Tofu: proper cleanup of device files when mcexec gets killed

Change-Id: I6cb0290f72d96682700f945b29585e132e525ac1

executer/kernel/mcctrl/arch/arm64/archdeps.c

@@ -2,6 +2,7 @@
 #include <linux/version.h>
 #include <linux/mm_types.h>
 #include <linux/kallsyms.h>
+#include <linux/delay.h>
 #if KERNEL_VERSION(4, 11, 0) <= LINUX_VERSION_CODE
 #include <linux/sched/task_stack.h>
 #endif /* LINUX_VERSION_CODE >= KERNEL_VERSION(4, 11, 0) */
@@ -27,6 +28,14 @@ void *vdso_end;
 static struct vm_special_mapping (*vdso_spec)[2];
 #endif
 
+/* Tofu CQ and barrier gate release functions */
+struct file_operations *mcctrl_tof_utofu_procfs_ops_cq;
+int (*mcctrl_tof_utofu_release_cq)(struct inode *inode,
+		struct file *filp);
+struct file_operations *mcctrl_tof_utofu_procfs_ops_bch;
+int (*mcctrl_tof_utofu_release_bch)(struct inode *inode,
+		struct file *filp);
+
 int arch_symbols_init(void)
 {
 #if LINUX_VERSION_CODE >= KERNEL_VERSION(4, 0, 0)
@@ -43,6 +52,26 @@ int arch_symbols_init(void)
 		return -EFAULT;
 #endif
 
+	mcctrl_tof_utofu_procfs_ops_cq =
+		(void *)kallsyms_lookup_name("tof_utofu_procfs_ops_cq");
+	if (WARN_ON(!mcctrl_tof_utofu_procfs_ops_cq))
+		return -EFAULT;
+
+	mcctrl_tof_utofu_procfs_ops_bch =
+		(void *)kallsyms_lookup_name("tof_utofu_procfs_ops_bch");
+	if (WARN_ON(!mcctrl_tof_utofu_procfs_ops_bch))
+		return -EFAULT;
+
+	mcctrl_tof_utofu_release_cq =
+		(void *)kallsyms_lookup_name("tof_utofu_release_cq");
+	if (WARN_ON(!mcctrl_tof_utofu_release_cq))
+		return -EFAULT;
+
+	mcctrl_tof_utofu_release_bch =
+		(void *)kallsyms_lookup_name("tof_utofu_release_bch");
+	if (WARN_ON(!mcctrl_tof_utofu_release_bch))
+		return -EFAULT;
+
 	return 0;
 }
 
