Save/restore epoll items during checkpoint/restart respectively.
kmalloc failures should be dealt with more kindly than just error-out
because epoll is made to poll many thousands of file descriptors.
Subsequent patches will change epoll c/r to "chunk" its output/input
respectively.
Signed-off-by: Matt Helsley <matthltc@xxxxxxxxxx>
Changelog:
v4: ckpt-v18
Use files_deferq as submitted by Dan Smith
Cleanup to only report >= 1 items when debugging.
v3: [unposted]
Removed most of the TODOs -- the remainder will be removed by
subsequent patches.
Fixed missing ep_file_collect() [Serge]
Rather than include checkpoint_hdr.h declare (but do not define)
the two structs needed in eventpoll.h [Oren]
Complain with ckpt_write_err() when we detect checkpoint obj
leaks. [Oren]
Remove redundant is_epoll_file() check in collect. [Oren]
Move epfile_objref lookup to simplify error handling. [Oren]
Simplify error handling with early return in
ep_eventpoll_checkpoint(). [Oren]
Cleaned up a comment. [Oren]
Shorten CKPT_HDR_FILE_EPOLL_ITEMS (-FILE) [Oren]
Renumbered to indicate that it follows the file table.
Renamed the epoll struct in checkpoint_hdr.h [Oren]
Also renamed substruct.
Fixup return of empty ep_file_restore(). [Oren]
Changed some error returns. [Oren]
Changed some tests to BUG_ON(). [Oren]
Factored out watch insert with epoll_ctl() into do_epoll_ctl().
[Cedric, Oren]
---
checkpoint/files.c | 21 +++-
checkpoint/restart.c | 2 +-
checkpoint/sys.c | 1 -
fs/eventpoll.c | 310 ++++++++++++++++++++++++++++++++++++----
include/linux/checkpoint.h | 1 +
include/linux/checkpoint_hdr.h | 14 ++
include/linux/eventpoll.h | 17 ++-
7 files changed, 331 insertions(+), 35 deletions(-)
diff --git a/checkpoint/files.c b/checkpoint/files.c
index eac5f3b..0c9bba2 100644
--- a/checkpoint/files.c
+++ b/checkpoint/files.c
@@ -22,6 +22,8 @@
#include <linux/deferqueue.h>
#include <linux/checkpoint.h>
#include <linux/checkpoint_hdr.h>
+#include <linux/deferqueue.h>
+#include <linux/eventpoll.h>
#include <net/sock.h>
@@ -311,9 +313,11 @@ static int do_checkpoint_file_table(struct ckpt_ctx *ctx,
}
ret = deferqueue_run(ctx->files_deferq);
- ckpt_debug("files_deferq ran %d entries\n", ret);
- if (ret > 0)
+ if (ret > 0) {
+ ckpt_debug("file checkpoint deferred %d work items\n", ret);
ret = 0;
+ }
+
out:
kfree(fdtable);
return ret;
@@ -604,6 +608,13 @@ static struct restore_file_ops restore_file_ops[] = {
.file_type = CKPT_FILE_TTY,
.restore = tty_file_restore,
},
+#ifdef CONFIG_EPOLL
+ {
+ .file_name = "EPOLL",
+ .file_type = CKPT_FILE_EPOLL,
+ .restore = ep_file_restore,
+ },
+#endif
};
static struct file *do_restore_file(struct ckpt_ctx *ctx)
@@ -731,9 +742,11 @@ static struct files_struct *do_restore_file_table(struct ckpt_ctx *ctx)
}
ret = deferqueue_run(ctx->files_deferq);
- ckpt_debug("files_deferq ran %d entries\n", ret);
- if (ret > 0)
+ if (ret > 0) {
+ ckpt_debug("file restore deferred %d work items\n", ret);
ret = 0;
+ }
+
out:
ckpt_hdr_put(ctx, h);
if (!ret) {
diff --git a/checkpoint/restart.c b/checkpoint/restart.c
index 543b380..61b4921 100644
--- a/checkpoint/restart.c
+++ b/checkpoint/restart.c
@@ -193,7 +193,7 @@ int _ckpt_read_string(struct ckpt_ctx *ctx, void *ptr, int len)
*
* Return: new buffer allocated on success, error pointer otherwise
*/
-static void *ckpt_read_obj(struct ckpt_ctx *ctx, int len, int max)
+void *ckpt_read_obj(struct ckpt_ctx *ctx, int len, int max)
{
struct ckpt_hdr hh;
struct ckpt_hdr *h;
diff --git a/checkpoint/sys.c b/checkpoint/sys.c
index 76a3fa9..b8be421 100644
--- a/checkpoint/sys.c
