netns/mbuf: add a per net namespace ring buffer

add a log per netns log ring buffer which's reading
side interface is:
/proc/net/twatcher/log

this is backport from tk3, do some clean-code tasks

Reviewed-by: kernelxing <kernelxing@tencent.com>
Signed-off-by: MengEn Sun <mengensun@tencent.com>

fs/proc/proc_net.c

@@ -63,6 +63,62 @@ static int seq_open_net(struct inode *inode, struct file *file)
 	return 0;
 }
 
+#ifdef CONFIG_NETNS_MBUF
+/* token from seq_open_net, all is same except the private is
+ * alloc by vmalloc, why?
+ *
+ * sameone may need a big private, wasting continuous phy mem
+ * they can use this function to use vmalloc private
+ *
+ * from now if you using this open abi place write a write
+ * fops like proc_simple_write we delete the pde->write check
+ */
+void *seq_open_net_large_private(struct inode *inode, struct file *file)
+{
+	struct net *net;
+	struct seq_file *seq;
+	struct seq_net_private *p;
+	int ret;
+
+	unsigned int state_size = PDE(inode)->state_size;
+
+	WARN_ON_ONCE(state_size < sizeof(struct seq_net_private));
+
+	net = get_proc_net(inode);
+	if (!net) {
+		ret = -ENXIO;
+		goto out;
+	}
+
+	p = vmalloc(state_size);
+	if (!p) {
+		ret = -ENOMEM;
+		goto put_out;
+	}
+	memset(p, 0, state_size);
+
+	ret = seq_open(file, PDE(inode)->seq_ops);
+	if (ret < 0)
+		goto free_out;
+
+	seq = file->private_data;
+	seq->private = (void *)p;
+
+#ifdef CONFIG_NET_NS
+	p->net = net;
+#endif
+	return p;
+
+free_out:
+	vfree(p);
+put_out:
+	put_net(net);
+out:
+	return ERR_PTR(ret);
+}
+EXPORT_SYMBOL(seq_open_net_large_private);
+#endif
+
 static void seq_file_net_put_net(struct seq_file *seq)
 {
 #ifdef CONFIG_NET_NS
@@ -83,6 +139,31 @@ static int seq_release_net(struct inode *ino, struct file *f)
 	return 0;
 }
 
+#ifdef CONFIG_NETNS_MBUF
+/* add a ext-abi to allow someone define the fops by themself, this is all
+ * alike proc_create_net_data except has a extra f_ops parameter
+ */
+struct proc_dir_entry *proc_create_net_data_ops(const char *name, umode_t mode,
+						struct proc_dir_entry *parent,
+						const struct seq_operations *seq_ops,
+						unsigned int state_size, void *data,
+						const struct proc_ops *proc_ops)
+{
+	struct proc_dir_entry *p;
+
+	p = proc_create_reg(name, mode, &parent, data);
+	if (!p)
+		return NULL;
+
+	pde_force_lookup(p);
+	p->proc_ops = proc_ops;
+	p->seq_ops = seq_ops;
+	p->state_size = state_size;
+	return proc_register(parent, p);
+}
+EXPORT_SYMBOL_GPL(proc_create_net_data_ops);
+#endif
+
 static const struct proc_ops proc_net_seq_ops = {
 	.proc_open	= seq_open_net,
 	.proc_read	= seq_read,

include/linux/mbuf.h

@@ -49,7 +49,7 @@ struct mbuf_slot {
 	seqlock_t slot_lock;
 	/* rate limit */
 	struct ratelimit_state ratelimit;
-	struct cgroup *owner;
+	void *owner;
 	const struct mbuf_operations *ops;
 	struct mbuf_ring *mring;
 };
@@ -62,7 +62,7 @@ struct mbuf_operations {
 	u32 (*next)(struct mbuf_ring *mring, u32 idx);
 
 	/* write message */
-	ssize_t (*write)(struct cgroup *cg, const char *fmt, va_list);
+	ssize_t (*write)(struct mbuf_slot *mbuf, const char *fmt, va_list args);
 } ____cacheline_aligned;
 