@@ -417,3 +446,106 @@ long arch_switch_ctx(struct uti_switch_ctx_desc *desc)
 out:
 	return rc;
 }
+
+
+/*
+ * Tofu CQ and BCH release handlers
+ */
+int __mcctrl_tof_utofu_release_cq(struct inode *inode, struct file *filp);
+int __mcctrl_tof_utofu_release_bch(struct inode *inode, struct file *filp);
+
+void mcctrl_tofu_hijack_release_handlers(void)
+{
+	mcctrl_tof_utofu_procfs_ops_cq->release =
+		__mcctrl_tof_utofu_release_cq;
+	mcctrl_tof_utofu_procfs_ops_bch->release =
+		__mcctrl_tof_utofu_release_bch;
+}
+
+void mcctrl_tofu_restore_release_handlers(void)
+{
+	mcctrl_tof_utofu_procfs_ops_cq->release =
+		mcctrl_tof_utofu_release_cq;
+	mcctrl_tof_utofu_procfs_ops_bch->release =
+		mcctrl_tof_utofu_release_bch;
+}
+
+int __mcctrl_tof_utofu_release_handler(struct inode *inode, struct file *filp,
+		int (*__release_func)(struct inode *inode, struct file *filp))
+{
+	struct mcctrl_usrdata *usrdata;
+	struct mcctrl_file_to_pidfd *f2pfd;
+	struct mcctrl_per_proc_data *ppd;
+	struct ikc_scd_packet isp;
+	int ret;
+
+	dprintk("%s: current PID: %d, comm: %s \n",
+			__func__, task_tgid_vnr(current), current->comm);
+
+	f2pfd = mcctrl_file_to_pidfd_hash_lookup(filp, current->group_leader);
+	if (!f2pfd) {
+		goto out;
+	}
+
+	dprintk("%s: current PID: %d, PID: %d, fd: %d ...\n",
+			__func__, task_tgid_vnr(current), f2pfd->pid, f2pfd->fd);
+	usrdata = ihk_host_os_get_usrdata(f2pfd->os);
+
+	/* Look up per-process structure */
+	ppd = mcctrl_get_per_proc_data(usrdata, f2pfd->pid);
+	if (!ppd) {
+		pr_err("%s: PID: %d, fd: %d no PPD\n",
+				__func__, f2pfd->pid, f2pfd->fd);
+		goto out;
+	}
+
+	dprintk("%s: PID: %d, fd: %d PPD OK\n",
+			__func__, f2pfd->pid, f2pfd->fd);
+
+	/*
+	 * We are in release() due to the process being killed,
+	 * or because the application didn't close the file properly.
+	 * Ask McKernel to clean up this fd.
+	 */
+	isp.msg = SCD_MSG_CLEANUP_FD;
+	isp.pid = f2pfd->pid;
+	isp.arg = f2pfd->fd;
+
+	ret = mcctrl_ikc_send_wait(f2pfd->os, ppd->ikc_target_cpu,
+			&isp, -20, NULL, NULL, 0);
+	if (ret != 0) {
+		dprintk("%s: WARNING: failed to send IKC msg: %d\n",
+				__func__, ret);
+	}
+
+	mcctrl_file_to_pidfd_hash_remove(filp, f2pfd->os,
+			current->group_leader, f2pfd->fd);
+
+	mcctrl_put_per_proc_data(ppd);
+
+	/* Do not call into Linux driver if timed out in SIGKILL.. */
+	if (ret == -ETIME && __fatal_signal_pending(current)) {
+		pr_err("%s: WARNING: failed to send IKC msg in SIGKILL: %d\n",
+				__func__, ret);
+		goto out_no_release;
+	}
+out:
+	dprintk("%s: current PID: %d, comm: %s -> calling release\n",
+			__func__, task_tgid_vnr(current), current->comm);
+	return __release_func(inode, filp);
+
+out_no_release:
+	return ret;
+}
+
+int __mcctrl_tof_utofu_release_cq(struct inode *inode, struct file *filp)
+{
+	return __mcctrl_tof_utofu_release_handler(inode, filp,
+			mcctrl_tof_utofu_release_cq);
+}
+
+int __mcctrl_tof_utofu_release_bch(struct inode *inode, struct file *filp)
+{
+	return __mcctrl_tof_utofu_release_handler(inode, filp,
+			mcctrl_tof_utofu_release_bch);
+}

executer/kernel/mcctrl/control.c

@@ -36,6 +36,7 @@
 #include <linux/semaphore.h>
 #include <linux/interrupt.h>
 #include <linux/cpumask.h>
+#include <linux/delay.h>
 #include <asm/uaccess.h>
 #include <asm/delay.h>
 #include <asm/io.h>
@@ -378,6 +379,7 @@ static void release_handler(ihk_os_t os, void *param)
 	int os_ind = ihk_host_os_get_index(os);
 	unsigned long flags;
 	struct host_thread *thread;
+	int ret;
 
 	/* Finalize FS switch for uti threads */ 
 	write_lock_irqsave(&host_thread_lock, flags);
@@ -399,7 +401,13 @@ static void release_handler(ihk_os_t os, void *param)
 
 	dprintk("%s: SCD_MSG_CLEANUP_PROCESS, info: %p, cpu: %d\n",
 			__FUNCTION__, info, info->cpu);
-	mcctrl_ikc_send(os, info->cpu, &isp);
+	ret = mcctrl_ikc_send_wait(os, info->cpu,
+			&isp, -20, NULL, NULL, 0);
+	if (ret != 0) {
+		printk("%s: WARNING: failed to send IKC msg: %d\n",
+				__func__, ret);
+	}
+
 	if (os_ind >= 0) {
 		delete_pid_entry(os_ind, info->pid);
 	}

executer/kernel/mcctrl/driver.c

@@ -50,6 +50,7 @@ extern void procfs_exit(int);
 extern void uti_attr_finalize(void);
 extern void binfmt_mcexec_init(void);
 extern void binfmt_mcexec_exit(void);
+extern void mcctrl_file_to_pidfd_hash_init(void);
 
 extern int mcctrl_os_read_cpu_register(ihk_os_t os, int cpu,
 		struct ihk_os_cpu_register *desc);
@@ -57,6 +58,11 @@ extern int mcctrl_os_write_cpu_register(ihk_os_t os, int cpu,
 		struct ihk_os_cpu_register *desc);
 extern int mcctrl_get_request_os_cpu(ihk_os_t os, int *cpu);
 
