From: James Simmons <jsimmons@infradead.org>
To: Andreas Dilger <adilger@whamcloud.com>,
Oleg Drokin <green@whamcloud.com>, NeilBrown <neilb@suse.de>
Cc: Lustre Development List <lustre-devel@lists.lustre.org>
Subject: [lustre-devel] [PATCH 01/22] lustre: llite: clear stale page's uptodate bit
Date: Sun, 20 Nov 2022 09:16:47 -0500 [thread overview]
Message-ID: <1668953828-10909-2-git-send-email-jsimmons@infradead.org> (raw)
In-Reply-To: <1668953828-10909-1-git-send-email-jsimmons@infradead.org>
From: Bobi Jam <bobijam@whamcloud.com>
With truncate_inode_page()->do_invalidatepage()->ll_invalidatepage()
call path before deleting vmpage from page cache, the page could be
possibly picked up by ll_read_ahead_page()->grab_cache_page_nowait().
If ll_invalidatepage()->cl_page_delete() does not clear the vmpage's
uptodate bit, the read ahead could pick it up and think it's already
uptodate wrongly.
In ll_fault()->vvp_io_fault_start()->vvp_io_kernel_fault(), the
filemap_fault() will call ll_readpage() to read vmpage and wait for
the unlock of the vmpage, and when ll_readpage() successfully read
the vmpage then unlock the vmpage, memory pressure or truncate can
get in and delete the cl_page, afterward filemap_fault() find that
the vmpage is not uptodate and VM_FAULT_SIGBUS got returned. To fix
this situation, this patch makes vvp_io_kernel_fault() restart
filemap_fault() to get uptodated vmpage again.
WC-bug-id: https://jira.whamcloud.com/browse/LU-16160
Lustre-commit: 5b911e03261c3de6b ("LU-16160 llite: clear stale page's uptodate bit")
Signed-off-by: Bobi Jam <bobijam@whamcloud.com>
Reviewed-on: https://review.whamcloud.com/c/fs/lustre-release/+/48607
Reviewed-by: Alex Zhuravlev <bzzz@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>
Signed-off-by: James Simmons <jsimmons@infradead.org>
---
fs/lustre/include/cl_object.h | 15 ++++-
fs/lustre/llite/rw.c | 10 +++-
fs/lustre/llite/vvp_io.c | 124 +++++++++++++++++++++++++++++++++++++++---
fs/lustre/llite/vvp_page.c | 5 ++
fs/lustre/obdclass/cl_page.c | 37 ++++++++++---
5 files changed, 172 insertions(+), 19 deletions(-)
diff --git a/fs/lustre/include/cl_object.h b/fs/lustre/include/cl_object.h
index 41ce0b0..8be58ff 100644
--- a/fs/lustre/include/cl_object.h
+++ b/fs/lustre/include/cl_object.h
@@ -768,7 +768,15 @@ struct cl_page {
enum cl_page_type cp_type:CP_TYPE_BITS;
unsigned int cp_defer_uptodate:1,
cp_ra_updated:1,
- cp_ra_used:1;
+ cp_ra_used:1,
+ /* fault page read grab extra referece */
+ cp_fault_ref:1,
+ /**
+ * if fault page got delete before returned to
+ * filemap_fault(), defer the vmpage detach/put
+ * until filemap_fault() has been handled.
+ */
+ cp_defer_detach:1;
/* which slab kmem index this memory allocated from */
short int cp_kmem_index;
@@ -2393,6 +2401,11 @@ int cl_io_lru_reserve(const struct lu_env *env, struct cl_io *io,
int cl_io_read_ahead(const struct lu_env *env, struct cl_io *io,
pgoff_t start, struct cl_read_ahead *ra);
+static inline int cl_io_is_pagefault(const struct cl_io *io)
+{
+ return io->ci_type == CIT_FAULT && !io->u.ci_fault.ft_mkwrite;
+}
+
/**
* True, if @io is an O_APPEND write(2).