+++ b/checkpoint/sys.c
@@ -251,7 +251,6 @@ static struct ckpt_ctx *ckpt_ctx_alloc(int fd, unsigned long uflags,
ctx->deferqueue = deferqueue_create();
if (!ctx->deferqueue)
goto err;
-
ctx->files_deferq = deferqueue_create();
if (!ctx->files_deferq)
goto err;
diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index 085c5c0..cf3f309 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -39,6 +39,12 @@
#include <asm/mman.h>
#include <asm/atomic.h>
+#ifdef CONFIG_CHECKPOINT
+#include <linux/checkpoint.h>
+#include <linux/checkpoint_hdr.h>
+#include <linux/deferqueue.h>
+#endif
+
/*
* LOCKING:
* There are three level of locking required by epoll :
@@ -671,10 +677,20 @@ static unsigned int ep_eventpoll_poll(struct file *file, poll_table *wait)
return pollflags != -1 ? pollflags : 0;
}
+#ifdef CONFIG_CHECKPOINT
+static int ep_eventpoll_checkpoint(struct ckpt_ctx *ctx, struct file *file);
+static int ep_file_collect(struct ckpt_ctx *ctx, struct file *file);
+#else
+#define ep_eventpoll_checkpoint NULL
+#define ep_file_collect NULL
+#endif
+
/* File callbacks that implement the eventpoll file behaviour */
static const struct file_operations eventpoll_fops = {
.release = ep_eventpoll_release,
- .poll = ep_eventpoll_poll
+ .poll = ep_eventpoll_poll,
+ .checkpoint = ep_eventpoll_checkpoint,
+ .collect = ep_file_collect,
};
/* Fast test to see if the file is an evenpoll file */
@@ -1226,35 +1242,18 @@ SYSCALL_DEFINE1(epoll_create, int, size)
* the eventpoll file that enables the insertion/removal/change of
* file descriptors inside the interest set.
*/
-SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
- struct epoll_event __user *, event)
+int do_epoll_ctl(int op, int fd,
+ struct file *file, struct file *tfile,
+ struct epoll_event *epds)
{
int error;
- struct file *file, *tfile;
struct eventpoll *ep;
struct epitem *epi;
- struct epoll_event epds;
-
- error = -EFAULT;
- if (ep_op_has_event(op) &&
- copy_from_user(&epds, event, sizeof(struct epoll_event)))
- goto error_return;
-
- /* Get the "struct file *" for the eventpoll file */
- error = -EBADF;
- file = fget(epfd);
- if (!file)
- goto error_return;
-
- /* Get the "struct file *" for the target file */
- tfile = fget(fd);
- if (!tfile)
- goto error_fput;
/* The target file descriptor must support poll */
error = -EPERM;
if (!tfile->f_op || !tfile->f_op->poll)
- goto error_tgt_fput;
+ return error;
/*
* We have to check that the file structure underneath the file descriptor
@@ -1263,7 +1262,7 @@ SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
*/
error = -EINVAL;
if (file == tfile || !is_file_epoll(file))
- goto error_tgt_fput;
+ return error;
/*
* At this point it is safe to assume that the "private_data" contains
@@ -1284,8 +1283,8 @@ SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
switch (op) {
case EPOLL_CTL_ADD:
if (!epi) {
- epds.events |= POLLERR | POLLHUP;
- error = ep_insert(ep, &epds, tfile, fd);
+ epds->events |= POLLERR | POLLHUP;
+ error = ep_insert(ep, epds, tfile, fd);
} else
error = -EEXIST;
break;
@@ -1297,15 +1296,46 @@ SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
break;
case EPOLL_CTL_MOD:
if (epi) {
- epds.events |= POLLERR | POLLHUP;
- error = ep_modify(ep, epi, &epds);
+ epds->events |= POLLERR | POLLHUP;
+ error = ep_modify(ep, epi, epds);
} else
error = -ENOENT;
break;
}
mutex_unlock(&ep->mtx);
-error_tgt_fput:
+ return error;
+}
+
+/*
+ * The following function implements the controller interface for
+ * the eventpoll file that enables the insertion/removal/change of
+ * file descriptors inside the interest set.
