| Message ID | 1586257192-58369-1-git-send-email-zhe.he@windriver.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/2] eventfd: Make wake counter work for single fd instead of all | expand |
On 4/7/20 3:59 AM, zhe.he@windriver.com wrote: > From: He Zhe <zhe.he@windriver.com> > > commit b5e683d5cab8 ("eventfd: track eventfd_signal() recursion depth") > introduces a percpu counter that tracks the percpu recursion depth and > warn if it greater than one, to avoid potential deadlock and stack > overflow. > > However sometimes different eventfds may be used in parallel. > Specifically, when high network load goes through kvm and vhost, working > as below, it would trigger the following call trace. > > - 100.00% > - 66.51% > ret_from_fork > kthread > - vhost_worker > - 33.47% handle_tx_kick > handle_tx > handle_tx_copy > vhost_tx_batch.isra.0 > vhost_add_used_and_signal_n > eventfd_signal > - 33.05% handle_rx_net > handle_rx > vhost_add_used_and_signal_n > eventfd_signal > - 33.49% > ioctl > entry_SYSCALL_64_after_hwframe > do_syscall_64 > __x64_sys_ioctl > ksys_ioctl > do_vfs_ioctl > kvm_vcpu_ioctl > kvm_arch_vcpu_ioctl_run > vmx_handle_exit > handle_ept_misconfig > kvm_io_bus_write > __kvm_io_bus_write > eventfd_signal > > 001: WARNING: CPU: 1 PID: 1503 at fs/eventfd.c:73 eventfd_signal+0x85/0xa0 > ---- snip ---- > 001: Call Trace: > 001: vhost_signal+0x15e/0x1b0 [vhost] > 001: vhost_add_used_and_signal_n+0x2b/0x40 [vhost] > 001: handle_rx+0xb9/0x900 [vhost_net] > 001: handle_rx_net+0x15/0x20 [vhost_net] > 001: vhost_worker+0xbe/0x120 [vhost] > 001: kthread+0x106/0x140 > 001: ? log_used.part.0+0x20/0x20 [vhost] > 001: ? kthread_park+0x90/0x90 > 001: ret_from_fork+0x35/0x40 > 001: ---[ end trace 0000000000000003 ]--- > > This patch moves the percpu counter into eventfd control structure and > does the clean-ups, so that eventfd can still be protected from deadlock > while allowing different ones to work in parallel. > > As to potential stack overflow, we might want to figure out a better > solution in the future to warn when the stack is about to overflow so it > can be better utilized, rather than break the working flow when just the > second one comes. This doesn't work for the infinite recursion case, the state has to be global, or per thread.
On 4/8/20 4:06 AM, Jens Axboe wrote: > On 4/7/20 3:59 AM, zhe.he@windriver.com wrote: >> From: He Zhe <zhe.he@windriver.com> >> >> commit b5e683d5cab8 ("eventfd: track eventfd_signal() recursion depth") >> introduces a percpu counter that tracks the percpu recursion depth and >> warn if it greater than one, to avoid potential deadlock and stack >> overflow. >> >> However sometimes different eventfds may be used in parallel. >> Specifically, when high network load goes through kvm and vhost, working >> as below, it would trigger the following call trace. >> >> - 100.00% >> - 66.51% >> ret_from_fork >> kthread >> - vhost_worker >> - 33.47% handle_tx_kick >> handle_tx >> handle_tx_copy >> vhost_tx_batch.isra.0 >> vhost_add_used_and_signal_n >> eventfd_signal >> - 33.05% handle_rx_net >> handle_rx >> vhost_add_used_and_signal_n >> eventfd_signal >> - 33.49% >> ioctl >> entry_SYSCALL_64_after_hwframe >> do_syscall_64 >> __x64_sys_ioctl >> ksys_ioctl >> do_vfs_ioctl >> kvm_vcpu_ioctl >> kvm_arch_vcpu_ioctl_run >> vmx_handle_exit >> handle_ept_misconfig >> kvm_io_bus_write >> __kvm_io_bus_write >> eventfd_signal >> >> 001: WARNING: CPU: 1 PID: 1503 at fs/eventfd.c:73 eventfd_signal+0x85/0xa0 >> ---- snip ---- >> 001: Call Trace: >> 001: vhost_signal+0x15e/0x1b0 [vhost] >> 001: vhost_add_used_and_signal_n+0x2b/0x40 [vhost] >> 001: handle_rx+0xb9/0x900 [vhost_net] >> 001: handle_rx_net+0x15/0x20 [vhost_net] >> 001: vhost_worker+0xbe/0x120 [vhost] >> 001: kthread+0x106/0x140 >> 001: ? log_used.part.0+0x20/0x20 [vhost] >> 001: ? kthread_park+0x90/0x90 >> 001: ret_from_fork+0x35/0x40 >> 001: ---[ end trace 0000000000000003 ]--- >> >> This patch moves the percpu counter into eventfd control structure and >> does the clean-ups, so that eventfd can still be protected from deadlock >> while allowing different ones to work in parallel. >> >> As to potential stack overflow, we might want to figure out a better >> solution in the future to warn when the stack is about to overflow so it >> can be better utilized, rather than break the working flow when just the >> second one comes. > This doesn't work for the infinite recursion case, the