+#ifdef ENABLE_TOFU
+extern void mcctrl_tofu_hijack_release_handlers(void);
+extern void mcctrl_tofu_restore_release_handlers(void);
+#endif
+
 static long mcctrl_ioctl(ihk_os_t os, unsigned int request, void *priv,
                          unsigned long arg, struct file *file)
 {
@@ -319,10 +325,15 @@ static int __init mcctrl_init(void)
 	}
 
 	binfmt_mcexec_init();
+	mcctrl_file_to_pidfd_hash_init();
 
 	if ((ret = symbols_init()))
 		goto error;
 
+#ifdef ENABLE_TOFU
+	mcctrl_tofu_hijack_release_handlers();
+#endif
+
 	if ((ret = ihk_host_register_os_notifier(&mcctrl_os_notifier)) != 0) {
 		printk("mcctrl: error: registering OS notifier\n");
 		goto error;
@@ -345,6 +356,9 @@ static void __exit mcctrl_exit(void)
 
 	binfmt_mcexec_exit();
 	uti_attr_finalize();
+#ifdef ENABLE_TOFU
+	mcctrl_tofu_restore_release_handlers();
+#endif
 
 	printk("mcctrl: unregistered.\n");
 }

executer/kernel/mcctrl/ikc.c

@@ -148,8 +148,29 @@ int mcctrl_ikc_send_wait(ihk_os_t os, int cpu, struct ikc_scd_packet *pisp,
 	}
 
 	if (timeout) {
-		ret = wait_event_interruptible_timeout(desc->wq,
+		/*
-			desc->status, timeout);
+		 * Negative timeout indicates busy waiting, which can be used
+		 * in situations where wait_event_interruptible_XXX() would
+		 * fail, e.g., in a signal handler, at the time the process
+		 * is being killed, etc.
+		 */
+		if (timeout < 0) {
+			unsigned long timeout_jiffies =
+				jiffies + msecs_to_jiffies(timeout * -1);
+			ret = -ETIME;
+
+			while (time_before(jiffies, timeout_jiffies)) {
+				schedule();
+				if (READ_ONCE(desc->status)) {
+					ret = 0;
+					break;
+				}
+			}
+		}
+		else {
+			ret = wait_event_interruptible_timeout(desc->wq,
+					desc->status, msecs_to_jiffies(timeout));
+		}
 	} else {
 		ret = wait_event_interruptible(desc->wq, desc->status);
 	}
@@ -211,6 +232,8 @@ static int syscall_packet_handler(struct ihk_ikc_channel_desc *c,
 	case SCD_MSG_PROCFS_ANSWER:
 	case SCD_MSG_REMOTE_PAGE_FAULT_ANSWER:
 	case SCD_MSG_CPU_RW_REG_RESP:
+	case SCD_MSG_CLEANUP_PROCESS_RESP:
+	case SCD_MSG_CLEANUP_FD_RESP:
 		mcctrl_wakeup_cb(__os, pisp);
 		break;
 

executer/kernel/mcctrl/mcctrl.h

@@ -58,7 +58,8 @@
 #define SCD_MSG_SEND_SIGNAL             0x7
 #define SCD_MSG_SEND_SIGNAL_ACK         0x8
 #define SCD_MSG_CLEANUP_PROCESS         0x9
-#define SCD_MSG_GET_VDSO_INFO           0xa
+#define SCD_MSG_CLEANUP_PROCESS_RESP    0xa
+#define SCD_MSG_GET_VDSO_INFO           0xb
 
 //#define SCD_MSG_GET_CPU_MAPPING         0xc
 //#define SCD_MSG_REPLY_GET_CPU_MAPPING   0xd
@@ -104,6 +105,8 @@
 
 #define SCD_MSG_CPU_RW_REG              0x52
 #define SCD_MSG_CPU_RW_REG_RESP         0x53
+#define SCD_MSG_CLEANUP_FD              0x54
+#define SCD_MSG_CLEANUP_FD_RESP         0x55
 