+ */
+SYSCALL_DEFINE4(epoll_ctl, int, epfd, int, op, int, fd,
+ struct epoll_event __user *, event)
+{
+ int error;
+ struct file *file, *tfile;
+ struct epoll_event epds;
+
+ error = -EFAULT;
+ if (ep_op_has_event(op) &&
+ copy_from_user(&epds, event, sizeof(struct epoll_event)))
+ goto error_return;
+
+ /* Get the "struct file *" for the eventpoll file */
+ error = -EBADF;
+ file = fget(epfd);
+ if (!file)
+ goto error_return;
+
+ /* Get the "struct file *" for the target file */
+ tfile = fget(fd);
+ if (!tfile)
+ goto error_fput;
+
+ error = do_epoll_ctl(op, fd, file, tfile, &epds);
fput(tfile);
error_fput:
fput(file);
@@ -1413,6 +1443,230 @@ SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
#endif /* HAVE_SET_RESTORE_SIGMASK */
+#ifdef CONFIG_CHECKPOINT
+static int ep_file_collect(struct ckpt_ctx *ctx, struct file *file)
+{
+ struct rb_node *rbp;
+ struct eventpoll *ep;
+ int ret = 0;
+
+ ep = file->private_data;
+ mutex_lock(&ep->mtx);
+ for (rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp)) {
+ struct epitem *epi;
+
+ epi = rb_entry(rbp, struct epitem, rbn);
+ ret = ckpt_obj_collect(ctx, epi->ffd.file, CKPT_OBJ_FILE);
+ if (ret < 0)
+ break;
+ }
+ mutex_unlock(&ep->mtx);
+ return ret;
+}
+
+struct epoll_deferq_entry {
+ struct ckpt_ctx *ctx;
+ struct file *epfile;
+};
+
+static int ep_items_checkpoint(void *data)
+{
+ struct epoll_deferq_entry *ep_dq_entry = data;
+ struct ckpt_ctx *ctx;
+ struct file *file;
+ struct ckpt_hdr_eventpoll_items *h;
+ struct rb_node *rbp;
+ struct eventpoll *ep;
+ __s32 epfile_objref;
+ int i, ret;
+
+ file = ep_dq_entry->epfile;
+ ctx = ep_dq_entry->ctx;
+
+ epfile_objref = ckpt_obj_lookup(ctx, file, CKPT_OBJ_FILE);
+ BUG_ON(epfile_objref <= 0);
+
+
+ ep = file->private_data;
+ mutex_lock(&ep->mtx);
+ for (i = 0, rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp), i++) {}
+ mutex_unlock(&ep->mtx);
+
+ h = ckpt_hdr_get_type(ctx, sizeof(*h) + i*sizeof(h->items[0]),
+ CKPT_HDR_EPOLL_ITEMS);
+ if (!h)
+ return -ENOMEM;
+
+ h->num_items = i;
+ h->epfile_objref = epfile_objref;
+
+ ret = 0;
+ mutex_lock(&ep->mtx);
+ for (i = 0, rbp = rb_first(&ep->rbr); rbp; rbp = rb_next(rbp), i++) {
+ struct epitem *epi;
+ int objref;
+
+ epi = rb_entry(rbp, struct epitem, rbn);
+ objref = ckpt_obj_lookup(ctx, epi->ffd.file, CKPT_OBJ_FILE);
+ if (objref <= 0) {
+ ret = -EBUSY; /* checkpoint obj leak */
+ break;
+ }
+ h->items[i].fd = epi->ffd.fd;
+ h->items[i].file_objref = objref;
+ h->items[i].events = epi->event.events;
+ h->items[i].data = epi->event.data;
+ }
+ mutex_unlock(&ep->mtx);
+ if (!ret && (i != h->num_items))
+ /*
+ * We raced with another thread between our first and second
+ * walks of the rbtree such that there weren't the same number
+ * of items. This means there is a checkpoint "leak".