 
@@ -70,9 +70,13 @@ void __init mbuf_bmap_init(void);
 void __init setup_mbuf(void);
 
 struct mbuf_slot *mbuf_slot_alloc(struct cgroup *cg);
+struct mbuf_slot *mbuf_slot_alloc_v2(void *owner, struct mbuf_operations *ops);
 void mbuf_free(struct cgroup *cg);
+
 ssize_t mbuf_print(struct cgroup *cgrp, const char *fmt, ...);
 void snapshot_mbuf(struct mbuf_slot *, struct mbuf_slot*, seqlock_t *);
 u32 get_mbuf_slot_len(void);
+void mbuf_free_slot(struct mbuf_slot *slot);
+void mbuf_reset(struct mbuf_slot *mbuf);
 #endif
 #endif

include/linux/proc_fs.h

@@ -161,6 +161,15 @@ int proc_pid_arch_status(struct seq_file *m, struct pid_namespace *ns,
 void arch_report_meminfo(struct seq_file *m);
 void arch_proc_pid_thread_features(struct seq_file *m, struct task_struct *task);
 
+#ifdef CONFIG_NETNS_MBUF
+void *seq_open_net_large_private(struct inode *inode, struct file *file);
+struct proc_dir_entry *proc_create_net_data_ops(const char *name, umode_t mode,
+					       struct proc_dir_entry *parent,
+					       const struct seq_operations *seq_ops,
+					       unsigned int state_size, void *data,
+					       const struct proc_ops *proc_ops);
+#endif
+
 #else /* CONFIG_PROC_FS */
 
 static inline void proc_root_init(void)

include/net/net_namespace.h

@@ -43,7 +43,9 @@
 #include <linux/skbuff.h>
 #include <linux/notifier.h>
 #include <linux/xarray.h>
-
+#ifdef CONFIG_NETNS_MBUF
+#include <net/netns_mbuf.h>
+#endif
 struct user_namespace;
 struct proc_dir_entry;
 struct net_device;
@@ -190,6 +192,9 @@ struct net {
 #if IS_ENABLED(CONFIG_SMC)
 	struct netns_smc	smc;
 #endif
+#ifdef CONFIG_NETNS_MBUF
+	struct net_mbuf		mbuf;
+#endif
 } __randomize_layout;
 
 #include <linux/seq_file_net.h>

include/net/netns_mbuf.h

@@ -0,0 +1,29 @@
+/* SPDX-License-Identifier: GPL-2.0-only
+ *
+ * make mbuf can be used by net namespace
+ *
+ * Author: mengensun <mengensun@tencent.com>
+ * Copyright (C) 2024 Tencent, Inc
+ */
+#ifndef __NETNS_MBUF
+#define __NETNS_MBUF
+
+#include<linux/proc_fs.h>
+#include<linux/mbuf.h>
+
+#ifdef CONFIG_NETNS_MBUF
+struct net_mbuf {
+	struct proc_dir_entry	*twatcher;
+	struct proc_dir_entry	*log;
+	struct mbuf_slot	*slot;
+};
+
+int inet_mbuf_init(void);
+void inet_mbuf_exit(void);
+ssize_t net_mbuf_print(struct net *net, const char *fmt, ...);
+#else
+static __always_inline int inet_mbuf_init(void) {return 0; }
+static __always_inline void inet_mbuf_exit(void) {}
+static __always_inline ssize_t net_mbuf_print(struct net *net, const char *fmt, ...) {return 0; };
+#endif
+#endif

kernel/cgroup/mbuf.c

@@ -252,7 +252,7 @@ static int mbuf_prepare(struct mbuf_ring *mring, u32 msg_size)
 }
 
 /* Write monitor buffer message */
-static ssize_t do_mbuf_write(struct cgroup *cg, char *buffer, size_t size)
+static ssize_t do_mbuf_write(struct mbuf_slot *mbuf, char *buffer, size_t size)
 {
 	struct mbuf_ring *mring;
 	struct mbuf_ring_desc *desc;
@@ -265,13 +265,13 @@ static ssize_t do_mbuf_write(struct cgroup *cg, char *buffer, size_t size)
 		return 0;
 	}
 