 #define SCD_MSG_FUTEX_WAKE              0x60
 
@@ -556,4 +559,29 @@ struct uti_futex_resp {
 	int done;
 	wait_queue_head_t wq;
 };
+
+/*
+ * Hash table to keep track of files and related processes
+ * and file descriptors.
+ * NOTE: Used for Tofu driver release handlers.
+ */
+#define MCCTRL_FILE_2_PIDFD_HASH_SHIFT 4
+#define MCCTRL_FILE_2_PIDFD_HASH_SIZE (1 << MCCTRL_FILE_2_PIDFD_HASH_SHIFT)
+#define MCCTRL_FILE_2_PIDFD_HASH_MASK (MCCTRL_FILE_2_PIDFD_HASH_SIZE - 1)
+
+struct mcctrl_file_to_pidfd {
+	struct file *filp;
+	ihk_os_t os;
+	struct task_struct *group_leader;
+	int pid;
+	int fd;
+	struct list_head hash;
+};
+
+int mcctrl_file_to_pidfd_hash_insert(struct file *filp,
+	ihk_os_t os, int pid, struct task_struct *group_leader, int fd);
+struct mcctrl_file_to_pidfd *mcctrl_file_to_pidfd_hash_lookup(
+	struct file *filp, struct task_struct *group_leader);
+int mcctrl_file_to_pidfd_hash_remove(struct file *filp,
+	ihk_os_t os, struct task_struct *group_leader, int fd);
 #endif

executer/kernel/mcctrl/syscall.c

@@ -1843,6 +1843,136 @@ static long pager_call(ihk_os_t os, struct syscall_request *req)
 	return ret;
 }
 
+struct list_head mcctrl_file_to_pidfd_hash[MCCTRL_FILE_2_PIDFD_HASH_SIZE];
+spinlock_t mcctrl_file_to_pidfd_hash_lock;
+
+void mcctrl_file_to_pidfd_hash_init(void)
+{
+	int hash;
+	spin_lock_init(&mcctrl_file_to_pidfd_hash_lock);
+
+	for (hash = 0; hash < MCCTRL_FILE_2_PIDFD_HASH_SIZE; ++hash) {
+		INIT_LIST_HEAD(&mcctrl_file_to_pidfd_hash[hash]);
+	}
+}
+
+int mcctrl_file_to_pidfd_hash_insert(struct file *filp,
+	ihk_os_t os, int pid, struct task_struct *group_leader, int fd)
+{
+	unsigned long irqflags;
+	struct mcctrl_file_to_pidfd *file2pidfd_iter;
+	struct mcctrl_file_to_pidfd *file2pidfd;
+	int hash = (int)((unsigned long)filp &
+			(unsigned long)MCCTRL_FILE_2_PIDFD_HASH_MASK);
+	int ret = 0;
+
+	file2pidfd = kmalloc(sizeof(*file2pidfd), GFP_ATOMIC);
+	if (!file2pidfd)
+		return -ENOMEM;
+
+	file2pidfd->filp = filp;
+	file2pidfd->os = os;
+	file2pidfd->pid = pid;
+	file2pidfd->group_leader = group_leader;
+	file2pidfd->fd = fd;
+
+	spin_lock_irqsave(&mcctrl_file_to_pidfd_hash_lock, irqflags);
+	list_for_each_entry(file2pidfd_iter,
+			&mcctrl_file_to_pidfd_hash[hash], hash) {
+		if (file2pidfd_iter->filp == filp) {
+			printk("%s: WARNING: filp: %p, pid: %d, fd: %d exists\n",
+					__func__, filp, pid, fd);
+			ret = -EBUSY;
+			goto free_out;
+		}
+	}
+
+	list_add_tail(&file2pidfd->hash,
+			&mcctrl_file_to_pidfd_hash[hash]);
+	dprintk("%s: filp: %p, pid: %d, fd: %d added\n",
+			__func__, filp, pid, fd);
+
+	spin_unlock_irqrestore(&mcctrl_file_to_pidfd_hash_lock, irqflags);
+	return ret;
+
+free_out:
+	kfree(file2pidfd);
+	spin_unlock_irqrestore(&mcctrl_file_to_pidfd_hash_lock, irqflags);
+	return ret;
+}
+
+/*
+ * XXX: lookup relies on group_leader to identify the process
+ * because PIDs might be different across name spaces (e.g.,
+ * when using Docker)
+ */
+struct mcctrl_file_to_pidfd *mcctrl_file_to_pidfd_hash_lookup(
+	struct file *filp, struct task_struct *group_leader)
+{
+	unsigned long irqflags;
+	struct mcctrl_file_to_pidfd *file2pidfd_iter;
+	struct mcctrl_file_to_pidfd *file2pidfd = NULL;
+	int hash = (int)((unsigned long)filp &
+			(unsigned long)MCCTRL_FILE_2_PIDFD_HASH_MASK);
+
+	spin_lock_irqsave(&mcctrl_file_to_pidfd_hash_lock, irqflags);
+	list_for_each_entry(file2pidfd_iter,
+			&mcctrl_file_to_pidfd_hash[hash], hash) {
+		if (file2pidfd_iter->filp == filp && 
+				file2pidfd_iter->group_leader == group_leader) {
+			file2pidfd = file2pidfd_iter;
+			dprintk("%s: filp: %p, pid: %d, fd: %d found\n",
+					__func__, filp, file2pidfd->pid, file2pidfd->fd);
+			break;
+		}
+	}
+
+	spin_unlock_irqrestore(&mcctrl_file_to_pidfd_hash_lock, irqflags);
+
+	return file2pidfd;
+}
+
+int mcctrl_file_to_pidfd_hash_remove(struct file *filp,
+	ihk_os_t os, struct task_struct *group_leader, int fd)
+{
+	unsigned long irqflags;
+	struct mcctrl_file_to_pidfd *file2pidfd_iter;
+	int hash = (int)((unsigned long)filp &
+			(unsigned long)MCCTRL_FILE_2_PIDFD_HASH_MASK);
+	int ret = 0;
+
+	spin_lock_irqsave(&mcctrl_file_to_pidfd_hash_lock, irqflags);
+	list_for_each_entry(file2pidfd_iter,
+			&mcctrl_file_to_pidfd_hash[hash], hash) {
+		if (file2pidfd_iter->filp != filp)
+			continue;
+
+		if (file2pidfd_iter->os != os)
+			continue;
+
+		if (file2pidfd_iter->group_leader != group_leader)
+			continue;
+
+		if (file2pidfd_iter->fd != fd)
+			continue;
+
+		list_del(&file2pidfd_iter->hash);
+		dprintk("%s: filp: %p, pid: %d, fd: %d removed\n",
+				__func__, filp, file2pidfd_iter->pid, fd);
+		kfree(file2pidfd_iter);
+		goto unlock_out;
+	}
+
+	dprintk("%s: filp: %p, pid: %d, fd: %d couldn't be found\n",
+			__func__, filp, pid, fd);
+	ret = -ENOENT;
+
+unlock_out:
+	spin_unlock_irqrestore(&mcctrl_file_to_pidfd_hash_lock, irqflags);
+	return ret;
+}
+
+
 void __return_syscall(ihk_os_t os, struct ikc_scd_packet *packet,
 		long ret, int stid)
 {
@@ -1874,18 +2004,14 @@ void __return_syscall(ihk_os_t os, struct ikc_scd_packet *packet,
 	res->stid = stid;
 
