+ cookie = cl_env_reenter();
+ ll_invalidatepage(page, 0);
+ cl_env_reexit(cookie);
+ return 1;
+}
+
+static int ll_set_page_dirty(struct page *vmpage)
+{
+#if 0
+ struct cl_page *page = vvp_vmpage_page_transient(vmpage);
+ struct vvp_object *obj = cl_inode2vvp(vmpage->mapping->host);
+ struct vvp_page *cpg;
+
+ /*
+ * XXX should page method be called here?
+ */
+ LASSERT(&obj->co_cl == page->cp_obj);
+ cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
+ /*
+ * XXX cannot do much here, because page is possibly not locked:
+ * sys_munmap()->...
+ * ->unmap_page_range()->zap_pte_range()->set_page_dirty().
+ */
+ vvp_write_pending(obj, cpg);
+#endif
+ RETURN(__set_page_dirty_nobuffers(vmpage));
+}
+
+#define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
+
+static inline int ll_get_user_pages(int rw, unsigned long user_addr,
+ size_t size, struct page ***pages)
+{
+ int result = -ENOMEM;
+ int page_count;
+
+ /* set an arbitrary limit to prevent arithmetic overflow */
+ if (size > MAX_DIRECTIO_SIZE) {
+ *pages = NULL;
+ return -EFBIG;
}
- LL_CDEBUG_PAGE(page, "io complete, unlocking\n");
+ page_count = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
+ page_count -= user_addr >> CFS_PAGE_SHIFT;
- unlock_page(page);
+ OBD_ALLOC_WAIT(*pages, page_count * sizeof(**pages));
+ if (*pages) {
+ down_read(¤t->mm->mmap_sem);
+ result = get_user_pages(current, current->mm, user_addr,
+ page_count, (rw == READ), 0, *pages,
+ NULL);
+ up_read(¤t->mm->mmap_sem);
+ if (result < 0)
+ OBD_FREE(*pages, page_count * sizeof(**pages));
+ }
+
+ return result;
+}
+
+/* ll_free_user_pages - tear down page struct array
+ * @pages: array of page struct pointers underlying target buffer */
+static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
+{
+ int i;
- if (0 && cmd == OBD_BRW_WRITE) {
- llap_write_complete(page->mapping->host, llap);
- ll_try_done_writing(page->mapping->host);
+ for (i = 0; i < npages; i++) {
+ if (do_dirty)
+ set_page_dirty_lock(pages[i]);
+ page_cache_release(pages[i]);
}
- page_cache_release(page);
+ OBD_FREE(pages, npages * sizeof(*pages));
}
-static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
+static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
+ int rw, struct inode *inode,
+ struct address_space *mapping,
+ size_t size, loff_t file_offset,
+ struct page **pages, int page_count)
{
- struct inode *inode = page->mapping->host;
- struct obd_export *exp;
- struct ll_async_page *llap;
- int rc;
+ struct cl_page *clp;
+ struct ccc_page *clup;
+ struct cl_2queue *queue;
+ struct cl_object *obj = io->ci_obj;
+ struct cl_sync_io *anchor = &ccc_env_info(env)->cti_sync_io;
+ int i;
+ ssize_t rc = 0;
+ ssize_t size_orig = size;
+ size_t page_size = cl_page_size(obj);
ENTRY;
- LASSERT(!PageDirty(page));
- LASSERT(PageLocked(page));
-
- exp = ll_i2obdexp(inode);
- if (exp == NULL)
- GOTO(out, rc = -EINVAL);
-
- llap = llap_from_page(page);
- if (IS_ERR(llap))
- GOTO(out, rc = PTR_ERR(llap));
- page_cache_get(page);
-
- if (llap->llap_write_queued) {
- LL_CDEBUG_PAGE(page, "marking urgent\n");
- rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
- llap->llap_cookie, ASYNC_READY |
- ASYNC_URGENT);
- } else {
- llap->llap_write_queued = 1;
- rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
- llap->llap_cookie, OBD_BRW_WRITE, 0, 0,
- OBD_BRW_CREATE, ASYNC_READY |
- ASYNC_URGENT);
- if (rc == 0)
- LL_CDEBUG_PAGE(page, "mmap write queued\n");
- else
- llap->llap_write_queued = 0;
+ cl_sync_io_init(anchor, page_count);
+
+ queue = &io->ci_queue;
+ cl_2queue_init(queue);
+ for (i = 0; i < page_count; i++) {
+ clp = cl_page_find(env, obj, cl_index(obj, file_offset),
+ pages[i], CPT_TRANSIENT);
+ if (IS_ERR(clp)) {
+ rc = PTR_ERR(clp);
+ break;
+ }
+
+ /* check the page type: if the page is a host page, then do
+ * write directly */
+ /*
+ * Very rare case that the host pages can be found for
+ * directIO case, since linux kernel truncated all covered
+ * pages before getting here. So, to make the OST happy(to
+ * write a contiguous region), all pages are issued
