-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* GPL HEADER START
*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
* GPL HEADER END
*/
/*
- * Copyright 2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2012, Whamcloud, Inc.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
* Lustre Lite I/O page cache routines shared by different kernel revs
*/
-#include <linux/autoconf.h>
#include <linux/kernel.h>
#include <linux/mm.h>
#include <linux/string.h>
#include <linux/stat.h>
#include <linux/errno.h>
-#include <linux/smp_lock.h>
#include <linux/unistd.h>
#include <linux/version.h>
#include <asm/system.h>
#include <asm/uaccess.h>
#include <linux/mm.h>
#include <linux/pagemap.h>
-#include <linux/smp_lock.h>
/* current_is_kswapd() */
#include <linux/swap.h>
#define DEBUG_SUBSYSTEM S_LLITE
-//#include <lustre_mdc.h>
#include <lustre_lite.h>
#include <obd_cksum.h>
#include "llite_internal.h"
* must be called under ->lli_size_sem */
void ll_truncate(struct inode *inode)
{
- struct ll_inode_info *lli = ll_i2info(inode);
- loff_t new_size;
ENTRY;
- CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
- inode->i_generation, inode, i_size_read(inode),
- i_size_read(inode));
-
- ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
- if (lli->lli_size_sem_owner != cfs_current()) {
- EXIT;
- return;
- }
-
- if (!lli->lli_smd) {
- CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
- inode->i_ino);
- GOTO(out_unlock, 0);
- }
- LASSERT_SEM_LOCKED(&lli->lli_size_sem);
-
- if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) &&
- (i_size_read(inode) & ~CFS_PAGE_MASK))) {
- /* If the truncate leaves a partial page, update its checksum */
- struct page *page = find_get_page(inode->i_mapping,
- i_size_read(inode) >>
- CFS_PAGE_SHIFT);
- if (page != NULL) {
-#if 0 /* XXX */
- struct ll_async_page *llap = llap_cast_private(page);
- if (llap != NULL) {
- char *kaddr = kmap_atomic(page, KM_USER0);
- llap->llap_checksum =
- init_checksum(OSC_DEFAULT_CKSUM);
- llap->llap_checksum =
- compute_checksum(llap->llap_checksum,
- kaddr, CFS_PAGE_SIZE,
- OSC_DEFAULT_CKSUM);
- kunmap_atomic(kaddr, KM_USER0);
- }
- page_cache_release(page);
-#endif
- }
- }
- new_size = i_size_read(inode);
- ll_inode_size_unlock(inode, 0);
+ CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino,
+ inode->i_generation, inode, i_size_read(inode));
EXIT;
return;
-
- out_unlock:
- ll_inode_size_unlock(inode, 0);
} /* ll_truncate */
/**
cio = ccc_env_io(env);
io = cio->cui_cl.cis_io;
if (io == NULL && create) {
- struct vvp_io *vio;
- loff_t pos;
+ struct inode *inode = vmpage->mapping->host;
+ loff_t pos;
+
+ if (mutex_trylock(&inode->i_mutex)) {
+ mutex_unlock(&(inode)->i_mutex);
+
+ /* this is too bad. Someone is trying to write the
+ * page w/o holding inode mutex. This means we can
+ * add dirty pages into cache during truncate */
+ CERROR("Proc %s is dirting page w/o inode lock, this"
+ "will break truncate.\n", cfs_current()->comm);
+ libcfs_debug_dumpstack(NULL);
+ LBUG();
+ return ERR_PTR(-EIO);
+ }
/*
* Loop-back driver calls ->prepare_write() and ->sendfile()
* methods directly, bypassing file system ->write() operation,
* so cl_io has to be created here.
*/
-
- io = &ccc_env_info(env)->cti_io;
- vio = vvp_env_io(env);
+ io = ccc_env_thread_io(env);
ll_io_init(io, file, 1);
/* No lock at all for this kind of IO - we can't do it because
lcc = ERR_PTR(result);
}
- CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %i %p %p\n",
+ CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n",
vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result,
env, io);
return lcc;
struct lu_env *env;
struct cl_io *io;
struct cl_page *page;
- int result;
+ int result = 0;
ENTRY;
lcc = ll_cl_get();
io = lcc->lcc_io;
LASSERT(cl_page_is_owned(page, io));
- result = cl_io_commit_write(env, io, page, from, to);
+ LASSERT(from <= to);
+ if (from != to) /* handle short write case. */
+ result = cl_io_commit_write(env, io, page, from, to);
if (cl_page_is_owned(page, io))
cl_page_unassume(env, io, page);
+
/*
- * Release reference acquired by cl_io_prepare_write().