*/
diff --git a/fs/lustre/llite/rw.c b/fs/lustre/llite/rw.c
index 2290b31..0283af4 100644
--- a/fs/lustre/llite/rw.c
+++ b/fs/lustre/llite/rw.c
@@ -1947,7 +1947,15 @@ int ll_readpage(struct file *file, struct page *vmpage)
unlock_page(vmpage);
result = 0;
}
- cl_page_put(env, page);
+ if (cl_io_is_pagefault(io) && result == 0) {
+ /**
+ * page fault, retain the cl_page reference until
+ * vvp_io_kernel_fault() release it.
+ */
+ page->cp_fault_ref = 1;
+ } else {
+ cl_page_put(env, page);
+ }
} else {
unlock_page(vmpage);
result = PTR_ERR(page);
diff --git a/fs/lustre/llite/vvp_io.c b/fs/lustre/llite/vvp_io.c
index ef7a3d92..be6f17f 100644
--- a/fs/lustre/llite/vvp_io.c
+++ b/fs/lustre/llite/vvp_io.c
@@ -1292,14 +1292,41 @@ static void vvp_io_rw_end(const struct lu_env *env,
trunc_sem_up_read(&lli->lli_trunc_sem);
}
-static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
+static void detach_and_deref_page(struct cl_page *clp, struct page *vmpage)
+{
+ if (!clp->cp_defer_detach)
+ return;
+
+ /**
+ * cl_page_delete0() took a vmpage reference, but not unlink the vmpage
+ * from its cl_page.
+ */
+ clp->cp_defer_detach = 0;
+ ClearPagePrivate(vmpage);
+ vmpage->private = 0;
+
+ put_page(vmpage);
+ refcount_dec(&clp->cp_ref);
+}
+
+static int vvp_io_kernel_fault(const struct lu_env *env,
+ struct vvp_fault_io *cfio)
{
struct vm_fault *vmf = cfio->ft_vmf;
+ struct file *vmff = cfio->ft_vma->vm_file;
+ struct address_space *mapping = vmff->f_mapping;
+ struct inode *inode = mapping->host;
+ struct page *vmpage = NULL;
+ struct cl_page *clp = NULL;
+ int rc = 0;
+ ll_inode_size_lock(inode);
+retry:
cfio->ft_flags = filemap_fault(vmf);
cfio->ft_flags_valid = 1;
if (vmf->page) {
+ /* success, vmpage is locked */
CDEBUG(D_PAGE,
"page %p map %p index %lu flags %lx count %u priv %0lx: got addr %p type NOPAGE\n",
vmf->page, vmf->page->mapping, vmf->page->index,
@@ -1311,24 +1338,105 @@ static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
}
cfio->ft_vmpage = vmf->page;
- return 0;
+
+ /**
+ * ll_filemap_fault()->ll_readpage() could get an extra cl_page
+ * reference. So we have to get the cl_page's to check its
+ * cp_fault_ref and drop the reference later.
+ */
+ clp = cl_vmpage_page(vmf->page, NULL);
+
+ goto unlock;
+ }
+
+ /* filemap_fault() fails, vmpage is not locked */
+ if (!clp) {
+ vmpage = find_get_page(mapping, vmf->pgoff);
+ if (vmpage) {
+ lock_page(vmpage);
+ clp = cl_vmpage_page(vmpage, NULL);
+ unlock_page(vmpage);
+ }
}
if (cfio->ft_flags & (VM_FAULT_SIGBUS | VM_FAULT_SIGSEGV)) {
+ pgoff_t max_idx;
+
+ /**
+ * ll_filemap_fault()->ll_readpage() could fill vmpage
+ * correctly, and unlock the vmpage, while memory pressure or
+ * truncate could detach cl_page from vmpage, and kernel
+ * filemap_fault() will wait_on_page_locked(vmpage) and find
+ * out that the vmpage has been cleared its uptodate bit,
+ * so it returns VM_FAULT_SIGBUS.
+ *
+ * In this case, we'd retry the filemap_fault()->ll_readpage()
+ * to rebuild the cl_page and fill vmpage with uptodated data.