 #ifdef ENABLE_TOFU
-	/* Record PDE_DATA after open()/ioctl() calls for Tofu driver */
+	/* Record PDE_DATA after open() calls for Tofu driver */
-	if ((packet->req.number == __NR_ioctl && ret == 0) ||
+	if (packet->req.number == __NR_openat && ret > 1) {
-			(packet->req.number == __NR_openat && ret > 1)) {
 		char *pathbuf, *fullpath;
 		struct fd f;
+		int fd;
 
-		if (packet->req.number == __NR_ioctl) {
+		fd = ret;
-			f = fdget(packet->req.args[0]);
+		f = fdget(fd);
-		}
-		else if (packet->req.number == __NR_openat) {
-			f = fdget(ret);
-		}
 
 		if (!f.file) {
 			goto out_notify;
@@ -1903,12 +2029,11 @@ void __return_syscall(ihk_os_t os, struct ikc_scd_packet *packet,
 
 		if (!strncmp("/proc/tofu/dev/", fullpath, 15)) {
 			res->pde_data = PDE_DATA(file_inode(f.file));
-			dprintk("%s: %s(): fd: %ld, path: %s, PDE_DATA: 0x%lx\n",
+			dprintk("%s: fd: %d, path: %s, PDE_DATA: 0x%lx\n",
-				__func__,
+					__func__,
-				packet->req.number == __NR_ioctl ? "ioctl" : "openat",
+					fd,
-				packet->req.args[0],
+					fullpath,
-				fullpath,
+					(unsigned long)res->pde_data);
-				(unsigned long)res->pde_data);
 			dprintk("%s: pgd_index: %ld, pmd_index: %ld, pte_index: %ld\n",
 				__func__,
 				pgd_index((unsigned long)res->pde_data),
@@ -1918,6 +2043,9 @@ void __return_syscall(ihk_os_t os, struct ikc_scd_packet *packet,
 			dprintk("CONFIG_ARM64_VA_BITS: %d, PGDIR_SHIFT: %d\n",
 				CONFIG_ARM64_VA_BITS, PGDIR_SHIFT);
 #endif
+			mcctrl_file_to_pidfd_hash_insert(f.file, os,
+					task_tgid_vnr(current),
+					current->group_leader, fd);
 		}
 
