| Message ID | 20210507140638.339-1-anand.a.khoje@oracle.com (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Leon Romanovsky |
| Headers | show |
| Series | IB/core: Obtain subnet_prefix from cache in IB devices | expand |
On Fri, May 07, 2021 at 07:36:38PM +0530, Anand Khoje wrote: > ib_query_port() calls device->ops.query_port() to get the port > attributes. The method of querying is device driver specific. > The same function calls device->ops.query_gid() to get the GID and > extract the subnet_prefix (gid_prefix). Please add blank lines to separate paragraphs. > The GID and subnet_prefix are stored in a cache. But they do not get > read from the cache if the device is an Infiniband device. The > following change takes advantage of the cached subnet_prefix. > Testing with RDBMS has shown a significant improvement in performance > with this change. Here too > The function ib_cache_is_initialised() is introduced because > ib_query_port() gets called early in the stage when the cache is not > built while reading port immutable property. > In that case, the default GID still gets read from HCA for IB link > layer. The shuffling of netdev_lock in struct ib_port_data is done > such that the size of struct ib_port_data remains the same after > adding flags. > > Fixes: fad61ad ("IB/core: Add subnet prefix to port info") > No blank line here. > Signed-off-by: Anand Khoje <anand.a.khoje@oracle.com> > Signed-off-by: Haakon Bugge <haakon.bugge@oracle.com> > --- > drivers/infiniband/core/cache.c | 7 ++++++- > drivers/infiniband/core/device.c | 11 +++++++++++ > include/rdma/ib_cache.h | 7 +++++++ > include/rdma/ib_verbs.h | 10 +++++++++- > 4 files changed, 33 insertions(+), 2 deletions(-) > > diff --git a/drivers/infiniband/core/cache.c b/drivers/infiniband/core/cache.c > index 3b0991f..b580c26 100644 > --- a/drivers/infiniband/core/cache.c > +++ b/drivers/infiniband/core/cache.c > @@ -1627,6 +1627,8 @@ int ib_cache_setup_one(struct ib_device *device) > err = ib_cache_update(device, p, true); > if (err) > return err; > + set_bit(IB_PORT_CACHE_INITIALIZED, > + &device->port_data[p].flags); > } > > return 0; > @@ -1642,8 +1644,11 @@ void ib_cache_release_one(struct ib_device *device) > * all the device's resources when the cache could no > * longer be accessed. > */ > - rdma_for_each_port (device, p) > + rdma_for_each_port (device, p) { > + clear_bit(IB_PORT_CACHE_INITIALIZED, > + &device->port_data[p].flags); > kfree(device->port_data[p].cache.pkey); > + } > > gid_table_release_one(device); > } > diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c > index c660cef..6d62023 100644 > --- a/drivers/infiniband/core/device.c > +++ b/drivers/infiniband/core/device.c > @@ -2064,6 +2064,17 @@ static int __ib_query_port(struct ib_device *device, > IB_LINK_LAYER_INFINIBAND) > return 0; > > + if (!ib_cache_is_initialised(device, port_num)) > + goto query_gid_from_device; > + > + err = ib_get_cached_subnet_prefix(device, port_num, > + &port_attr->subnet_prefix); Can you please make ib_get_cached_subnet_prefix() void and move port check to the security.c? > + if (err) > + goto query_gid_from_device; > + > + return 0; > + > +query_gid_from_device: > err = device->ops.query_gid(device, port_num, 0, &gid); > if (err) > return err; > diff --git a/include/rdma/ib_cache.h b/include/rdma/ib_cache.h > index 226ae37..bebeb94 100644 > --- a/include/rdma/ib_cache.h > +++ b/include/rdma/ib_cache.h > @@ -114,4 +114,11 @@ ssize_t rdma_query_gid_table(struct ib_device *device, > struct ib_uverbs_gid_entry *entries, > size_t max_entries); > > +static inline bool ib_cache_is_initialised(struct ib_device *device, > + u8 port_num) > +{ > + return test_bit(IB_PORT_CACHE_INITIALIZED, > + &device->port_data[port_num].flags); > +} > + > #endif /* _IB_CACHE_H */ > diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h > index 7e2f369..ad2a55e 100644 > --- a/include/rdma/ib_verbs.h > +++ b/include/rdma/ib_verbs.h > @@ -2169,17 +2169,25 @@ struct ib_port_immutable { > u32 max_mad_size; > }; > > +enum ib_port_data_flags { > + IB_PORT_CACHE_INITIALIZED = 1 << 0, > +}; > + > struct ib_port_data { > struct ib_device *ib_dev; > > struct ib_port_immutable immutable; > > spinlock_t pkey_list_lock; > + > + spinlock_t netdev_lock; This change is unrelated. > + > + unsigned long flags; > + > struct list_head pkey_list; > > struct ib_port_cache cache; > > - spinlock_t netdev_lock; > struct net_device __rcu *netdev; > struct hlist_node ndev_hash_link; > struct rdma_port_counter port_counter; > -- > 1.8.3.1 >