-	mring = cg->mbuf->mring;
+	mring = mbuf->mring;
 	len = sizeof(struct mbuf_ring_desc) + size;
 
-	write_seqlock_irqsave(&cg->mbuf->slot_lock, flags);
+	write_seqlock_irqsave(&mbuf->slot_lock, flags);
 
 	if (mbuf_prepare(mring, len)) {
-		write_sequnlock_irqrestore(&cg->mbuf->slot_lock, flags);
+		write_sequnlock_irqrestore(&mbuf->slot_lock, flags);
 		pr_err("mbuf: Can not find enough space.\n");
 		return 0;
 	}
@@ -290,20 +290,23 @@ static ssize_t do_mbuf_write(struct cgroup *cg, char *buffer, size_t size)
 	mring->next_idx += desc->len;
 	mring->next_seq++;
 
-	write_sequnlock_irqrestore(&cg->mbuf->slot_lock, flags);
+	write_sequnlock_irqrestore(&mbuf->slot_lock, flags);
 
 	return size;
 }
 
-void mbuf_reset(struct mbuf_ring *mring)
+void mbuf_reset(struct mbuf_slot *mbuf)
 {
-	mring->first_idx = mring->base_idx;
-	mring->first_seq = 0;
-	mring->next_idx = mring->base_idx;
-	mring->next_seq = 0;
+	write_seqlock(&mbuf->slot_lock);
+	mbuf->mring->first_idx = mbuf->mring->base_idx;
+	mbuf->mring->first_seq = 0;
+	mbuf->mring->next_idx = mbuf->mring->base_idx;
+	mbuf->mring->next_seq = 0;
+	write_sequnlock(&mbuf->slot_lock);
 }
+EXPORT_SYMBOL(mbuf_reset);
 
-static ssize_t mbuf_write(struct cgroup *cg, const char *fmt, va_list args)
+static ssize_t mbuf_write(struct mbuf_slot *mbuf, const char *fmt, va_list args)
 {
 	static char buf[MBUF_MSG_LEN_MAX];
 	char *text = buf;
@@ -313,7 +316,7 @@ static ssize_t mbuf_write(struct cgroup *cg, const char *fmt, va_list args)
 	t_len = vscnprintf(text, sizeof(buf), fmt, args);
 
 	/* Write string to mbuf */
-	ret = do_mbuf_write(cg, text, t_len);
+	ret = do_mbuf_write(mbuf, text, t_len);
 
 	return ret;
 }
@@ -335,11 +338,17 @@ static int get_next_mbuf_id(unsigned long *addr, u32 start)
 	return index;
 }
 
-static void mbuf_slot_init(struct mbuf_slot *mb, struct cgroup *cg, u32 index)
+static void mbuf_slot_init(struct mbuf_slot *mb,
+			   void *owner, u32 index, struct mbuf_operations *ops)
 {
-	mb->owner = cg;
+	mb->owner = owner;
 	mb->idx = index;
+
+	if (!ops)
 		mb->ops = &mbuf_ops;
+	else
+		mb->ops = ops;
+
 	seqlock_init(&mb->slot_lock);
 	ratelimit_state_init(&mb->ratelimit, 5 * HZ, 50);
 
@@ -349,10 +358,10 @@ static void mbuf_slot_init(struct mbuf_slot *mb, struct cgroup *cg, u32 index)
 				+ sizeof(struct mbuf_ring);
 	mb->mring->end_idx = (index + 1) * g_mbuf.mbuf_size_per_cg - 1;
 
-	mbuf_reset(mb->mring);
+	mbuf_reset(mb);
 }
 
-struct mbuf_slot *mbuf_slot_alloc(struct cgroup *cg)
+struct mbuf_slot *mbuf_slot_alloc_v2(void *owner, struct mbuf_operations *ops)
 {
 	struct mbuf_slot *mb;
 	u32 index = 0;
@@ -401,26 +410,38 @@ again:
 	g_mbuf.mbuf_next_id = index;
 
 	mb = (struct mbuf_slot *)(g_mbuf.mbuf + index * g_mbuf.mbuf_size_per_cg);
-	mbuf_slot_init(mb, cg, index);
+	mbuf_slot_init(mb, owner, index, ops);
 	g_mbuf.mbuf_frees--;
 
 	spin_unlock_irqrestore(&g_mbuf.mbuf_lock, flags);
 
 	return mb;
 }
+EXPORT_SYMBOL(mbuf_slot_alloc_v2);
 
