X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fllite%2Frw26.c;h=b0eaf8efdcc0b652144ab644f3933403cb1551da;hp=a85c6f2833bb7e87e5d25ed420de86f9db106a6b;hb=ca58021173bbfe465a13cd34cc42b12fe5af31e4;hpb=f5fb215365791507ec98796ea624abeafa31901c diff --git a/lustre/llite/rw26.c b/lustre/llite/rw26.c index a85c6f2..b0eaf8e 100644 --- a/lustre/llite/rw26.c +++ b/lustre/llite/rw26.c @@ -27,7 +27,7 @@ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2012, Whamcloud, Inc. + * Copyright (c) 2011, 2012, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -38,21 +38,22 @@ * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version */ -#ifndef AUTOCONF_INCLUDED -#include -#endif #include #include #include #include #include #include -#include -#include #include +#ifdef HAVE_MIGRATE_H +#include +#elif defined(HAVE_MIGRATE_MODE_H) +#include +#endif #include #include +#include #include #include #include @@ -61,7 +62,6 @@ #define DEBUG_SUBSYSTEM S_LLITE -#include #include "llite_internal.h" #include @@ -75,7 +75,13 @@ * aligned truncate). Lustre leaves partially truncated page in the cache, * relying on struct inode::i_size to limit further accesses. */ -static void ll_invalidatepage(struct page *vmpage, unsigned long offset) +static void ll_invalidatepage(struct page *vmpage, +#ifdef HAVE_INVALIDATE_RANGE + unsigned int offset, unsigned int length +#else + unsigned long offset +#endif + ) { struct inode *inode; struct lu_env *env; @@ -87,12 +93,16 @@ static void ll_invalidatepage(struct page *vmpage, unsigned long offset) LASSERT(PageLocked(vmpage)); LASSERT(!PageWriteback(vmpage)); - /* - * 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 - */ - if (offset == 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 + */ +#ifdef HAVE_INVALIDATE_RANGE + if (offset == 0 && length == PAGE_CACHE_SIZE) { +#else + if (offset == 0) { +#endif env = cl_env_get(&refcheck); if (!IS_ERR(env)) { inode = vmpage->mapping->host; @@ -100,11 +110,7 @@ static void ll_invalidatepage(struct page *vmpage, unsigned long offset) 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); - lu_ref_del(&page->cp_reference, - "delete", vmpage); cl_page_put(env, page); } } else @@ -121,71 +127,61 @@ static void ll_invalidatepage(struct page *vmpage, unsigned long offset) #endif static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask) { - struct cl_env_nest nest; - struct lu_env *env; - struct cl_object *obj; - struct cl_page *page; - struct address_space *mapping; - int result; - - LASSERT(PageLocked(vmpage)); - if (PageWriteback(vmpage) || PageDirty(vmpage)) - return 0; - - mapping = vmpage->mapping; - if (mapping == NULL) - return 1; - - obj = ll_i2info(mapping->host)->lli_clob; - if (obj == NULL) - return 1; - - /* 1 for page allocator, 1 for cl_page and 1 for page cache */ - if (page_count(vmpage) > 3) - return 0; - - /* TODO: determine what gfp should be used by @gfp_mask. */ - env = cl_env_nested_get(&nest); - if (IS_ERR(env)) - /* If we can't allocate an env we won't call cl_page_put() - * later on which further means it's impossible to drop - * page refcount by cl_page, so ask kernel to not free - * this page. */ - return 0; - - page = cl_vmpage_page(vmpage, obj); - result = page == NULL; - if (page != NULL) { - if (cfs_atomic_read(&page->cp_ref) == 1) { - result = 1; - cl_page_delete(env, page); - } - cl_page_put(env, page); - } - cl_env_nested_put(&nest, env); - return result; -} + struct lu_env *env; + void *cookie; + struct cl_object *obj; + struct cl_page *page; + struct address_space *mapping; + int result = 0; + + LASSERT(PageLocked(vmpage)); + if (PageWriteback(vmpage) || PageDirty(vmpage)) + return 0; + + mapping = vmpage->mapping; + if (mapping == NULL) + return 1; + + obj = ll_i2info(mapping->host)->lli_clob; + if (obj == NULL) + return 1; + + /* 1 for caller, 1 for cl_page and 1 for page cache */ + if (page_count(vmpage) > 3) + return 0; + + page = cl_vmpage_page(vmpage, obj); + if (page == NULL) + return 1; + + cookie = cl_env_reenter(); + env = cl_env_percpu_get(); + LASSERT(!IS_ERR(env)); + + if (!cl_page_in_use(page)) { + result = 1; + cl_page_delete(env, page); + } -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)); + /* To use percpu env array, the call path can not be rescheduled; + * otherwise percpu array will be messed if ll_releaspage() called + * again on the same CPU. + * + * If this page holds the last refc of cl_object, the following + * call path may cause reschedule: + * cl_page_put -> cl_page_free -> cl_object_put -> + * lu_object_put -> lu_object_free -> lov_delete_raid0. + * + * However, the kernel can't get rid of this inode until all pages have + * been cleaned up. Now that we hold page lock here, it's pretty safe + * that we won't get into object delete path. + */ + LASSERT(cl_object_refc(obj) > 1); + cl_page_put(env, page); + + cl_env_percpu_put(env); + cl_env_reexit(cookie); + return result; } #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL @@ -202,8 +198,9 @@ static inline int ll_get_user_pages(int rw, unsigned long user_addr, return -EFBIG; } - *max_pages = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT; - *max_pages -= user_addr >> CFS_PAGE_SHIFT; + *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >> + PAGE_CACHE_SHIFT; + *max_pages -= user_addr >> PAGE_CACHE_SHIFT; OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages)); if (*pages) { @@ -240,19 +237,19 @@ ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io, int rw, struct inode *inode, struct ll_dio_pages *pv) { - struct cl_page *clp; - struct cl_2queue *queue; - struct cl_object *obj = io->ci_obj; - int i; - ssize_t rc = 0; - loff_t file_offset = pv->ldp_start_offset; - long size = pv->ldp_size; - int page_count = pv->ldp_nr; - struct page **pages = pv->ldp_pages; - long page_size = cl_page_size(obj); - bool do_io; - int io_pages = 0; - ENTRY; + struct cl_page *clp; + struct cl_2queue *queue; + struct cl_object *obj = io->ci_obj; + int i; + ssize_t rc = 0; + loff_t file_offset = pv->ldp_start_offset; + size_t size = pv->ldp_size; + int page_count = pv->ldp_nr; + struct page **pages = pv->ldp_pages; + size_t page_size = cl_page_size(obj); + bool do_io; + int io_pages = 0; + ENTRY; queue = &io->ci_queue; cl_2queue_init(queue); @@ -280,20 +277,20 @@ ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io, /* check the page type: if the page is a host page, then do * write directly */ 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; + struct page *vmpage = cl_page_vmpage(clp); + struct page *src_page; + struct page *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); + src = ll_kmap_atomic(src_page, KM_USER0); + dst = ll_kmap_atomic(dst_page, KM_USER1); memcpy(dst, src, min(page_size, size)); - kunmap_atomic(dst, KM_USER1); - kunmap_atomic(src, KM_USER0); + ll_kunmap_atomic(dst, KM_USER1); + ll_kunmap_atomic(src, KM_USER0); /* make sure page will be added to the transfer by * cl_io_submit()->...