Hi Leon, If I understand your comment " Can you please make ib_get_cached_subnet_prefix() void and move port check to the security.c? " correctly, you want the check rdma_is_port_valid() to be moved to get_pkey_and_subnet_prefix() and ib_security_pkey_access() before the call to ib_get_cached_pkey()? If yes, do you want to move that check from inside ib_get_cached_pkey() as well? Thanks, Anand A. Khoje -----Original Message----- From: Leon Romanovsky <leon@kernel.org> Sent: Sunday, May 9, 2021 4:38 PM To: Anand Khoje <anand.a.khoje@oracle.com> Cc: dledford@redhat.com; jgg@ziepe.ca; avihaih@nvidia.com; liangwenpeng@huawei.com; jackm@dev.mellanox.co.il; galpress@amazon.com; kamalheib1@gmail.com; mbloch@nvidia.com; lee.jones@linaro.org; maorg@mellanox.com; maxg@mellanox.com; parav@nvidia.com; eli@mellanox.com; ogerlitz@mellanox.com; linux-rdma@vger.kernel.org; linux-kernel@vger.kernel.org; Haakon Bugge <haakon.bugge@oracle.com> Subject: Re: [PATCH] IB/core: Obtain subnet_prefix from cache in IB devices On Fri, May 07, 2021 at 07:36:38PM +0530, Anand Khoje wrote: > ib_query_port() calls device->ops.query_port() to get the port > attributes. The method of querying is device driver specific. > The same function calls device->ops.query_gid() to get the GID and > extract the subnet_prefix (gid_prefix). Please add blank lines to separate paragraphs. > The GID and subnet_prefix are stored in a cache. But they do not get > read from the cache if the device is an Infiniband device. The > following change takes advantage of the cached subnet_prefix. > Testing with RDBMS has shown a significant improvement in performance > with this change. Here too > The function ib_cache_is_initialised() is introduced because > ib_query_port() gets called early in the stage when the cache is not > built while reading port immutable property. > In that case, the default GID still gets read from HCA for IB link > layer. The shuffling of netdev_lock in struct ib_port_data is done > such that the size of struct ib_port_data remains the same after > adding flags. > > Fixes: fad61ad ("IB/core: Add subnet prefix to port info") > No blank line here. > Signed-off-by: Anand Khoje <anand.a.khoje@oracle.com> > Signed-off-by: Haakon Bugge <haakon.bugge@oracle.com> > --- > drivers/infiniband/core/cache.c | 7 ++++++- > drivers/infiniband/core/device.c | 11 +++++++++++ > include/rdma/ib_cache.h | 7 +++++++ > include/rdma/ib_verbs.h | 10 +++++++++- > 4 files changed, 33 insertions(+), 2 deletions(-) > > diff --git a/drivers/infiniband/core/cache.c > b/drivers/infiniband/core/cache.c index 3b0991f..b580c26 100644 > --- a/drivers/infiniband/core/cache.c > +++ b/drivers/infiniband/core/cache.c > @@ -1627,6 +1627,8 @@ int ib_cache_setup_one(struct ib_device *device) > err = ib_cache_update(device, p, true); > if (err) > return err; > + set_bit(IB_PORT_CACHE_INITIALIZED, > + &device->port_data[p].flags); > } > > return 0; > @@ -1642,8 +1644,11 @@ void ib_cache_release_one(struct ib_device *device) > * all the device's resources when the cache could no > * longer be accessed. > */ > - rdma_for_each_port (device, p) > + rdma_for_each_port (device, p) { > + clear_bit(IB_PORT_CACHE_INITIALIZED, > + &device->port_data[p].flags); > kfree(device->port_data[p].cache.pkey); > + } > > gid_table_release_one(device); > } > diff --git a/drivers/infiniband/core/device.c > b/drivers/infiniband/core/device.c > index c660cef..6d62023 100644 > --- a/drivers/infiniband/core/device.c > +++ b/drivers/infiniband/core/device.c > @@ -2064,6 +2064,17 @@ static int __ib_query_port(struct ib_device *device, > IB_LINK_LAYER_INFINIBAND) > return 0; > > + if (!ib_cache_is_initialised(device, port_num)) > + goto query_gid_from_device; > + > + err = ib_get_cached_subnet_prefix(device, port_num, > + &port_attr->subnet_prefix); Can you please make ib_get_cached_subnet_prefix() void and move port check to the security.c? > + if (err) > + goto query_gid_from_device; > + > + return 0; > + > +query_gid_from_device: > err = device->ops.query_gid(device, port_num, 0, &gid); > if (err) > return err; > diff --git a/include/rdma/ib_cache.h b/include/rdma/ib_cache.h index > 226ae37..bebeb94 100644 > --- a/include/rdma/ib_cache.h > +++ b/include/rdma/ib_cache.h > @@ -114,4 +114,11 @@ ssize_t rdma_query_gid_table(struct ib_device *device, > struct ib_uverbs_gid_entry *entries, > size_t max_entries); > > +static inline bool ib_cache_is_initialised(struct ib_device *device, > + u8 port_num) > +{ > + return test_bit(IB_PORT_CACHE_INITIALIZED, > + &device->port_data[port_num].flags); > +} > + > #endif /* _IB_CACHE_H */ > diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h index > 7e2f369..ad2a55e 100644 > --- a/include/rdma/ib_verbs.h > +++ b/include/rdma/ib_verbs.h > @@ -2169,17 +2169,25 @@ struct ib_port_immutable { > u32 max_mad_size; > }; > > +enum ib_port_data_flags { > + IB_PORT_CACHE_INITIALIZED = 1 << 0, > +}; > + > struct ib_port_data { > struct ib_device *ib_dev; > > struct ib_port_immutable immutable; > > spinlock_t pkey_list_lock; > + > + spinlock_t netdev_lock; This change is unrelated. > + > + unsigned long flags; > + > struct list_head pkey_list; > > struct ib_port_cache cache; > > - spinlock_t netdev_lock; > struct net_device __rcu *netdev; > struct hlist_node ndev_hash_link; > struct rdma_port_counter port_counter; > -- > 1.8.3.1 >
On Fri, May 14, 2021 at 01:42:42PM +0000, Anand Khoje wrote: > Hi Leon, > > If I understand your comment " Can you please make ib_get_cached_subnet_prefix() void and move port check to the security.c? " correctly, > you want the check rdma_is_port_valid() to be moved to get_pkey_and_subnet_prefix() and ib_security_pkey_access() before the call to ib_get_cached_pkey()? Right now, no one really carry about returned value from ib_get_cached_subnet_prefix() and the reason to it is because "port_num" is valid. In device.c, we call to ib_get_cached_subnet_prefix() in the rdma_for_each_port() loop. And in security.c, we checked port validity in advance too. We didn't do it in explicit way but most likely we don't need it. > If yes, do you want to move that check from inside ib_get_cached_pkey() as well? It depends. > > Thanks, > Anand A. Khoje > > > -----Original Message----- > From: Leon Romanovsky <leon@kernel.org> > Sent: Sunday, May 9, 2021 4:38 PM > To: Anand Khoje <anand.a.khoje@oracle.com> > Cc: dledford@redhat.com; jgg@ziepe.ca; avihaih@nvidia.com; liangwenpeng@huawei.com; jackm@dev.mellanox.co.il; galpress@amazon.com; kamalheib1@gmail.com; mbloch@nvidia.com; lee.jones@linaro.org; maorg@mellanox.com; maxg@mellanox.com; parav@nvidia.com; eli@mellanox.com; ogerlitz@mellanox.com; linux-rdma@vger.kernel.org; linux-kernel@vger.kernel.org; Haakon Bugge <haakon.bugge@oracle.com> > Subject: Re: [PATCH] IB/core: Obtain subnet_prefix from cache in IB devices > > On Fri, May 07, 2021 at 07:36:38PM +0530, Anand Khoje wrote: > > ib_query_port() calls device->ops.query_port() to get the port > > attributes. The method of querying is device driver specific. > > The same function calls device->ops.query_gid() to get the GID and > > extract the subnet_prefix (gid_prefix). > > Please add blank lines to separate paragraphs. > > > The GID and subnet_prefix are stored in a cache. But they do not get > > read from the cache if the device is an Infiniband device. The > > following change takes advantage of the cached subnet_prefix. > > Testing with RDBMS has shown a significant improvement in performance > > with this change. > > Here too > > > The function ib_cache_is_initialised() is introduced because > > ib_query_port() gets called early in the stage when the cache is not > > built while reading port immutable property. > > In that case, the default GID still gets read from HCA for IB link > > layer. The shuffling of netdev_lock in struct ib_port_data is done > > such that the size of struct ib_port_data remains the same after > > adding flags. > > > > Fixes: fad61ad ("IB/core: Add subnet prefix to port info") > > > > No blank line here. > > > Signed-off-by: Anand Khoje <anand.a.khoje@oracle.com> > > Signed-off-by: Haakon Bugge <haakon.bugge@oracle.com> > > --- > > drivers/infiniband/core/cache.c | 7 ++++++- > > drivers/infiniband/core/device.c | 11 +++++++++++ > > include/rdma/ib_cache.h | 7 +++++++ > > include/rdma/ib_verbs.h | 10 +++++++++- > > 4 files changed, 33 insertions(+), 2 deletions(-) > > > > diff --git a/drivers/infiniband/core/cache.c > > b/drivers/infiniband/core/cache.c index 3b0991f..b580c26 100644 > > --- a/drivers/infiniband/core/cache.c > > +++ b/drivers/infiniband/core/cache.c > > @@ -1627,6 +1627,8 @@ int ib_cache_setup_one(struct ib_device *device) > > err = ib_cache_update(device, p, true); > > if (err) > > return err; > > + set_bit(IB_PORT_CACHE_INITIALIZED, > > + &device->port_data[p].flags); > > } > > > > return 0; > > @@ -1642,8 +1644,11 @@ void ib_cache_release_one(struct ib_device *device) > > * all the device's resources when the cache could no > > * longer be accessed. > > */ > > - rdma_for_each_port (device, p) > > + rdma_for_each_port (device, p) { > > + clear_bit(IB_PORT_CACHE_INITIALIZED, > > + &device->port_data[p].flags); > > kfree(device->port_data[p].cache.pkey); > > + } > > > > gid_table_release_one(device); > > } > > diff --git a/drivers/infiniband/core/device.c > > b/drivers/infiniband/core/device.c > > index c660cef..6d62023 100644 > > --- a/drivers/infiniband/core/device.c > > +++ b/drivers/infiniband/core/device.c > > @@ -2064,6 +2064,17 @@ static int __ib_query_port(struct ib_device *device, > > IB_LINK_LAYER_INFINIBAND) > > return 0; > > > > + if (!ib_cache_is_initialised(device, port_num)) > > + goto query_gid_from_device; > > + > > + err = ib_get_cached_subnet_prefix(device, port_num, > > + &port_attr->subnet_prefix); > > Can you please make ib_get_cached_subnet_prefix() void and move port check to the security.c? > > > + if (err) > > + goto query_gid_from_device; > > + > > + return 0; > > + > > +query_gid_from_device: > > err = device->ops.query_gid(device, port_num, 0, &gid); > > if (err) > > return err; > > diff --git a/include/rdma/ib_cache.h b/include/rdma/ib_cache.h index > > 226ae37..bebeb94 100644 > > --- a/include/rdma/ib_cache.h > > +++ b/include/rdma/ib_cache.h > > @@ -114,4 +114,11 @@ ssize_t rdma_query_gid_table(struct ib_device *device, > > struct ib_uverbs_gid_entry *entries, > > size_t max_entries); > > > > +static inline bool ib_cache_is_initialised(struct ib_device *device, > > + u8 port_num) > > +{ > > + return test_bit(IB_PORT_CACHE_INITIALIZED, > > + &device->port_data[port_num].flags); > > +} > > + > > #endif /* _IB_CACHE_H */ > > diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h index > > 7e2f369..ad2a55e 100644 > > --- a/include/rdma/ib_verbs.h > > +++ b/include/rdma/ib_verbs.h > > @@ -2169,17 +2169,25 @@ struct ib_port_immutable { > > u32 max_mad_size; > > }; > > > > +enum ib_port_data_flags { > > + IB_PORT_CACHE_INITIALIZED = 1 << 0, > > +}; > > + > > struct ib_port_data { > > struct ib_device *ib_dev; > > > > struct ib_port_immutable immutable; > > > > spinlock_t pkey_list_lock; > > + > > + spinlock_t netdev_lock; > > This change is unrelated. > > > + > > + unsigned long flags; > > + > > struct list_head pkey_list; > > > > struct ib_port_cache cache; > > > > - spinlock_t netdev_lock; > > struct net_device __rcu *netdev; > > struct hlist_node ndev_hash_link; > > struct rdma_port_counter port_counter; > > -- > > 1.8.3.1 > >