+ */
+ if (likely(vmpage)) {
+ bool need_retry = false;
+
+ if (clp) {
+ if (clp->cp_defer_detach) {
+ detach_and_deref_page(clp, vmpage);
+ /**
+ * check i_size to make sure it's not
+ * over EOF, we don't want to call
+ * filemap_fault() repeatedly since it
+ * returns VM_FAULT_SIGBUS without even
+ * trying if vmf->pgoff is over EOF.
+ */
+ max_idx = DIV_ROUND_UP(i_size_read(inode),
+ PAGE_SIZE);
+ if (vmf->pgoff < max_idx)
+ need_retry = true;
+ }
+ if (clp->cp_fault_ref) {
+ clp->cp_fault_ref = 0;
+ /* ref not released in ll_readpage() */
+ cl_page_put(env, clp);
+ }
+ if (need_retry)
+ goto retry;
+ }
+ }
+
CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", (void *)vmf->address);
- return -EFAULT;
+ rc = -EFAULT;
+ goto unlock;
}
if (cfio->ft_flags & VM_FAULT_OOM) {
CDEBUG(D_PAGE, "got addr %p - OOM\n", (void *)vmf->address);
- return -ENOMEM;
+ rc = -ENOMEM;
+ goto unlock;
}
- if (cfio->ft_flags & VM_FAULT_RETRY)
- return -EAGAIN;
+ if (cfio->ft_flags & VM_FAULT_RETRY) {
+ rc = -EAGAIN;
+ goto unlock;
+ }
CERROR("Unknown error in page fault %d!\n", cfio->ft_flags);
- return -EINVAL;
+ rc = -EINVAL;
+unlock:
+ ll_inode_size_unlock(inode);
+ if (clp) {
+ if (clp->cp_defer_detach && vmpage)
+ detach_and_deref_page(clp, vmpage);
+
+ /* additional cl_page ref has been taken in ll_readpage() */
+ if (clp->cp_fault_ref) {
+ clp->cp_fault_ref = 0;
+ /* ref not released in ll_readpage() */
+ cl_page_put(env, clp);
+ }
+ /* ref taken in this function */
+ cl_page_put(env, clp);
+ }
+ if (vmpage)
+ put_page(vmpage);
+ return rc;
}
static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
@@ -1368,7 +1476,7 @@ static int vvp_io_fault_start(const struct lu_env *env,
LASSERT(cfio->ft_vmpage);
lock_page(cfio->ft_vmpage);
} else {
- result = vvp_io_kernel_fault(cfio);
+ result = vvp_io_kernel_fault(env, cfio);
if (result != 0)
return result;
}
diff --git a/fs/lustre/llite/vvp_page.c b/fs/lustre/llite/vvp_page.c
index f359596..9e8c158 100644
--- a/fs/lustre/llite/vvp_page.c
+++ b/fs/lustre/llite/vvp_page.c
@@ -104,6 +104,11 @@ static void vvp_page_completion_read(const struct lu_env *env,
ll_ra_count_put(ll_i2sbi(inode), 1);
if (ioret == 0) {
+ /**
+ * cp_defer_uptodate is used for readahead page, and the
+ * vmpage Uptodate bit is deferred to set in ll_readpage/
+ * ll_io_read_page.
+ */
if (!cp->cp_defer_uptodate)
SetPageUptodate(vmpage);
} else if (cp->cp_defer_uptodate) {
diff --git a/fs/lustre/obdclass/cl_page.c b/fs/lustre/obdclass/cl_page.c
index 7011235..3bc1a9b 100644
--- a/fs/lustre/obdclass/cl_page.c
+++ b/fs/lustre/obdclass/cl_page.c
@@ -725,16 +725,35 @@ static void __cl_page_delete(const struct lu_env *env, struct cl_page *cp)
LASSERT(PageLocked(vmpage));
LASSERT((struct cl_page *)vmpage->private == cp);
- /* Drop the reference count held in vvp_page_init */
- refcount_dec(&cp->cp_ref);
- ClearPagePrivate(vmpage);
- vmpage->private = 0;
-
- /*
- * The reference from vmpage to cl_page is removed,
- * but the reference back is still here. It is removed
- * later in cl_page_free().