 out_free:
@@ -2330,6 +2458,26 @@ int __do_in_kernel_syscall(ihk_os_t os, struct ikc_scd_packet *packet)
 
 	dprintk("%s: system call: %lx\n", __FUNCTION__, sc->args[0]);
 	switch (sc->number) {
+	case __NR_close: {
+		struct fd f;
+		int fd;
+
+		fd = (int)sc->args[0];
+		if (fd > 2) {
+			f = fdget(fd);
+
+			if (f.file) {
+				mcctrl_file_to_pidfd_hash_remove(f.file, os,
+						current->group_leader, fd);
+				fdput(f);
+			}
+		}
+
+		error = -ENOSYS;
+		goto out;
+
+		break;
+	}
 	case __NR_mmap:
 		ret = pager_call(os, sc);
 		break;

kernel/host.c

@@ -778,12 +778,36 @@ out_remote_pf:
 		ret = 0;
 		break;
 
-	case SCD_MSG_CLEANUP_PROCESS:
+	case SCD_MSG_CLEANUP_PROCESS: {
+		extern int process_cleanup_before_terminate(int pid);
 		dkprintf("SCD_MSG_CLEANUP_PROCESS pid=%d, thread=0x%llx\n",
 				packet->pid, packet->arg);
+
+		pckt.msg = SCD_MSG_CLEANUP_PROCESS_RESP;
+		pckt.err = process_cleanup_before_terminate(packet->pid);
+		pckt.ref = packet->ref;
+		pckt.arg = packet->arg;
+		pckt.reply = packet->reply;
+		syscall_channel_send(resp_channel, &pckt);
 		terminate_host(packet->pid, (struct thread *)packet->arg);
 		ret = 0;
 		break;
+	}
+
+	case SCD_MSG_CLEANUP_FD: {
+		extern int process_cleanup_fd(int pid, int fd);
+		pckt.msg = SCD_MSG_CLEANUP_FD_RESP;
+		pckt.err = process_cleanup_fd(packet->pid, packet->arg);
+		dkprintf("SCD_MSG_CLEANUP_FD pid=%d, fd=%d -> err: %d\n",
+				packet->pid, packet->arg, pckt.err);
+
+		pckt.ref = packet->ref;
+		pckt.arg = packet->arg;
+		pckt.reply = packet->reply;
+		syscall_channel_send(resp_channel, &pckt);
+		ret = 0;
+		break;
+	}
 
 	case SCD_MSG_DEBUG_LOG:
 		dkprintf("SCD_MSG_DEBUG_LOG code=%lx\n", packet->arg);

kernel/include/syscall.h

@@ -39,7 +39,8 @@
 #define SCD_MSG_SEND_SIGNAL             0x7
 #define SCD_MSG_SEND_SIGNAL_ACK         0x8
 #define SCD_MSG_CLEANUP_PROCESS         0x9
-#define SCD_MSG_GET_VDSO_INFO           0xa
+#define SCD_MSG_CLEANUP_PROCESS_RESP    0xa
+#define SCD_MSG_GET_VDSO_INFO           0xb
 
 #define SCD_MSG_GET_CPU_MAPPING         0xc
 #define SCD_MSG_REPLY_GET_CPU_MAPPING   0xd
@@ -84,6 +85,8 @@
 
 #define SCD_MSG_CPU_RW_REG              0x52
 #define SCD_MSG_CPU_RW_REG_RESP         0x53
+#define SCD_MSG_CLEANUP_FD              0x54
+#define SCD_MSG_CLEANUP_FD_RESP         0x55
 
 #define SCD_MSG_FUTEX_WAKE              0x60
 

kernel/syscall.c

@@ -1303,10 +1303,8 @@ void terminate(int rc, int sig)
 			if (proc->enable_tofu && proc->fd_pde_data[fd]) {
 				extern void tof_utofu_release_fd(struct process *proc, int fd);
 
-				if (proc->fd_path[fd]) {
+				dkprintf("%s: -> tof_utofu_release_fd() @ fd: %d (%s)\n",
-					dkprintf("%s: -> tof_utofu_release_fd() @ fd: %d (%s)\n",
+						__func__, fd, proc->fd_path[fd]);
-							__func__, fd, proc->fd_path[fd]);
-				}
 				tof_utofu_release_fd(proc, fd);
 				proc->fd_pde_data[fd] = NULL;
 			}
@@ -1487,6 +1485,67 @@ void terminate(int rc, int sig)
 	panic("panic");
 }
 