state has to be > global, or per thread. Thanks, but I'm not very clear about why the counter has to be global or per thread. If the recursion happens on the same eventfd, the attempt to re-grab the same ctx->wqh.lock would be blocked by the fd-specific counter in this patch. If the recursion happens with a chain of different eventfds, that might lead to a stack overflow issue. The issue should be handled but it seems unnecessary to stop the just the second ring(when the counter is going to be 2) of the chain. Specifically in the vhost case, it runs very likely with heavy network load which generates loads of eventfd_signal. Delaying the eventfd_signal to worker threads will still end up violating the global counter later and failing as above. So we might want to take care of the potential overflow later, hopefully with a measurement that can tell us if it's about to overflow. Zhe >
On 4/9/20 3:37 AM, He Zhe wrote: > > > On 4/8/20 4:06 AM, Jens Axboe wrote: >> On 4/7/20 3:59 AM, zhe.he@windriver.com wrote: >>> From: He Zhe <zhe.he@windriver.com> >>> >>> commit b5e683d5cab8 ("eventfd: track eventfd_signal() recursion depth") >>> introduces a percpu counter that tracks the percpu recursion depth and >>> warn if it greater than one, to avoid potential deadlock and stack >>> overflow. >>> >>> However sometimes different eventfds may be used in parallel. >>> Specifically, when high network load goes through kvm and vhost, working >>> as below, it would trigger the following call trace. >>> >>> - 100.00% >>> - 66.51% >>> ret_from_fork >>> kthread >>> - vhost_worker >>> - 33.47% handle_tx_kick >>> handle_tx >>> handle_tx_copy >>> vhost_tx_batch.isra.0 >>> vhost_add_used_and_signal_n >>> eventfd_signal >>> - 33.05% handle_rx_net >>> handle_rx >>> vhost_add_used_and_signal_n >>> eventfd_signal >>> - 33.49% >>> ioctl >>> entry_SYSCALL_64_after_hwframe >>> do_syscall_64 >>> __x64_sys_ioctl >>> ksys_ioctl >>> do_vfs_ioctl >>> kvm_vcpu_ioctl >>> kvm_arch_vcpu_ioctl_run >>> vmx_handle_exit >>> handle_ept_misconfig >>> kvm_io_bus_write >>> __kvm_io_bus_write >>> eventfd_signal >>> >>> 001: WARNING: CPU: 1 PID: 1503 at fs/eventfd.c:73 eventfd_signal+0x85/0xa0 >>> ---- snip ---- >>> 001: Call Trace: >>> 001: vhost_signal+0x15e/0x1b0 [vhost] >>> 001: vhost_add_used_and_signal_n+0x2b/0x40 [vhost] >>> 001: handle_rx+0xb9/0x900 [vhost_net] >>> 001: handle_rx_net+0x15/0x20 [vhost_net] >>> 001: vhost_worker+0xbe/0x120 [vhost] >>> 001: kthread+0x106/0x140 >>> 001: ? log_used.part.0+0x20/0x20 [vhost] >>> 001: ? kthread_park+0x90/0x90 >>> 001: ret_from_fork+0x35/0x40 >>> 001: ---[ end trace 0000000000000003 ]--- >>> >>> This patch moves the percpu counter into eventfd control structure and >>> does the clean-ups, so that eventfd can still be protected from deadlock >>> while allowing different ones to work in parallel. >>> >>> As to potential stack overflow, we might want to figure out a better >>> solution in the future to warn when the stack is about to overflow so it >>> can be better utilized, rather than break the working flow when just the >>> second one comes. >> This doesn't work for the infinite recursion case, the state has to be >> global, or per thread. > > Thanks, but I'm not very clear about why the counter has to be global > or per thread. > > If the recursion happens on the same eventfd, the attempt to re-grab > the same ctx->wqh.lock would be blocked by the fd-specific counter in > this patch. > > If the recursion happens with a chain of different eventfds, that > might lead to a stack overflow issue. The issue should be handled but > it seems unnecessary to stop the just the second ring(when the counter > is going to be 2) of the chain. > > Specifically in the vhost case, it runs very likely with heavy network > load which generates loads of eventfd_signal. Delaying the > eventfd_signal to worker threads will still end up violating the > global counter later and failing as above. > > So we might want to take care of the potential overflow later, > hopefully with a measurement that can tell us if it's about to > overflow. The worry is different eventfds, recursion on a single one could be detected by keeping state in the ctx itself. And yeah, I agree that one level isn't very deep, but wakeup chains can be deep and we can't allow a whole lot more. I'm sure folks would be open to increasing it, if some worst case kind of data was collected to prove it's fine to go deeper.