+ */
+ ret = -EBUSY;
+ if (ret == -EBUSY)
+ ckpt_write_err(ctx, "ep_items_checkpoint(): checkpoint leak detected.\n", "");
+ else if (!ret)
+ ret = ckpt_write_obj(ctx, &h->h);
+ ckpt_hdr_put(ctx, &h->h);
+ return ret;
+}
+
+static int ep_eventpoll_checkpoint(struct ckpt_ctx *ctx, struct file *file)
+{
+ struct ckpt_hdr_file *h;
+ struct epoll_deferq_entry ep_dq_entry;
+ int ret = -ENOMEM;
+
+ h = ckpt_hdr_get_type(ctx, sizeof(*h), CKPT_HDR_FILE);
+ if (!h)
+ return -ENOMEM;
+ h->f_type = CKPT_FILE_EPOLL;
+ ret = checkpoint_file_common(ctx, file, h);
+ if (ret < 0)
+ goto out;
+ ret = ckpt_write_obj(ctx, &h->h);
+ if (ret < 0)
+ goto out;
+
+ /*
+ * Defer saving the epoll items until all of the ffd.file pointers
+ * have an objref; after the file table has been checkpointed.
+ */
+ ep_dq_entry.ctx = ctx;
+ ep_dq_entry.epfile = file;
+ ret = deferqueue_add(ctx->files_deferq, &ep_dq_entry,
+ sizeof(ep_dq_entry), ep_items_checkpoint, NULL);
+out:
+ ckpt_hdr_put(ctx, h);
+ return ret;
+}
+
+static int ep_items_restore(void *data)
+{
+ struct ckpt_ctx *ctx = deferqueue_data_ptr(data);
+ struct ckpt_hdr_eventpoll_items *h;
+ struct eventpoll *ep;
+ struct file *epfile = NULL;
+ int ret, i = 0, remaining_watches;
+
+ h = ckpt_read_obj(ctx, 0,
+ sizeof(*h) + max_user_watches*sizeof(h->items[0]));
+ if (IS_ERR(h))
+ return PTR_ERR(h);
+
+ ret = -EINVAL;
+ if ((h->h.type != CKPT_HDR_EPOLL_ITEMS) ||
+ (h->h.len < sizeof(*h)))
+ goto out;
+
+ /* Make sure the items match the size we expect */
+ if (h->num_items != ((h->h.len - sizeof(*h)) / sizeof(h->items[0])))
+ goto out;
+
+ epfile = ckpt_obj_fetch(ctx, h->epfile_objref, CKPT_OBJ_FILE);
+ BUG_ON(IS_ERR(epfile));
+ BUG_ON(!is_file_epoll(epfile));
+
+ /* Make sure there are enough watches left. */
+ ret = -ENOSPC;
+ ep = epfile->private_data;
+ remaining_watches = (max_user_watches -
+ atomic_read(&ep->user->epoll_watches));
+ if (h->num_items > remaining_watches)
+ goto out;
+
+ ret = 0;
+ /* Restore the epoll items/watches */
+ for (i = 0; !ret && i < h->num_items; i++) {
+ struct epoll_event epev;
+ struct file *tfile;
+
+ /* Get the file* for the target file */
+ if (h->items[i].file_objref <= 0) {
+ ret = -EINVAL;
+ break;
+ }
+ tfile = ckpt_obj_fetch(ctx, h->items[i].file_objref,
+ CKPT_OBJ_FILE);
+ if (IS_ERR(tfile)) {
+ ret = PTR_ERR(tfile);
+ break;
+ }
+
+ epev.events = h->items[i].events;
+ epev.data = h->items[i].data;
+
+ ret = do_epoll_ctl(EPOLL_CTL_ADD, h->items[i].fd,
+ epfile, tfile, &epev);
+ }
+out:
+ ckpt_hdr_put(ctx, h);
+ return ret;
+}
+
+struct file* ep_file_restore(struct ckpt_ctx *ctx,
+ struct ckpt_hdr_file *h)
+{
+ struct file *epfile;
+ int epfd, ret;
+
+ if (h->h.type != CKPT_HDR_FILE ||
+ h->h.len != sizeof(*h) ||
+ h->f_type != CKPT_FILE_EPOLL)
+ return ERR_PTR(-EINVAL);
+
+ epfd = sys_epoll_create1(h->f_flags & EPOLL_CLOEXEC);
+ if (epfd < 0)
+ return ERR_PTR(epfd);
+ epfile = fget(epfd);
+ BUG_ON(!epfile);
+
+ /*
+ * Needed before we can properly restore the watches and enforce the
+ * limit on watch numbers.
+ */
+ ret = restore_file_common(ctx, epfile, h);
+ if (ret < 0)
+ goto fput_out;
+
+ /*
+ * Defer restoring the epoll items until the file table is
+ * fully restored. Ensures that valid file objrefs will resolve.
