From a32a8813d0173163ba44d8f9556e0d89fdc4fb46 Mon Sep 17 00:00:00 2001 From: Rusty Russell Date: Fri, 12 Jun 2009 22:27:02 -0600 Subject: lguest: improve interrupt handling, speed up stream networking lguest never checked for pending interrupts when enabling interrupts, and things still worked. However, it makes a significant difference to TCP performance, so it's time we fixed it by introducing a pending_irq flag and checking it on irq_restore and irq_enable. These two routines are now too big to patch into the 8/10 bytes patch space, so we drop that code. Note: The high latency on interrupt delivery had a very curious effect: once everything else was optimized, networking without GSO was faster than networking with GSO, since more interrupts were sent and hence a greater chance of one getting through to the Guest! Note2: (Almost) Closing the same loophole for iret doesn't have any measurable effect, so I'm leaving that patch for the moment. Before: 1GB tcpblast Guest->Host: 30.7 seconds 1GB tcpblast Guest->Host (no GSO): 76.0 seconds After: 1GB tcpblast Guest->Host: 6.8 seconds 1GB tcpblast Guest->Host (no GSO): 27.8 seconds Signed-off-by: Rusty Russell --- include/linux/lguest.h | 4 ++++ 1 file changed, 4 insertions(+) (limited to 'include') diff --git a/include/linux/lguest.h b/include/linux/lguest.h index 175e63f4a8c0..7bc1440fc473 100644 --- a/include/linux/lguest.h +++ b/include/linux/lguest.h @@ -30,6 +30,10 @@ struct lguest_data /* Wallclock time set by the Host. */ struct timespec time; + /* Interrupt pending set by the Host. The Guest should do a hypercall + * if it re-enables interrupts and sees this set (to X86_EFLAGS_IF). */ + int irq_pending; + /* Async hypercall ring. Instead of directly making hypercalls, we can * place them in here for processing the next time the Host wants. * This batching can be quite efficient. */ -- cgit v1.2.3 From df60aeef4f4fe0645d9a195a7689005520422de5 Mon Sep 17 00:00:00 2001 From: Rusty Russell Date: Fri, 12 Jun 2009 22:27:09 -0600 Subject: lguest: use eventfds for device notification Currently, when a Guest wants to perform I/O it calls LHCALL_NOTIFY with an address: the main Launcher process returns with this address, and figures out what device to run. A far nicer model is to let processes bind an eventfd to an address: if we find one, we simply signal the eventfd. Signed-off-by: Rusty Russell Cc: Davide Libenzi --- drivers/lguest/Kconfig | 2 +- drivers/lguest/core.c | 8 ++-- drivers/lguest/lg.h | 13 ++++++ drivers/lguest/lguest_user.c | 98 ++++++++++++++++++++++++++++++++++++++++- include/linux/lguest_launcher.h | 1 + 5 files changed, 116 insertions(+), 6 deletions(-) (limited to 'include') diff --git a/drivers/lguest/Kconfig b/drivers/lguest/Kconfig index 8f63845db830..0aaa0597a622 100644 --- a/drivers/lguest/Kconfig +++ b/drivers/lguest/Kconfig @@ -1,6 +1,6 @@ config LGUEST tristate "Linux hypervisor example code" - depends on X86_32 && EXPERIMENTAL && FUTEX + depends on X86_32 && EXPERIMENTAL && EVENTFD select HVC_DRIVER ---help--- This is a very simple module which allows you to run diff --git a/drivers/lguest/core.c b/drivers/lguest/core.c index d0298dc45d97..508569c9571a 100644 --- a/drivers/lguest/core.c +++ b/drivers/lguest/core.c @@ -198,9 +198,11 @@ int run_guest(struct lg_cpu *cpu, unsigned long __user *user) /* It's possible the Guest did a NOTIFY hypercall to the * Launcher, in which case we return from the read() now. */ if (cpu->pending_notify) { - if (put_user(cpu->pending_notify, user)) - return -EFAULT; - return sizeof(cpu->pending_notify); + if (!send_notify_to_eventfd(cpu)) { + if (put_user(cpu->pending_notify, user)) + return -EFAULT; + return sizeof(cpu->pending_notify); + } } /* Check for signals */ diff --git a/drivers/lguest/lg.h b/drivers/lguest/lg.h index 040cb70780e7..32fefdc6ad3e 100644 --- a/drivers/lguest/lg.h +++ b/drivers/lguest/lg.h @@ -82,6 +82,16 @@ struct lg_cpu { struct lg_cpu_arch arch; }; +struct lg_eventfd { + unsigned long addr; + struct file *event; +}; + +struct lg_eventfd_map { + unsigned int num; + struct lg_eventfd map[]; +}; + /* The private info the thread maintains about the guest. */ struct lguest { @@ -102,6 +112,8 @@ struct lguest unsigned int stack_pages; u32 tsc_khz; + struct lg_eventfd_map *eventfds; + /* Dead? */ const char *dead; }; @@ -154,6 +166,7 @@ void setup_default_idt_entries(struct lguest_ro_state *state, void copy_traps(const struct lg_cpu *cpu, struct desc_struct *idt, const unsigned long *def); void guest_set_clockevent(struct lg_cpu *cpu, unsigned long delta); +bool send_notify_to_eventfd(struct lg_cpu *cpu); void init_clockdev(struct lg_cpu *cpu); bool check_syscall_vector(struct lguest *lg); int init_interrupts(void); diff --git a/drivers/lguest/lguest_user.c b/drivers/lguest/lguest_user.c index 1982b45bd935..f6bf255f1837 100644 --- a/drivers/lguest/lguest_user.c +++ b/drivers/lguest/lguest_user.c @@ -7,6 +7,8 @@ #include #include #include +#include +#include #include "lg.h" /*L:055 When something happens, the Waker process needs a way to stop the @@ -35,6 +37,81 @@ static int break_guest_out(struct lg_cpu *cpu, const unsigned long __user*input) } } +bool send_notify_to_eventfd(struct lg_cpu *cpu) +{ + unsigned int i; + struct lg_eventfd_map *map; + + /* lg->eventfds is RCU-protected */ + rcu_read_lock(); + map = rcu_dereference(cpu->lg->eventfds); + for (i = 0; i < map->num; i++) { + if (map->map[i].addr == cpu->pending_notify) { + eventfd_signal(map->map[i].event, 1); + cpu->pending_notify = 0; + break; + } + } + rcu_read_unlock(); + return cpu->pending_notify == 0; +} + +static int add_eventfd(struct lguest *lg, unsigned long addr, int fd) +{ + struct lg_eventfd_map *new, *old = lg->eventfds; + + if (!addr) + return -EINVAL; + + /* Replace the old array with the new one, carefully: others can + * be accessing it at the same time */ + new = kmalloc(sizeof(*new) + sizeof(new->map[0]) * (old->num + 1), + GFP_KERNEL); + if (!new) + return -ENOMEM; + + /* First make identical copy. */ + memcpy(new->map, old->map, sizeof(old->map[0]) * old->num); + new->num = old->num; + + /* Now append new entry. */ + new->map[new->num].addr = addr; + new->map[new->num].event = eventfd_fget(fd); + if (IS_ERR(new->map[new->num].event)) { + kfree(new); + return PTR_ERR(new->map[new->num].event); + } + new->num++; + + /* Now put new one in place. */ + rcu_assign_pointer(lg->eventfds, new); + + /* We're not in a big hurry. Wait until noone's looking at old + * version, then delete it. */ + synchronize_rcu(); + kfree(old); + + return 0; +} + +static int attach_eventfd(struct lguest *lg, const unsigned long __user *input) +{ + unsigned long addr, fd; + int err; + + if (get_user(addr, input) != 0) + return -EFAULT; + input++; + if (get_user(fd, input) != 0) + return -EFAULT; + + mutex_lock(&lguest_lock); + err = add_eventfd(lg, addr, fd); + mutex_unlock(&lguest_lock); + + return 0; +} + /*L:050 Sending an