-void mbuf_free(struct cgroup *cg)
+struct mbuf_slot *mbuf_slot_alloc(struct cgroup *cg)
+{
+	return mbuf_slot_alloc_v2((void *)cg, NULL);
+}
+EXPORT_SYMBOL(mbuf_slot_alloc);
+
+void mbuf_free_slot(struct mbuf_slot *slot)
 {
 	unsigned long flags;
 
 	spin_lock_irqsave(&g_mbuf.mbuf_lock, flags);
-
 	/* Make current idx the next available buffer */
-	g_mbuf.mbuf_next_id = cg->mbuf->idx;
+	g_mbuf.mbuf_next_id = slot->idx;
 	__clear_bit(g_mbuf.mbuf_next_id, g_mbuf.mbuf_bitmap);
-
 	g_mbuf.mbuf_frees++;
 	spin_unlock_irqrestore(&g_mbuf.mbuf_lock, flags);
+
+}
+EXPORT_SYMBOL(mbuf_free_slot);
+
+void mbuf_free(struct cgroup *cg)
+{
+	mbuf_free_slot(cg->mbuf);
 }
 
 static u32 rd_mbuf_next(struct mbuf_ring *mring, u32 curr_idx)

net/Kconfig

@@ -124,6 +124,16 @@ source "net/mptcp/Kconfig"
 
 endif # if INET
 
+config NETNS_MBUF
+	bool "attach a mbuf to net namespace"
+	default y
+	depends on RQM && INET && PROC_FS
+	help
+	  this allows attach a mbuf to each net namespace. mbuf is a ring
+	  buffer of log, you can used to print log to it.
+
+	  if you are unsure how to answer this question, answer N.
+
 config NETWORK_SECMARK
 	bool "Security Marking"
 	help

net/core/Makefile

@@ -40,3 +40,4 @@ obj-$(CONFIG_NET_SOCK_MSG) += skmsg.o
 obj-$(CONFIG_BPF_SYSCALL) += sock_map.o
 obj-$(CONFIG_BPF_SYSCALL) += bpf_sk_storage.o
 obj-$(CONFIG_OF)	+= of_net.o
+obj-$(CONFIG_NETNS_MBUF) += netns_mbuf.o