+ * here. -jay */
+ if (clp->cp_type == CPT_CACHEABLE) {
+ cfs_page_t *vmpage = cl_page_vmpage(env, clp);
+ cfs_page_t *src_page;
+ cfs_page_t *dst_page;
+ void *src;
+ void *dst;
+
+ src_page = (rw == WRITE) ? pages[i] : vmpage;
+ dst_page = (rw == WRITE) ? vmpage : pages[i];
+
+ src = kmap_atomic(src_page, KM_USER0);
+ dst = kmap_atomic(dst_page, KM_USER1);
+ memcpy(dst, (const void *)src, min(page_size, size));
+ kunmap_atomic(dst, KM_USER1);
+ kunmap_atomic(src, KM_USER0);
+
+ /* make sure page will be added to the transfer by
+ * cl_io_submit()->...->vvp_page_prep_write(). */
+ if (rw == WRITE)
+ set_page_dirty(vmpage);
+ /*
+ * If direct-io read finds up-to-date page in the
+ * cache, just copy it to the user space. Page will be
+ * filtered out by vvp_page_prep_read(). This
+ * preserves an invariant, that page is read at most
+ * once, see cl_page_flags::CPF_READ_COMPLETED.
+ */
+ }
+
+ rc = cl_page_own(env, io, clp);
+ if (rc) {
+ LASSERT(clp->cp_state == CPS_FREEING);
+ cl_page_put(env, clp);
+ break;
+ }
+
+ clup = cl2ccc_page(cl_page_at(clp, &vvp_device_type));
+ clup->cpg_sync_io = anchor;
+ cl_2queue_add(queue, clp);
+
+ /* drop the reference count for cl_page_find, so that the page
+ * will be freed in cl_2queue_fini. */
+ cl_page_put(env, clp);
+ /*
+ * Set page clip to tell transfer formation engine that page
+ * has to be sent even if it is beyond KMS.
+ */
+ cl_page_clip(env, clp, 0, min(size, page_size));
+ size -= page_size;
+ file_offset += page_size;
}
- if (rc)
- page_cache_release(page);
-out:
- if (rc)
- unlock_page(page);
+
+ if (rc == 0) {
+ rc = cl_io_submit_rw(env, io, rw == READ ? CRT_READ : CRT_WRITE,
+ queue);
+ if (rc == 0) {
+ /*
+ * If some pages weren't sent for any reason (e.g.,
+ * direct-io read found up-to-date pages in the
+ * cache), count them as completed to avoid infinite
+ * wait.
+ */
+ cl_page_list_for_each(clp, &queue->c2_qin)
+ cl_sync_io_note(anchor, +1);
+ /* wait for the IO to be finished. */
+ rc = cl_sync_io_wait(env, io, &queue->c2_qout,
+ anchor) ?: size_orig;
+ }
+ }
+
+ cl_2queue_discard(env, io, queue);
+ cl_2queue_disown(env, io, queue);
+ cl_2queue_fini(env, queue);
RETURN(rc);
}
+/* This is the maximum size of a single O_DIRECT request, based on a 128kB
+ * kmalloc limit. We need to fit all of the brw_page structs, each one
+ * representing PAGE_SIZE worth of user data, into a single buffer, and
+ * then truncate this to be a full-sized RPC. This is 22MB for 4kB pages. */
+#define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
+ ~(PTLRPC_MAX_BRW_SIZE - 1))
+static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
+ const struct iovec *iov, loff_t file_offset,
+ unsigned long nr_segs)
+{
+ struct lu_env *env;
+ struct cl_io *io;
+ struct file *file = iocb->ki_filp;
+ struct inode *inode = file->f_mapping->host;
+ struct ccc_object *obj = cl_inode2ccc(inode);
+ ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
+ struct ll_inode_info *lli = ll_i2info(inode);
+ struct lov_stripe_md *lsm = lli->lli_smd;
+ unsigned long seg = 0;
+ size_t size = MAX_DIO_SIZE;
+ int refcheck;
+ ENTRY;
+
+ if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
+ RETURN(-EBADF);
+
+ /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
+ if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
+ RETURN(-EINVAL);
+
+ CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
+ "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
+ inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
+ file_offset, file_offset, count >> CFS_PAGE_SHIFT,
+ MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
+
+ /* Check that all user buffers are aligned as well */
+ for (seg = 0; seg < nr_segs; seg++) {
+ if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
+ (iov[seg].iov_len & ~CFS_PAGE_MASK))
+ RETURN(-EINVAL);
+ }
+
+ env = cl_env_get(&refcheck);
+ LASSERT(!IS_ERR(env));
+ io = ccc_env_io(env)->cui_cl.cis_io;
+ LASSERT(io != NULL);
+
+ /* 0. Need locking between buffered and direct access. and race with
+ *size changing by concurrent truncates and writes.