->vvp_page_prep_write(). */ @@ -369,8 +366,8 @@ static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io, * representing PAGE_SIZE worth of user data, into a single buffer, and * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */ -#define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \ - ~(PTLRPC_MAX_BRW_SIZE - 1)) +#define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \ + ~(DT_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) @@ -379,7 +376,6 @@ static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb, struct cl_io *io; struct file *file = iocb->ki_filp; struct inode *inode = file->f_mapping->host; - struct ccc_object *obj = cl_inode2ccc(inode); long count = iov_length(iov, nr_segs); long tot_bytes = 0, result = 0; struct ll_inode_info *lli = ll_i2info(inode); @@ -395,11 +391,11 @@ static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb, if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK)) RETURN(-EINVAL); - CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), " - "offset=%lld=%llx, pages %lu (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); + CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%lu (max %lu), " + "offset=%lld=%llx, pages %lu (max %lu)\n", + PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE, + file_offset, file_offset, count >> PAGE_CACHE_SHIFT, + MAX_DIO_SIZE >> PAGE_CACHE_SHIFT); /* Check that all user buffers are aligned as well */ for (seg = 0; seg < nr_segs; seg++) { @@ -413,13 +409,6 @@ static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb, 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++) { long iov_left = iov[seg].iov_len; unsigned long user_addr = (unsigned long)iov[seg].iov_base; @@ -441,7 +430,7 @@ static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb, &pages, &max_pages); if (likely(page_count > 0)) { if (unlikely(page_count < max_pages)) - bytes = page_count << CFS_PAGE_SHIFT; + bytes = page_count << PAGE_CACHE_SHIFT; result = ll_direct_IO_26_seg(env, io, rw, inode, file->f_mapping, bytes, file_offset, @@ -459,8 +448,8 @@ static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb, * We should always be able to kmalloc for a * page worth of page pointers = 4MB on i386. */ if (result == -ENOMEM && - size > (CFS_PAGE_SIZE / sizeof(*pages)) * - CFS_PAGE_SIZE) { + size > (PAGE_CACHE_SIZE / sizeof(*pages)) * + PAGE_CACHE_SIZE) { size = ((((size / 2) - 1) | ~CFS_PAGE_MASK) + 1) & CFS_PAGE_MASK; @@ -479,70 +468,221 @@ static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb, } } out: - LASSERT(obj->cob_transient_pages == 0); - if (rw == READ) - UNLOCK_INODE_MUTEX(inode); - if (tot_bytes > 0) { - if (rw == WRITE) { - struct lov_stripe_md *lsm; - - lsm = ccc_inode_lsm_get(inode); - LASSERT(lsm != NULL); - lov_stripe_lock(lsm); - obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0); - lov_stripe_unlock(lsm); - ccc_inode_lsm_put(inode, lsm); - } + struct ccc_io *cio = ccc_env_io(env); + + /* no commit async for direct IO */ + cio->u.write.cui_written += tot_bytes; } cl_env_put(env, &refcheck); - RETURN(tot_bytes ? : result); + RETURN(tot_bytes ? tot_bytes : result); +} + +/** + * Prepare partially written-to page for a write. + */ +static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io, + struct cl_page *pg) +{ + struct cl_attr *attr = ccc_env_thread_attr(env); + struct cl_object *obj = io->ci_obj; + struct ccc_page *cp = cl_object_page_slice(obj, pg); + loff_t offset = cl_offset(obj, ccc_index(cp)); + int result; + + cl_object_attr_lock(obj); + result = cl_object_attr_get(env, obj, attr); + cl_object_attr_unlock(obj); + if (result == 0) { + /* + * If are writing to a new page, no need to read old data. + * The extent locking will have updated the KMS, and for our + * purposes here we can treat it like i_size. + */ + if (attr->cat_kms <= offset) { + char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0); + + memset(kaddr, 0, cl_page_size(obj)); + ll_kunmap_atomic(kaddr, KM_USER0); + } else if (cp->cpg_defer_uptodate) + cp->cpg_ra_used = 1; + else + result = ll_page_sync_io(env, io, pg, CRT_READ); + } + return result; } -#if