net/core/netns_mbuf.c

@@ -0,0 +1,275 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/* make mbuf can be used by net namespace
+ *
+ * Author: mengensun <mengensun@tencent.com>
+ * Author: yuehongwu <yuehongwu@tencent.com>
+ * Copyright (C) 2024 Tencent, Inc
+ */
+#include<linux/cgroup.h>
+#include<linux/mbuf.h>
+#include<linux/proc_fs.h>
+
+#include<net/net_namespace.h>
+#include<net/netns/generic.h>
+
+struct mbuf_seq_data {
+	struct seq_net_private snp;
+	struct mbuf_user_desc udesc;
+	struct mbuf_slot snapshot[];
+};
+
+static inline struct mbuf_slot *get_net_mbuf(struct net *net)
+{
+	return net->mbuf.slot;
+}
+
+/* not controlled by sysctl_qos_mbuf_enable because we will
+ * have a /proc/net/ipv4/netlat/enable in later patch
+ */
+ssize_t net_mbuf_print(struct net *net, const char *fmt, ...)
+{
+	va_list args;
+	struct mbuf_slot *slot;
+
+	slot = net->mbuf.slot;
+	if (!slot || !__ratelimit(&slot->ratelimit))
+		goto out;
+
+	va_start(args, fmt);
+	slot->ops->write(slot, fmt, args);
+	va_end(args);
+out:
+	return 0;
+}
+EXPORT_SYMBOL(net_mbuf_print);
+
+/* udesc is the user side interface, used to get data from mbuf,
+ * we can alloc a udesc per user, not to alloc a udesc and bind
+ * to mbuf when user accessing mbuf.
+ *
+ * seq file private data is the ideal place to hold the udesc
+ * if we put udesc in seq file private data all things is simple
+ */
+static void *netns_mbuf_start(struct seq_file *s, loff_t *pos)
+{
+	u32 index;
+	struct mbuf_user_desc *udesc;
+	struct mbuf_seq_data *pd;
+
+	pd = s->private;
+	udesc = &pd->udesc;
+	index = *pos;
+
+	/* why: see seq_mbuf_open */
+	if (!pd->snapshot->mring)
+		return NULL;
+
+	/* If already reach end, just return */
+	if (index && index == pd->snapshot->mring->next_idx)
+		return NULL;
+
+	udesc->user_idx = pd->snapshot->mring->first_idx;
+	udesc->user_seq = pd->snapshot->mring->first_seq;
+
+	/* Maybe reach end or empty */
+	if (udesc->user_idx == pd->snapshot->mring->next_idx)
+		return NULL;
+	return udesc;
+}
+
+static void *netns_mbuf_next(struct seq_file *s, void *v, loff_t *pos)
+{
+	struct mbuf_seq_data *pd;
+	struct mbuf_user_desc *udesc = v;
+
+	pd = s->private;
+
+	/* why: see seq_mbuf_open */
+	if (!pd->snapshot->mring)
+		return NULL;
+
+	udesc->user_idx = pd->snapshot->ops->next(pd->snapshot->mring,
+			udesc->user_idx);
+	*pos = udesc->user_idx;
+	if (udesc->user_idx == pd->snapshot->mring->next_idx)
+		return NULL;
+
+	return udesc;
+}
+
+static void netns_mbuf_stop(struct seq_file *s, void *v) { }
+
+static int netns_mbuf_show(struct seq_file *s, void *v)
+{
+	ssize_t ret;
+	struct mbuf_seq_data *pd;
+	struct mbuf_user_desc *udesc = (struct mbuf_user_desc *)v;
+
+	pd = s->private;
+
+	/* why: see seq_mbuf_open */
+	if (!pd->snapshot->mring)
+		return 0;
+
+	memset(udesc->buf, 0, sizeof(udesc->buf));
+	ret = pd->snapshot->ops->read(pd->snapshot, udesc);
+	if (ret > 0)
+		seq_printf(s, "%s", udesc->buf);
+	return 0;
+}
+
+static int seq_mbuf_open(struct inode *inode, struct file *file)
+{
+	struct mbuf_seq_data *p;
+	struct mbuf_slot *mbuf;
+
+	p = seq_open_net_large_private(inode, file);
+
+	if (IS_ERR(p))
+		return PTR_ERR(p);
+
+	mbuf = get_net_mbuf(p->snp.net);
+	/* netns may have no mbuf attached, because the mbuf
+	 * pool has a max num
+	 * here we let file open success, so, seq_ops must
+	 * check mring point
+	 *
+	 * btw: we memzerod the private in
+	 * seq_open_net_large_private
+	 */
+	if (!mbuf)
+		return 0;
+
+	snapshot_mbuf(p->snapshot, mbuf, &mbuf->slot_lock);
+	return 0;
+}
+
+/* this function is token from seq_release_net, all is the
+ * same except for using **vfree** to free the private
+ */
