Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753079AbdHOCdK (ORCPT ); Mon, 14 Aug 2017 22:33:10 -0400 Received: from aserp1040.oracle.com ([141.146.126.69]:27676 "EHLO aserp1040.oracle.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752873AbdHOCdJ (ORCPT ); Mon, 14 Aug 2017 22:33:09 -0400 Subject: Re: [PATCH v3 09/13] xen/pvcalls: implement sendmsg To: Stefano Stabellini , xen-devel@lists.xen.org Cc: linux-kernel@vger.kernel.org, jgross@suse.com, Stefano Stabellini References: <1501541855-7354-1-git-send-email-sstabellini@kernel.org> <1501541855-7354-9-git-send-email-sstabellini@kernel.org> From: Boris Ostrovsky Message-ID: <4d87fa71-65fd-f4c6-7242-e3e356d1f875@oracle.com> Date: Mon, 14 Aug 2017 22:31:53 -0400 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.2.1 MIME-Version: 1.0 In-Reply-To: <1501541855-7354-9-git-send-email-sstabellini@kernel.org> Content-Type: text/plain; charset=utf-8; format=flowed Content-Language: en-US Content-Transfer-Encoding: 7bit X-Source-IP: aserv0021.oracle.com [141.146.126.233] Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 5877 Lines: 196 On 07/31/2017 06:57 PM, Stefano Stabellini wrote: > Send data to an active socket by copying data to the "out" ring. Take > the active socket out_mutex so that only one function can access the > ring at any given time. > > If not enough room is available on the ring, rather than returning > immediately or sleep-waiting, spin for up to 5000 cycles. This small > optimization turns out to improve performance significantly. > > Signed-off-by: Stefano Stabellini > CC: boris.ostrovsky@oracle.com > CC: jgross@suse.com > --- > drivers/xen/pvcalls-front.c | 109 ++++++++++++++++++++++++++++++++++++++++++++ > drivers/xen/pvcalls-front.h | 3 ++ > 2 files changed, 112 insertions(+) > > diff --git a/drivers/xen/pvcalls-front.c b/drivers/xen/pvcalls-front.c > index f83b910..369acde 100644 > --- a/drivers/xen/pvcalls-front.c > +++ b/drivers/xen/pvcalls-front.c > @@ -29,6 +29,7 @@ > #define PVCALLS_INVALID_ID UINT_MAX > #define PVCALLS_RING_ORDER XENBUS_MAX_RING_GRANT_ORDER > #define PVCALLS_NR_REQ_PER_RING __CONST_RING_SIZE(xen_pvcalls, XEN_PAGE_SIZE) > +#define PVCALLS_FRONT_MAX_SPIN 5000 > > struct pvcalls_bedata { > struct xen_pvcalls_front_ring ring; > @@ -88,6 +89,22 @@ static inline int get_request(struct pvcalls_bedata *bedata, int *req_id) > return 0; > } > > +static int pvcalls_front_write_todo(struct sock_mapping *map) > +{ > + struct pvcalls_data_intf *intf = map->active.ring; > + RING_IDX cons, prod, size = XEN_FLEX_RING_SIZE(intf->ring_order); > + int32_t error; > + > + cons = intf->out_cons; > + prod = intf->out_prod; > + error = intf->out_error; > + if (error == -ENOTCONN) > + return 0; > + if (error != 0) > + return error; > + return size - pvcalls_queued(prod, cons, size); Do you ever look at actual return value except whether it's zero or not? Both here and in the poll patch you look for !=0 and never check for an error. > +} > + > static irqreturn_t pvcalls_front_event_handler(int irq, void *dev_id) > { > struct xenbus_device *dev = dev_id; > @@ -325,6 +342,98 @@ int pvcalls_front_connect(struct socket *sock, struct sockaddr *addr, > return ret; > } > > +static int __write_ring(struct pvcalls_data_intf *intf, > + struct pvcalls_data *data, > + struct iov_iter *msg_iter, > + size_t len) > +{ > + RING_IDX cons, prod, size, masked_prod, masked_cons; > + RING_IDX array_size = XEN_FLEX_RING_SIZE(intf->ring_order); > + int32_t error; > + > + cons = intf->out_cons; > + prod = intf->out_prod; > + error = intf->out_error; > + /* read indexes before continuing */ > + virt_mb(); > + > + if (error < 0) > + return error; This can be done before the barrier. (In fact, this can be done first thing, before cons and prod are read). > + > + size = pvcalls_queued(prod, cons, array_size); > + if (size >= array_size) > + return 0; > + if (len > array_size - size) > + len = array_size - size; > + > + masked_prod = pvcalls_mask(prod, array_size); > + masked_cons = pvcalls_mask(cons, array_size); > + > + if (masked_prod < masked_cons) { > + copy_from_iter(data->out + masked_prod, len, msg_iter); > + } else { > + if (len > array_size - masked_prod) { > + copy_from_iter(data->out + masked_prod, > + array_size - masked_prod, msg_iter); > + copy_from_iter(data->out, > + len - (array_size - masked_prod), > + msg_iter); > + } else { > + copy_from_iter(data->out + masked_prod, len, msg_iter); > + } > + } > + /* write to ring before updating pointer */ > + virt_wmb(); > + intf->out_prod += len; > + > + return len; You are returning size_t. I actually was going to ask in one of the previous patches whether using int for sizes was correct. Unfortunately I can't remember which struct/function I was looking at ;-( Of course, you are also possibly returning a (negative) error here. > +} > + > +int pvcalls_front_sendmsg(struct socket *sock, struct msghdr *msg, > + size_t len) > +{ > + struct pvcalls_bedata *bedata; > + struct sock_mapping *map; > + int sent = 0, tot_sent = 0; 'sent' doesn't need to be initialized. > + int count = 0, flags; > + > + if (!pvcalls_front_dev) > + return -ENOTCONN; > + bedata = dev_get_drvdata(&pvcalls_front_dev->dev); > + > + map = (struct sock_mapping *) READ_ONCE(sock->sk->sk_send_head); > + if (!map) > + return -ENOTSOCK; IIRC the error value for sk_send_head being zero is inconsistent across patches. > + > + flags = msg->msg_flags; > + if (flags & (MSG_CONFIRM|MSG_DONTROUTE|MSG_EOR|MSG_OOB)) > + return -EOPNOTSUPP; > + > + mutex_lock(&map->active.out_mutex); > + if ((flags & MSG_DONTWAIT) && !pvcalls_front_write_todo(map)) { > + mutex_unlock(&map->active.out_mutex); > + return -EAGAIN; > + } > + > +again: > + count++; > + sent = __write_ring(map->active.ring, > + &map->active.data, &msg->msg_iter, > + len); > + if (sent > 0) { > + len -= sent; > + tot_sent += sent; > + notify_remote_via_irq(map->active.irq); > + } > + if (sent >= 0 && len > 0 && count < PVCALLS_FRONT_MAX_SPIN) > + goto again; > + if (sent < 0) > + tot_sent = sent; What does it mean when an error is detected on the interface? Does it need to be somehow propagated to the caller? -boris > + > + mutex_unlock(&map->active.out_mutex); > + return tot_sent; > +} > + > int pvcalls_front_bind(struct socket *sock, struct sockaddr *addr, int addr_len) > { > struct pvcalls_bedata *bedata; > diff --git a/drivers/xen/pvcalls-front.h b/drivers/xen/pvcalls-front.h > index ab4f1da..d937c24 100644 > --- a/drivers/xen/pvcalls-front.h > +++ b/drivers/xen/pvcalls-front.h > @@ -13,5 +13,8 @@ int pvcalls_front_bind(struct socket *sock, > int pvcalls_front_accept(struct socket *sock, > struct socket *newsock, > int flags); > +int pvcalls_front_sendmsg(struct socket *sock, > + struct msghdr *msg, > + size_t len); > > #endif >