+ */
+ ret = deferqueue_add_ptr(ctx->files_deferq, ctx, ep_items_restore, NULL);
+ if (ret < 0) {
+fput_out:
+ fput(epfile);
+ epfile = ERR_PTR(ret);
+ }
+ sys_close(epfd); /* harmless even if an error occured */
+ return epfile;
+}
+
+#endif /* CONFIG_CHECKPOINT */
+
static int __init eventpoll_init(void)
{
struct sysinfo si;
diff --git a/include/linux/checkpoint.h b/include/linux/checkpoint.h
index e00dd70..a8594cc 100644
--- a/include/linux/checkpoint.h
+++ b/include/linux/checkpoint.h
@@ -72,6 +72,7 @@ extern int _ckpt_read_obj_type(struct ckpt_ctx *ctx,
void *ptr, int len, int type);
extern int _ckpt_read_buffer(struct ckpt_ctx *ctx, void *ptr, int len);
extern int _ckpt_read_string(struct ckpt_ctx *ctx, void *ptr, int len);
+extern void *ckpt_read_obj(struct ckpt_ctx *ctx, int len, int max);
extern void *ckpt_read_obj_type(struct ckpt_ctx *ctx, int len, int type);
extern void *ckpt_read_buf_type(struct ckpt_ctx *ctx, int len, int type);
extern int ckpt_read_payload(struct ckpt_ctx *ctx,
diff --git a/include/linux/checkpoint_hdr.h b/include/linux/checkpoint_hdr.h
index 2ed523f..48736bd 100644
--- a/include/linux/checkpoint_hdr.h
+++ b/include/linux/checkpoint_hdr.h
@@ -85,6 +85,7 @@ enum {
CKPT_HDR_PIPE_BUF,
CKPT_HDR_TTY,
CKPT_HDR_TTY_LDISC,
+ CKPT_HDR_EPOLL_ITEMS = 391, /* Follows file-table */
CKPT_HDR_MM = 401,
CKPT_HDR_VMA,
@@ -380,6 +381,7 @@ enum file_type {
CKPT_FILE_FIFO,
CKPT_FILE_SOCKET,
CKPT_FILE_TTY,
+ CKPT_FILE_EPOLL,
CKPT_FILE_MAX
};
@@ -475,6 +477,18 @@ struct ckpt_hdr_file_socket {
__s32 sock_objref;
} __attribute__((aligned(8)));
+struct ckpt_hdr_eventpoll_items {
+ struct ckpt_hdr h;
+ __s32 epfile_objref;
+ __u32 num_items;
+ struct ckpt_eventpoll_item {
+ __u64 data;
+ __u32 fd;
+ __s32 file_objref;
+ __u32 events;
+ } items[0];
+} __attribute__((aligned(8)));
+
/* memory layout */
struct ckpt_hdr_mm {
struct ckpt_hdr h;
diff --git a/include/linux/eventpoll.h b/include/linux/eventpoll.h
index f6856a5..34538be 100644
--- a/include/linux/eventpoll.h
+++ b/include/linux/eventpoll.h
@@ -56,6 +56,9 @@ struct file;
#ifdef CONFIG_EPOLL
+struct ckpt_ctx;
+struct ckpt_hdr_file;
+
/* Used to initialize the epoll bits inside the "struct file" */
static inline void eventpoll_init_file(struct file *file)
@@ -95,11 +98,23 @@ static inline void eventpoll_release(struct file *file)
eventpoll_release_file(file);
}
-#else
+#ifdef CONFIG_CHECKPOINT
+extern struct file* ep_file_restore(struct ckpt_ctx *ctx,
+ struct ckpt_hdr_file *h);
+#endif
+#else
+/* !defined(CONFIG_EPOLL) */
static inline void eventpoll_init_file(struct file *file) {}
static inline void eventpoll_release(struct file *file) {}
+#ifdef CONFIG_CHECKPOINT
+static inline struct file* ep_file_restore(struct ckpt_ctx *ctx,
+ struct ckpt_hdr_file *ptr)
+{
+ return ERR_PTR(-ENOSYS);
+}
+#endif
#endif
#endif /* #ifdef __KERNEL__ */
--
1.5.6.3
_______________________________________________
Containers mailing list
Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linux-foundation.org/mailman/listinfo/containers