+static int seq_mbuf_release(struct inode *ino, struct file *f)
+{
+	struct seq_file *seq = f->private_data;
+
+	put_net(seq_file_net(seq));
+	vfree(seq->private);
+	seq->private = NULL;
+	seq_release(ino, f);
+	return 0;
+}
+
+/* when write clear the data */
+ssize_t seq_mbuf_write(struct file *f, const char __user *ubuf,
+		       size_t size, loff_t *_pos)
+{
+	struct seq_file *seq = f->private_data;
+	struct mbuf_seq_data *p;
+	struct mbuf_slot *mb;
+
+	p = seq->private;
+	mb = get_net_mbuf(p->snp.net);
+
+	/* the netns not attached mbuf */
+	if (!mb)
+		return size;
+
+	mbuf_reset(mb);
+	return size;
+}
+
+/* seq_read have a mutex lock hold when called thoes function
+ * while the mutex lock is bind to struct file, not to inode,
+ * that mutex lock can control mutex access to mbuf among tasks
+ * which have the same file object (eg: muti-threads of
+ * a process)
+ *
+ * if there are muti-process access the mbuf, there have no
+ * mutex accessing.
+ */
+static const struct seq_operations mbuf_seq_ops = {
+	.show = netns_mbuf_show,
+	.start = netns_mbuf_start,
+	.next = netns_mbuf_next,
+	.stop = netns_mbuf_stop,
+};
+
+static const struct proc_ops mbuf_seq_fops = {
+	.proc_open	= seq_mbuf_open,
+	.proc_read	= seq_read,
+	.proc_write	= seq_mbuf_write,
+	.proc_lseek	= seq_lseek,
+	.proc_release	= seq_mbuf_release,
+};
+
+static int __net_init net_mbuf_init(struct net *net)
+{
+	int ret = 0;
+
+	/* if mbuf alloc failed, make the netns create success
+	 *
+	 * returning error here will put a limit on max netns
+	 * can be created on current system
+	 *
+	 * btw: mbuf_slot has a max num 1024 for now, if mbuf_slot
+	 * is all used, more allocing may failed, what we can do
+	 * is make usr interface not changed, and make netlat
+	 * `speak nothing`
+	 * cgroup is used for kabi
+	 */
+	net->mbuf.slot = mbuf_slot_alloc_v2((void *)net, NULL);
+	if (!net->mbuf.slot)
+		pr_err("fail alloc mbuf");
+
+	net->mbuf.twatcher = proc_net_mkdir(net, "twatcher", net->proc_net);
+	if (!net->mbuf.twatcher) {
+		ret = -ENOMEM;
+		goto free_mbuf;
+	}
+
+	net->mbuf.log = proc_create_net_data_ops("log", S_IFREG | 0644,
+						 net->mbuf.twatcher,
+						 &mbuf_seq_ops,
+						 sizeof(struct mbuf_seq_data) + get_mbuf_slot_len(),
+						 NULL, &mbuf_seq_fops);
+	if (!net->mbuf.log) {
+		ret = -ENOMEM;
+		goto remove_watcher;
+	}
+	return ret;
+
+remove_watcher:
+	remove_proc_entry("twatcher", net->proc_net);
+
+free_mbuf:
+	if (net->mbuf.slot)
+		mbuf_free_slot(net->mbuf.slot);
+	return ret;
+}
+
+static void __net_exit net_mbuf_exit(struct net *net)
+{
+	remove_proc_entry("log", net->mbuf.log);
+	remove_proc_entry("twatcher", net->mbuf.twatcher);
+
+	/* if mbuf allocate failed, no need to free */
+	if (!net->mbuf.slot)
+		return;
+	mbuf_free_slot(net->mbuf.slot);
+}
+
+static struct pernet_operations net_mbuf_ops = {
+	.init = net_mbuf_init,
+	.exit = net_mbuf_exit,
+};
+
+int  inet_mbuf_init(void)
+{
+	return register_pernet_subsys(&net_mbuf_ops);
+}
+EXPORT_SYMBOL(inet_mbuf_init);
+
+void inet_mbuf_exit(void)
+{
+	unregister_pernet_subsys(&net_mbuf_ops);
+}
+EXPORT_SYMBOL(inet_mbuf_exit);

net/ipv4/af_inet.c

@@ -114,6 +114,7 @@
 #include <net/xfrm.h>
 #include <net/net_namespace.h>
 #include <net/secure_seq.h>
+#include <net/netns_mbuf.h>
 #ifdef CONFIG_IP_MROUTE
 #include <linux/mroute.h>
 #endif
@@ -2102,11 +2103,15 @@ static int __init ipv4_proc_init(void)
 		goto out_udp;
 	if (ping_proc_init())
 		goto out_ping;
+	if (inet_mbuf_init())
+		goto out_mbuf;
 	if (ip_misc_proc_init())
 		goto out_misc;
 out:
 	return rc;
 out_misc:
+	inet_mbuf_exit();
+out_mbuf:
 	ping_proc_exit();
 out_ping:
 	udp4_proc_exit();