defined(HAVE_KERNEL_WRITE_BEGIN_END) || defined(MS_HAS_NEW_AOPS) static int ll_write_begin(struct file *file, struct address_space *mapping, - loff_t pos, unsigned len, unsigned flags, - struct page **pagep, void **fsdata) + loff_t pos, unsigned len, unsigned flags, + struct page **pagep, void **fsdata) { - pgoff_t index = pos >> PAGE_CACHE_SHIFT; - struct page *page; - int rc; - unsigned from = pos & (PAGE_CACHE_SIZE - 1); - ENTRY; + struct ll_cl_context *lcc; + const struct lu_env *env; + struct cl_io *io; + struct cl_page *page; + + struct cl_object *clob = ll_i2info(mapping->host)->lli_clob; + pgoff_t index = pos >> PAGE_CACHE_SHIFT; + struct page *vmpage = NULL; + unsigned from = pos & (PAGE_CACHE_SIZE - 1); + unsigned to = from + len; + int result = 0; + ENTRY; + + CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len); + + lcc = ll_cl_find(file); + if (lcc == NULL) + GOTO(out, result = -EIO); + + env = lcc->lcc_env; + io = lcc->lcc_io; + + /* To avoid deadlock, try to lock page first. */ + vmpage = grab_cache_page_nowait(mapping, index); + + if (unlikely(vmpage == NULL || + PageDirty(vmpage) || PageWriteback(vmpage))) { + struct ccc_io *cio = ccc_env_io(env); + struct cl_page_list *plist = &cio->u.write.cui_queue; + + /* if the page is already in dirty cache, we have to commit + * the pages right now; otherwise, it may cause deadlock + * because it holds page lock of a dirty page and request for + * more grants. It's okay for the dirty page to be the first + * one in commit page list, though. */ + if (vmpage != NULL && plist->pl_nr > 0) { + unlock_page(vmpage); + page_cache_release(vmpage); + vmpage = NULL; + } - page = grab_cache_page_write_begin(mapping, index, flags); - if (!page) - RETURN(-ENOMEM); + /* commit pages and then wait for page lock */ + result = vvp_io_write_commit(env, io); + if (result < 0) + GOTO(out, result); - *pagep = page; + if (vmpage == NULL) { + vmpage = grab_cache_page_write_begin(mapping, index, + flags); + if (vmpage == NULL) + GOTO(out, result = -ENOMEM); + } + } - rc = ll_prepare_write(file, page, from, from + len); - if (rc) { - unlock_page(page); - page_cache_release(page); - } - RETURN(rc); + page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE); + if (IS_ERR(page)) + GOTO(out, result = PTR_ERR(page)); + + lcc->lcc_page = page; + lu_ref_add(&page->cp_reference, "cl_io", io); + + cl_page_assume(env, io, page); + if (!PageUptodate(vmpage)) { + /* + * We're completely overwriting an existing page, + * so _don't_ set it up to date until commit_write + */ + if (from == 0 && to == PAGE_SIZE) { + CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n"); + POISON_PAGE(vmpage, 0x11); + } else { + /* TODO: can be optimized at OSC layer to check if it + * is a lockless IO. In that case, it's not necessary + * to read the data. */ + result = ll_prepare_partial_page(env, io, page); + if (result == 0) + SetPageUptodate(vmpage); + } + } + if (result < 0) + cl_page_unassume(env, io, page); + EXIT; +out: + if (result < 0) { + if (vmpage != NULL) { + unlock_page(vmpage); + page_cache_release(vmpage); + } + } else { + *pagep = vmpage; + *fsdata = lcc; + } + RETURN(result); } static int ll_write_end(struct file *file, struct address_space *mapping, - loff_t pos, unsigned len, unsigned copied, - struct page *page, void *fsdata) + loff_t pos, unsigned len, unsigned copied, + struct page *vmpage, void *fsdata) { - unsigned from = pos & (PAGE_CACHE_SIZE - 1); - int rc; - - rc = ll_commit_write(file, page, from, from + copied); - unlock_page(page); - page_cache_release(page); + struct ll_cl_context *lcc = fsdata; + const struct lu_env *env; + struct cl_io *io; + struct ccc_io *cio; + struct cl_page *page; + unsigned from = pos & (PAGE_CACHE_SIZE - 1); + bool unplug = false; + int result = 0; + ENTRY; + + page_cache_release(vmpage); + + LASSERT(lcc != NULL); + env = lcc->lcc_env; + page = lcc->lcc_page; + io = lcc->lcc_io; + cio = ccc_env_io(env); + + LASSERT(cl_page_is_owned(page, io)); + if (copied > 0) { + struct cl_page_list *plist = &cio->u.write.cui_queue; + + lcc->lcc_page = NULL; /* page will be queued */ + + /* Add it into write queue */ + cl_page_list_add(plist, page); + if (plist->pl_nr == 1) /* first page */ + cio->u.write.cui_from = from; + else + LASSERT(from == 0); + cio->u.write.cui_to = from + copied; + + /* To address the deadlock in balance_dirty_pages() where + * this dirty page may be written back in the same thread. */ + if (PageDirty(vmpage)) + unplug = true; + + /* We may have one full RPC, commit it soon */ + if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES) + unplug = true; + + CL_PAGE_DEBUG(D_VFSTRACE, env, page, + "queued page: %d.\n", plist->pl_nr); + } else { + cl_page_disown(env, io, page); + + lcc->lcc_page = NULL; + lu_ref_del(&page->cp_reference, "cl_io", io); + cl_page_put(env, page); + + /* page list is not contiguous now, commit it now */ + unplug = true; + } + if (unplug || + file->f_flags & O_SYNC || IS_SYNC(file->f_dentry->d_inode)) + result = vvp_io_write_commit(env, io); - return rc ?: copied; + RETURN(result >= 0 ? copied : result); } -#endif #ifdef CONFIG_MIGRATION -int ll_migratepage(struct address_space *mapping, - struct page *newpage, struct page *page) +static int ll_migratepage(struct address_space *mapping, + struct page *newpage, struct page *page +#ifdef HAVE_MIGRATEPAGE_4ARGS + , enum migrate_mode mode +#endif + ) { /* Always fail page migration until we have a proper implementation */ return -EIO; @@ -550,44 +690,33 @@ int ll_migratepage(struct address_space *mapping, #endif #ifndef MS_HAS_NEW_AOPS -struct address_space_operations ll_aops = { - .readpage = ll_readpage, -// .readpages = ll_readpages, +const struct address_space_operations ll_aops = { + .readpage = ll_readpage, .direct_IO = ll_direct_IO_26, .writepage = ll_writepage, .writepages = ll_writepages, - .set_page_dirty = ll_set_page_dirty, -#ifdef HAVE_KERNEL_WRITE_BEGIN_END + .set_page_dirty = __set_page_dirty_nobuffers, .write_begin = ll_write_begin, .write_end = ll_write_end, -#else - .prepare_write = ll_prepare_write, - .commit_write = ll_commit_write, -#endif .invalidatepage = ll_invalidatepage, .releasepage = (void *)ll_releasepage, #ifdef CONFIG_MIGRATION .migratepage = ll_migratepage, #endif - .bmap = NULL }; #else -struct address_space_operations_ext ll_aops = { - .orig_aops.readpage = ll_readpage, -// .orig_aops.readpages = ll_readpages, - .orig_aops.direct_IO = ll_direct_IO_26, - .orig_aops.writepage = ll_writepage, - .orig_aops.writepages = ll_writepages, - .orig_aops.set_page_dirty = ll_set_page_dirty, - .orig_aops.prepare_write = ll_prepare_write, - .orig_aops.commit_write = ll_commit_write, - .orig_aops.invalidatepage = ll_invalidatepage, - .orig_aops.releasepage = ll_releasepage, +const struct address_space_operations_ext ll_aops = { + .orig_aops.readpage = ll_readpage, + .orig_aops.direct_IO = ll_direct_IO_26, + .orig_aops.writepage = ll_writepage, + .orig_aops.writepages = ll_writepages, + .orig_aops.set_page_dirty = __set_page_dirty_nobuffers, + .orig_aops.invalidatepage = ll_invalidatepage, + .orig_aops.releasepage = ll_releasepage, #ifdef CONFIG_MIGRATION - .orig_aops.migratepage = ll_migratepage, + .orig_aops.migratepage = ll_migratepage, #endif - .orig_aops.bmap = NULL, - .write_begin = ll_write_begin, - .write_end = ll_write_end + .write_begin = ll_write_begin, + .write_end = ll_write_end }; #endif