+ * 1. Need inode sem to operate transient pages. */
+ if (rw == READ)
+ LOCK_INODE_MUTEX(inode);
+
+ LASSERT(obj->cob_transient_pages == 0);
+ for (seg = 0; seg < nr_segs; seg++) {
+ size_t iov_left = iov[seg].iov_len;
+ unsigned long user_addr = (unsigned long)iov[seg].iov_base;
+
+ if (rw == READ) {
+ if (file_offset >= inode->i_size)
+ break;
+ if (file_offset + iov_left > inode->i_size)
+ iov_left = inode->i_size - file_offset;
+ }
+
+ while (iov_left > 0) {
+ struct page **pages;
+ int page_count;
+ ssize_t result;
+
+ page_count = ll_get_user_pages(rw, user_addr,
+ min(size, iov_left),
+ &pages);
+ LASSERT(page_count != 0);
+ if (page_count > 0) {
+ result = ll_direct_IO_26_seg(env, io, rw, inode,
+ file->f_mapping,
+ min(size,iov_left),
+ file_offset, pages,
+ page_count);
+ ll_free_user_pages(pages, page_count, rw==READ);
+ } else {
+ result = 0;
+ }
+ if (page_count < 0 || result <= 0) {
+ /* If we can't allocate a large enough buffer
+ * for the request, shrink it to a smaller
+ * PAGE_SIZE multiple and try again.
+ * We should always be able to kmalloc for a
+ * page worth of page pointers = 4MB on i386. */
+ if ((page_count == -ENOMEM||result == -ENOMEM)&&
+ size > (CFS_PAGE_SIZE / sizeof(*pages)) *
+ CFS_PAGE_SIZE) {
+ size = ((((size / 2) - 1) |
+ ~CFS_PAGE_MASK) + 1) &
+ CFS_PAGE_MASK;
+ CDEBUG(D_VFSTRACE, "DIO size now %u\n",
+ (int)size);
+ continue;
+ }
+
+ if (tot_bytes <= 0)
+ tot_bytes = page_count < 0 ? page_count : result;
+ GOTO(out, tot_bytes);
+ }
+
+ tot_bytes += result;
+ file_offset += result;
+ iov_left -= result;
+ user_addr += result;
+ }
+ }
+out:
+ LASSERT(obj->cob_transient_pages == 0);
+ if (rw == READ)
+ UNLOCK_INODE_MUTEX(inode);
+
+ if (tot_bytes > 0) {
+ if (rw == WRITE) {
+ lov_stripe_lock(lsm);
+ obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
+ lov_stripe_unlock(lsm);
+ }
+ }
+
+ cl_env_put(env, &refcheck);
+ RETURN(tot_bytes);
+}
+
struct address_space_operations ll_aops = {
- readpage: ll_readpage,
-// readpages: ll_readpages,
-// direct_IO: ll_direct_IO_26,
- writepage: ll_writepage_26,
- writepages: generic_writepages,
- set_page_dirty: __set_page_dirty_nobuffers,
- sync_page: ll_sync_page,
- prepare_write: ll_prepare_write,
- commit_write: ll_commit_write,
- removepage: ll_removepage,
- bmap: NULL
+ .readpage = ll_readpage,
+// .readpages = ll_readpages,
+ .direct_IO = ll_direct_IO_26,
+ .writepage = ll_writepage,
+ .writepages = generic_writepages,
+ .set_page_dirty = ll_set_page_dirty,
+ .sync_page = NULL,
+ .prepare_write = ll_prepare_write,
+ .commit_write = ll_commit_write,
+ .invalidatepage = ll_invalidatepage,
+ .releasepage = (void *)ll_releasepage,
+ .bmap = NULL
};