interrupt is done by writing LHREQ_IRQ and an interrupt * number to /dev/lguest. */ static int user_send_irq(struct lg_cpu *cpu, const unsigned long __user *input) @@ -184,6 +261,13 @@ static int initialize(struct file *file, const unsigned long __user *input) goto unlock; } + lg->eventfds = kmalloc(sizeof(*lg->eventfds), GFP_KERNEL); + if (!lg->eventfds) { + err = -ENOMEM; + goto free_lg; + } + lg->eventfds->num = 0; + /* Populate the easy fields of our "struct lguest" */ lg->mem_base = (void __user *)args[0]; lg->pfn_limit = args[1]; @@ -191,7 +275,7 @@ static int initialize(struct file *file, const unsigned long __user *input) /* This is the first cpu (cpu 0) and it will start booting at args[2] */ err = lg_cpu_start(&lg->cpus[0], 0, args[2]); if (err) - goto release_guest; + goto free_eventfds; /* Initialize the Guest's shadow page tables, using the toplevel * address the Launcher gave us. This allocates memory, so can fail. */ @@ -210,7 +294,9 @@ static int initialize(struct file *file, const unsigned long __user *input) free_regs: /* FIXME: This should be in free_vcpu */ free_page(lg->cpus[0].regs_page); -release_guest: +free_eventfds: + kfree(lg->eventfds); +free_lg: kfree(lg); unlock: mutex_unlock(&lguest_lock); @@ -260,6 +346,8 @@ static ssize_t write(struct file *file, const char __user *in, return user_send_irq(cpu, input); case LHREQ_BREAK: return break_guest_out(cpu, input); + case LHREQ_EVENTFD: + return attach_eventfd(lg, input); default: return -EINVAL; } @@ -297,6 +385,12 @@ static int close(struct inode *inode, struct file *file) * the Launcher's memory management structure. */ mmput(lg->cpus[i].mm); } + + /* Release any eventfds they registered. */ + for (i = 0; i < lg->eventfds->num; i++) + fput(lg->eventfds->map[i].event); + kfree(lg->eventfds); + /* If lg->dead doesn't contain an error code it will be NULL or a * kmalloc()ed string, either of which is ok to hand to kfree(). */ if (!IS_ERR(lg->dead)) diff --git a/include/linux/lguest_launcher.h b/include/linux/lguest_launcher.h index a53407a4165c..9de964b90586 100644 --- a/include/linux/lguest_launcher.h +++ b/include/linux/lguest_launcher.h @@ -58,6 +58,7 @@ enum lguest_req LHREQ_GETDMA, /* No longer used */ LHREQ_IRQ, /* + irq */ LHREQ_BREAK, /* + on/off flag (on blocks until someone does off) */ + LHREQ_EVENTFD, /* + address, fd. */ }; /* The alignment to use between consumer and producer parts of vring. -- cgit v1.2.3 From 5dac051bc6030963181b69faddd9e0ad04f85fa8 Mon Sep 17 00:00:00 2001 From: Rusty Russell Date: Fri, 12 Jun 2009 22:27:10 -0600 Subject: lguest: remove obsolete LHREQ_BREAK call We no longer need an efficient mechanism to force the Guest back into host userspace, as each device is serviced without bothering the main Guest process (aka. the Launcher). Signed-off-by: Rusty Russell --- drivers/lguest/core.c | 11 +++-------- drivers/lguest/lg.h | 4 +--- drivers/lguest/lguest_user.c | 31 ------------------------------- include/linux/lguest_launcher.h | 2 +- 4 files changed, 5 insertions(+), 43 deletions(-) (limited to 'include') diff --git a/drivers/lguest/core.c b/drivers/lguest/core.c index 508569c9571a..a6974e9b8ebf 100644 --- a/drivers/lguest/core.c +++ b/drivers/lguest/core.c @@ -209,10 +209,6 @@ int run_guest(struct lg_cpu *cpu, unsigned long __user *user) if (signal_pending(current)) return -ERESTARTSYS; - /* If Waker set break_out, return to Launcher. */ - if (cpu->break_out) - return -EAGAIN; - /* Check if there