+ * Release reference acquired by ll_prepare_write().
*/
lu_ref_del(&page->cp_reference, "prepare_write", cfs_current());
cl_page_put(env, page);
static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
-/* WARNING: This algorithm is used to reduce the contention on
- * sbi->ll_lock. It should work well if the ra_max_pages is much
- * greater than the single file's read-ahead window.
+/**
+ * Get readahead pages from the filesystem readahead pool of the client for a
+ * thread.
*
- * TODO: There may exist a `global sync problem' in this implementation.
- * Considering the global ra window is 100M, and each file's ra window is 10M,
- * there are over 10 files trying to get its ra budget and reach
- * ll_ra_count_get at the exactly same time. All of them will get a zero ra
- * window, although the global window is 100M. -jay
- */
-static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
+ * /param sbi superblock for filesystem readahead state ll_ra_info
+ * /param ria per-thread readahead state
+ * /param pages number of pages requested for readahead for the thread.
+ *
+ * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
+ * It should work well if the ra_max_pages is much greater than the single
+ * file's read-ahead window, and not too many threads contending for
+ * these readahead pages.
+ *
+ * TODO: There may be a 'global sync problem' if many threads are trying
+ * to get an ra budget that is larger than the remaining readahead pages
+ * and reach here at exactly the same time. They will compute /a ret to
+ * consume the remaining pages, but will fail at atomic_add_return() and
+ * get a zero ra window, although there is still ra space remaining. - Jay */
+
+static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
+ struct ra_io_arg *ria,
+ unsigned long pages)
{
struct ll_ra_info *ra = &sbi->ll_ra_info;
- unsigned long ret;
+ long ret;
ENTRY;
- /**
- * If read-ahead pages left are less than 1M, do not do read-ahead,
+ /* If read-ahead pages left are less than 1M, do not do read-ahead,
* otherwise it will form small read RPC(< 1M), which hurt server
- * performance a lot.
- */
- ret = min(ra->ra_max_pages - cfs_atomic_read(&ra->ra_cur_pages), len);
- if ((int)ret < 0 || ret < min((unsigned long)PTLRPC_MAX_BRW_PAGES, len))
+ * performance a lot. */
+ ret = min(ra->ra_max_pages - cfs_atomic_read(&ra->ra_cur_pages), pages);
+ if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
GOTO(out, ret = 0);
+ /* If the non-strided (ria_pages == 0) readahead window
+ * (ria_start + ret) has grown across an RPC boundary, then trim
+ * readahead size by the amount beyond the RPC so it ends on an
+ * RPC boundary. If the readahead window is already ending on
+ * an RPC boundary (beyond_rpc == 0), or smaller than a full
+ * RPC (beyond_rpc < ret) the readahead size is unchanged.
+ * The (beyond_rpc != 0) check is skipped since the conditional
+ * branch is more expensive than subtracting zero from the result.
+ *
+ * Strided read is left unaligned to avoid small fragments beyond
+ * the RPC boundary from needing an extra read RPC. */
+ if (ria->ria_pages == 0) {
+ long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES;
+ if (/* beyond_rpc != 0 && */ beyond_rpc < ret)
+ ret -= beyond_rpc;
+ }
+
if (cfs_atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
cfs_atomic_sub(ret, &ra->ra_cur_pages);
ret = 0;
}
+
out:
RETURN(ret);
}
*/
static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
struct cl_page_list *queue,
- int index, struct address_space *mapping)
+ pgoff_t index, struct address_space *mapping)
{
struct page *vmpage;
struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
#ifdef __GFP_NOWARN
gfp_mask |= __GFP_NOWARN;
#endif
- vmpage = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
+ vmpage = grab_cache_page_nowait(mapping, index);
if (vmpage != NULL) {
/* Check if vmpage was truncated or reclaimed */
if (vmpage->mapping == mapping) {
stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
unsigned long off, unsigned long length)
{
- unsigned long start = off > st_off ? off - st_off : 0;
- unsigned long end = off + length > st_off ? off + length - st_off : 0;
+ __u64 start = off > st_off ? off - st_off : 0;
+ __u64 end = off + length > st_off ? off + length - st_off : 0;
unsigned long start_left = 0;
unsigned long end_left = 0;
unsigned long pg_count;
if (end_left > st_pgs)
end_left = st_pgs;
- CDEBUG(D_READA, "start %lu, end %lu start_left %lu end_left %lu \n",
+ CDEBUG(D_READA, "start "LPU64", end "LPU64" start_left %lu end_left %lu \n",
start, end, start_left, end_left);
if (start == end)
* For stride I/O mode, just check whether the idx is inside
* the ria_pages. */
return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
- (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages;
+ (idx >= ria->ria_stoff && (idx - ria->ria_stoff) %
+ ria->ria_length < ria->ria_pages);
}
static int ll_read_ahead_pages(const struct lu_env *env,
/* FIXME: This assertion only is valid when it is for
* forward read-ahead, it will be fixed when backward
* read-ahead is implemented */
- LASSERTF(page_idx > ria->ria_stoff, "since %lu in the"
- " gap of ra window,it should bigger than stride"
- " offset %lu \n", page_idx, ria->ria_stoff);
-
+ LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu"
+ "rs %lu re %lu ro %lu rl %lu rp %lu\n", page_idx,
+ ria->ria_start, ria->ria_end, ria->ria_stoff,
+ ria->ria_length, ria->ria_pages);
offset = page_idx - ria->ria_stoff;
offset = offset % (ria->ria_length);
if (offset > ria->ria_pages) {
end = ras->ras_window_start + ras->ras_window_len - 1;
}
if (end != 0) {
+ unsigned long rpc_boundary;
+ /*
+ * Align RA window to an optimal boundary.