+int __process_cleanup_fd(struct process *proc, int fd)
+{
+#ifdef ENABLE_TOFU
+	/* Tofu? */
+	if (proc->enable_tofu) {
+		extern void tof_utofu_release_fd(struct process *proc, int fd);
+
+		dkprintf("%s: -> tof_utofu_release_fd() @ fd: %d (%s)\n",
+				__func__, fd, proc->fd_path[fd]);
+		tof_utofu_release_fd(proc, fd);
+		proc->fd_pde_data[fd] = NULL;
+
+		if (proc->fd_path[fd]) {
+			kfree(proc->fd_path[fd]);
+			proc->fd_path[fd] = NULL;
+		}
+	}
+#endif
+	return 0;
+}
+
+int process_cleanup_fd(int pid, int fd)
+{
+	struct process *proc;
+	struct mcs_rwlock_node_irqsave lock;
+
+	proc = find_process(pid, &lock);
+	if (!proc) {
+		/* This is normal behavior */
+		dkprintf("%s: PID %d couldn't be found\n", __func__, pid);
+		return 0;
+	}
+
+	__process_cleanup_fd(proc, fd);
+
+	process_unlock(proc, &lock);
+	return 0;
+}
+
+int process_cleanup_before_terminate(int pid)
+{
+	struct process *proc;
+	struct mcs_rwlock_node_irqsave lock;
+	int fd;
+
+	proc = find_process(pid, &lock);
+	if (!proc) {
+		/* This is normal behavior */
+		return 0;
+	}
+
+	/* Clean up PDE file descriptors */
+	for (fd = 2; fd < MAX_FD_PDE; ++fd) {
+		__process_cleanup_fd(proc, fd);
+	}
+
+	process_unlock(proc, &lock);
+	return 0;
+}
+
+
 void
 terminate_host(int pid, struct thread *thread)
 {
@@ -4032,15 +4091,14 @@ SYSCALL_DECLARE(close)
 
 #ifdef ENABLE_TOFU
 	/* Clear path and PDE data */
-	if (fd >= 0 && fd < MAX_FD_PDE) {
+	if (thread->proc->enable_tofu &&
+			fd >= 0 && fd < MAX_FD_PDE) {
 		/* Tofu? */
 		if (thread->proc->fd_pde_data[fd]) {
 			extern void tof_utofu_release_fd(struct process *proc, int fd);
 
-			if (thread->proc->fd_path[fd]) {
+			dkprintf("%s: -> tof_utofu_release_fd() @ fd: %d (%s)\n",
-				dkprintf("%s: -> tof_utofu_release_fd() @ fd: %d (%s)\n",
+					__func__, fd, thread->proc->fd_path[fd]);
-						__func__, fd, thread->proc->fd_path[fd]);
-			}
 			tof_utofu_release_fd(thread->proc, fd);
 			thread->proc->fd_pde_data[fd] = NULL;
 		}

kernel/tofu/tof_utofu_main.c

@@ -896,16 +896,18 @@ static void tof_utofu_trans_update(struct tof_utofu_cq *ucq, int stag, uintptr_t
 		struct tof_trans_table ent;
 		uint64_t atomic;
 	} tmp;
+	unsigned long flags;
+
 	tmp.ent.steering.bits.start = start >> PAGE_SHIFT;
 	tmp.ent.steering.bits.len = len >> PAGE_SHIFT;
 	tmp.ent.steering.bits.ps_code = (pgszbits == PAGE_SHIFT)? TOF_STAG_TRANS_PS_CODE_64KB:TOF_STAG_TRANS_PS_CODE_2MB;
 	//atomic64_set((atomic64_t *)&table[stag], tmp.atomic);
 	ihk_atomic64_set((ihk_atomic64_t *)&table[stag], tmp.atomic);
 
-	linux_spin_lock(&ucq->trans.mru_lock);
+	linux_spin_lock_irqsave(&ucq->trans.mru_lock, flags);
 	tof_utofu_trans_mru_delete(ucq, stag);
 	tof_utofu_trans_mru_insert(ucq, stag, pgszbits, mbpt);
-	linux_spin_unlock(&ucq->trans.mru_lock);
+	linux_spin_unlock_irqrestore(&ucq->trans.mru_lock, flags);
 }
 
