| Message ID | 5523CEE4.5060901@profitbricks.com |
|---|---|
| State | Not Applicable, archived |
| Delegated to: | David Miller |
| Headers | show |
On 4/7/2015 8:34 AM, Michael Wang wrote: > /** > + * cap_read_multi_sge - Check if the port of device has the capability > + * RDMA Read Multiple Scatter-Gather Entries. > + * > + * @device: Device to be checked > + * @port_num: Port number of the device > + * > + * Return 0 when port of the device don't support > + * RDMA Read Multiple Scatter-Gather Entries. > + */ > +static inline int cap_read_multi_sge(struct ib_device *device, u8 port_num) > +{ > + return !rdma_transport_iwarp(device, port_num); > +} This just papers over the issue we discussed earlier. How *many* entries does the device support? If a device supports one, or two, is that enough? How does the upper layer know the limit? This needs an explicit device attribute, to be fixed properly. > + > +/** > * cap_ipoib - Check if the port of device has the capability > * IP over Infiniband. > * > diff --git a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c > index e011027..604d035 100644 > --- a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c > +++ b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c > @@ -118,8 +118,8 @@ static void rdma_build_arg_xdr(struct svc_rqst *rqstp, > > static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count) > { > - if (rdma_node_get_transport(xprt->sc_cm_id->device->node_type) == > - RDMA_TRANSPORT_IWARP) > + if (!cap_read_multi_sge(xprt->sc_cm_id->device, > + xprt->sc_cm_id->port_num)) > return 1; > else > return min_t(int, sge_count, xprt->sc_max_sge); This is incorrect. The RDMA Read max is not at all the same as the max_sge. It is a different operation, with a different set of work request parameters. In other words, the above same comment applies. > diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c > index 4e61880..e75175d 100644 > --- a/net/sunrpc/xprtrdma/svc_rdma_transport.c > +++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c > @@ -979,8 +979,8 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) > /* > * Determine if a DMA MR is required and if so, what privs are required > */ > - switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) { > - case RDMA_TRANSPORT_IWARP: > + if (rdma_transport_iwarp(newxprt->sc_cm_id->device, > + newxprt->sc_cm_id->port_num)) { > newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; Do I read this correctly that it is forcing the "read with invalidate" capability to "on" for all iWARP devices? I don't think that is correct, for the legacy devices you're also supporting. > @@ -992,8 +992,8 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) > dma_mr_acc = IB_ACCESS_LOCAL_WRITE; > } else > need_dma_mr = 0; > - break; > - case RDMA_TRANSPORT_IB: > + } else if (rdma_ib_mgmt(newxprt->sc_cm_id->device, > + newxprt->sc_cm_id->port_num)) { > if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { > need_dma_mr = 1; > dma_mr_acc = IB_ACCESS_LOCAL_WRITE; Now I'm even more confused. How is the presence of IB management related to needing a privileged lmr? -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi, Tom Thanks for the comments :-) On 04/07/2015 05:46 PM, Tom Talpey wrote: > On 4/7/2015 8:34 AM, Michael Wang wrote: >> /** >> + * cap_read_multi_sge - Check if the port of device has the capability >> + * RDMA Read Multiple Scatter-Gather Entries. >> + * >> + * @device: Device to be checked >> + * @port_num: Port number of the device >> + * >> + * Return 0 when port of the device don't support >> + * RDMA Read Multiple Scatter-Gather Entries. >> + */ >> +static inline int cap_read_multi_sge(struct ib_device *device, u8 port_num) >> +{ >> + return !rdma_transport_iwarp(device, port_num); >> +} > > This just papers over the issue we discussed earlier. How *many* > entries does the device support? If a device supports one, or two, > is that enough? How does the upper layer know the limit? > > This needs an explicit device attribute, to be fixed properly. This is the prototype to expose the problem we have in here, I would prefer some one good at this part to extending the API in future, basing on the right logical. Currently this just inherit from the legacy, it implemented in order to be compatible with the current code. > >> + >> +/** >> * cap_ipoib - Check if the port of device has the capability >> * IP over Infiniband. >> * >> diff --git a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c >> index e011027..604d035 100644 >> --- a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c >> +++ b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c >> @@ -118,8 +118,8 @@ static void rdma_build_arg_xdr(struct svc_rqst *rqstp, >> >> static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count) >> { >> - if (rdma_node_get_transport(xprt->sc_cm_id->device->node_type) == >> - RDMA_TRANSPORT_IWARP) >> + if (!cap_read_multi_sge(xprt->sc_cm_id->device, >> + xprt->sc_cm_id->port_num)) >> return 1; >> else >> return min_t(int, sge_count, xprt->sc_max_sge); > > This is incorrect. The RDMA Read max is not at all the same as the > max_sge. It is a different operation, with a different set of work > request parameters. > > In other words, the above same comment applies. Any idea on how to improve this part? Again, all these helpers just inherit the old logical, if it's wrong, let's correct it ;-) And if we don't know how to correct, we can leave this as a signpost and waiting for someone good at this particular part to fix it. > > >> diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c >> index 4e61880..e75175d 100644 >> --- a/net/sunrpc/xprtrdma/svc_rdma_transport.c >> +++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c >> @@ -979,8 +979,8 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) >> /* >> * Determine if a DMA MR is required and if so, what privs are required >> */ >> - switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) { >> - case RDMA_TRANSPORT_IWARP: >> + if (rdma_transport_iwarp(newxprt->sc_cm_id->device, >> + newxprt->sc_cm_id->port_num)) { >> newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; > > Do I read this correctly that it is forcing the "read with invalidate" > capability to "on" for all iWARP devices? I don't think that is correct, > for the legacy devices you're also supporting. Hmm.. but that's exactly same as the old logical, correct? Or do you mean the old logical is wrong? > > >> @@ -992,8 +992,8 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) >> dma_mr_acc = IB_ACCESS_LOCAL_WRITE; >> } else >> need_dma_mr = 0; >> - break; >> - case RDMA_TRANSPORT_IB: >> + } else if (rdma_ib_mgmt(newxprt->sc_cm_id->device, >> + newxprt->sc_cm_id->port_num)) { >> if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { >> need_dma_mr = 1; >> dma_mr_acc = IB_ACCESS_LOCAL_WRITE; > > Now I'm even more confused. How is the presence of IB management > related to needing a privileged lmr? I think you actually mean we need some more wrapper here with the right name, correct? I'm not good at this part, any suggestions? Regards, Michael Wang > > -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, Apr 07, 2015 at 11:46:57AM -0400, Tom Talpey wrote: > On 4/7/2015 8:34 AM, Michael Wang wrote: > > /** > >+ * cap_read_multi_sge - Check if the port of device has the capability > >+ * RDMA Read Multiple Scatter-Gather Entries. > >+ * > >+ * @device: Device to be checked > >+ * @port_num: Port number of the device > >+ * > >+ * Return 0 when port of the device don't support > >+ * RDMA Read Multiple Scatter-Gather Entries. > >+ */ > >+static inline int cap_read_multi_sge(struct ib_device *device, u8 port_num) > >+{ > >+ return !rdma_transport_iwarp(device, port_num); > >+} > > This just papers over the issue we discussed earlier. How *many* > entries does the device support? If a device supports one, or two, > is that enough? How does the upper layer know the limit? I think Michael is fine to just make this one mechanical change. The kernel only supports two kinds of devices today, ones with 1 read SGE and ones where READ SGE == WRITE SGE == SEND SGE. If someone makes another variation then it is up to them to propose a better fix. > > static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count) > > { > >- if (rdma_node_get_transport(xprt->sc_cm_id->device->node_type) == > >- RDMA_TRANSPORT_IWARP) > >+ if (!cap_read_multi_sge(xprt->sc_cm_id->device, > >+ xprt->sc_cm_id->port_num)) > > return 1; > > else > > return min_t(int, sge_count, xprt->sc_max_sge); > > This is incorrect. The RDMA Read max is not at all the same as the > max_sge. It is a different operation, with a different set of work > request parameters. The algorithm looks OK to me, newxprt->sc_max_sge = min((size_t)devattr.max_sge, (size_t)RPCSVC_MAXPAGES); So it returns 1 or the number of sge entries per WR, and max_sge is for READ/WRITE/SEND in every case except when cap_read_multi_sge == 1 > > /* > > * Determine if a DMA MR is required and if so, what privs are required > > */ > >- switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) { > >- case RDMA_TRANSPORT_IWARP: > >+ if (rdma_transport_iwarp(newxprt->sc_cm_id->device, > >+ newxprt->sc_cm_id->port_num)) { > > newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; > > Do I read this correctly that it is forcing the "read with invalidate" > capability to "on" for all iWARP devices? I don't think that is correct, > for the legacy devices you're also supporting. No idea here, this logic was added in: commit 3a5c63803d0552a3ad93b85c262f12cd86471443 Author: Tom Tucker <tom@opengridcomputing.com> Date: Tue Sep 30 13:46:13 