are any interrupts which can be delivered now: * if so, this sets up the hander to be executed when we next * run the Guest. */ @@ -231,13 +227,12 @@ int run_guest(struct lg_cpu *cpu, unsigned long __user *user) break; /* If the Guest asked to be stopped, we sleep. The Guest's - * clock timer or LHREQ_BREAK from the Waker will wake us. */ + * clock timer will wake us. */ if (cpu->halted) { set_current_state(TASK_INTERRUPTIBLE); - /* Just before we sleep, make sure nothing snuck in + /* Just before we sleep, make sure no interrupt snuck in * which we should be doing. */ - if (interrupt_pending(cpu, &more) < LGUEST_IRQS - || cpu->break_out) + if (interrupt_pending(cpu, &more) < LGUEST_IRQS) set_current_state(TASK_RUNNING); else schedule(); diff --git a/drivers/lguest/lg.h b/drivers/lguest/lg.h index 32fefdc6ad3e..d4e8979735cb 100644 --- a/drivers/lguest/lg.h +++ b/drivers/lguest/lg.h @@ -71,9 +71,7 @@ struct lg_cpu { /* Virtual clock device */ struct hrtimer hrt; - /* Do we need to stop what we're doing and return to userspace? */ - int break_out; - wait_queue_head_t break_wq; + /* Did the Guest tell us to halt? */ int halted; /* Pending virtual interrupts */ diff --git a/drivers/lguest/lguest_user.c b/drivers/lguest/lguest_user.c index f6bf255f1837..32e297121058 100644 --- a/drivers/lguest/lguest_user.c +++ b/drivers/lguest/lguest_user.c @@ -11,32 +11,6 @@ #include #include "lg.h" -/*L:055 When something happens, the Waker process needs a way to stop the - * kernel running the Guest and return to the Launcher. So the Waker writes - * LHREQ_BREAK and the value "1" to /dev/lguest to do this. Once the Launcher - * has done whatever needs attention, it writes LHREQ_BREAK and "0" to release - * the Waker. */ -static int break_guest_out(struct lg_cpu *cpu, const unsigned long __user*input) -{ - unsigned long on; - - /* Fetch whether they're turning break on or off. */ - if (get_user(on, input) != 0) - return -EFAULT; - - if (on) { - cpu->break_out = 1; - if (!wake_up_process(cpu->tsk)) - kick_process(cpu->tsk); - /* Wait for them to reset it */ - return wait_event_interruptible(cpu->break_wq, !cpu->break_out); - } else { - cpu->break_out = 0; - wake_up(&cpu->break_wq); - return 0; - } -} - bool send_notify_to_eventfd(struct lg_cpu *cpu) { unsigned int i; @@ -202,9 +176,6 @@ static int lg_cpu_start(struct lg_cpu *cpu, unsigned id, unsigned long start_ip) * address. */ lguest_arch_setup_regs(cpu, start_ip); - /* Initialize the queue for the Waker to wait on */ - init_waitqueue_head(&cpu->break_wq); - /* We keep a pointer to the Launcher task (ie. current task) for when * other Guests want to wake this one (eg. console input). */ cpu->tsk = current; @@ -344,8 +315,6 @@ static ssize_t write(struct file *file, const char __user *in, return initialize(file, input); case LHREQ_IRQ: return user_send_irq(cpu, input); - case LHREQ_BREAK: - return break_guest_out(cpu, input); case LHREQ_EVENTFD: return attach_eventfd(lg, input); default: diff --git a/include/linux/lguest_launcher.h b/include/linux/lguest_launcher.h index 9de964b90586..bfefbdf7498a 100644 --- a/include/linux/lguest_launcher.h +++ b/include/linux/lguest_launcher.h @@ -57,7 +57,7 @@ enum lguest_req LHREQ_INITIALIZE, /* + base, pfnlimit, start */ LHREQ_GETDMA, /* No longer used */ LHREQ_IRQ, /* + irq */ - LHREQ_BREAK, /* + on/off flag (on blocks until someone does off) */ + LHREQ_BREAK, /* No longer used */ LHREQ_EVENTFD, /* + address, fd. */ }; -- cgit v1.2.3