| Message ID | 20210616154509.1047-4-anand.a.khoje@oracle.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Jason Gunthorpe |
| Headers | show |
| Series | IB/core: Obtaining subnet_prefix from cache in | expand |
On Wed, Jun 16, 2021 at 09:15:09PM +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). > > 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 devices. > > In the situation of an event causing cache update, the subnet_prefix > will get retrieved from newly updated GID cache in ib_cache_update(), > so that we do not end up reading a stale value from cache via > ib_query_port(). > > Fixes: fad61ad ("IB/core: Add subnet prefix to port info") > Suggested-by: Leon Romanovsky <leonro@nvidia.com> > Suggested-by: Aru Kolappan <aru.kolappan@oracle.com> > Signed-off-by: Anand Khoje <anand.a.khoje@oracle.com> > Signed-off-by: Haakon Bugge <haakon.bugge@oracle.com> > --- > > v1 -> v2: > - Split the v1 patch in 3 patches as per Leon's suggestion. > > v2 -> v3: > - Added changes as per Mark Zhang's suggestion of clearing > flags in git_table_cleanup_one(). > v3 -> v4: > - Removed the enum ib_port_data_flags and 8 byte flags from > struct ib_port_data, and the set_bit()/clear_bit() API > used to update this flag as that was not necessary. > Done to keep the code simple. > - Added code to read subnet_prefix from updated GID cache in the > event of cache update. Prior to this change, ib_cache_update > was reading the value for subnet_prefix via ib_query_port(), > due to this patch, we ended up reading a stale cached value of > subnet_prefix. > v4 -> v5: > - Removed the code to reset cache_is_initialised bit from cleanup > as per Leon's suggestion. > - Removed ib_cache_is_initialised() function. > > --- > drivers/infiniband/core/cache.c | 14 ++++++++++++-- > drivers/infiniband/core/device.c | 9 +++++++++ > include/rdma/ib_verbs.h | 1 + > 3 files changed, 22 insertions(+), 2 deletions(-) > Thanks, Reviewed-by: Leon Romanovsky <leonro@nvidia.com>
On Wed, Jun 16, 2021 at 09:15:09PM +0530, Anand Khoje wrote: > > @@ -1523,13 +1524,21 @@ static int config_non_roce_gid_cache(struct ib_device *device, > device->port_data[port].cache.lmc = tprops->lmc; > device->port_data[port].cache.port_state = tprops->state; > > - device->port_data[port].cache.subnet_prefix = tprops->subnet_prefix; > + ret = rdma_query_gid(device, port, 0, &gid); > + if (ret) { This is quite a bit different than just calling ops.query_gid() - why are you changing it? I'm not sure all the additional tests will pass, the 0 gid entry is not required to be valid.. > @@ -1629,6 +1638,7 @@ int ib_cache_setup_one(struct ib_device *device) > err = ib_cache_update(device, p, true, true, true); > if (err) > return err; > + device->port_data[p].cache_is_initialized = 1; > } And I would much prefer things be re-organized so the cache can be valid sooner to adding this variable. What is the earlier call that is motivating this? Jason
On 6/22/2021 5:19 AM, Jason Gunthorpe wrote: > On Wed, Jun 16, 2021 at 09:15:09PM +0530, Anand Khoje wrote: >> >> @@ -1523,13 +1524,21 @@ static int config_non_roce_gid_cache(struct ib_device *device, >> device->port_data[port].cache.lmc = tprops->lmc; >> device->port_data[port].cache.port_state = tprops->state; >> >> - device->port_data[port].cache.subnet_prefix = tprops->subnet_prefix; >> + ret = rdma_query_gid(device, port, 0, &gid); >> + if (ret) { > > This is quite a bit different than just calling ops.query_gid() - why > are you changing it? I'm not sure all the additional tests will pass, > the 0 gid entry is not required to be valid.. > Hi Jason, We have opted for rdma_query_gid(), as during ib_cache_update() the code calls ops.query_gid() earlier in config_non_roce_gid_cache(), thereby updating the value of GID in cache. We utilize this updated value, instead of calling ops->query_gid() again. I'm not sure all the additional tests will pass, the 0 gid entry is not required to be valid.. To get subnet_prefix __ib_query_port() does indeed obtain zero index GID. https://elixir.bootlin.com/linux/v5.13-rc5/source/drivers/infiniband/core/device.c#L2067 >> @@ -1629,6 +1638,7 @@ int ib_cache_setup_one(struct ib_device *device) >> err = ib_cache_update(device, p, true, true, true); >> if (err) >> return err; >> + device->port_data[p].cache_is_initialized = 1; >> } > > And I would much prefer things be re-organized so the cache can be > valid sooner to adding this variable. What is the earlier call that is > motivating this? > > Jason > During device load and when cache is yet to be updated, ib_query_port() should have a mechanism to identify if the cache entry is valid or invalid (uninitialized), we have added this variable just to ensure the validity of cache. Thanks, Anand