2008 -0500 svcrdma: Query device for Fast Reg support during connection setup Query the device capabilities in the svc_rdma_accept function to determine what advanced memory management capabilities are supported by the device. Based on the query, select the most secure model available given the requirements of the transport and capabilities of the adapter. Signed-off-by: Tom Tucker <tom@opengridcomputing.com> > >@@ -992,8 +992,8 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) > > dma_mr_acc = IB_ACCESS_LOCAL_WRITE; > > } else > > need_dma_mr = 0; > >- break; > >- case RDMA_TRANSPORT_IB: > >+ } else if (rdma_ib_mgmt(newxprt->sc_cm_id->device, > >+ newxprt->sc_cm_id->port_num)) { > > if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { > > need_dma_mr = 1; > > dma_mr_acc = IB_ACCESS_LOCAL_WRITE; > > Now I'm even more confused. How is the presence of IB management > related to needing a privileged lmr? Agree, this needs to be someone else. I think the test is probably based on this comment: * NB: iWARP requires remote write access for the data sink * of an RDMA_READ. IB does not. So the if should be: if (cap_rdma_read_needs_write(..) && !(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { need_dma_mr = 1; dma_mr_acc = (IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE); And the identical if blocks merged. Plus the if (rdma_transport_iwarp(newxprt->sc_cm_id->device, newxprt->sc_cm_id->port_num)) newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV Jason -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h index 9db8966..cae6f2d 100644 --- a/include/rdma/ib_verbs.h +++ b/include/rdma/ib_verbs.h @@ -1849,6 +1849,21 @@ static inline int cap_ib_mcast(struct ib_device *device, u8 port_num) } /** + * cap_read_multi_sge - Check if the port of device has the capability + * RDMA Read Multiple Scatter-Gather Entries. + * + * @device: Device to be checked + * @port_num: Port number of the device + * + * Return 0 when port of the device don't support + * RDMA Read Multiple Scatter-Gather Entries. + */ +static inline int cap_read_multi_sge(struct ib_device *device, u8 port_num) +{ + return !rdma_transport_iwarp(device, port_num); +} + +/** * cap_ipoib - Check if the port of device has the capability * IP over Infiniband. * diff --git a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c index e011027..604d035 100644 --- a/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c +++ b/net/sunrpc/xprtrdma/svc_rdma_recvfrom.c @@ -118,8 +118,8 @@ static void rdma_build_arg_xdr(struct svc_rqst *rqstp, static int rdma_read_max_sge(struct svcxprt_rdma *xprt, int sge_count) { - if (rdma_node_get_transport(xprt->sc_cm_id->device->node_type) == - RDMA_TRANSPORT_IWARP) + if (!cap_read_multi_sge(xprt->sc_cm_id->device, + xprt->sc_cm_id->port_num)) return 1; else return min_t(int, sge_count, xprt->sc_max_sge); diff --git a/net/sunrpc/xprtrdma/svc_rdma_transport.c b/net/sunrpc/xprtrdma/svc_rdma_transport.c index 4e61880..e75175d 100644 --- a/net/sunrpc/xprtrdma/svc_rdma_transport.c +++ b/net/sunrpc/xprtrdma/svc_rdma_transport.c @@ -979,8 +979,8 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) /* * Determine if a DMA MR is required and if so, what privs are required */ - switch (rdma_node_get_transport(newxprt->sc_cm_id->device->node_type)) { - case RDMA_TRANSPORT_IWARP: + if (rdma_transport_iwarp(newxprt->sc_cm_id->device, + newxprt->sc_cm_id->port_num)) { newxprt->sc_dev_caps |= SVCRDMA_DEVCAP_READ_W_INV; if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { need_dma_mr = 1; @@ -992,8 +992,8 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) dma_mr_acc = IB_ACCESS_LOCAL_WRITE; } else need_dma_mr = 0; - break; - case RDMA_TRANSPORT_IB: + } else if (rdma_ib_mgmt(newxprt->sc_cm_id->device, + newxprt->sc_cm_id->port_num)) { if (!(newxprt->sc_dev_caps & SVCRDMA_DEVCAP_FAST_REG)) { need_dma_mr = 1; dma_mr_acc = IB_ACCESS_LOCAL_WRITE; @@ -1003,10 +1003,8 @@ static struct svc_xprt *svc_rdma_accept(struct svc_xprt *xprt) dma_mr_acc = IB_ACCESS_LOCAL_WRITE; } else need_dma_mr = 0; - break; - default: + } else goto errout; - } /* Create the DMA MR if needed, otherwise, use the DMA LKEY */ if (need_dma_mr) {
Introduce helper cap_read_multi_sge() to help us check if the port of an IB device support RDMA Read Multiple Scatter-Gather Entries. Reform svc_rdma_accept() to adopt management helpers. Cc: Tom Talpey <tom@talpey.com> Cc: Jason Gunthorpe <jgunthorpe@obsidianresearch.com> Cc: Doug Ledford <dledford@redhat.com> Cc: Ira Weiny <ira.weiny@intel.com> Cc: Sean Hefty <sean.hefty@intel.com> Signed-off-by: Michael Wang <yun.wang@profitbricks.com> --- include/rdma/ib_verbs.h | 15 +++++++++++++++ net/sunrpc/xprtrdma/svc_rdma_recvfrom.c | 4 ++-- net/sunrpc/xprtrdma/svc_rdma_transport.c | 12 +++++------- 3 files changed, 22 insertions(+), 9 deletions(-)