diff --git a/drivers/infiniband/core/cache.c b/drivers/infiniband/core/cache.c index 3b0991f..b580c26 100644 --- a/drivers/infiniband/core/cache.c +++ b/drivers/infiniband/core/cache.c @@ -1627,6 +1627,8 @@ int ib_cache_setup_one(struct ib_device *device) err = ib_cache_update(device, p, true); if (err) return err; + set_bit(IB_PORT_CACHE_INITIALIZED, + &device->port_data[p].flags); } return 0; @@ -1642,8 +1644,11 @@ void ib_cache_release_one(struct ib_device *device) * all the device's resources when the cache could no * longer be accessed. */ - rdma_for_each_port (device, p) + rdma_for_each_port (device, p) { + clear_bit(IB_PORT_CACHE_INITIALIZED, + &device->port_data[p].flags); kfree(device->port_data[p].cache.pkey); + } gid_table_release_one(device); } diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c index c660cef..6d62023 100644 --- a/drivers/infiniband/core/device.c +++ b/drivers/infiniband/core/device.c @@ -2064,6 +2064,17 @@ static int __ib_query_port(struct ib_device *device, IB_LINK_LAYER_INFINIBAND) return 0; + if (!ib_cache_is_initialised(device, port_num)) + goto query_gid_from_device; + + err = ib_get_cached_subnet_prefix(device, port_num, + &port_attr->subnet_prefix); + if (err) + goto query_gid_from_device; + + return 0; + +query_gid_from_device: err = device->ops.query_gid(device, port_num, 0, &gid); if (err) return err; diff --git a/include/rdma/ib_cache.h b/include/rdma/ib_cache.h index 226ae37..bebeb94 100644 --- a/include/rdma/ib_cache.h +++ b/include/rdma/ib_cache.h @@ -114,4 +114,11 @@ ssize_t rdma_query_gid_table(struct ib_device *device, struct ib_uverbs_gid_entry *entries, size_t max_entries); +static inline bool ib_cache_is_initialised(struct ib_device *device, + u8 port_num) +{ + return test_bit(IB_PORT_CACHE_INITIALIZED, + &device->port_data[port_num].flags); +} + #endif /* _IB_CACHE_H */ diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h index 7e2f369..ad2a55e 100644 --- a/include/rdma/ib_verbs.h +++ b/include/rdma/ib_verbs.h @@ -2169,17 +2169,25 @@ struct ib_port_immutable { u32 max_mad_size; }; +enum ib_port_data_flags { + IB_PORT_CACHE_INITIALIZED = 1 << 0, +}; + struct ib_port_data { struct ib_device *ib_dev; struct ib_port_immutable immutable; spinlock_t pkey_list_lock; + + spinlock_t netdev_lock; + + unsigned long flags; + struct list_head pkey_list; struct ib_port_cache cache; - spinlock_t netdev_lock; struct net_device __rcu *netdev; struct hlist_node ndev_hash_link; struct rdma_port_counter port_counter;
ib_query_port() calls device->ops.query_port() to get the port attributes. The method of querying is device driver specific. The same function calls device->ops.query_gid() to get the GID and extract the subnet_prefix (gid_prefix). The GID and subnet_prefix are stored in a cache. But they do not get read from the cache if the device is an Infiniband device. The following change takes advantage of the cached subnet_prefix. Testing with RDBMS has shown a significant improvement in performance with this change. The function ib_cache_is_initialised() is introduced because ib_query_port() gets called early in the stage when the cache is not built while reading port immutable property. In that case, the default GID still gets read from HCA for IB link layer. The shuffling of netdev_lock in struct ib_port_data is done such that the size of struct ib_port_data remains the same after adding flags. Fixes: fad61ad ("IB/core: Add subnet prefix to port info") Signed-off-by: Anand Khoje <anand.a.khoje@oracle.com> Signed-off-by: Haakon Bugge <haakon.bugge@oracle.com> --- drivers/infiniband/core/cache.c | 7 ++++++- drivers/infiniband/core/device.c | 11 +++++++++++ include/rdma/ib_cache.h | 7 +++++++ include/rdma/ib_verbs.h | 10 +++++++++- 4 files changed, 33 insertions(+), 2 deletions(-)