On 4/9/20 11:44 PM, Jens Axboe wrote: > On 4/9/20 3:37 AM, He Zhe wrote: >> >> On 4/8/20 4:06 AM, Jens Axboe wrote: >>> On 4/7/20 3:59 AM, zhe.he@windriver.com wrote: >>>> From: He Zhe <zhe.he@windriver.com> >>>> >>>> commit b5e683d5cab8 ("eventfd: track eventfd_signal() recursion depth") >>>> introduces a percpu counter that tracks the percpu recursion depth and >>>> warn if it greater than one, to avoid potential deadlock and stack >>>> overflow. >>>> >>>> However sometimes different eventfds may be used in parallel. >>>> Specifically, when high network load goes through kvm and vhost, working >>>> as below, it would trigger the following call trace. >>>> >>>> - 100.00% >>>> - 66.51% >>>> ret_from_fork >>>> kthread >>>> - vhost_worker >>>> - 33.47% handle_tx_kick >>>> handle_tx >>>> handle_tx_copy >>>> vhost_tx_batch.isra.0 >>>> vhost_add_used_and_signal_n >>>> eventfd_signal >>>> - 33.05% handle_rx_net >>>> handle_rx >>>> vhost_add_used_and_signal_n >>>> eventfd_signal >>>> - 33.49% >>>> ioctl >>>> entry_SYSCALL_64_after_hwframe >>>> do_syscall_64 >>>> __x64_sys_ioctl >>>> ksys_ioctl >>>> do_vfs_ioctl >>>> kvm_vcpu_ioctl >>>> kvm_arch_vcpu_ioctl_run >>>> vmx_handle_exit >>>> handle_ept_misconfig >>>> kvm_io_bus_write >>>> __kvm_io_bus_write >>>> eventfd_signal >>>> >>>> 001: WARNING: CPU: 1 PID: 1503 at fs/eventfd.c:73 eventfd_signal+0x85/0xa0 >>>> ---- snip ---- >>>> 001: Call Trace: >>>> 001: vhost_signal+0x15e/0x1b0 [vhost] >>>> 001: vhost_add_used_and_signal_n+0x2b/0x40 [vhost] >>>> 001: handle_rx+0xb9/0x900 [vhost_net] >>>> 001: handle_rx_net+0x15/0x20 [vhost_net] >>>> 001: vhost_worker+0xbe/0x120 [vhost] >>>> 001: kthread+0x106/0x140 >>>> 001: ? log_used.part.0+0x20/0x20 [vhost] >>>> 001: ? kthread_park+0x90/0x90 >>>> 001: ret_from_fork+0x35/0x40 >>>> 001: ---[ end trace 0000000000000003 ]--- >>>> >>>> This patch moves the percpu counter into eventfd control structure and >>>> does the clean-ups, so that eventfd can still be protected from deadlock >>>> while allowing different ones to work in parallel. >>>> >>>> As to potential stack overflow, we might want to figure out a better >>>> solution in the future to warn when the stack is about to overflow so it >>>> can be better utilized, rather than break the working flow when just the >>>> second one comes. >>> This doesn't work for the infinite recursion case, the state has to be >>> global, or per thread. >> Thanks, but I'm not very clear about why the counter has to be global >> or per thread. >> >> If the recursion happens on the same eventfd, the attempt to re-grab >> the same ctx->wqh.lock would be blocked by the fd-specific counter in >> this patch. >> >> If the recursion happens with a chain of different eventfds, that >> might lead to a stack overflow issue. The issue should be handled but >> it seems unnecessary to stop the just the second ring(when the counter >> is going to be 2) of the chain. >> >> Specifically in the vhost case, it runs very likely with heavy network >> load which generates loads of eventfd_signal. Delaying the >> eventfd_signal to worker threads will still end up violating the >> global counter later and failing as above. >> >> So we might want to take care of the potential overflow later, >> hopefully with a measurement that can tell us if it's about to >> overflow. > The worry is different eventfds, recursion on a single one could be > detected by keeping state in the ctx itself. And yeah, I agree that one > level isn't very deep, but wakeup chains can be deep and we can't allow > a whole lot more. I'm sure folks would be open to increasing it, if some > worst case kind of data was collected to prove it's fine to go deeper. OK, thanks. v2 will be sent. Zhe >
diff --git a/fs/eventfd.c b/fs/eventfd.c index 78e41c7..bb4385a 100644 --- a/fs/eventfd.c +++ b/fs/eventfd.c @@ -24,26 +24,8 @@ #include <linux/seq_file.h> #include <linux/idr.h> -DEFINE_PER_CPU(int, eventfd_wake_count); - static DEFINE_IDA(eventfd_ida); -struct eventfd_ctx { - struct kref kref; - wait_queue_head_t wqh; - /* - * Every time that a write(2) is performed on an eventfd, the - * value of the __u64 being written is added to "count" and a - * wakeup is performed on "wqh". A read(2) will return the "count" - * value to userspace, and will reset "count" to zero. The kernel - * side eventfd_signal() also, adds to the "count" counter and - * issue a wakeup. - */ - __u64 count; - unsigned int flags; - int id; -}; - /** * eventfd_signal - Adds @n to the eventfd counter. * @ctx: [in] Pointer to the eventfd context. @@ -70,17 +52,17 @@ __u64 eventfd_signal(struct eventfd_ctx *ctx, __u64 n) * it returns true, the eventfd_signal() call should be deferred to a * safe context. */ - if (WARN_ON_ONCE(this_cpu_read(eventfd_wake_count))) + if (WARN_ON_ONCE(this_cpu_read(*ctx->wake_count))) return 0; spin_lock_irqsave(&ctx->wqh.lock, flags); - this_cpu_inc(eventfd_wake_count); + this_cpu_inc(*ctx->wake_count); if (ULLONG_MAX - ctx->count < n) n = ULLONG_MAX - ctx->count; ctx->count += n; if (waitqueue_active(&ctx->wqh)) wake_up_locked_poll(&ctx->wqh, EPOLLIN); - this_cpu_dec(eventfd_wake_count); + this_cpu_dec(*ctx->wake_count); spin_unlock_irqrestore(&ctx->wqh.lock, flags); return n; @@ -406,7 +388,7 @@ EXPORT_SYMBOL_GPL(eventfd_ctx_fileget); static int do_eventfd(unsigned int count, int flags) { struct eventfd_ctx *ctx; - int fd; + int fd, cpu; /* Check the EFD_* constants for consistency. */ BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC); @@ -424,6 +406,11 @@ static int do_eventfd(unsigned int count, int flags) ctx->count = count; ctx->flags = flags; ctx->id = ida_simple_get(&eventfd_ida, 0, 0, GFP_KERNEL); + ctx->wake_count = alloc_percpu(typeof(*ctx->wake_count)); + if (!ctx->wake_count) + return -ENOMEM; + for_each_possible_cpu(cpu) + *per_cpu_ptr(ctx->wake_count, cpu) = 0; fd = anon_inode_getfd("[eventfd]", &eventfd_fops, ctx, O_RDWR | (flags & EFD_SHARED_FCNTL_FLAGS)); diff --git a/include/linux/eventfd.h b/include/linux/eventfd.h index dc4fd8a..b0d0f44 100644 --- a/include/linux/eventfd.h +++ b/include/linux/eventfd.h @@ -29,7 +29,23 @@ #define EFD_SHARED_FCNTL_FLAGS (O_CLOEXEC | O_NONBLOCK) #define EFD_FLAGS_SET (EFD_SHARED_FCNTL_FLAGS | EFD_SEMAPHORE) -struct eventfd_ctx; +struct eventfd_ctx { + struct kref kref; + wait_queue_head_t wqh; + /* + * Every time that a write(2) is performed on an eventfd, the + * value of the __u64 being written is added to "count" and a + * wakeup is performed on "wqh". A read(2) will return the "count" + * value to userspace, and will reset "count" to zero. The kernel + * side eventfd_signal() also, adds to the "count" counter and + * issue a wakeup. + */ + __u64 count; + unsigned int flags; + int id; + int __percpu *wake_count; +}; + struct file; #ifdef CONFIG_EVENTFD @@ -42,11 +58,9 @@ __u64 eventfd_signal(struct eventfd_ctx *ctx, __u64 n); int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx, wait_queue_entry_t *wait, __u64 *cnt); -DECLARE_PER_CPU(int, eventfd_wake_count); - -static inline bool eventfd_signal_count(void) +static inline bool eventfd_signal_count(struct eventfd_ctx *ctx) { - return this_cpu_read(eventfd_wake_count); + return this_cpu_read(*ctx->wake_count); } #else /* CONFIG_EVENTFD */