+ *
+ * XXX This would be better to align to cl_max_pages_per_rpc
+ * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may
+ * be aligned to the RAID stripe size in the future and that
+ * is more important than the RPC size.
+ */
+ /* Note: we only trim the RPC, instead of extending the RPC
+ * to the boundary, so to avoid reading too much pages during
+ * random reading. */
+ rpc_boundary = ((end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)));
+ if (rpc_boundary > 0)
+ rpc_boundary--;
+
+ if (rpc_boundary > start)
+ end = rpc_boundary;
+
/* Truncate RA window to end of file */
end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
+
ras->ras_next_readahead = max(end, end + 1);
RAS_CDEBUG(ras);
}
if (len == 0)
RETURN(0);
- reserved = ll_ra_count_get(ll_i2sbi(inode), len);
-
+ reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len);
if (reserved < len)
ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
- CDEBUG(D_READA, "reserved page %lu \n", reserved);
+ CDEBUG(D_READA, "reserved page %lu ra_cur %d ra_max %lu\n", reserved,
+ cfs_atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
+ ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
ret = ll_read_ahead_pages(env, io, queue,
ria, &reserved, mapping, &ra_end);
int zero = 0, stride_detect = 0, ra_miss = 0;
ENTRY;
- cfs_spin_lock(&sbi->ll_lock);
cfs_spin_lock(&ras->ras_lock);
ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
ras->ras_consecutive_pages++;
ras->ras_last_readpage = index;
ras_set_start(ras, index);
- ras->ras_next_readahead = max(ras->ras_window_start,
- ras->ras_next_readahead);
+
+ if (stride_io_mode(ras))
+ /* Since stride readahead is sentivite to the offset
+ * of read-ahead, so we use original offset here,
+ * instead of ras_window_start, which is 1M aligned*/
+ ras->ras_next_readahead = max(index,
+ ras->ras_next_readahead);
+ else
+ ras->ras_next_readahead = max(ras->ras_window_start,
+ ras->ras_next_readahead);
RAS_CDEBUG(ras);
/* Trigger RA in the mmap case where ras_consecutive_requests
RAS_CDEBUG(ras);
ras->ras_request_index++;
cfs_spin_unlock(&ras->ras_lock);
- cfs_spin_unlock(&sbi->ll_lock);
return;
}
-int ll_writepage(struct page *vmpage, struct writeback_control *unused)
+int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
{
- struct inode *inode = vmpage->mapping->host;
+ struct inode *inode = vmpage->mapping->host;
+ struct ll_inode_info *lli = ll_i2info(inode);
struct lu_env *env;
struct cl_io *io;
struct cl_page *page;
struct cl_object *clob;
- struct cl_2queue *queue;
struct cl_env_nest nest;
+ bool redirtied = false;
+ bool unlocked = false;
int result;
ENTRY;
LASSERT(PageLocked(vmpage));
LASSERT(!PageWriteback(vmpage));
- if (ll_i2dtexp(inode) == NULL)
- RETURN(-EINVAL);
+ LASSERT(ll_i2dtexp(inode) != NULL);
- env = cl_env_nested_get(&nest);
- if (IS_ERR(env))
- RETURN(PTR_ERR(env));
+ env = cl_env_nested_get(&nest);
+ if (IS_ERR(env))
+ GOTO(out, result = PTR_ERR(env));
- io = &ccc_env_info(env)->cti_io;
- queue = &vvp_env_info(env)->vti_queue;
clob = ll_i2info(inode)->lli_clob;
LASSERT(clob != NULL);
+ io = ccc_env_thread_io(env);
io->ci_obj = clob;
+ io->ci_ignore_layout = 1;
result = cl_io_init(env, io, CIT_MISC, clob);
if (result == 0) {
page = cl_page_find(env, clob, vmpage->index,
lu_ref_add(&page->cp_reference, "writepage",
cfs_current());
cl_page_assume(env, io, page);
- /*
- * Mark page dirty, because this is what
- * ->vio_submit()->cpo_prep_write() assumes.