+ /**
+ * clear vmpage uptodate bit, since ll_read_ahead_pages()->
+ * ll_read_ahead_page() could pick up this stale vmpage and
+ * take it as uptodated.
*/
+ ClearPageUptodate(vmpage);
+ /**
+ * vvp_io_kernel_fault()->ll_readpage() set cp_fault_ref
+ * and need it to check cl_page to retry the page fault read.
+ */
+ if (cp->cp_fault_ref) {
+ cp->cp_defer_detach = 1;
+ /**
+ * get a vmpage reference, so that filemap_fault()
+ * won't free it from pagecache.
+ */
+ get_page(vmpage);
+ } else {
+ /* Drop the reference count held in vvp_page_init */
+ refcount_dec(&cp->cp_ref);
+ ClearPagePrivate(vmpage);
+ vmpage->private = 0;
+
+ /*
+ * The reference from vmpage to cl_page is removed,
+ * but the reference back is still here. It is removed
+ * later in cl_page_free().
+ */
+ }
}
}
--
1.8.3.1
_______________________________________________
lustre-devel mailing list
lustre-devel@lists.lustre.org
http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org
next prev parent reply other threads:[~2022-11-20 14:17 UTC|newest]
Thread overview: 23+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-11-20 14:16 [lustre-devel] [PATCH 00/22] lustre: backport OpenSFS work as of Nov 20, 2022 James Simmons
2022-11-20 14:16 ` James Simmons [this message]
2022-11-20 14:16 ` [lustre-devel] [PATCH 02/22] lustre: osc: Remove oap lock James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 03/22] lnet: Don't modify uptodate peer with temp NI James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 04/22] lustre: llite: Explicitly support .splice_write James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 05/22] lnet: o2iblnd: add verbose debug prints for rx/tx events James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 06/22] lnet: use Netlink to support old and new NI APIs James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 07/22] lustre: obdclass: improve precision of wakeups for mod_rpcs James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 08/22] lnet: allow ping packet to contain large nids James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 09/22] lustre: llog: skip bad records in llog James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 10/22] lnet: fix build issue when IPv6 is disabled James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 11/22] lustre: obdclass: fill jobid in a safe way James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 12/22] lustre: llite: remove linefeed from LDLM_DEBUG James Simmons
2022-11-20 14:16 ` [lustre-devel] [PATCH 13/22] lnet: selftest: migrate LNet selftest session handling to Netlink James Simmons
2022-11-20 14:17 ` [lustre-devel] [PATCH 14/22] lustre: clio: append to non-existent component James Simmons
2022-11-20 14:17 ` [lustre-devel] [PATCH 15/22] lnet: fix debug message in lnet_discovery_event_reply James Simmons
2022-11-20 14:17 ` [lustre-devel] [PATCH 16/22] lustre: ldlm: group lock unlock fix James Simmons
2022-11-20 14:17 ` [lustre-devel] [PATCH 17/22] lnet: Signal completion on ping send failure James Simmons
2022-11-20 14:17 ` [lustre-devel] [PATCH 18/22] lnet: extend lnet_is_nid_in_ping_info() James Simmons
2022-11-20 14:17 ` [lustre-devel] [PATCH 19/22] lnet: find correct primary for peer James Simmons
2022-11-20 14:17 ` [lustre-devel] [PATCH 20/22] lnet: change lnet_notify() to take struct lnet_nid James Simmons
2022-11-20 14:17 ` [lustre-devel] [PATCH 21/22] lnet: discard lnet_nid2ni_*() James Simmons
2022-11-20 14:17 ` [lustre-devel] [PATCH 22/22] lnet: change lnet_debug_peer() to struct lnet_nid James Simmons
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=1668953828-10909-2-git-send-email-jsimmons@infradead.org \
--to=jsimmons@infradead.org \
--cc=adilger@whamcloud.com \
--cc=green@whamcloud.com \
--cc=lustre-devel@lists.lustre.org \
--cc=neilb@suse.de \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).