 
@@ -1009,6 +1011,7 @@ static int tof_utofu_ioctl_alloc_stag(struct tof_utofu_device *dev, unsigned lon
 	uint8_t pgszbits;
 	size_t pgsz;
 	int ret = -ENOTSUPP;
+	unsigned long irqflags;
 
 	ucq = container_of(dev, struct tof_utofu_cq, common);
 	if(!ucq->common.enabled){
@@ -1054,9 +1057,9 @@ static int tof_utofu_ioctl_alloc_stag(struct tof_utofu_device *dev, unsigned lon
 #if 1
 		/* normal stag */
 		int stag;
-		linux_spin_lock(&ucq->trans.mru_lock);
+		linux_spin_lock_irqsave(&ucq->trans.mru_lock, irqflags);
 		stag = tof_utofu_trans_search(ucq, start, end, pgszbits, readonly);
-		linux_spin_unlock(&ucq->trans.mru_lock);
+		linux_spin_unlock_irqrestore(&ucq->trans.mru_lock, irqflags);
 		if(stag < 0){
 			struct tof_utofu_mbpt *mbpt = NULL;
 			stag = tof_utofu_reserve_stag(ucq, readonly);
@@ -1304,6 +1307,7 @@ static int tof_utofu_ioctl_free_stags(struct tof_utofu_device *dev, unsigned lon
 	struct tof_free_stags req;
 	int i, no_free_cnt = 0, ret;
 	int stags[1024];
+	unsigned long irqflags;
 
 	ucq = container_of(dev, struct tof_utofu_cq, common);
 
@@ -1325,9 +1329,9 @@ static int tof_utofu_ioctl_free_stags(struct tof_utofu_device *dev, unsigned lon
 	}
 
 	for(i = 0; i < req.num; i++){
-		linux_spin_lock(&ucq->trans.mru_lock);
+		linux_spin_lock_irqsave(&ucq->trans.mru_lock, irqflags);
 		ret = tof_utofu_free_stag(ucq, stags[i]);
-		linux_spin_unlock(&ucq->trans.mru_lock);
+		linux_spin_unlock_irqrestore(&ucq->trans.mru_lock, irqflags);
 		if(ret == 0){
 			stags[i] = -1;
 		}
@@ -1372,6 +1376,7 @@ void tof_utofu_release_cq(void *pde_data)
 	struct tof_utofu_cq *ucq;
 	int stag;
 	struct tof_utofu_device *dev;
+	unsigned long irqflags;
 
 	dev = (struct tof_utofu_device *)pde_data;
 	ucq = container_of(dev, struct tof_utofu_cq, common);
@@ -1383,9 +1388,9 @@ void tof_utofu_release_cq(void *pde_data)
 	}
 
 	for (stag = 0; stag < TOF_UTOFU_NUM_STAG(ucq->num_stag); stag++) {
-		linux_spin_lock(&ucq->trans.mru_lock);
+		linux_spin_lock_irqsave(&ucq->trans.mru_lock, irqflags);
 		tof_utofu_free_stag(ucq, stag);
-		linux_spin_unlock(&ucq->trans.mru_lock);
+		linux_spin_unlock_irqrestore(&ucq->trans.mru_lock, irqflags);
 	}
 
 	dkprintf("%s: UCQ (pde: %p) TNI %d, CQ %d\n",
@@ -1911,6 +1916,8 @@ static int tof_utofu_disable_bch(struct tof_utofu_bg *ubg){
 	//tof_smmu_release_ipa_bg(ubg->tni, ubg->bgid, ubg->bch.iova, TOF_ICC_BCH_DMA_ALIGN);
 	//put_page(ubg->bch.page);
 	ubg->bch.enabled = false;
+	smp_mb();
+	dkprintf("%s: tni=%d bgid=%d\n", __func__, ubg->tni, ubg->bgid);
 	return 0;
 }
 
@@ -2084,11 +2091,15 @@ void tof_utofu_release_fd(struct process *proc, int fd)
 		return;
 	}
 
-	if (strstr((const char *)proc->fd_path, "cq")) {
+	if (strstr((const char *)proc->fd_path[fd], "cq")) {
+		dkprintf("%s: PID: %d, fd: %d -> release CQ\n",
+			__func__, proc->pid, fd);
 		tof_utofu_release_cq(proc->fd_pde_data[fd]);
 	}
 
-	else if (strstr((const char *)proc->fd_path, "bch")) {
+	else if (strstr((const char *)proc->fd_path[fd], "bch")) {
+		dkprintf("%s: PID: %d, fd: %d -> release BCH\n",
+			__func__, proc->pid, fd);
 		tof_utofu_release_bch(proc->fd_pde_data[fd]);
 	}
 }