- *
- * XXX better solution is to detect this from within
- * cl_io_submit_rw() somehow.
- */
- set_page_dirty(vmpage);
- cl_2queue_init_page(queue, page);
- result = cl_io_submit_rw(env, io, CRT_WRITE,
- queue, CRP_NORMAL);
- cl_page_list_disown(env, io, &queue->c2_qin);
- if (result != 0) {
- /*
- * There is no need to clear PG_writeback, as
- * cl_io_submit_rw() calls completion callback
- * on failure.
- */
- /*
- * Re-dirty page on error so it retries write,
- * but not in case when IO has actually
- * occurred and completed with an error.
- */
- if (!PageError(vmpage))
- set_page_dirty(vmpage);
- }
- LASSERT(!cl_page_is_owned(page, io));
+ result = cl_page_flush(env, io, page);
+ if (result != 0) {
+ /*
+ * Re-dirty page on error so it retries write,
+ * but not in case when IO has actually
+ * occurred and completed with an error.
+ */
+ if (!PageError(vmpage)) {
+ redirty_page_for_writepage(wbc, vmpage);
+ result = 0;
+ redirtied = true;
+ }
+ }
+ cl_page_disown(env, io, page);
+ unlocked = true;
lu_ref_del(&page->cp_reference,
"writepage", cfs_current());
cl_page_put(env, page);
- cl_2queue_fini(env, queue);
- }
+ } else {
+ result = PTR_ERR(page);
+ }
}
cl_io_fini(env, io);
+
+ if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
+ loff_t offset = cl_offset(clob, vmpage->index);
+
+ /* Flush page failed because the extent is being written out.
+ * Wait for the write of extent to be finished to avoid
+ * breaking kernel which assumes ->writepage should mark
+ * PageWriteback or clean the page. */
+ result = cl_sync_file_range(inode, offset,
+ offset + CFS_PAGE_SIZE - 1,
+ CL_FSYNC_LOCAL);
+ if (result > 0) {
+ /* actually we may have written more than one page.
+ * decreasing this page because the caller will count
+ * it. */
+ wbc->nr_to_write -= result - 1;
+ result = 0;
+ }
+ }
+
cl_env_nested_put(&nest, env);
- RETURN(result);
+ GOTO(out, result);
+
+out:
+ if (result < 0) {
+ if (!lli->lli_async_rc)
+ lli->lli_async_rc = result;
+ SetPageError(vmpage);
+ if (!unlocked)
+ unlock_page(vmpage);
+ }
+ return result;
+}
+
+int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
+{
+ struct inode *inode = mapping->host;
+ loff_t start;
+ loff_t end;
+ enum cl_fsync_mode mode;
+ int range_whole = 0;
+ int result;
+ ENTRY;
+
+ if (wbc->range_cyclic) {
+ start = mapping->writeback_index << CFS_PAGE_SHIFT;
+ end = OBD_OBJECT_EOF;
+ } else {
+ start = wbc->range_start;
+ end = wbc->range_end;
+ if (end == LLONG_MAX) {
+ end = OBD_OBJECT_EOF;
+ range_whole = start == 0;
+ }
+ }
+
+ mode = CL_FSYNC_NONE;
+ if (wbc->sync_mode == WB_SYNC_ALL)
+ mode = CL_FSYNC_LOCAL;
+
+ result = cl_sync_file_range(inode, start, end, mode);
+ if (result > 0) {
+ wbc->nr_to_write -= result;
+ result = 0;
+ }
+
+ if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
+ if (end == OBD_OBJECT_EOF)
+ end = i_size_read(inode);
+ mapping->writeback_index = (end >> CFS_PAGE_SHIFT) + 1;
+ }
+ RETURN(result);
}
int ll_readpage(struct file *file, struct page *vmpage)
result = cl_io_read_page(env, io, page);
} else {
/* Page from a non-object file. */
- LASSERT(!ll_i2info(vmpage->mapping->host)->lli_smd);
+ LASSERT(!ll_i2info(vmpage->mapping->host)->lli_has_smd);
unlock_page(vmpage);
result = 0;
}
ll_cl_fini(lcc);
} else {
+ unlock_page(vmpage);
result = PTR_ERR(lcc);
}
RETURN(result);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff