| Message ID | 20190219100903.15408-4-bmt@zurich.ibm.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Jason Gunthorpe |
| Headers | show |
| Series | SIW: Request for Comments | expand |
On Tue, Feb 19, 2019 at 11:08:53AM +0100, Bernard Metzler wrote: > Signed-off-by: Bernard Metzler <bmt@zurich.ibm.com> > drivers/infiniband/sw/siw/siw_main.c | 813 +++++++++++++++++++++++++++ > 1 file changed, 813 insertions(+) > create mode 100644 drivers/infiniband/sw/siw/siw_main.c > > diff --git a/drivers/infiniband/sw/siw/siw_main.c b/drivers/infiniband/sw/siw/siw_main.c > new file mode 100644 > index 000000000000..9758b7291dae > +++ b/drivers/infiniband/sw/siw/siw_main.c > @@ -0,0 +1,813 @@ > +// SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause > +/* > + * Software iWARP device driver > + * > + * Authors: Bernard Metzler <bmt@zurich.ibm.com> > + * > + * Copyright (c) 2008-2019, IBM Corporation > + * > + * This software is available to you under a choice of one of two > + * licenses. You may choose to be licensed under the terms of the GNU > + * General Public License (GPL) Version 2, available from the file > + * COPYING in the main directory of this source tree, or the > + * BSD license below: > + * > + * Redistribution and use in source and binary forms, with or > + * without modification, are permitted provided that the following > + * conditions are met: > + * > + * - Redistributions of source code must retain the above copyright notice, > + * this list of conditions and the following disclaimer. > + * > + * - Redistributions in binary form must reproduce the above copyright > + * notice, this list of conditions and the following disclaimer in the > + * documentation and/or other materials provided with the distribution. > + * > + * - Neither the name of IBM nor the names of its contributors may be > + * used to endorse or promote products derived from this software without > + * specific prior written permission. > + * > + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, > + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF > + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND > + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS > + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN > + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN > + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE > + * SOFTWARE. A general comment for all the files, once use SPDX is used the license boiler plate should not be included, just the copyright info. Authors is recorded in git and is redundant. Jason
On Tue, Feb 19, 2019 at 11:08:53AM +0100, Bernard Metzler wrote: > +static const struct ib_device_ops siw_device_ops = { > + .query_device = siw_query_device, > + .query_port = siw_query_port, > + .get_port_immutable = siw_get_port_immutable, > + .get_netdev = siw_get_netdev, This driver uses ib_device_set_netdev so get_netdev should not be set. > + .query_qp = siw_query_qp, > + .modify_port = siw_modify_port, > + .query_pkey = siw_query_pkey, > + .query_gid = siw_query_gid, > + .alloc_ucontext = siw_alloc_ucontext, > + .dealloc_ucontext = siw_dealloc_ucontext, > + .mmap = siw_mmap, > + .alloc_pd = siw_alloc_pd, > + .dealloc_pd = siw_dealloc_pd, > + .create_qp = siw_create_qp, > + .modify_qp = siw_verbs_modify_qp, > + .destroy_qp = siw_destroy_qp, > + .create_cq = siw_create_cq, > + .destroy_cq = siw_destroy_cq, > + .poll_cq = siw_poll_cq, > + .get_dma_mr = siw_get_dma_mr, > + .reg_user_mr = siw_reg_user_mr, > + .dereg_mr = siw_dereg_mr, > + .alloc_mr = siw_alloc_mr, > + .map_mr_sg = siw_map_mr_sg, > + .create_srq = siw_create_srq, > + .modify_srq = siw_modify_srq, > + .query_srq = siw_query_srq, > + .destroy_srq = siw_destroy_srq, > + .post_srq_recv = siw_post_srq_recv, > + .req_notify_cq = siw_req_notify_cq, > + .post_send = siw_post_send, > + .post_recv = siw_post_receive, > + .drain_sq = siw_verbs_sq_flush, > + .drain_rq = siw_verbs_rq_flush, > + .dealloc_driver = siw_device_cleanup, Keep lists of things sorted. Here and everywhere, makefiles, kconfig, etc. Jason
On Tue, Feb 19, 2019 at 11:08:53AM +0100, Bernard Metzler wrote: > +static struct siw_device *siw_device_create(struct net_device *netdev) > +{ > + struct siw_device *sdev; > + struct ib_device *base_dev; > + struct device *parent = netdev->dev.parent; > + int rv; > + > + sdev = ib_alloc_device(siw_device, base_dev); > + if (!sdev) > + goto out; > + > + base_dev = &sdev->base_dev; > + > + if (!parent) { > + /* > + * The loopback device has no parent device, > + * so it appears as a top-level device. To support > + * loopback device connectivity, take this device > + * as the parent device. Skip all other devices > + * w/o parent device. > + */ > + if (netdev->type != ARPHRD_LOOPBACK) { > + pr_warn("siw: device %s skipped (no parent dev)\n", > + netdev->name); > + ib_dealloc_device(base_dev); > + sdev = NULL; This should have the ib_dealloc_device in the goto error section > + goto out; > + } > + parent = &netdev->dev; > + } > + base_dev->iwcm = kmalloc(sizeof(struct iw_cm_verbs), GFP_KERNEL); > + if (!base_dev->iwcm) { > + ib_dealloc_device(base_dev); > + sdev = NULL; > + goto out; > + } > + > + sdev->netdev = netdev; > + > + memset(&base_dev->node_guid, 0, sizeof(base_dev->node_guid)); base_dev is already zero'd > + ib_set_device_ops(base_dev, &siw_device_ops); > + rv = ib_device_set_netdev(base_dev, netdev, 1); > + if (rv) { > + kfree(base_dev->iwcm); > + ib_dealloc_device(base_dev); > + sdev = NULL; > + goto out; Same comment > + } > + base_dev->iwcm->connect = siw_connect; > + base_dev->iwcm->accept = siw_accept; > + base_dev->iwcm->reject = siw_reject; > + base_dev->iwcm->create_listen = siw_create_listen; > + base_dev->iwcm->destroy_listen = siw_destroy_listen; > + base_dev->iwcm->add_ref = siw_qp_get_ref; > + base_dev->iwcm->rem_ref = siw_qp_put_ref; > + base_dev->iwcm->get_qp = siw_get_base_qp; > + > + /* Disable TCP port mapper service */ > + base_dev->iwcm->driver_flags = IW_F_NO_PORT_MAP; > + > + memcpy(base_dev->iwcm->ifname, netdev->name, > + sizeof(base_dev->iwcm->ifname)); > + > + sdev->attrs.max_qp = SIW_MAX_QP; > + sdev->attrs.max_qp_wr = SIW_MAX_QP_WR; > + sdev->attrs.max_ord = SIW_MAX_ORD_QP; > + sdev->attrs.max_ird = SIW_MAX_IRD_QP; > + sdev->attrs.max_sge = SIW_MAX_SGE; > + sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD; > + sdev->attrs.max_cq = SIW_MAX_CQ; > + sdev->attrs.max_cqe = SIW_MAX_CQE; > + sdev->attrs.max_mr = SIW_MAX_MR; > + sdev->attrs.max_pd = SIW_MAX_PD; > + sdev->attrs.max_mw = SIW_MAX_MW; > + sdev->attrs.max_fmr = SIW_MAX_FMR; > + sdev->attrs.max_srq = SIW_MAX_SRQ; > + sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR; > + sdev->attrs.max_srq_sge = SIW_MAX_SGE; > + > + siw_idr_init(sdev); Why idr_init/release functions? Seems unnecessary > +/* > + * Network link becomes unavailable. Mark all > + * affected QP's accordingly. > + */ > +static void siw_netdev_down(struct work_struct *work) > +{ > + struct siw_device *sdev = container_of(work, struct siw_device, > + netdev_down); > + > + struct siw_qp_attrs qp_attrs; > + struct list_head *pos, *tmp; > + > + memset(&qp_attrs, 0, sizeof(qp_attrs)); > + qp_attrs.state = SIW_QP_STATE_ERROR; > + > + list_for_each_safe(pos, tmp, &sdev->qp_list) { > + struct siw_qp *qp = list_entry(pos, struct siw_qp, devq); > + > + down_write(&qp->state_lock); > + (void) siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE); I always wonder if '(void) ' casts really want to be WARN_ONs.. > + up_write(&qp->state_lock); > + } > + ib_device_put(&sdev->base_dev); Not super excited to see the device refount held into a work .. This could create deadlock if you are careless... > +static void siw_device_goes_down(struct siw_device *sdev) > +{ > + if (ib_device_try_get(&sdev->base_dev)) { > + INIT_WORK(&sdev->netdev_down, siw_netdev_down); > + schedule_work(&sdev->netdev_down); .. ie here it should be the system unbound wq, but that would deadlock with ib_unregister_device_queued .. > + } > +} > + > +static int siw_netdev_event(struct notifier_block *nb, unsigned long event, > + void *arg) > +{ > + struct net_device *netdev = netdev_notifier_info_to_dev(arg); > + struct ib_device *base_dev; > + struct siw_device *sdev; > + > + dev_dbg(&netdev->dev, "siw: event %lu\n", event); > + > + if (dev_net(netdev) != &init_net) > + return NOTIFY_OK; > + > + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); > + if (!base_dev) > + return NOTIFY_OK; > + > + sdev = to_siw_dev(base_dev); > + > + switch (event) { > + > + case NETDEV_UP: > + sdev->state = IB_PORT_ACTIVE; > + siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE); > + break; > + > + case NETDEV_GOING_DOWN: > + siw_device_goes_down(sdev); > + break; This probably wants to be return as there is a double put otherwise. Never tested? > +static int siw_newlink(const char *basedev_name, struct net_device *netdev) > +{ > + struct ib_device *base_dev; > + struct siw_device *sdev; > + int rv; > + > + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); > + if (base_dev) { > + ib_device_put(base_dev); > + rv = -EEXIST; > + goto out; > + } > + if (!siw_dev_qualified(netdev)) { > + rv = -EINVAL; > + goto out; > + } > + sdev = siw_device_create(netdev); > + if (sdev) { > + dev_dbg(&netdev->dev, "siw: new device\n"); > + > + if (netif_running(netdev) && netif_carrier_ok(netdev)) > + sdev->state = IB_PORT_ACTIVE; > + else > + sdev->state = IB_PORT_DOWN; > + > + rv = siw_device_register(sdev, basedev_name); > + if (rv) { > + siw_device_cleanup(&sdev->base_dev); > + ib_dealloc_device(&sdev->base_dev); > + goto out; consistently use goto unwinds everywhere? > +static void __exit siw_exit_module(void) > +{ > + int nr_cpu; > + > + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) { > + if (siw_tx_thread[nr_cpu]) { > + siw_stop_tx_thread(nr_cpu); > + siw_tx_thread[nr_cpu] = NULL; > + } > + } Make sure this doesn't deadlock with ib_unregister_driver.. Is forward progress of unregister guarenteed if the tx threads are stopped? > + unregister_netdevice_notifier(&siw_netdev_nb); > + rdma_link_unregister(&siw_link_ops); > + ib_unregister_driver(RDMA_DRIVER_SIW); > + > + siw_cm_exit(); > + > + if (siw_crypto_shash) > + crypto_free_shash(siw_crypto_shash); > + > + siw_destroy_cpulist(); > + > + pr_info("SoftiWARP detached\n"); > +} > + > +module_init(siw_init_module); > +module_exit(siw_exit_module); > + > +MODULE_ALIAS_RDMA_LINK("siw");
On 2/21/2019 6:02 PM, Jason Gunthorpe wrote: > On Tue, Feb 19, 2019 at 11:08:53AM +0100, Bernard Metzler wrote: > >> +static const struct ib_device_ops siw_device_ops = { >> + .query_device = siw_query_device, >> + .query_port = siw_query_port, >> + .get_port_immutable = siw_get_port_immutable, >> + .get_netdev = siw_get_netdev, > This driver uses ib_device_set_netdev so get_netdev should not be set. Should ib_device_set_netdev() have a WARN() for this perhaps? >> + .query_qp = siw_query_qp, >> + .modify_port = siw_modify_port, >> + .query_pkey = siw_query_pkey, >> + .query_gid = siw_query_gid, >> + .alloc_ucontext = siw_alloc_ucontext, >> + .dealloc_ucontext = siw_dealloc_ucontext, >> + .mmap = siw_mmap, >> + .alloc_pd = siw_alloc_pd, >> + .dealloc_pd = siw_dealloc_pd, >> + .create_qp = siw_create_qp, >> + .modify_qp = siw_verbs_modify_qp, >> + .destroy_qp = siw_destroy_qp, >> + .create_cq = siw_create_cq, >> + .destroy_cq = siw_destroy_cq, >> + .poll_cq = siw_poll_cq, >> + .get_dma_mr = siw_get_dma_mr, >> + .reg_user_mr = siw_reg_user_mr, >> + .dereg_mr = siw_dereg_mr, >> + .alloc_mr = siw_alloc_mr, >> + .map_mr_sg = siw_map_mr_sg, >> + .create_srq = siw_create_srq, >> + .modify_srq = siw_modify_srq, >> + .query_srq = siw_query_srq, >> + .destroy_srq = siw_destroy_srq, >> + .post_srq_recv = siw_post_srq_recv, >> + .req_notify_cq = siw_req_notify_cq, >> + .post_send = siw_post_send, >> + .post_recv = siw_post_receive, >> + .drain_sq = siw_verbs_sq_flush, >> + .drain_rq = siw_verbs_rq_flush, >> + .dealloc_driver = siw_device_cleanup, > Keep lists of things sorted. Here and everywhere, makefiles, kconfig, > etc. > > Jason
-----"Jason Gunthorpe" <jgg@ziepe.ca> wrote: ----- >To: "Bernard Metzler" <bmt@zurich.ibm.com> >From: "Jason Gunthorpe" <jgg@ziepe.ca> >Date: 02/22/2019 01:00AM >Cc: linux-rdma@vger.kernel.org >Subject: Re: [PATCH v5 03/13] SIW network and RDMA core interface > >On Tue, Feb 19, 2019 at 11:08:53AM +0100, Bernard Metzler wrote: >> Signed-off-by: Bernard Metzler <bmt@zurich.ibm.com> >> drivers/infiniband/sw/siw/siw_main.c | 813 >+++++++++++++++++++++++++++ >> 1 file changed, 813 insertions(+) >> create mode 100644 drivers/infiniband/sw/siw/siw_main.c >> >> diff --git a/drivers/infiniband/sw/siw/siw_main.c >b/drivers/infiniband/sw/siw/siw_main.c >> new file mode 100644 >> index 000000000000..9758b7291dae >> +++ b/drivers/infiniband/sw/siw/siw_main.c >> @@ -0,0 +1,813 @@ >> +// SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause >> +/* >> + * Software iWARP device driver >> + * >> + * Authors: Bernard Metzler <bmt@zurich.ibm.com> >> + * >> + * Copyright (c) 2008-2019, IBM Corporation >> + * >> + * This software is available to you under a choice of one of two >> + * licenses. You may choose to be licensed under the terms of the >GNU >> + * General Public License (GPL) Version 2, available from the file >> + * COPYING in the main directory of this source tree, or the >> + * BSD license below: >> + * >> + * Redistribution and use in source and binary forms, with or >> + * without modification, are permitted provided that the >following >> + * conditions are met: >> + * >> + * - Redistributions of source code must retain the above >copyright notice, >> + * this list of conditions and the following disclaimer. >> + * >> + * - Redistributions in binary form must reproduce the above >copyright >> + * notice, this list of conditions and the following >disclaimer in the >> + * documentation and/or other materials provided with the >distribution. >> + * >> + * - Neither the name of IBM nor the names of its contributors >may be >> + * used to endorse or promote products derived from this >software without >> + * specific prior written permission. >> + * >> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, >> + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES >OF >> + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND >> + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT >HOLDERS >> + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN >AN >> + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR >IN >> + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN >THE >> + * SOFTWARE. > >A general comment for all the files, once use SPDX is used the >license boiler plate should not be included, just the copyright >info. Authors is recorded in git and is redundant. > >Jason > OK let me go through legal. I hope it turns out to be a simple change ;)
-----"Jason Gunthorpe" <jgg@ziepe.ca> wrote: ----- >To: "Bernard Metzler" <bmt@zurich.ibm.com> >From: "Jason Gunthorpe" <jgg@ziepe.ca> >Date: 02/22/2019 01:02AM >Cc: linux-rdma@vger.kernel.org >Subject: Re: [PATCH v5 03/13] SIW network and RDMA core interface > >On Tue, Feb 19, 2019 at 11:08:53AM +0100, Bernard Metzler wrote: > >> +static const struct ib_device_ops siw_device_ops = { >> + .query_device = siw_query_device, >> + .query_port = siw_query_port, >> + .get_port_immutable = siw_get_port_immutable, >> + .get_netdev = siw_get_netdev, > >This driver uses ib_device_set_netdev so get_netdev should not be >set. > >> + .query_qp = siw_query_qp, >> + .modify_port = siw_modify_port, >> + .query_pkey = siw_query_pkey, >> + .query_gid = siw_query_gid, >> + .alloc_ucontext = siw_alloc_ucontext, >> + .dealloc_ucontext = siw_dealloc_ucontext, >> + .mmap = siw_mmap, >> + .alloc_pd = siw_alloc_pd, >> + .dealloc_pd = siw_dealloc_pd, >> + .create_qp = siw_create_qp, >> + .modify_qp = siw_verbs_modify_qp, >> + .destroy_qp = siw_destroy_qp, >> + .create_cq = siw_create_cq, >> + .destroy_cq = siw_destroy_cq, >> + .poll_cq = siw_poll_cq, >> + .get_dma_mr = siw_get_dma_mr, >> + .reg_user_mr = siw_reg_user_mr, >> + .dereg_mr = siw_dereg_mr, >> + .alloc_mr = siw_alloc_mr, >> + .map_mr_sg = siw_map_mr_sg, >> + .create_srq = siw_create_srq, >> + .modify_srq = siw_modify_srq, >> + .query_srq = siw_query_srq, >> + .destroy_srq = siw_destroy_srq, >> + .post_srq_recv = siw_post_srq_recv, >> + .req_notify_cq = siw_req_notify_cq, >> + .post_send = siw_post_send, >> + .post_recv = siw_post_receive, >> + .drain_sq = siw_verbs_sq_flush, >> + .drain_rq = siw_verbs_rq_flush, >> + .dealloc_driver = siw_device_cleanup, > >Keep lists of things sorted. Here and everywhere, makefiles, kconfig, >etc. I *tried* to do so - by grouping methods here according to the resource it affects. But you are right, this is badly sorted, even under that premise! Will fix that (but let's not order alphabetical here, right?). > >Jason > >
-----"Jason Gunthorpe" <jgg@ziepe.ca> wrote: ----- >To: "Bernard Metzler" <bmt@zurich.ibm.com> >From: "Jason Gunthorpe" <jgg@ziepe.ca> >Date: 02/22/2019 01:14AM >Cc: linux-rdma@vger.kernel.org >Subject: Re: [PATCH v5 03/13] SIW network and RDMA core interface > >On Tue, Feb 19, 2019 at 11:08:53AM +0100, Bernard Metzler wrote: >> +static struct siw_device *siw_device_create(struct net_device >*netdev) >> +{ >> + struct siw_device *sdev; >> + struct ib_device *base_dev; >> + struct device *parent = netdev->dev.parent; >> + int rv; >> + >> + sdev = ib_alloc_device(siw_device, base_dev); >> + if (!sdev) >> + goto out; >> + >> + base_dev = &sdev->base_dev; >> + >> + if (!parent) { >> + /* >> + * The loopback device has no parent device, >> + * so it appears as a top-level device. To support >> + * loopback device connectivity, take this device >> + * as the parent device. Skip all other devices >> + * w/o parent device. >> + */ >> + if (netdev->type != ARPHRD_LOOPBACK) { >> + pr_warn("siw: device %s skipped (no parent dev)\n", >> + netdev->name); >> + ib_dealloc_device(base_dev); >> + sdev = NULL; > >This should have the ib_dealloc_device in the goto error section Right, thanks. Also true for other cases in that function, as you point out below. Will simplify accordingly. > >> + goto out; >> + } >> + parent = &netdev->dev; >> + } >> + base_dev->iwcm = kmalloc(sizeof(struct iw_cm_verbs), GFP_KERNEL); >> + if (!base_dev->iwcm) { >> + ib_dealloc_device(base_dev); >> + sdev = NULL; >> + goto out; >> + } >> + >> + sdev->netdev = netdev; >> + >> + memset(&base_dev->node_guid, 0, sizeof(base_dev->node_guid)); > >base_dev is already zero'd Okay, will remove. > >> + ib_set_device_ops(base_dev, &siw_device_ops); >> + rv = ib_device_set_netdev(base_dev, netdev, 1); >> + if (rv) { >> + kfree(base_dev->iwcm); >> + ib_dealloc_device(base_dev); >> + sdev = NULL; >> + goto out; > >Same comment > >> + } >> + base_dev->iwcm->connect = siw_connect; >> + base_dev->iwcm->accept = siw_accept; >> + base_dev->iwcm->reject = siw_reject; >> + base_dev->iwcm->create_listen = siw_create_listen; >> + base_dev->iwcm->destroy_listen = siw_destroy_listen; >> + base_dev->iwcm->add_ref = siw_qp_get_ref; >> + base_dev->iwcm->rem_ref = siw_qp_put_ref; >> + base_dev->iwcm->get_qp = siw_get_base_qp; >> + >> + /* Disable TCP port mapper service */ >> + base_dev->iwcm->driver_flags = IW_F_NO_PORT_MAP; >> + >> + memcpy(base_dev->iwcm->ifname, netdev->name, >> + sizeof(base_dev->iwcm->ifname)); >> + >> + sdev->attrs.max_qp = SIW_MAX_QP; >> + sdev->attrs.max_qp_wr = SIW_MAX_QP_WR; >> + sdev->attrs.max_ord = SIW_MAX_ORD_QP; >> + sdev->attrs.max_ird = SIW_MAX_IRD_QP; >> + sdev->attrs.max_sge = SIW_MAX_SGE; >> + sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD; >> + sdev->attrs.max_cq = SIW_MAX_CQ; >> + sdev->attrs.max_cqe = SIW_MAX_CQE; >> + sdev->attrs.max_mr = SIW_MAX_MR; >> + sdev->attrs.max_pd = SIW_MAX_PD; >> + sdev->attrs.max_mw = SIW_MAX_MW; >> + sdev->attrs.max_fmr = SIW_MAX_FMR; >> + sdev->attrs.max_srq = SIW_MAX_SRQ; >> + sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR; >> + sdev->attrs.max_srq_sge = SIW_MAX_SGE; >> + >> + siw_idr_init(sdev); > >Why idr_init/release functions? Seems unnecessary idr structs are allocated for each dynamically allocated siw device (one device per link). So each device has its own id space (e.g. for STags/Keys). So I think I need to initialize/destroy the idr resources dynamically as well. > >> +/* >> + * Network link becomes unavailable. Mark all >> + * affected QP's accordingly. >> + */ >> +static void siw_netdev_down(struct work_struct *work) >> +{ >> + struct siw_device *sdev = container_of(work, struct siw_device, >> + netdev_down); >> + >> + struct siw_qp_attrs qp_attrs; >> + struct list_head *pos, *tmp; >> + >> + memset(&qp_attrs, 0, sizeof(qp_attrs)); >> + qp_attrs.state = SIW_QP_STATE_ERROR; >> + >> + list_for_each_safe(pos, tmp, &sdev->qp_list) { >> + struct siw_qp *qp = list_entry(pos, struct siw_qp, devq); >> + >> + down_write(&qp->state_lock); >> + (void) siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE); > >I always wonder if '(void) ' casts really want to be WARN_ONs.. > Sounds good. Will do. >> + up_write(&qp->state_lock); >> + } >> + ib_device_put(&sdev->base_dev); > >Not super excited to see the device refount held into a work .. This >could create deadlock if you are careless... > >> +static void siw_device_goes_down(struct siw_device *sdev) >> +{ >> + if (ib_device_try_get(&sdev->base_dev)) { >> + INIT_WORK(&sdev->netdev_down, siw_netdev_down); >> + schedule_work(&sdev->netdev_down); > >.. ie here it should be the system unbound wq, but that would >deadlock >with ib_unregister_device_queued .. > Hmm, can you please help me understanding that concern? I agree I should check if schedule_work() succeeds (I have to fix that, but it should never happen), but otherwise? How can it deadlock? >> + } >> +} >> + >> +static int siw_netdev_event(struct notifier_block *nb, unsigned >long event, >> + void *arg) >> +{ >> + struct net_device *netdev = netdev_notifier_info_to_dev(arg); >> + struct ib_device *base_dev; >> + struct siw_device *sdev; >> + >> + dev_dbg(&netdev->dev, "siw: event %lu\n", event); >> + >> + if (dev_net(netdev) != &init_net) >> + return NOTIFY_OK; >> + >> + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); >> + if (!base_dev) >> + return NOTIFY_OK; >> + >> + sdev = to_siw_dev(base_dev); >> + >> + switch (event) { >> + >> + case NETDEV_UP: >> + sdev->state = IB_PORT_ACTIVE; >> + siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE); >> + break; >> + >> + case NETDEV_GOING_DOWN: >> + siw_device_goes_down(sdev); >> + break; > >This probably wants to be return as there is a double put >otherwise. Never tested? > I tested bringing down interfaces, also while there is siw caused traffic on it. siw_device_goes_down() calls ib_device_try_get(), which schedules work only if it got that additional reference (which then get's put in the work handler). So I don't see a potential double put case (yet)...? >> +static int siw_newlink(const char *basedev_name, struct net_device >*netdev) >> +{ >> + struct ib_device *base_dev; >> + struct siw_device *sdev; >> + int rv; >> + >> + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); >> + if (base_dev) { >> + ib_device_put(base_dev); >> + rv = -EEXIST; >> + goto out; >> + } >> + if (!siw_dev_qualified(netdev)) { >> + rv = -EINVAL; >> + goto out; >> + } >> + sdev = siw_device_create(netdev); >> + if (sdev) { >> + dev_dbg(&netdev->dev, "siw: new device\n"); >> + >> + if (netif_running(netdev) && netif_carrier_ok(netdev)) >> + sdev->state = IB_PORT_ACTIVE; >> + else >> + sdev->state = IB_PORT_DOWN; >> + >> + rv = siw_device_register(sdev, basedev_name); >> + if (rv) { >> + siw_device_cleanup(&sdev->base_dev); >> + ib_dealloc_device(&sdev->base_dev); >> + goto out; > >consistently use goto unwinds everywhere? OK, let's do that. > >> +static void __exit siw_exit_module(void) >> +{ >> + int nr_cpu; >> + >> + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) { >> + if (siw_tx_thread[nr_cpu]) { >> + siw_stop_tx_thread(nr_cpu); >> + siw_tx_thread[nr_cpu] = NULL; >> + } >> + } > >Make sure this doesn't deadlock with ib_unregister_driver.. Is >forward >progress of unregister guarenteed if the tx threads are stopped? I thought yes, but better let me think through it again and test that. > >> + unregister_netdevice_notifier(&siw_netdev_nb); >> + rdma_link_unregister(&siw_link_ops); >> + ib_unregister_driver(RDMA_DRIVER_SIW); >> + >> + siw_cm_exit(); >> + >> + if (siw_crypto_shash) >> + crypto_free_shash(siw_crypto_shash); >> + >> + siw_destroy_cpulist(); >> + >> + pr_info("SoftiWARP detached\n"); >> +} >> + >> +module_init(siw_init_module); >> +module_exit(siw_exit_module); >> + >> +MODULE_ALIAS_RDMA_LINK("siw"); > >
On 19-Feb-19 12:08, Bernard Metzler wrote: > Signed-off-by: Bernard Metzler <bmt@zurich.ibm.com> > --- > drivers/infiniband/sw/siw/siw_main.c | 813 +++++++++++++++++++++++++++ > 1 file changed, 813 insertions(+) > create mode 100644 drivers/infiniband/sw/siw/siw_main.c > > diff --git a/drivers/infiniband/sw/siw/siw_main.c b/drivers/infiniband/sw/siw/siw_main.c > new file mode 100644 > index 000000000000..9758b7291dae > --- /dev/null > +++ b/drivers/infiniband/sw/siw/siw_main.c > @@ -0,0 +1,813 @@ > +// SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause > +/* > + * Software iWARP device driver > + * > + * Authors: Bernard Metzler <bmt@zurich.ibm.com> > + * > + * Copyright (c) 2008-2019, IBM Corporation > + * > + * This software is available to you under a choice of one of two > + * licenses. You may choose to be licensed under the terms of the GNU > + * General Public License (GPL) Version 2, available from the file > + * COPYING in the main directory of this source tree, or the > + * BSD license below: > + * > + * Redistribution and use in source and binary forms, with or > + * without modification, are permitted provided that the following > + * conditions are met: > + * > + * - Redistributions of source code must retain the above copyright notice, > + * this list of conditions and the following disclaimer. > + * > + * - Redistributions in binary form must reproduce the above copyright > + * notice, this list of conditions and the following disclaimer in the > + * documentation and/or other materials provided with the distribution. > + * > + * - Neither the name of IBM nor the names of its contributors may be > + * used to endorse or promote products derived from this software without > + * specific prior written permission. > + * > + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, > + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF > + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND > + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS > + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN > + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN > + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE > + * SOFTWARE. > + */ > + > +#include <linux/init.h> > +#include <linux/errno.h> > +#include <linux/netdevice.h> > +#include <linux/inetdevice.h> > +#include <net/net_namespace.h> > +#include <linux/rtnetlink.h> > +#include <linux/if_arp.h> > +#include <linux/list.h> > +#include <linux/kernel.h> > +#include <linux/dma-mapping.h> > + > +#include <rdma/ib_verbs.h> > +#include <rdma/ib_smi.h> > +#include <rdma/ib_user_verbs.h> > +#include <rdma/rdma_netlink.h> > +#include <linux/kthread.h> > + > +#include "siw.h" > +#include "siw_obj.h" > +#include "siw_cm.h" > +#include "siw_verbs.h" > + > +MODULE_AUTHOR("Bernard Metzler"); > +MODULE_DESCRIPTION("Software iWARP Driver"); > +MODULE_LICENSE("Dual BSD/GPL"); > +MODULE_VERSION("0.2"); Are we still using module versions? > + > +/* transmit from user buffer, if possible */ > +const bool zcopy_tx = true; > + > +/* Restrict usage of GSO, if hardware peer iwarp is unable to process > + * large packets. gso_seg_limit = 1 lets siw send only packets up to > + * one real MTU in size, but severly limits siw maximum bandwidth. > + * gso_seg_limit = 0 makes use of GSO for large transfers. > + */ > +const u8 gso_seg_limit = 1; > + > +/* Attach siw also with loopback devices */ > +const bool loopback_enabled = true; > + > +/* We try to negotiate CRC on, if true */ > +const bool mpa_crc_required; > + > +/* MPA CRC on/off enforced */ > +const bool mpa_crc_strict; > + > +/* Control TCP_NODELAY socket option */ > +const bool siw_tcp_nagle; > + > +/* Select MPA version to be used during connection setup */ > +u_char mpa_version = MPA_REVISION_2; > + > +/* Selects MPA P2P mode (additional handshake during connection > + * setup, if true. > + */ > +const bool peer_to_peer; > + > +struct task_struct *siw_tx_thread[NR_CPUS]; > +struct crypto_shash *siw_crypto_shash; > + > +static ssize_t sw_version_show(struct device *dev, > + struct device_attribute *attr, char *buf) > +{ > + return sprintf(buf, "%x\n", VERSION_ID_SOFTIWARP); > +} > + > +static DEVICE_ATTR_RO(sw_version); > + > +static ssize_t parent_show(struct device *device, > + struct device_attribute *attr, char *buf) > +{ > + struct siw_device *sdev = > + rdma_device_to_drv_device(device, struct siw_device, base_dev); > + > + return snprintf(buf, 16, "%s\n", sdev->netdev->name); > +} > + > +static DEVICE_ATTR_RO(parent); > + > +static struct attribute *siw_dev_attributes[] = { > + &dev_attr_sw_version.attr, > + &dev_attr_parent.attr, > + NULL, > +}; > + > +static const struct attribute_group siw_attr_group = { > + .attrs = siw_dev_attributes, > +}; > + > +static int siw_modify_port(struct ib_device *base_dev, u8 port, int mask, > + struct ib_port_modify *props) > +{ > + return -EOPNOTSUPP; > +} Why return -EOPNOTSUPP, just don't implement the function. > + > +static int siw_device_register(struct siw_device *sdev, const char *name) > +{ > + struct ib_device *base_dev = &sdev->base_dev; > + static int dev_id = 1; > + int rv; > + > + base_dev->driver_id = RDMA_DRIVER_SIW; > + rdma_set_device_sysfs_group(base_dev, &siw_attr_group); > + > + rv = ib_register_device(base_dev, name); > + if (rv) { > + pr_warn("siw: device registration error %d\n", rv); > + return rv; > + } > + sdev->vendor_part_id = dev_id++; > + > + siw_dbg(sdev, "HWaddr=%02x.%02x.%02x.%02x.%02x.%02x\n", > + *(u8 *)sdev->netdev->dev_addr, > + *((u8 *)sdev->netdev->dev_addr + 1), > + *((u8 *)sdev->netdev->dev_addr + 2), > + *((u8 *)sdev->netdev->dev_addr + 3), > + *((u8 *)sdev->netdev->dev_addr + 4), > + *((u8 *)sdev->netdev->dev_addr + 5)); %pM is your friend here :). > + > + return 0; > +} > + > +static void siw_device_cleanup(struct ib_device *base_dev) > +{ > + struct siw_device *sdev = to_siw_dev(base_dev); > + > + siw_dbg(sdev, "Cleanup device\n"); > + > + if (atomic_read(&sdev->num_ctx) || atomic_read(&sdev->num_srq) || > + atomic_read(&sdev->num_mr) || atomic_read(&sdev->num_cep) || > + atomic_read(&sdev->num_qp) || atomic_read(&sdev->num_cq) || > + atomic_read(&sdev->num_pd)) { > + pr_warn("siw at %s: orphaned resources!\n", > + sdev->netdev->name); > + pr_warn(" CTX %d, SRQ %d, QP %d, CQ %d, MEM %d, CEP %d, PD %d\n", > + atomic_read(&sdev->num_ctx), > + atomic_read(&sdev->num_srq), > + atomic_read(&sdev->num_qp), > + atomic_read(&sdev->num_cq), > + atomic_read(&sdev->num_mr), > + atomic_read(&sdev->num_cep), > + atomic_read(&sdev->num_pd)); > + } > + while (!list_empty(&sdev->cep_list)) { > + struct siw_cep *cep = list_entry(sdev->cep_list.next, > + struct siw_cep, devq); > + list_del(&cep->devq); > + pr_warn("siw: at %s: free orphaned CEP 0x%p, state %d\n", > + sdev->base_dev.name, cep, cep->state); > + kfree(cep); > + } > + siw_idr_release(sdev); > + kfree(sdev->base_dev.iwcm); > +} > + > +static struct net_device *siw_get_netdev(struct ib_device *base_dev, u8 port) > +{ > + struct siw_device *sdev = to_siw_dev(base_dev); > + > + if (!sdev->netdev) > + return NULL; > + > + dev_hold(sdev->netdev); > + > + return sdev->netdev; > +} > + > +static int siw_create_tx_threads(void) > +{ > + int cpu, rv, assigned = 0; > + > + for_each_online_cpu(cpu) { > + /* Skip HT cores */ > + if (cpu % cpumask_weight(topology_sibling_cpumask(cpu))) { > + siw_tx_thread[cpu] = NULL; > + continue; > + } > + siw_tx_thread[cpu] = kthread_create(siw_run_sq, > + (unsigned long *)(long)cpu, > + "siw_tx/%d", cpu); > + if (IS_ERR(siw_tx_thread[cpu])) { > + rv = PTR_ERR(siw_tx_thread[cpu]); > + siw_tx_thread[cpu] = NULL; > + pr_info("Creating TX thread for CPU %d failed", cpu); > + continue; > + } > + kthread_bind(siw_tx_thread[cpu], cpu); > + > + wake_up_process(siw_tx_thread[cpu]); > + assigned++; > + } > + return assigned; > +} > + > +static int siw_dev_qualified(struct net_device *netdev) > +{ > + /* > + * Additional hardware support can be added here > + * (e.g. ARPHRD_FDDI, ARPHRD_ATM, ...) - see > + * <linux/if_arp.h> for type identifiers. > + */ > + if (netdev->type == ARPHRD_ETHER || > + netdev->type == ARPHRD_IEEE802 || > + (netdev->type == ARPHRD_LOOPBACK && loopback_enabled)) > + return 1; > + > + return 0; > +} > + > +static DEFINE_PER_CPU(atomic_t, use_cnt = ATOMIC_INIT(0)); > + > +static struct { > + struct cpumask **tx_valid_cpus; > + int num_nodes; > +} siw_cpu_info; > + > +static int siw_init_cpulist(void) > +{ > + int i, num_nodes; > + > + num_nodes = num_possible_nodes(); > + siw_cpu_info.num_nodes = num_nodes; > + > + siw_cpu_info.tx_valid_cpus = kcalloc(num_nodes, sizeof(void *), > + GFP_KERNEL); > + if (!siw_cpu_info.tx_valid_cpus) { > + siw_cpu_info.num_nodes = 0; > + return -ENOMEM; > + } > + > + for (i = 0; i < siw_cpu_info.num_nodes; i++) { > + siw_cpu_info.tx_valid_cpus[i] = kzalloc(sizeof(struct cpumask), > + GFP_KERNEL); > + if (!siw_cpu_info.tx_valid_cpus[i]) > + goto out_err; > + > + cpumask_clear(siw_cpu_info.tx_valid_cpus[i]); > + } > + for_each_possible_cpu(i) > + cpumask_set_cpu(i, siw_cpu_info.tx_valid_cpus[cpu_to_node(i)]); > + > + return 0; > + > +out_err: > + siw_cpu_info.num_nodes = 0; > + while (i) { > + kfree(siw_cpu_info.tx_valid_cpus[i]); > + siw_cpu_info.tx_valid_cpus[i--] = NULL; > + } > + kfree(siw_cpu_info.tx_valid_cpus); > + siw_cpu_info.tx_valid_cpus = NULL; > + > + return -ENOMEM; > +} > + > +static void siw_destroy_cpulist(void) > +{ > + int i = 0; > + > + while (i < siw_cpu_info.num_nodes) > + kfree(siw_cpu_info.tx_valid_cpus[i++]); > + > + kfree(siw_cpu_info.tx_valid_cpus); > +} > + > +/* > + * Choose CPU with least number of active QP's from NUMA node of > + * TX interface. > + */ > +int siw_get_tx_cpu(struct siw_device *sdev) > +{ > + const struct cpumask *tx_cpumask; > + int i, num_cpus, cpu, min_use, > + node = sdev->numa_node, > + tx_cpu = -1; > + > + if (node < 0) > + tx_cpumask = cpu_online_mask; > + else > + tx_cpumask = siw_cpu_info.tx_valid_cpus[node]; > + > + num_cpus = cpumask_weight(tx_cpumask); > + if (!num_cpus) { > + /* no CPU on this NUMA node */ > + tx_cpumask = cpu_online_mask; > + num_cpus = cpumask_weight(tx_cpumask); > + } > + if (!num_cpus) > + goto out; > + > + cpu = cpumask_first(tx_cpumask); > + > + for (i = 0, min_use = SIW_MAX_QP; i < num_cpus; > + i++, cpu = cpumask_next(cpu, tx_cpumask)) { > + int usage; > + > + /* Skip any cores which have no TX thread */ > + if (!siw_tx_thread[cpu]) > + continue; > + > + usage = atomic_read(&per_cpu(use_cnt, cpu)); > + if (usage <= min_use) { > + tx_cpu = cpu; > + min_use = usage; > + } > + } > + siw_dbg(sdev, "tx cpu %d, node %d, %d qp's\n", > + tx_cpu, node, min_use); > + > +out: > + if (tx_cpu >= 0) > + atomic_inc(&per_cpu(use_cnt, tx_cpu)); > + else > + pr_warn("siw: no tx cpu found\n"); > + > + return tx_cpu; > +} > + > +void siw_put_tx_cpu(int cpu) > +{ > + atomic_dec(&per_cpu(use_cnt, cpu)); > +} > + > +static void siw_verbs_sq_flush(struct ib_qp *base_qp) > +{ > + struct siw_qp *qp = to_siw_qp(base_qp); > + > + down_write(&qp->state_lock); > + siw_sq_flush(qp); > + up_write(&qp->state_lock); > +} > + > +static void siw_verbs_rq_flush(struct ib_qp *base_qp) > +{ > + struct siw_qp *qp = to_siw_qp(base_qp); > + > + down_write(&qp->state_lock); > + siw_rq_flush(qp); > + up_write(&qp->state_lock); > +} > + > +static const struct ib_device_ops siw_device_ops = { > + .query_device = siw_query_device, > + .query_port = siw_query_port, > + .get_port_immutable = siw_get_port_immutable, > + .get_netdev = siw_get_netdev, > + .query_qp = siw_query_qp, > + .modify_port = siw_modify_port, > + .query_pkey = siw_query_pkey, > + .query_gid = siw_query_gid, > + .alloc_ucontext = siw_alloc_ucontext, > + .dealloc_ucontext = siw_dealloc_ucontext, > + .mmap = siw_mmap, > + .alloc_pd = siw_alloc_pd, > + .dealloc_pd = siw_dealloc_pd, > + .create_qp = siw_create_qp, > + .modify_qp = siw_verbs_modify_qp, > + .destroy_qp = siw_destroy_qp, > + .create_cq = siw_create_cq, > + .destroy_cq = siw_destroy_cq, > + .poll_cq = siw_poll_cq, > + .get_dma_mr = siw_get_dma_mr, > + .reg_user_mr = siw_reg_user_mr, > + .dereg_mr = siw_dereg_mr, > + .alloc_mr = siw_alloc_mr, > + .map_mr_sg = siw_map_mr_sg, > + .create_srq = siw_create_srq, > + .modify_srq = siw_modify_srq, > + .query_srq = siw_query_srq, > + .destroy_srq = siw_destroy_srq, > + .post_srq_recv = siw_post_srq_recv, > + .req_notify_cq = siw_req_notify_cq, > + .post_send = siw_post_send, > + .post_recv = siw_post_receive, > + .drain_sq = siw_verbs_sq_flush, > + .drain_rq = siw_verbs_rq_flush, > + .dealloc_driver = siw_device_cleanup, > + INIT_RDMA_OBJ_SIZE(ib_pd, siw_pd, base_pd), > +}; > + > +static struct siw_device *siw_device_create(struct net_device *netdev) > +{ > + struct siw_device *sdev; > + struct ib_device *base_dev; > + struct device *parent = netdev->dev.parent; > + int rv; > + > + sdev = ib_alloc_device(siw_device, base_dev); > + if (!sdev) > + goto out; > + > + base_dev = &sdev->base_dev; > + > + if (!parent) { > + /* > + * The loopback device has no parent device, > + * so it appears as a top-level device. To support > + * loopback device connectivity, take this device > + * as the parent device. Skip all other devices > + * w/o parent device. > + */ > + if (netdev->type != ARPHRD_LOOPBACK) { > + pr_warn("siw: device %s skipped (no parent dev)\n", > + netdev->name); > + ib_dealloc_device(base_dev); > + sdev = NULL; > + goto out; > + } > + parent = &netdev->dev; > + } > + base_dev->iwcm = kmalloc(sizeof(struct iw_cm_verbs), GFP_KERNEL); > + if (!base_dev->iwcm) { > + ib_dealloc_device(base_dev); > + sdev = NULL; > + goto out; > + } > + > + sdev->netdev = netdev; > + > + memset(&base_dev->node_guid, 0, sizeof(base_dev->node_guid)); > + > + if (netdev->type != ARPHRD_LOOPBACK) { > + memcpy(&base_dev->node_guid, netdev->dev_addr, 6); > + } else { > + /* > + * The loopback device does not have a HW address, > + * but connection mangagement lib expects gid != 0 > + */ > + size_t gidlen = min_t(size_t, strlen(base_dev->name), 6); > + > + memcpy(&base_dev->node_guid, base_dev->name, gidlen); > + } > + base_dev->owner = THIS_MODULE; > + > + base_dev->uverbs_cmd_mask = > + (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) | > + (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) | > + (1ull << IB_USER_VERBS_CMD_QUERY_PORT) | > + (1ull << IB_USER_VERBS_CMD_ALLOC_PD) | > + (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) | > + (1ull << IB_USER_VERBS_CMD_REG_MR) | > + (1ull << IB_USER_VERBS_CMD_DEREG_MR) | > + (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) | > + (1ull << IB_USER_VERBS_CMD_CREATE_CQ) | > + (1ull << IB_USER_VERBS_CMD_POLL_CQ) | > + (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) | > + (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) | > + (1ull << IB_USER_VERBS_CMD_CREATE_QP) | > + (1ull << IB_USER_VERBS_CMD_QUERY_QP) | > + (1ull << IB_USER_VERBS_CMD_MODIFY_QP) | > + (1ull << IB_USER_VERBS_CMD_DESTROY_QP) | > + (1ull << IB_USER_VERBS_CMD_POST_SEND) | > + (1ull << IB_USER_VERBS_CMD_POST_RECV) | > + (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) | > + (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) | > + (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) | > + (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) | > + (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV); > + > + base_dev->node_type = RDMA_NODE_RNIC; > + memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON, > + sizeof(SIW_NODE_DESC_COMMON)); > + > + /* > + * Current model (one-to-one device association): > + * One Softiwarp device per net_device or, equivalently, > + * per physical port. > + */ > + base_dev->phys_port_cnt = 1; > + base_dev->dev.parent = parent; > + base_dev->dev.dma_ops = &dma_virt_ops; > + base_dev->num_comp_vectors = num_possible_cpus(); > + > + ib_set_device_ops(base_dev, &siw_device_ops); > + rv = ib_device_set_netdev(base_dev, netdev, 1); > + if (rv) { > + kfree(base_dev->iwcm); > + ib_dealloc_device(base_dev); > + sdev = NULL; > + goto out; > + } > + base_dev->iwcm->connect = siw_connect; > + base_dev->iwcm->accept = siw_accept; > + base_dev->iwcm->reject = siw_reject; > + base_dev->iwcm->create_listen = siw_create_listen; > + base_dev->iwcm->destroy_listen = siw_destroy_listen; > + base_dev->iwcm->add_ref = siw_qp_get_ref; > + base_dev->iwcm->rem_ref = siw_qp_put_ref; > + base_dev->iwcm->get_qp = siw_get_base_qp; > + > + /* Disable TCP port mapper service */ > + base_dev->iwcm->driver_flags = IW_F_NO_PORT_MAP; > + > + memcpy(base_dev->iwcm->ifname, netdev->name, > + sizeof(base_dev->iwcm->ifname)); > + > + sdev->attrs.max_qp = SIW_MAX_QP; > + sdev->attrs.max_qp_wr = SIW_MAX_QP_WR; > + sdev->attrs.max_ord = SIW_MAX_ORD_QP; > + sdev->attrs.max_ird = SIW_MAX_IRD_QP; > + sdev->attrs.max_sge = SIW_MAX_SGE; > + sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD; > + sdev->attrs.max_cq = SIW_MAX_CQ; > + sdev->attrs.max_cqe = SIW_MAX_CQE; > + sdev->attrs.max_mr = SIW_MAX_MR; > + sdev->attrs.max_pd = SIW_MAX_PD; > + sdev->attrs.max_mw = SIW_MAX_MW; > + sdev->attrs.max_fmr = SIW_MAX_FMR; > + sdev->attrs.max_srq = SIW_MAX_SRQ; > + sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR; > + sdev->attrs.max_srq_sge = SIW_MAX_SGE; > + > + siw_idr_init(sdev); > + INIT_LIST_HEAD(&sdev->cep_list); > + INIT_LIST_HEAD(&sdev->qp_list); > + INIT_LIST_HEAD(&sdev->mr_list); > + > + atomic_set(&sdev->num_ctx, 0); > + atomic_set(&sdev->num_srq, 0); > + atomic_set(&sdev->num_qp, 0); > + atomic_set(&sdev->num_cq, 0); > + atomic_set(&sdev->num_mr, 0); > + atomic_set(&sdev->num_pd, 0); > + atomic_set(&sdev->num_cep, 0); > + > + sdev->numa_node = dev_to_node(parent); > +out: > + return sdev; > +} > + > +/* > + * Network link becomes unavailable. Mark all > + * affected QP's accordingly. > + */ > +static void siw_netdev_down(struct work_struct *work) > +{ > + struct siw_device *sdev = container_of(work, struct siw_device, > + netdev_down); > + > + struct siw_qp_attrs qp_attrs; > + struct list_head *pos, *tmp; > + > + memset(&qp_attrs, 0, sizeof(qp_attrs)); > + qp_attrs.state = SIW_QP_STATE_ERROR; > + > + list_for_each_safe(pos, tmp, &sdev->qp_list) { > + struct siw_qp *qp = list_entry(pos, struct siw_qp, devq); > + > + down_write(&qp->state_lock); > + (void) siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE); > + up_write(&qp->state_lock); > + } > + ib_device_put(&sdev->base_dev); > +} > + > +static void siw_device_goes_down(struct siw_device *sdev) > +{ > + if (ib_device_try_get(&sdev->base_dev)) { > + INIT_WORK(&sdev->netdev_down, siw_netdev_down); > + schedule_work(&sdev->netdev_down); > + } > +} > + > +static int siw_netdev_event(struct notifier_block *nb, unsigned long event, > + void *arg) > +{ > + struct net_device *netdev = netdev_notifier_info_to_dev(arg); > + struct ib_device *base_dev; > + struct siw_device *sdev; > + > + dev_dbg(&netdev->dev, "siw: event %lu\n", event); > + > + if (dev_net(netdev) != &init_net) > + return NOTIFY_OK; > + > + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); > + if (!base_dev) > + return NOTIFY_OK; > + > + sdev = to_siw_dev(base_dev); > + > + switch (event) { > + > + case NETDEV_UP: > + sdev->state = IB_PORT_ACTIVE; > + siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE); > + break; > + > + case NETDEV_GOING_DOWN: > + siw_device_goes_down(sdev); > + break; > + > + case NETDEV_DOWN: > + sdev->state = IB_PORT_DOWN; > + siw_port_event(sdev, 1, IB_EVENT_PORT_ERR); > + break; > + > + case NETDEV_REGISTER: > + /* > + * Device registration now handled only by > + * rdma netlink commands. So it shall be impossible > + * to end up here with a valid siw device. > + */ > + siw_dbg(sdev, "unexpected NETDEV_REGISTER event\n"); > + break; > + > + case NETDEV_UNREGISTER: > + ib_unregister_device_queued(&sdev->base_dev); > + break; > + > + case NETDEV_CHANGEADDR: > + siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE); > + break; > + /* > + * Todo: Below netdev events are currently not handled. > + */ > + case NETDEV_CHANGEMTU: > + case NETDEV_CHANGE: > + > + break; > + > + default: > + break; > + } > + ib_device_put(&sdev->base_dev); > + > + return NOTIFY_OK; > +} > + > +static struct notifier_block siw_netdev_nb = { > + .notifier_call = siw_netdev_event, > +}; > + > +static int siw_newlink(const char *basedev_name, struct net_device *netdev) > +{ > + struct ib_device *base_dev; > + struct siw_device *sdev; > + int rv; > + > + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); > + if (base_dev) { > + ib_device_put(base_dev); > + rv = -EEXIST; > + goto out; > + } > + if (!siw_dev_qualified(netdev)) { > + rv = -EINVAL; > + goto out; > + } > + sdev = siw_device_create(netdev); > + if (sdev) { > + dev_dbg(&netdev->dev, "siw: new device\n"); > + > + if (netif_running(netdev) && netif_carrier_ok(netdev)) > + sdev->state = IB_PORT_ACTIVE; > + else > + sdev->state = IB_PORT_DOWN; > + > + rv = siw_device_register(sdev, basedev_name); > + if (rv) { > + siw_device_cleanup(&sdev->base_dev); > + ib_dealloc_device(&sdev->base_dev); > + goto out; > + } > + } else { > + rv = -ENOMEM; > + } > +out: > + return rv; > +} > + > +static struct rdma_link_ops siw_link_ops = { > + .type = "siw", > + .newlink = siw_newlink, > +}; > + > +/* > + * siw_init_module - Initialize Softiwarp module and register with netdev > + * subsystem to create Softiwarp devices per net_device > + */ > +static __init int siw_init_module(void) > +{ > + int rv; > + int nr_cpu; > + > + if (SENDPAGE_THRESH < SIW_MAX_INLINE) { > + pr_info("siw: sendpage threshold too small: %u\n", > + (int)SENDPAGE_THRESH); > + rv = EINVAL; -EINVAL? > + goto out_error; > + } > + rv = siw_init_cpulist(); > + if (rv) > + goto out_error; > + > + rv = siw_cm_init(); > + if (rv) > + goto out_error; > + > + /* > + * Allocate CRC SHASH object. Fail loading siw only, if CRC is > + * required by kernel module > + */ > + siw_crypto_shash = crypto_alloc_shash("crc32c", 0, 0); > + if (IS_ERR(siw_crypto_shash)) { > + pr_info("siw: Loading CRC32c failed: %ld\n", > + PTR_ERR(siw_crypto_shash)); > + siw_crypto_shash = NULL; > + if (mpa_crc_required == true) Bool tests don't need comparisons. > + goto out_error; > + } > + rv = register_netdevice_notifier(&siw_netdev_nb); > + if (rv) > + goto out_error; > + > + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) > + siw_tx_thread[nr_cpu] = NULL; > + > + if (!siw_create_tx_threads()) { > + pr_info("siw: Could not start any TX thread\n"); > + unregister_netdevice_notifier(&siw_netdev_nb); > + goto out_error; > + } > + rdma_link_register(&siw_link_ops); > + > + pr_info("SoftiWARP attached\n"); > + return 0; > + > +out_error: > + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) { > + if (siw_tx_thread[nr_cpu]) { > + siw_stop_tx_thread(nr_cpu); > + siw_tx_thread[nr_cpu] = NULL; > + } > + } > + if (siw_crypto_shash) > + crypto_free_shash(siw_crypto_shash); > + > + pr_info("SoftIWARP attach failed. Error: %d\n", rv); > + > + siw_cm_exit(); > + siw_destroy_cpulist(); > + > + return rv; > +} > + > +static void __exit siw_exit_module(void) > +{ > + int nr_cpu; > + > + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) { > + if (siw_tx_thread[nr_cpu]) { > + siw_stop_tx_thread(nr_cpu); > + siw_tx_thread[nr_cpu] = NULL; > + } > + } > + unregister_netdevice_notifier(&siw_netdev_nb); > + rdma_link_unregister(&siw_link_ops); > + ib_unregister_driver(RDMA_DRIVER_SIW); > + > + siw_cm_exit(); > + > + if (siw_crypto_shash) > + crypto_free_shash(siw_crypto_shash); > + > + siw_destroy_cpulist(); > + > + pr_info("SoftiWARP detached\n"); > +} > + > +module_init(siw_init_module); > +module_exit(siw_exit_module); > + > +MODULE_ALIAS_RDMA_LINK("siw"); > Reviewed-by: Gal Pressman <galpress@amazon.com>
-----"Gal Pressman" <galpress@amazon.com> wrote: ----- >To: "Bernard Metzler" <bmt@zurich.ibm.com>, ><linux-rdma@vger.kernel.org> >From: "Gal Pressman" <galpress@amazon.com> >Date: 02/28/2019 03:59PM >Subject: Re: [PATCH v5 03/13] SIW network and RDMA core interface > >On 19-Feb-19 12:08, Bernard Metzler wrote: >> Signed-off-by: Bernard Metzler <bmt@zurich.ibm.com> >> --- >> drivers/infiniband/sw/siw/siw_main.c | 813 >+++++++++++++++++++++++++++ >> 1 file changed, 813 insertions(+) >> create mode 100644 drivers/infiniband/sw/siw/siw_main.c >> >> diff --git a/drivers/infiniband/sw/siw/siw_main.c >b/drivers/infiniband/sw/siw/siw_main.c >> new file mode 100644 >> index 000000000000..9758b7291dae >> --- /dev/null >> +++ b/drivers/infiniband/sw/siw/siw_main.c >> @@ -0,0 +1,813 @@ >> +// SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause >> +/* >> + * Software iWARP device driver >> + * >> + * Authors: Bernard Metzler <bmt@zurich.ibm.com> >> + * >> + * Copyright (c) 2008-2019, IBM Corporation >> + * >> + * This software is available to you under a choice of one of two >> + * licenses. You may choose to be licensed under the terms of the >GNU >> + * General Public License (GPL) Version 2, available from the file >> + * COPYING in the main directory of this source tree, or the >> + * BSD license below: >> + * >> + * Redistribution and use in source and binary forms, with or >> + * without modification, are permitted provided that the >following >> + * conditions are met: >> + * >> + * - Redistributions of source code must retain the above >copyright notice, >> + * this list of conditions and the following disclaimer. >> + * >> + * - Redistributions in binary form must reproduce the above >copyright >> + * notice, this list of conditions and the following >disclaimer in the >> + * documentation and/or other materials provided with the >distribution. >> + * >> + * - Neither the name of IBM nor the names of its contributors >may be >> + * used to endorse or promote products derived from this >software without >> + * specific prior written permission. >> + * >> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, >> + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES >OF >> + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND >> + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT >HOLDERS >> + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN >AN >> + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR >IN >> + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN >THE >> + * SOFTWARE. >> + */ >> + >> +#include <linux/init.h> >> +#include <linux/errno.h> >> +#include <linux/netdevice.h> >> +#include <linux/inetdevice.h> >> +#include <net/net_namespace.h> >> +#include <linux/rtnetlink.h> >> +#include <linux/if_arp.h> >> +#include <linux/list.h> >> +#include <linux/kernel.h> >> +#include <linux/dma-mapping.h> >> + >> +#include <rdma/ib_verbs.h> >> +#include <rdma/ib_smi.h> >> +#include <rdma/ib_user_verbs.h> >> +#include <rdma/rdma_netlink.h> >> +#include <linux/kthread.h> >> + >> +#include "siw.h" >> +#include "siw_obj.h" >> +#include "siw_cm.h" >> +#include "siw_verbs.h" >> + >> +MODULE_AUTHOR("Bernard Metzler"); >> +MODULE_DESCRIPTION("Software iWARP Driver"); >> +MODULE_LICENSE("Dual BSD/GPL"); >> +MODULE_VERSION("0.2"); > >Are we still using module versions? ? I don't know. Let me remove it. > >> + >> +/* transmit from user buffer, if possible */ >> +const bool zcopy_tx = true; >> + >> +/* Restrict usage of GSO, if hardware peer iwarp is unable to >process >> + * large packets. gso_seg_limit = 1 lets siw send only packets up >to >> + * one real MTU in size, but severly limits siw maximum bandwidth. >> + * gso_seg_limit = 0 makes use of GSO for large transfers. >> + */ >> +const u8 gso_seg_limit = 1; >> + >> +/* Attach siw also with loopback devices */ >> +const bool loopback_enabled = true; >> + >> +/* We try to negotiate CRC on, if true */ >> +const bool mpa_crc_required; >> + >> +/* MPA CRC on/off enforced */ >> +const bool mpa_crc_strict; >> + >> +/* Control TCP_NODELAY socket option */ >> +const bool siw_tcp_nagle; >> + >> +/* Select MPA version to be used during connection setup */ >> +u_char mpa_version = MPA_REVISION_2; >> + >> +/* Selects MPA P2P mode (additional handshake during connection >> + * setup, if true. >> + */ >> +const bool peer_to_peer; >> + >> +struct task_struct *siw_tx_thread[NR_CPUS]; >> +struct crypto_shash *siw_crypto_shash; >> + >> +static ssize_t sw_version_show(struct device *dev, >> + struct device_attribute *attr, char *buf) >> +{ >> + return sprintf(buf, "%x\n", VERSION_ID_SOFTIWARP); >> +} >> + >> +static DEVICE_ATTR_RO(sw_version); >> + >> +static ssize_t parent_show(struct device *device, >> + struct device_attribute *attr, char *buf) >> +{ >> + struct siw_device *sdev = >> + rdma_device_to_drv_device(device, struct siw_device, base_dev); >> + >> + return snprintf(buf, 16, "%s\n", sdev->netdev->name); >> +} >> + >> +static DEVICE_ATTR_RO(parent); >> + >> +static struct attribute *siw_dev_attributes[] = { >> + &dev_attr_sw_version.attr, >> + &dev_attr_parent.attr, >> + NULL, >> +}; >> + >> +static const struct attribute_group siw_attr_group = { >> + .attrs = siw_dev_attributes, >> +}; >> + >> +static int siw_modify_port(struct ib_device *base_dev, u8 port, >int mask, >> + struct ib_port_modify *props) >> +{ >> + return -EOPNOTSUPP; >> +} > >Why return -EOPNOTSUPP, just don't implement the function. > yes. this seem to have changed. in earlier days I got a crash dump dereferencing a NULL function. >> + >> +static int siw_device_register(struct siw_device *sdev, const char >*name) >> +{ >> + struct ib_device *base_dev = &sdev->base_dev; >> + static int dev_id = 1; >> + int rv; >> + >> + base_dev->driver_id = RDMA_DRIVER_SIW; >> + rdma_set_device_sysfs_group(base_dev, &siw_attr_group); >> + >> + rv = ib_register_device(base_dev, name); >> + if (rv) { >> + pr_warn("siw: device registration error %d\n", rv); >> + return rv; >> + } >> + sdev->vendor_part_id = dev_id++; >> + >> + siw_dbg(sdev, "HWaddr=%02x.%02x.%02x.%02x.%02x.%02x\n", >> + *(u8 *)sdev->netdev->dev_addr, >> + *((u8 *)sdev->netdev->dev_addr + 1), >> + *((u8 *)sdev->netdev->dev_addr + 2), >> + *((u8 *)sdev->netdev->dev_addr + 3), >> + *((u8 *)sdev->netdev->dev_addr + 4), >> + *((u8 *)sdev->netdev->dev_addr + 5)); > >%pM is your friend here :). > yes, looks better... >> + >> + return 0; >> +} >> + >> +static void siw_device_cleanup(struct ib_device *base_dev) >> +{ >> + struct siw_device *sdev = to_siw_dev(base_dev); >> + >> + siw_dbg(sdev, "Cleanup device\n"); >> + >> + if (atomic_read(&sdev->num_ctx) || atomic_read(&sdev->num_srq) || >> + atomic_read(&sdev->num_mr) || atomic_read(&sdev->num_cep) || >> + atomic_read(&sdev->num_qp) || atomic_read(&sdev->num_cq) || >> + atomic_read(&sdev->num_pd)) { >> + pr_warn("siw at %s: orphaned resources!\n", >> + sdev->netdev->name); >> + pr_warn(" CTX %d, SRQ %d, QP %d, CQ %d, MEM %d, CEP >%d, PD %d\n", >> + atomic_read(&sdev->num_ctx), >> + atomic_read(&sdev->num_srq), >> + atomic_read(&sdev->num_qp), >> + atomic_read(&sdev->num_cq), >> + atomic_read(&sdev->num_mr), >> + atomic_read(&sdev->num_cep), >> + atomic_read(&sdev->num_pd)); >> + } >> + while (!list_empty(&sdev->cep_list)) { >> + struct siw_cep *cep = list_entry(sdev->cep_list.next, >> + struct siw_cep, devq); >> + list_del(&cep->devq); >> + pr_warn("siw: at %s: free orphaned CEP 0x%p, state %d\n", >> + sdev->base_dev.name, cep, cep->state); >> + kfree(cep); >> + } >> + siw_idr_release(sdev); >> + kfree(sdev->base_dev.iwcm); >> +} >> + >> +static struct net_device *siw_get_netdev(struct ib_device >*base_dev, u8 port) >> +{ >> + struct siw_device *sdev = to_siw_dev(base_dev); >> + >> + if (!sdev->netdev) >> + return NULL; >> + >> + dev_hold(sdev->netdev); >> + >> + return sdev->netdev; >> +} >> + >> +static int siw_create_tx_threads(void) >> +{ >> + int cpu, rv, assigned = 0; >> + >> + for_each_online_cpu(cpu) { >> + /* Skip HT cores */ >> + if (cpu % cpumask_weight(topology_sibling_cpumask(cpu))) { >> + siw_tx_thread[cpu] = NULL; >> + continue; >> + } >> + siw_tx_thread[cpu] = kthread_create(siw_run_sq, >> + (unsigned long *)(long)cpu, >> + "siw_tx/%d", cpu); >> + if (IS_ERR(siw_tx_thread[cpu])) { >> + rv = PTR_ERR(siw_tx_thread[cpu]); >> + siw_tx_thread[cpu] = NULL; >> + pr_info("Creating TX thread for CPU %d failed", cpu); >> + continue; >> + } >> + kthread_bind(siw_tx_thread[cpu], cpu); >> + >> + wake_up_process(siw_tx_thread[cpu]); >> + assigned++; >> + } >> + return assigned; >> +} >> + >> +static int siw_dev_qualified(struct net_device *netdev) >> +{ >> + /* >> + * Additional hardware support can be added here >> + * (e.g. ARPHRD_FDDI, ARPHRD_ATM, ...) - see >> + * <linux/if_arp.h> for type identifiers. >> + */ >> + if (netdev->type == ARPHRD_ETHER || >> + netdev->type == ARPHRD_IEEE802 || >> + (netdev->type == ARPHRD_LOOPBACK && loopback_enabled)) >> + return 1; >> + >> + return 0; >> +} >> + >> +static DEFINE_PER_CPU(atomic_t, use_cnt = ATOMIC_INIT(0)); >> + >> +static struct { >> + struct cpumask **tx_valid_cpus; >> + int num_nodes; >> +} siw_cpu_info; >> + >> +static int siw_init_cpulist(void) >> +{ >> + int i, num_nodes; >> + >> + num_nodes = num_possible_nodes(); >> + siw_cpu_info.num_nodes = num_nodes; >> + >> + siw_cpu_info.tx_valid_cpus = kcalloc(num_nodes, sizeof(void *), >> + GFP_KERNEL); >> + if (!siw_cpu_info.tx_valid_cpus) { >> + siw_cpu_info.num_nodes = 0; >> + return -ENOMEM; >> + } >> + >> + for (i = 0; i < siw_cpu_info.num_nodes; i++) { >> + siw_cpu_info.tx_valid_cpus[i] = kzalloc(sizeof(struct cpumask), >> + GFP_KERNEL); >> + if (!siw_cpu_info.tx_valid_cpus[i]) >> + goto out_err; >> + >> + cpumask_clear(siw_cpu_info.tx_valid_cpus[i]); >> + } >> + for_each_possible_cpu(i) >> + cpumask_set_cpu(i, siw_cpu_info.tx_valid_cpus[cpu_to_node(i)]); >> + >> + return 0; >> + >> +out_err: >> + siw_cpu_info.num_nodes = 0; >> + while (i) { >> + kfree(siw_cpu_info.tx_valid_cpus[i]); >> + siw_cpu_info.tx_valid_cpus[i--] = NULL; >> + } >> + kfree(siw_cpu_info.tx_valid_cpus); >> + siw_cpu_info.tx_valid_cpus = NULL; >> + >> + return -ENOMEM; >> +} >> + >> +static void siw_destroy_cpulist(void) >> +{ >> + int i = 0; >> + >> + while (i < siw_cpu_info.num_nodes) >> + kfree(siw_cpu_info.tx_valid_cpus[i++]); >> + >> + kfree(siw_cpu_info.tx_valid_cpus); >> +} >> + >> +/* >> + * Choose CPU with least number of active QP's from NUMA node of >> + * TX interface. >> + */ >> +int siw_get_tx_cpu(struct siw_device *sdev) >> +{ >> + const struct cpumask *tx_cpumask; >> + int i, num_cpus, cpu, min_use, >> + node = sdev->numa_node, >> + tx_cpu = -1; >> + >> + if (node < 0) >> + tx_cpumask = cpu_online_mask; >> + else >> + tx_cpumask = siw_cpu_info.tx_valid_cpus[node]; >> + >> + num_cpus = cpumask_weight(tx_cpumask); >> + if (!num_cpus) { >> + /* no CPU on this NUMA node */ >> + tx_cpumask = cpu_online_mask; >> + num_cpus = cpumask_weight(tx_cpumask); >> + } >> + if (!num_cpus) >> + goto out; >> + >> + cpu = cpumask_first(tx_cpumask); >> + >> + for (i = 0, min_use = SIW_MAX_QP; i < num_cpus; >> + i++, cpu = cpumask_next(cpu, tx_cpumask)) { >> + int usage; >> + >> + /* Skip any cores which have no TX thread */ >> + if (!siw_tx_thread[cpu]) >> + continue; >> + >> + usage = atomic_read(&per_cpu(use_cnt, cpu)); >> + if (usage <= min_use) { >> + tx_cpu = cpu; >> + min_use = usage; >> + } >> + } >> + siw_dbg(sdev, "tx cpu %d, node %d, %d qp's\n", >> + tx_cpu, node, min_use); >> + >> +out: >> + if (tx_cpu >= 0) >> + atomic_inc(&per_cpu(use_cnt, tx_cpu)); >> + else >> + pr_warn("siw: no tx cpu found\n"); >> + >> + return tx_cpu; >> +} >> + >> +void siw_put_tx_cpu(int cpu) >> +{ >> + atomic_dec(&per_cpu(use_cnt, cpu)); >> +} >> + >> +static void siw_verbs_sq_flush(struct ib_qp *base_qp) >> +{ >> + struct siw_qp *qp = to_siw_qp(base_qp); >> + >> + down_write(&qp->state_lock); >> + siw_sq_flush(qp); >> + up_write(&qp->state_lock); >> +} >> + >> +static void siw_verbs_rq_flush(struct ib_qp *base_qp) >> +{ >> + struct siw_qp *qp = to_siw_qp(base_qp); >> + >> + down_write(&qp->state_lock); >> + siw_rq_flush(qp); >> + up_write(&qp->state_lock); >> +} >> + >> +static const struct ib_device_ops siw_device_ops = { >> + .query_device = siw_query_device, >> + .query_port = siw_query_port, >> + .get_port_immutable = siw_get_port_immutable, >> + .get_netdev = siw_get_netdev, >> + .query_qp = siw_query_qp, >> + .modify_port = siw_modify_port, >> + .query_pkey = siw_query_pkey, >> + .query_gid = siw_query_gid, >> + .alloc_ucontext = siw_alloc_ucontext, >> + .dealloc_ucontext = siw_dealloc_ucontext, >> + .mmap = siw_mmap, >> + .alloc_pd = siw_alloc_pd, >> + .dealloc_pd = siw_dealloc_pd, >> + .create_qp = siw_create_qp, >> + .modify_qp = siw_verbs_modify_qp, >> + .destroy_qp = siw_destroy_qp, >> + .create_cq = siw_create_cq, >> + .destroy_cq = siw_destroy_cq, >> + .poll_cq = siw_poll_cq, >> + .get_dma_mr = siw_get_dma_mr, >> + .reg_user_mr = siw_reg_user_mr, >> + .dereg_mr = siw_dereg_mr, >> + .alloc_mr = siw_alloc_mr, >> + .map_mr_sg = siw_map_mr_sg, >> + .create_srq = siw_create_srq, >> + .modify_srq = siw_modify_srq, >> + .query_srq = siw_query_srq, >> + .destroy_srq = siw_destroy_srq, >> + .post_srq_recv = siw_post_srq_recv, >> + .req_notify_cq = siw_req_notify_cq, >> + .post_send = siw_post_send, >> + .post_recv = siw_post_receive, >> + .drain_sq = siw_verbs_sq_flush, >> + .drain_rq = siw_verbs_rq_flush, >> + .dealloc_driver = siw_device_cleanup, >> + INIT_RDMA_OBJ_SIZE(ib_pd, siw_pd, base_pd), >> +}; >> + >> +static struct siw_device *siw_device_create(struct net_device >*netdev) >> +{ >> + struct siw_device *sdev; >> + struct ib_device *base_dev; >> + struct device *parent = netdev->dev.parent; >> + int rv; >> + >> + sdev = ib_alloc_device(siw_device, base_dev); >> + if (!sdev) >> + goto out; >> + >> + base_dev = &sdev->base_dev; >> + >> + if (!parent) { >> + /* >> + * The loopback device has no parent device, >> + * so it appears as a top-level device. To support >> + * loopback device connectivity, take this device >> + * as the parent device. Skip all other devices >> + * w/o parent device. >> + */ >> + if (netdev->type != ARPHRD_LOOPBACK) { >> + pr_warn("siw: device %s skipped (no parent dev)\n", >> + netdev->name); >> + ib_dealloc_device(base_dev); >> + sdev = NULL; >> + goto out; >> + } >> + parent = &netdev->dev; >> + } >> + base_dev->iwcm = kmalloc(sizeof(struct iw_cm_verbs), GFP_KERNEL); >> + if (!base_dev->iwcm) { >> + ib_dealloc_device(base_dev); >> + sdev = NULL; >> + goto out; >> + } >> + >> + sdev->netdev = netdev; >> + >> + memset(&base_dev->node_guid, 0, sizeof(base_dev->node_guid)); >> + >> + if (netdev->type != ARPHRD_LOOPBACK) { >> + memcpy(&base_dev->node_guid, netdev->dev_addr, 6); >> + } else { >> + /* >> + * The loopback device does not have a HW address, >> + * but connection mangagement lib expects gid != 0 >> + */ >> + size_t gidlen = min_t(size_t, strlen(base_dev->name), 6); >> + >> + memcpy(&base_dev->node_guid, base_dev->name, gidlen); >> + } >> + base_dev->owner = THIS_MODULE; >> + >> + base_dev->uverbs_cmd_mask = >> + (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) | >> + (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) | >> + (1ull << IB_USER_VERBS_CMD_QUERY_PORT) | >> + (1ull << IB_USER_VERBS_CMD_ALLOC_PD) | >> + (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) | >> + (1ull << IB_USER_VERBS_CMD_REG_MR) | >> + (1ull << IB_USER_VERBS_CMD_DEREG_MR) | >> + (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) | >> + (1ull << IB_USER_VERBS_CMD_CREATE_CQ) | >> + (1ull << IB_USER_VERBS_CMD_POLL_CQ) | >> + (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) | >> + (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) | >> + (1ull << IB_USER_VERBS_CMD_CREATE_QP) | >> + (1ull << IB_USER_VERBS_CMD_QUERY_QP) | >> + (1ull << IB_USER_VERBS_CMD_MODIFY_QP) | >> + (1ull << IB_USER_VERBS_CMD_DESTROY_QP) | >> + (1ull << IB_USER_VERBS_CMD_POST_SEND) | >> + (1ull << IB_USER_VERBS_CMD_POST_RECV) | >> + (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) | >> + (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) | >> + (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) | >> + (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) | >> + (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV); >> + >> + base_dev->node_type = RDMA_NODE_RNIC; >> + memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON, >> + sizeof(SIW_NODE_DESC_COMMON)); >> + >> + /* >> + * Current model (one-to-one device association): >> + * One Softiwarp device per net_device or, equivalently, >> + * per physical port. >> + */ >> + base_dev->phys_port_cnt = 1; >> + base_dev->dev.parent = parent; >> + base_dev->dev.dma_ops = &dma_virt_ops; >> + base_dev->num_comp_vectors = num_possible_cpus(); >> + >> + ib_set_device_ops(base_dev, &siw_device_ops); >> + rv = ib_device_set_netdev(base_dev, netdev, 1); >> + if (rv) { >> + kfree(base_dev->iwcm); >> + ib_dealloc_device(base_dev); >> + sdev = NULL; >> + goto out; >> + } >> + base_dev->iwcm->connect = siw_connect; >> + base_dev->iwcm->accept = siw_accept; >> + base_dev->iwcm->reject = siw_reject; >> + base_dev->iwcm->create_listen = siw_create_listen; >> + base_dev->iwcm->destroy_listen = siw_destroy_listen; >> + base_dev->iwcm->add_ref = siw_qp_get_ref; >> + base_dev->iwcm->rem_ref = siw_qp_put_ref; >> + base_dev->iwcm->get_qp = siw_get_base_qp; >> + >> + /* Disable TCP port mapper service */ >> + base_dev->iwcm->driver_flags = IW_F_NO_PORT_MAP; >> + >> + memcpy(base_dev->iwcm->ifname, netdev->name, >> + sizeof(base_dev->iwcm->ifname)); >> + >> + sdev->attrs.max_qp = SIW_MAX_QP; >> + sdev->attrs.max_qp_wr = SIW_MAX_QP_WR; >> + sdev->attrs.max_ord = SIW_MAX_ORD_QP; >> + sdev->attrs.max_ird = SIW_MAX_IRD_QP; >> + sdev->attrs.max_sge = SIW_MAX_SGE; >> + sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD; >> + sdev->attrs.max_cq = SIW_MAX_CQ; >> + sdev->attrs.max_cqe = SIW_MAX_CQE; >> + sdev->attrs.max_mr = SIW_MAX_MR; >> + sdev->attrs.max_pd = SIW_MAX_PD; >> + sdev->attrs.max_mw = SIW_MAX_MW; >> + sdev->attrs.max_fmr = SIW_MAX_FMR; >> + sdev->attrs.max_srq = SIW_MAX_SRQ; >> + sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR; >> + sdev->attrs.max_srq_sge = SIW_MAX_SGE; >> + >> + siw_idr_init(sdev); >> + INIT_LIST_HEAD(&sdev->cep_list); >> + INIT_LIST_HEAD(&sdev->qp_list); >> + INIT_LIST_HEAD(&sdev->mr_list); >> + >> + atomic_set(&sdev->num_ctx, 0); >> + atomic_set(&sdev->num_srq, 0); >> + atomic_set(&sdev->num_qp, 0); >> + atomic_set(&sdev->num_cq, 0); >> + atomic_set(&sdev->num_mr, 0); >> + atomic_set(&sdev->num_pd, 0); >> + atomic_set(&sdev->num_cep, 0); >> + >> + sdev->numa_node = dev_to_node(parent); >> +out: >> + return sdev; >> +} >> + >> +/* >> + * Network link becomes unavailable. Mark all >> + * affected QP's accordingly. >> + */ >> +static void siw_netdev_down(struct work_struct *work) >> +{ >> + struct siw_device *sdev = container_of(work, struct siw_device, >> + netdev_down); >> + >> + struct siw_qp_attrs qp_attrs; >> + struct list_head *pos, *tmp; >> + >> + memset(&qp_attrs, 0, sizeof(qp_attrs)); >> + qp_attrs.state = SIW_QP_STATE_ERROR; >> + >> + list_for_each_safe(pos, tmp, &sdev->qp_list) { >> + struct siw_qp *qp = list_entry(pos, struct siw_qp, devq); >> + >> + down_write(&qp->state_lock); >> + (void) siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE); >> + up_write(&qp->state_lock); >> + } >> + ib_device_put(&sdev->base_dev); >> +} >> + >> +static void siw_device_goes_down(struct siw_device *sdev) >> +{ >> + if (ib_device_try_get(&sdev->base_dev)) { >> + INIT_WORK(&sdev->netdev_down, siw_netdev_down); >> + schedule_work(&sdev->netdev_down); >> + } >> +} >> + >> +static int siw_netdev_event(struct notifier_block *nb, unsigned >long event, >> + void *arg) >> +{ >> + struct net_device *netdev = netdev_notifier_info_to_dev(arg); >> + struct ib_device *base_dev; >> + struct siw_device *sdev; >> + >> + dev_dbg(&netdev->dev, "siw: event %lu\n", event); >> + >> + if (dev_net(netdev) != &init_net) >> + return NOTIFY_OK; >> + >> + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); >> + if (!base_dev) >> + return NOTIFY_OK; >> + >> + sdev = to_siw_dev(base_dev); >> + >> + switch (event) { >> + >> + case NETDEV_UP: >> + sdev->state = IB_PORT_ACTIVE; >> + siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE); >> + break; >> + >> + case NETDEV_GOING_DOWN: >> + siw_device_goes_down(sdev); >> + break; >> + >> + case NETDEV_DOWN: >> + sdev->state = IB_PORT_DOWN; >> + siw_port_event(sdev, 1, IB_EVENT_PORT_ERR); >> + break; >> + >> + case NETDEV_REGISTER: >> + /* >> + * Device registration now handled only by >> + * rdma netlink commands. So it shall be impossible >> + * to end up here with a valid siw device. >> + */ >> + siw_dbg(sdev, "unexpected NETDEV_REGISTER event\n"); >> + break; >> + >> + case NETDEV_UNREGISTER: >> + ib_unregister_device_queued(&sdev->base_dev); >> + break; >> + >> + case NETDEV_CHANGEADDR: >> + siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE); >> + break; >> + /* >> + * Todo: Below netdev events are currently not handled. >> + */ >> + case NETDEV_CHANGEMTU: >> + case NETDEV_CHANGE: >> + >> + break; >> + >> + default: >> + break; >> + } >> + ib_device_put(&sdev->base_dev); >> + >> + return NOTIFY_OK; >> +} >> + >> +static struct notifier_block siw_netdev_nb = { >> + .notifier_call = siw_netdev_event, >> +}; >> + >> +static int siw_newlink(const char *basedev_name, struct net_device >*netdev) >> +{ >> + struct ib_device *base_dev; >> + struct siw_device *sdev; >> + int rv; >> + >> + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); >> + if (base_dev) { >> + ib_device_put(base_dev); >> + rv = -EEXIST; >> + goto out; >> + } >> + if (!siw_dev_qualified(netdev)) { >> + rv = -EINVAL; >> + goto out; >> + } >> + sdev = siw_device_create(netdev); >> + if (sdev) { >> + dev_dbg(&netdev->dev, "siw: new device\n"); >> + >> + if (netif_running(netdev) && netif_carrier_ok(netdev)) >> + sdev->state = IB_PORT_ACTIVE; >> + else >> + sdev->state = IB_PORT_DOWN; >> + >> + rv = siw_device_register(sdev, basedev_name); >> + if (rv) { >> + siw_device_cleanup(&sdev->base_dev); >> + ib_dealloc_device(&sdev->base_dev); >> + goto out; >> + } >> + } else { >> + rv = -ENOMEM; >> + } >> +out: >> + return rv; >> +} >> + >> +static struct rdma_link_ops siw_link_ops = { >> + .type = "siw", >> + .newlink = siw_newlink, >> +}; >> + >> +/* >> + * siw_init_module - Initialize Softiwarp module and register with >netdev >> + * subsystem to create Softiwarp devices per >net_device >> + */ >> +static __init int siw_init_module(void) >> +{ >> + int rv; >> + int nr_cpu; >> + >> + if (SENDPAGE_THRESH < SIW_MAX_INLINE) { >> + pr_info("siw: sendpage threshold too small: %u\n", >> + (int)SENDPAGE_THRESH); >> + rv = EINVAL; > >-EINVAL? right, thanks. > >> + goto out_error; >> + } >> + rv = siw_init_cpulist(); >> + if (rv) >> + goto out_error; >> + >> + rv = siw_cm_init(); >> + if (rv) >> + goto out_error; >> + >> + /* >> + * Allocate CRC SHASH object. Fail loading siw only, if CRC is >> + * required by kernel module >> + */ >> + siw_crypto_shash = crypto_alloc_shash("crc32c", 0, 0); >> + if (IS_ERR(siw_crypto_shash)) { >> + pr_info("siw: Loading CRC32c failed: %ld\n", >> + PTR_ERR(siw_crypto_shash)); >> + siw_crypto_shash = NULL; >> + if (mpa_crc_required == true) > >Bool tests don't need comparisons. > true ;) >> + goto out_error; >> + } >> + rv = register_netdevice_notifier(&siw_netdev_nb); >> + if (rv) >> + goto out_error; >> + >> + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) >> + siw_tx_thread[nr_cpu] = NULL; >> + >> + if (!siw_create_tx_threads()) { >> + pr_info("siw: Could not start any TX thread\n"); >> + unregister_netdevice_notifier(&siw_netdev_nb); >> + goto out_error; >> + } >> + rdma_link_register(&siw_link_ops); >> + >> + pr_info("SoftiWARP attached\n"); >> + return 0; >> + >> +out_error: >> + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) { >> + if (siw_tx_thread[nr_cpu]) { >> + siw_stop_tx_thread(nr_cpu); >> + siw_tx_thread[nr_cpu] = NULL; >> + } >> + } >> + if (siw_crypto_shash) >> + crypto_free_shash(siw_crypto_shash); >> + >> + pr_info("SoftIWARP attach failed. Error: %d\n", rv); >> + >> + siw_cm_exit(); >> + siw_destroy_cpulist(); >> + >> + return rv; >> +} >> + >> +static void __exit siw_exit_module(void) >> +{ >> + int nr_cpu; >> + >> + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) { >> + if (siw_tx_thread[nr_cpu]) { >> + siw_stop_tx_thread(nr_cpu); >> + siw_tx_thread[nr_cpu] = NULL; >> + } >> + } >> + unregister_netdevice_notifier(&siw_netdev_nb); >> + rdma_link_unregister(&siw_link_ops); >> + ib_unregister_driver(RDMA_DRIVER_SIW); >> + >> + siw_cm_exit(); >> + >> + if (siw_crypto_shash) >> + crypto_free_shash(siw_crypto_shash); >> + >> + siw_destroy_cpulist(); >> + >> + pr_info("SoftiWARP detached\n"); >> +} >> + >> +module_init(siw_init_module); >> +module_exit(siw_exit_module); >> + >> +MODULE_ALIAS_RDMA_LINK("siw"); >> > >Reviewed-by: Gal Pressman <galpress@amazon.com> > >
On Tue, Feb 26, 2019 at 04:03:39PM -0600, Steve Wise wrote: > > On 2/21/2019 6:02 PM, Jason Gunthorpe wrote: > > On Tue, Feb 19, 2019 at 11:08:53AM +0100, Bernard Metzler wrote: > > > >> +static const struct ib_device_ops siw_device_ops = { > >> + .query_device = siw_query_device, > >> + .query_port = siw_query_port, > >> + .get_port_immutable = siw_get_port_immutable, > >> + .get_netdev = siw_get_netdev, > > This driver uses ib_device_set_netdev so get_netdev should not be set. > > > Should ib_device_set_netdev() have a WARN() for this perhaps? It is a good idea.. But ib_device_set_netdev is potentially called before ops is set.. Jason
On Wed, Feb 27, 2019 at 12:11:30PM +0000, Bernard Metzler wrote: > > >To: "Bernard Metzler" <bmt@zurich.ibm.com> > >From: "Jason Gunthorpe" <jgg@ziepe.ca> > >Date: 02/22/2019 01:02AM > >Cc: linux-rdma@vger.kernel.org > >Subject: Re: [PATCH v5 03/13] SIW network and RDMA core interface > > > >On Tue, Feb 19, 2019 at 11:08:53AM +0100, Bernard Metzler wrote: > > > >> +static const struct ib_device_ops siw_device_ops = { > >> + .query_device = siw_query_device, > >> + .query_port = siw_query_port, > >> + .get_port_immutable = siw_get_port_immutable, > >> + .get_netdev = siw_get_netdev, > > > >This driver uses ib_device_set_netdev so get_netdev should not be > >set. > > > >> + .query_qp = siw_query_qp, > >> + .modify_port = siw_modify_port, > >> + .query_pkey = siw_query_pkey, > >> + .query_gid = siw_query_gid, > >> + .alloc_ucontext = siw_alloc_ucontext, > >> + .dealloc_ucontext = siw_dealloc_ucontext, > >> + .mmap = siw_mmap, > >> + .alloc_pd = siw_alloc_pd, > >> + .dealloc_pd = siw_dealloc_pd, > >> + .create_qp = siw_create_qp, > >> + .modify_qp = siw_verbs_modify_qp, > >> + .destroy_qp = siw_destroy_qp, > >> + .create_cq = siw_create_cq, > >> + .destroy_cq = siw_destroy_cq, > >> + .poll_cq = siw_poll_cq, > >> + .get_dma_mr = siw_get_dma_mr, > >> + .reg_user_mr = siw_reg_user_mr, > >> + .dereg_mr = siw_dereg_mr, > >> + .alloc_mr = siw_alloc_mr, > >> + .map_mr_sg = siw_map_mr_sg, > >> + .create_srq = siw_create_srq, > >> + .modify_srq = siw_modify_srq, > >> + .query_srq = siw_query_srq, > >> + .destroy_srq = siw_destroy_srq, > >> + .post_srq_recv = siw_post_srq_recv, > >> + .req_notify_cq = siw_req_notify_cq, > >> + .post_send = siw_post_send, > >> + .post_recv = siw_post_receive, > >> + .drain_sq = siw_verbs_sq_flush, > >> + .drain_rq = siw_verbs_rq_flush, > >> + .dealloc_driver = siw_device_cleanup, > > > >Keep lists of things sorted. Here and everywhere, makefiles, kconfig, > >etc. > > I *tried* to do so - by grouping methods here according to the > resource it affects. But you are right, this is badly sorted, > even under that premise! Will fix that (but let's not order > alphabetical here, right?). Alphabetical is the way all other drivers do it. Otherwise it is a royal pain and subject to interpretation. Jason
On Wed, Feb 27, 2019 at 01:17:51PM +0000, Bernard Metzler wrote: > >> + sdev->attrs.max_qp = SIW_MAX_QP; > >> + sdev->attrs.max_qp_wr = SIW_MAX_QP_WR; > >> + sdev->attrs.max_ord = SIW_MAX_ORD_QP; > >> + sdev->attrs.max_ird = SIW_MAX_IRD_QP; > >> + sdev->attrs.max_sge = SIW_MAX_SGE; > >> + sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD; > >> + sdev->attrs.max_cq = SIW_MAX_CQ; > >> + sdev->attrs.max_cqe = SIW_MAX_CQE; > >> + sdev->attrs.max_mr = SIW_MAX_MR; > >> + sdev->attrs.max_pd = SIW_MAX_PD; > >> + sdev->attrs.max_mw = SIW_MAX_MW; > >> + sdev->attrs.max_fmr = SIW_MAX_FMR; > >> + sdev->attrs.max_srq = SIW_MAX_SRQ; > >> + sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR; > >> + sdev->attrs.max_srq_sge = SIW_MAX_SGE; > >> + > >> + siw_idr_init(sdev); > > > >Why idr_init/release functions? Seems unnecessary > > idr structs are allocated for each dynamically allocated > siw device (one device per link). So each device has its own > id space (e.g. for STags/Keys). So I think I need to > initialize/destroy the idr resources dynamically as well. I mean why the function siw_idr_init called from exactly one place that has a couple of lines in it? > >> +static void siw_device_goes_down(struct siw_device *sdev) > >> +{ > >> + if (ib_device_try_get(&sdev->base_dev)) { > >> + INIT_WORK(&sdev->netdev_down, siw_netdev_down); > >> + schedule_work(&sdev->netdev_down); > > > >.. ie here it should be the system unbound wq, but that would > >deadlock > >with ib_unregister_device_queued .. > > > > Hmm, can you please help me understanding that > concern? I agree I should check if schedule_work() > succeeds (I have to fix that, but it should never happen), > but otherwise? > How can it deadlock? The work queue could be busy with work that is waiting for the refcount to go to 0 and not executing the work that takes the refcount to zero. > >> +static int siw_netdev_event(struct notifier_block *nb, unsigned > >long event, > >> + void *arg) > >> +{ > >> + struct net_device *netdev = netdev_notifier_info_to_dev(arg); > >> + struct ib_device *base_dev; > >> + struct siw_device *sdev; > >> + > >> + dev_dbg(&netdev->dev, "siw: event %lu\n", event); > >> + > >> + if (dev_net(netdev) != &init_net) > >> + return NOTIFY_OK; > >> + > >> + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); > >> + if (!base_dev) > >> + return NOTIFY_OK; > >> + > >> + sdev = to_siw_dev(base_dev); > >> + > >> + switch (event) { > >> + > >> + case NETDEV_UP: > >> + sdev->state = IB_PORT_ACTIVE; > >> + siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE); > >> + break; > >> + > >> + case NETDEV_GOING_DOWN: > >> + siw_device_goes_down(sdev); > >> + break; > > > >This probably wants to be return as there is a double put > >otherwise. Never tested? > > > I tested bringing down interfaces, also while there is siw caused > traffic on it. > siw_device_goes_down() calls ib_device_try_get(), which > schedules work only if it got that additional reference > (which then get's put in the work handler). So I don't see a > potential double put case (yet)...? It is not such a great idea to call ib_device_try_get when you already have a get, as in that case it cannot fail and things just get confusing. Jason
diff --git a/drivers/infiniband/sw/siw/siw_main.c b/drivers/infiniband/sw/siw/siw_main.c new file mode 100644 index 000000000000..9758b7291dae --- /dev/null +++ b/drivers/infiniband/sw/siw/siw_main.c @@ -0,0 +1,813 @@ +// SPDX-License-Identifier: GPL-2.0 or BSD-3-Clause +/* + * Software iWARP device driver + * + * Authors: Bernard Metzler <bmt@zurich.ibm.com> + * + * Copyright (c) 2008-2019, IBM Corporation + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above copyright notice, + * this list of conditions and the following disclaimer. + * + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * - Neither the name of IBM nor the names of its contributors may be + * used to endorse or promote products derived from this software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include <linux/init.h> +#include <linux/errno.h> +#include <linux/netdevice.h> +#include <linux/inetdevice.h> +#include <net/net_namespace.h> +#include <linux/rtnetlink.h> +#include <linux/if_arp.h> +#include <linux/list.h> +#include <linux/kernel.h> +#include <linux/dma-mapping.h> + +#include <rdma/ib_verbs.h> +#include <rdma/ib_smi.h> +#include <rdma/ib_user_verbs.h> +#include <rdma/rdma_netlink.h> +#include <linux/kthread.h> + +#include "siw.h" +#include "siw_obj.h" +#include "siw_cm.h" +#include "siw_verbs.h" + +MODULE_AUTHOR("Bernard Metzler"); +MODULE_DESCRIPTION("Software iWARP Driver"); +MODULE_LICENSE("Dual BSD/GPL"); +MODULE_VERSION("0.2"); + +/* transmit from user buffer, if possible */ +const bool zcopy_tx = true; + +/* Restrict usage of GSO, if hardware peer iwarp is unable to process + * large packets. gso_seg_limit = 1 lets siw send only packets up to + * one real MTU in size, but severly limits siw maximum bandwidth. + * gso_seg_limit = 0 makes use of GSO for large transfers. + */ +const u8 gso_seg_limit = 1; + +/* Attach siw also with loopback devices */ +const bool loopback_enabled = true; + +/* We try to negotiate CRC on, if true */ +const bool mpa_crc_required; + +/* MPA CRC on/off enforced */ +const bool mpa_crc_strict; + +/* Control TCP_NODELAY socket option */ +const bool siw_tcp_nagle; + +/* Select MPA version to be used during connection setup */ +u_char mpa_version = MPA_REVISION_2; + +/* Selects MPA P2P mode (additional handshake during connection + * setup, if true. + */ +const bool peer_to_peer; + +struct task_struct *siw_tx_thread[NR_CPUS]; +struct crypto_shash *siw_crypto_shash; + +static ssize_t sw_version_show(struct device *dev, + struct device_attribute *attr, char *buf) +{ + return sprintf(buf, "%x\n", VERSION_ID_SOFTIWARP); +} + +static DEVICE_ATTR_RO(sw_version); + +static ssize_t parent_show(struct device *device, + struct device_attribute *attr, char *buf) +{ + struct siw_device *sdev = + rdma_device_to_drv_device(device, struct siw_device, base_dev); + + return snprintf(buf, 16, "%s\n", sdev->netdev->name); +} + +static DEVICE_ATTR_RO(parent); + +static struct attribute *siw_dev_attributes[] = { + &dev_attr_sw_version.attr, + &dev_attr_parent.attr, + NULL, +}; + +static const struct attribute_group siw_attr_group = { + .attrs = siw_dev_attributes, +}; + +static int siw_modify_port(struct ib_device *base_dev, u8 port, int mask, + struct ib_port_modify *props) +{ + return -EOPNOTSUPP; +} + +static int siw_device_register(struct siw_device *sdev, const char *name) +{ + struct ib_device *base_dev = &sdev->base_dev; + static int dev_id = 1; + int rv; + + base_dev->driver_id = RDMA_DRIVER_SIW; + rdma_set_device_sysfs_group(base_dev, &siw_attr_group); + + rv = ib_register_device(base_dev, name); + if (rv) { + pr_warn("siw: device registration error %d\n", rv); + return rv; + } + sdev->vendor_part_id = dev_id++; + + siw_dbg(sdev, "HWaddr=%02x.%02x.%02x.%02x.%02x.%02x\n", + *(u8 *)sdev->netdev->dev_addr, + *((u8 *)sdev->netdev->dev_addr + 1), + *((u8 *)sdev->netdev->dev_addr + 2), + *((u8 *)sdev->netdev->dev_addr + 3), + *((u8 *)sdev->netdev->dev_addr + 4), + *((u8 *)sdev->netdev->dev_addr + 5)); + + return 0; +} + +static void siw_device_cleanup(struct ib_device *base_dev) +{ + struct siw_device *sdev = to_siw_dev(base_dev); + + siw_dbg(sdev, "Cleanup device\n"); + + if (atomic_read(&sdev->num_ctx) || atomic_read(&sdev->num_srq) || + atomic_read(&sdev->num_mr) || atomic_read(&sdev->num_cep) || + atomic_read(&sdev->num_qp) || atomic_read(&sdev->num_cq) || + atomic_read(&sdev->num_pd)) { + pr_warn("siw at %s: orphaned resources!\n", + sdev->netdev->name); + pr_warn(" CTX %d, SRQ %d, QP %d, CQ %d, MEM %d, CEP %d, PD %d\n", + atomic_read(&sdev->num_ctx), + atomic_read(&sdev->num_srq), + atomic_read(&sdev->num_qp), + atomic_read(&sdev->num_cq), + atomic_read(&sdev->num_mr), + atomic_read(&sdev->num_cep), + atomic_read(&sdev->num_pd)); + } + while (!list_empty(&sdev->cep_list)) { + struct siw_cep *cep = list_entry(sdev->cep_list.next, + struct siw_cep, devq); + list_del(&cep->devq); + pr_warn("siw: at %s: free orphaned CEP 0x%p, state %d\n", + sdev->base_dev.name, cep, cep->state); + kfree(cep); + } + siw_idr_release(sdev); + kfree(sdev->base_dev.iwcm); +} + +static struct net_device *siw_get_netdev(struct ib_device *base_dev, u8 port) +{ + struct siw_device *sdev = to_siw_dev(base_dev); + + if (!sdev->netdev) + return NULL; + + dev_hold(sdev->netdev); + + return sdev->netdev; +} + +static int siw_create_tx_threads(void) +{ + int cpu, rv, assigned = 0; + + for_each_online_cpu(cpu) { + /* Skip HT cores */ + if (cpu % cpumask_weight(topology_sibling_cpumask(cpu))) { + siw_tx_thread[cpu] = NULL; + continue; + } + siw_tx_thread[cpu] = kthread_create(siw_run_sq, + (unsigned long *)(long)cpu, + "siw_tx/%d", cpu); + if (IS_ERR(siw_tx_thread[cpu])) { + rv = PTR_ERR(siw_tx_thread[cpu]); + siw_tx_thread[cpu] = NULL; + pr_info("Creating TX thread for CPU %d failed", cpu); + continue; + } + kthread_bind(siw_tx_thread[cpu], cpu); + + wake_up_process(siw_tx_thread[cpu]); + assigned++; + } + return assigned; +} + +static int siw_dev_qualified(struct net_device *netdev) +{ + /* + * Additional hardware support can be added here + * (e.g. ARPHRD_FDDI, ARPHRD_ATM, ...) - see + * <linux/if_arp.h> for type identifiers. + */ + if (netdev->type == ARPHRD_ETHER || + netdev->type == ARPHRD_IEEE802 || + (netdev->type == ARPHRD_LOOPBACK && loopback_enabled)) + return 1; + + return 0; +} + +static DEFINE_PER_CPU(atomic_t, use_cnt = ATOMIC_INIT(0)); + +static struct { + struct cpumask **tx_valid_cpus; + int num_nodes; +} siw_cpu_info; + +static int siw_init_cpulist(void) +{ + int i, num_nodes; + + num_nodes = num_possible_nodes(); + siw_cpu_info.num_nodes = num_nodes; + + siw_cpu_info.tx_valid_cpus = kcalloc(num_nodes, sizeof(void *), + GFP_KERNEL); + if (!siw_cpu_info.tx_valid_cpus) { + siw_cpu_info.num_nodes = 0; + return -ENOMEM; + } + + for (i = 0; i < siw_cpu_info.num_nodes; i++) { + siw_cpu_info.tx_valid_cpus[i] = kzalloc(sizeof(struct cpumask), + GFP_KERNEL); + if (!siw_cpu_info.tx_valid_cpus[i]) + goto out_err; + + cpumask_clear(siw_cpu_info.tx_valid_cpus[i]); + } + for_each_possible_cpu(i) + cpumask_set_cpu(i, siw_cpu_info.tx_valid_cpus[cpu_to_node(i)]); + + return 0; + +out_err: + siw_cpu_info.num_nodes = 0; + while (i) { + kfree(siw_cpu_info.tx_valid_cpus[i]); + siw_cpu_info.tx_valid_cpus[i--] = NULL; + } + kfree(siw_cpu_info.tx_valid_cpus); + siw_cpu_info.tx_valid_cpus = NULL; + + return -ENOMEM; +} + +static void siw_destroy_cpulist(void) +{ + int i = 0; + + while (i < siw_cpu_info.num_nodes) + kfree(siw_cpu_info.tx_valid_cpus[i++]); + + kfree(siw_cpu_info.tx_valid_cpus); +} + +/* + * Choose CPU with least number of active QP's from NUMA node of + * TX interface. + */ +int siw_get_tx_cpu(struct siw_device *sdev) +{ + const struct cpumask *tx_cpumask; + int i, num_cpus, cpu, min_use, + node = sdev->numa_node, + tx_cpu = -1; + + if (node < 0) + tx_cpumask = cpu_online_mask; + else + tx_cpumask = siw_cpu_info.tx_valid_cpus[node]; + + num_cpus = cpumask_weight(tx_cpumask); + if (!num_cpus) { + /* no CPU on this NUMA node */ + tx_cpumask = cpu_online_mask; + num_cpus = cpumask_weight(tx_cpumask); + } + if (!num_cpus) + goto out; + + cpu = cpumask_first(tx_cpumask); + + for (i = 0, min_use = SIW_MAX_QP; i < num_cpus; + i++, cpu = cpumask_next(cpu, tx_cpumask)) { + int usage; + + /* Skip any cores which have no TX thread */ + if (!siw_tx_thread[cpu]) + continue; + + usage = atomic_read(&per_cpu(use_cnt, cpu)); + if (usage <= min_use) { + tx_cpu = cpu; + min_use = usage; + } + } + siw_dbg(sdev, "tx cpu %d, node %d, %d qp's\n", + tx_cpu, node, min_use); + +out: + if (tx_cpu >= 0) + atomic_inc(&per_cpu(use_cnt, tx_cpu)); + else + pr_warn("siw: no tx cpu found\n"); + + return tx_cpu; +} + +void siw_put_tx_cpu(int cpu) +{ + atomic_dec(&per_cpu(use_cnt, cpu)); +} + +static void siw_verbs_sq_flush(struct ib_qp *base_qp) +{ + struct siw_qp *qp = to_siw_qp(base_qp); + + down_write(&qp->state_lock); + siw_sq_flush(qp); + up_write(&qp->state_lock); +} + +static void siw_verbs_rq_flush(struct ib_qp *base_qp) +{ + struct siw_qp *qp = to_siw_qp(base_qp); + + down_write(&qp->state_lock); + siw_rq_flush(qp); + up_write(&qp->state_lock); +} + +static const struct ib_device_ops siw_device_ops = { + .query_device = siw_query_device, + .query_port = siw_query_port, + .get_port_immutable = siw_get_port_immutable, + .get_netdev = siw_get_netdev, + .query_qp = siw_query_qp, + .modify_port = siw_modify_port, + .query_pkey = siw_query_pkey, + .query_gid = siw_query_gid, + .alloc_ucontext = siw_alloc_ucontext, + .dealloc_ucontext = siw_dealloc_ucontext, + .mmap = siw_mmap, + .alloc_pd = siw_alloc_pd, + .dealloc_pd = siw_dealloc_pd, + .create_qp = siw_create_qp, + .modify_qp = siw_verbs_modify_qp, + .destroy_qp = siw_destroy_qp, + .create_cq = siw_create_cq, + .destroy_cq = siw_destroy_cq, + .poll_cq = siw_poll_cq, + .get_dma_mr = siw_get_dma_mr, + .reg_user_mr = siw_reg_user_mr, + .dereg_mr = siw_dereg_mr, + .alloc_mr = siw_alloc_mr, + .map_mr_sg = siw_map_mr_sg, + .create_srq = siw_create_srq, + .modify_srq = siw_modify_srq, + .query_srq = siw_query_srq, + .destroy_srq = siw_destroy_srq, + .post_srq_recv = siw_post_srq_recv, + .req_notify_cq = siw_req_notify_cq, + .post_send = siw_post_send, + .post_recv = siw_post_receive, + .drain_sq = siw_verbs_sq_flush, + .drain_rq = siw_verbs_rq_flush, + .dealloc_driver = siw_device_cleanup, + INIT_RDMA_OBJ_SIZE(ib_pd, siw_pd, base_pd), +}; + +static struct siw_device *siw_device_create(struct net_device *netdev) +{ + struct siw_device *sdev; + struct ib_device *base_dev; + struct device *parent = netdev->dev.parent; + int rv; + + sdev = ib_alloc_device(siw_device, base_dev); + if (!sdev) + goto out; + + base_dev = &sdev->base_dev; + + if (!parent) { + /* + * The loopback device has no parent device, + * so it appears as a top-level device. To support + * loopback device connectivity, take this device + * as the parent device. Skip all other devices + * w/o parent device. + */ + if (netdev->type != ARPHRD_LOOPBACK) { + pr_warn("siw: device %s skipped (no parent dev)\n", + netdev->name); + ib_dealloc_device(base_dev); + sdev = NULL; + goto out; + } + parent = &netdev->dev; + } + base_dev->iwcm = kmalloc(sizeof(struct iw_cm_verbs), GFP_KERNEL); + if (!base_dev->iwcm) { + ib_dealloc_device(base_dev); + sdev = NULL; + goto out; + } + + sdev->netdev = netdev; + + memset(&base_dev->node_guid, 0, sizeof(base_dev->node_guid)); + + if (netdev->type != ARPHRD_LOOPBACK) { + memcpy(&base_dev->node_guid, netdev->dev_addr, 6); + } else { + /* + * The loopback device does not have a HW address, + * but connection mangagement lib expects gid != 0 + */ + size_t gidlen = min_t(size_t, strlen(base_dev->name), 6); + + memcpy(&base_dev->node_guid, base_dev->name, gidlen); + } + base_dev->owner = THIS_MODULE; + + base_dev->uverbs_cmd_mask = + (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) | + (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) | + (1ull << IB_USER_VERBS_CMD_QUERY_PORT) | + (1ull << IB_USER_VERBS_CMD_ALLOC_PD) | + (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) | + (1ull << IB_USER_VERBS_CMD_REG_MR) | + (1ull << IB_USER_VERBS_CMD_DEREG_MR) | + (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) | + (1ull << IB_USER_VERBS_CMD_CREATE_CQ) | + (1ull << IB_USER_VERBS_CMD_POLL_CQ) | + (1ull << IB_USER_VERBS_CMD_REQ_NOTIFY_CQ) | + (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) | + (1ull << IB_USER_VERBS_CMD_CREATE_QP) | + (1ull << IB_USER_VERBS_CMD_QUERY_QP) | + (1ull << IB_USER_VERBS_CMD_MODIFY_QP) | + (1ull << IB_USER_VERBS_CMD_DESTROY_QP) | + (1ull << IB_USER_VERBS_CMD_POST_SEND) | + (1ull << IB_USER_VERBS_CMD_POST_RECV) | + (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) | + (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) | + (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) | + (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ) | + (1ull << IB_USER_VERBS_CMD_POST_SRQ_RECV); + + base_dev->node_type = RDMA_NODE_RNIC; + memcpy(base_dev->node_desc, SIW_NODE_DESC_COMMON, + sizeof(SIW_NODE_DESC_COMMON)); + + /* + * Current model (one-to-one device association): + * One Softiwarp device per net_device or, equivalently, + * per physical port. + */ + base_dev->phys_port_cnt = 1; + base_dev->dev.parent = parent; + base_dev->dev.dma_ops = &dma_virt_ops; + base_dev->num_comp_vectors = num_possible_cpus(); + + ib_set_device_ops(base_dev, &siw_device_ops); + rv = ib_device_set_netdev(base_dev, netdev, 1); + if (rv) { + kfree(base_dev->iwcm); + ib_dealloc_device(base_dev); + sdev = NULL; + goto out; + } + base_dev->iwcm->connect = siw_connect; + base_dev->iwcm->accept = siw_accept; + base_dev->iwcm->reject = siw_reject; + base_dev->iwcm->create_listen = siw_create_listen; + base_dev->iwcm->destroy_listen = siw_destroy_listen; + base_dev->iwcm->add_ref = siw_qp_get_ref; + base_dev->iwcm->rem_ref = siw_qp_put_ref; + base_dev->iwcm->get_qp = siw_get_base_qp; + + /* Disable TCP port mapper service */ + base_dev->iwcm->driver_flags = IW_F_NO_PORT_MAP; + + memcpy(base_dev->iwcm->ifname, netdev->name, + sizeof(base_dev->iwcm->ifname)); + + sdev->attrs.max_qp = SIW_MAX_QP; + sdev->attrs.max_qp_wr = SIW_MAX_QP_WR; + sdev->attrs.max_ord = SIW_MAX_ORD_QP; + sdev->attrs.max_ird = SIW_MAX_IRD_QP; + sdev->attrs.max_sge = SIW_MAX_SGE; + sdev->attrs.max_sge_rd = SIW_MAX_SGE_RD; + sdev->attrs.max_cq = SIW_MAX_CQ; + sdev->attrs.max_cqe = SIW_MAX_CQE; + sdev->attrs.max_mr = SIW_MAX_MR; + sdev->attrs.max_pd = SIW_MAX_PD; + sdev->attrs.max_mw = SIW_MAX_MW; + sdev->attrs.max_fmr = SIW_MAX_FMR; + sdev->attrs.max_srq = SIW_MAX_SRQ; + sdev->attrs.max_srq_wr = SIW_MAX_SRQ_WR; + sdev->attrs.max_srq_sge = SIW_MAX_SGE; + + siw_idr_init(sdev); + INIT_LIST_HEAD(&sdev->cep_list); + INIT_LIST_HEAD(&sdev->qp_list); + INIT_LIST_HEAD(&sdev->mr_list); + + atomic_set(&sdev->num_ctx, 0); + atomic_set(&sdev->num_srq, 0); + atomic_set(&sdev->num_qp, 0); + atomic_set(&sdev->num_cq, 0); + atomic_set(&sdev->num_mr, 0); + atomic_set(&sdev->num_pd, 0); + atomic_set(&sdev->num_cep, 0); + + sdev->numa_node = dev_to_node(parent); +out: + return sdev; +} + +/* + * Network link becomes unavailable. Mark all + * affected QP's accordingly. + */ +static void siw_netdev_down(struct work_struct *work) +{ + struct siw_device *sdev = container_of(work, struct siw_device, + netdev_down); + + struct siw_qp_attrs qp_attrs; + struct list_head *pos, *tmp; + + memset(&qp_attrs, 0, sizeof(qp_attrs)); + qp_attrs.state = SIW_QP_STATE_ERROR; + + list_for_each_safe(pos, tmp, &sdev->qp_list) { + struct siw_qp *qp = list_entry(pos, struct siw_qp, devq); + + down_write(&qp->state_lock); + (void) siw_qp_modify(qp, &qp_attrs, SIW_QP_ATTR_STATE); + up_write(&qp->state_lock); + } + ib_device_put(&sdev->base_dev); +} + +static void siw_device_goes_down(struct siw_device *sdev) +{ + if (ib_device_try_get(&sdev->base_dev)) { + INIT_WORK(&sdev->netdev_down, siw_netdev_down); + schedule_work(&sdev->netdev_down); + } +} + +static int siw_netdev_event(struct notifier_block *nb, unsigned long event, + void *arg) +{ + struct net_device *netdev = netdev_notifier_info_to_dev(arg); + struct ib_device *base_dev; + struct siw_device *sdev; + + dev_dbg(&netdev->dev, "siw: event %lu\n", event); + + if (dev_net(netdev) != &init_net) + return NOTIFY_OK; + + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); + if (!base_dev) + return NOTIFY_OK; + + sdev = to_siw_dev(base_dev); + + switch (event) { + + case NETDEV_UP: + sdev->state = IB_PORT_ACTIVE; + siw_port_event(sdev, 1, IB_EVENT_PORT_ACTIVE); + break; + + case NETDEV_GOING_DOWN: + siw_device_goes_down(sdev); + break; + + case NETDEV_DOWN: + sdev->state = IB_PORT_DOWN; + siw_port_event(sdev, 1, IB_EVENT_PORT_ERR); + break; + + case NETDEV_REGISTER: + /* + * Device registration now handled only by + * rdma netlink commands. So it shall be impossible + * to end up here with a valid siw device. + */ + siw_dbg(sdev, "unexpected NETDEV_REGISTER event\n"); + break; + + case NETDEV_UNREGISTER: + ib_unregister_device_queued(&sdev->base_dev); + break; + + case NETDEV_CHANGEADDR: + siw_port_event(sdev, 1, IB_EVENT_LID_CHANGE); + break; + /* + * Todo: Below netdev events are currently not handled. + */ + case NETDEV_CHANGEMTU: + case NETDEV_CHANGE: + + break; + + default: + break; + } + ib_device_put(&sdev->base_dev); + + return NOTIFY_OK; +} + +static struct notifier_block siw_netdev_nb = { + .notifier_call = siw_netdev_event, +}; + +static int siw_newlink(const char *basedev_name, struct net_device *netdev) +{ + struct ib_device *base_dev; + struct siw_device *sdev; + int rv; + + base_dev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_SIW); + if (base_dev) { + ib_device_put(base_dev); + rv = -EEXIST; + goto out; + } + if (!siw_dev_qualified(netdev)) { + rv = -EINVAL; + goto out; + } + sdev = siw_device_create(netdev); + if (sdev) { + dev_dbg(&netdev->dev, "siw: new device\n"); + + if (netif_running(netdev) && netif_carrier_ok(netdev)) + sdev->state = IB_PORT_ACTIVE; + else + sdev->state = IB_PORT_DOWN; + + rv = siw_device_register(sdev, basedev_name); + if (rv) { + siw_device_cleanup(&sdev->base_dev); + ib_dealloc_device(&sdev->base_dev); + goto out; + } + } else { + rv = -ENOMEM; + } +out: + return rv; +} + +static struct rdma_link_ops siw_link_ops = { + .type = "siw", + .newlink = siw_newlink, +}; + +/* + * siw_init_module - Initialize Softiwarp module and register with netdev + * subsystem to create Softiwarp devices per net_device + */ +static __init int siw_init_module(void) +{ + int rv; + int nr_cpu; + + if (SENDPAGE_THRESH < SIW_MAX_INLINE) { + pr_info("siw: sendpage threshold too small: %u\n", + (int)SENDPAGE_THRESH); + rv = EINVAL; + goto out_error; + } + rv = siw_init_cpulist(); + if (rv) + goto out_error; + + rv = siw_cm_init(); + if (rv) + goto out_error; + + /* + * Allocate CRC SHASH object. Fail loading siw only, if CRC is + * required by kernel module + */ + siw_crypto_shash = crypto_alloc_shash("crc32c", 0, 0); + if (IS_ERR(siw_crypto_shash)) { + pr_info("siw: Loading CRC32c failed: %ld\n", + PTR_ERR(siw_crypto_shash)); + siw_crypto_shash = NULL; + if (mpa_crc_required == true) + goto out_error; + } + rv = register_netdevice_notifier(&siw_netdev_nb); + if (rv) + goto out_error; + + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) + siw_tx_thread[nr_cpu] = NULL; + + if (!siw_create_tx_threads()) { + pr_info("siw: Could not start any TX thread\n"); + unregister_netdevice_notifier(&siw_netdev_nb); + goto out_error; + } + rdma_link_register(&siw_link_ops); + + pr_info("SoftiWARP attached\n"); + return 0; + +out_error: + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) { + if (siw_tx_thread[nr_cpu]) { + siw_stop_tx_thread(nr_cpu); + siw_tx_thread[nr_cpu] = NULL; + } + } + if (siw_crypto_shash) + crypto_free_shash(siw_crypto_shash); + + pr_info("SoftIWARP attach failed. Error: %d\n", rv); + + siw_cm_exit(); + siw_destroy_cpulist(); + + return rv; +} + +static void __exit siw_exit_module(void) +{ + int nr_cpu; + + for (nr_cpu = 0; nr_cpu < nr_cpu_ids; nr_cpu++) { + if (siw_tx_thread[nr_cpu]) { + siw_stop_tx_thread(nr_cpu); + siw_tx_thread[nr_cpu] = NULL; + } + } + unregister_netdevice_notifier(&siw_netdev_nb); + rdma_link_unregister(&siw_link_ops); + ib_unregister_driver(RDMA_DRIVER_SIW); + + siw_cm_exit(); + + if (siw_crypto_shash) + crypto_free_shash(siw_crypto_shash); + + siw_destroy_cpulist(); + + pr_info("SoftiWARP detached\n"); +} + +module_init(siw_init_module); +module_exit(siw_exit_module); + +MODULE_ALIAS_RDMA_LINK("siw");
Signed-off-by: Bernard Metzler <bmt@zurich.ibm.com> --- drivers/infiniband/sw/siw/siw_main.c | 813 +++++++++++++++++++++++++++ 1 file changed, 813 insertions(+) create mode 100644 drivers/infiniband/sw/siw/siw_main.c