On Wed, Jun 23, 2021 at 06:33:32PM +0530, Anand Khoje wrote: > On 6/22/2021 5:19 AM, Jason Gunthorpe wrote: > > On Wed, Jun 16, 2021 at 09:15:09PM +0530, Anand Khoje wrote: > > > @@ -1523,13 +1524,21 @@ static int config_non_roce_gid_cache(struct ib_device *device, > > > device->port_data[port].cache.lmc = tprops->lmc; > > > device->port_data[port].cache.port_state = tprops->state; > > > - device->port_data[port].cache.subnet_prefix = tprops->subnet_prefix; > > > + ret = rdma_query_gid(device, port, 0, &gid); > > > + if (ret) { > > > > This is quite a bit different than just calling ops.query_gid() - why > > are you changing it? I'm not sure all the additional tests will pass, > > the 0 gid entry is not required to be valid.. > > > Hi Jason, > > We have opted for rdma_query_gid(), as during ib_cache_update() the code > calls ops.query_gid() earlier in config_non_roce_gid_cache(), thereby > updating the value of GID in cache. We utilize this updated value, instead > of calling ops->query_gid() again. Uhhhh, so just store the subnet prefix at that point then? diff --git a/drivers/infiniband/core/cache.c b/drivers/infiniband/core/cache.c index c9e9fc81447e89..5c554ebd000e89 100644 --- a/drivers/infiniband/core/cache.c +++ b/drivers/infiniband/core/cache.c @@ -1428,8 +1428,8 @@ int rdma_read_gid_l2_fields(const struct ib_gid_attr *attr, } EXPORT_SYMBOL(rdma_read_gid_l2_fields); -static int config_non_roce_gid_cache(struct ib_device *device, - u32 port, int gid_tbl_len) +static int config_non_roce_gid_cache(struct ib_device *device, u32 port, + struct ib_port_attr *tprops) { struct ib_gid_attr gid_attr = {}; struct ib_gid_table *table; @@ -1441,7 +1441,7 @@ static int config_non_roce_gid_cache(struct ib_device *device, table = rdma_gid_table(device, port); mutex_lock(&table->lock); - for (i = 0; i < gid_tbl_len; ++i) { + for (i = 0; i < tprops->gid_tbl_len; ++i) { if (!device->ops.query_gid) continue; ret = device->ops.query_gid(device, port, i, &gid_attr.gid); @@ -1452,6 +1452,8 @@ static int config_non_roce_gid_cache(struct ib_device *device, goto err; } gid_attr.index = i; + tprops->subnet_prefix = + be64_to_cpu(gid_attr.global.subnet_prefix); add_modify_gid(table, &gid_attr); } err: @@ -1484,7 +1486,7 @@ ib_cache_update(struct ib_device *device, u32 port, bool update_gids, if (!rdma_protocol_roce(device, port) && update_gids) { ret = config_non_roce_gid_cache(device, port, - tprops->gid_tbl_len); + tprops); if (ret) goto err; } > > And I would much prefer things be re-organized so the cache can be > > valid sooner to adding this variable. What is the earlier call that is > > motivating this? > > During device load and when cache is yet to be updated, ib_query_port() > should have a mechanism to identify if the cache entry is valid or invalid > (uninitialized), we have added this variable just to ensure the validity of > cache. Unless there is an actual user of ib_query_port() before config_non_roce_gid_cache() that I can't see, don't bother, returning 0 is fine. Jason
On 6/24/2021 11:24 PM, Jason Gunthorpe wrote: > On Wed, Jun 23, 2021 at 06:33:32PM +0530, Anand Khoje wrote: >> On 6/22/2021 5:19 AM, Jason Gunthorpe wrote: >>> On Wed, Jun 16, 2021 at 09:15:09PM +0530, Anand Khoje wrote: >>>> @@ -1523,13 +1524,21 @@ static int config_non_roce_gid_cache(struct ib_device *device, >>>> device->port_data[port].cache.lmc = tprops->lmc; >>>> device->port_data[port].cache.port_state = tprops->state; >>>> - device->port_data[port].cache.subnet_prefix = tprops->subnet_prefix; >>>> + ret = rdma_query_gid(device, port, 0, &gid); >>>> + if (ret) { >>> >>> This is quite a bit different than just calling ops.query_gid() - why >>> are you changing it? I'm not sure all the additional tests will pass, >>> the 0 gid entry is not required to be valid.. >>> >> Hi Jason, >> >> We have opted for rdma_query_gid(), as during ib_cache_update() the code >> calls ops.query_gid() earlier in config_non_roce_gid_cache(), thereby >> updating the value of GID in cache. We utilize this updated value, instead >> of calling ops->query_gid() again. > > Uhhhh, so just store the subnet prefix at that point then? > > diff --git a/drivers/infiniband/core/cache.c b/drivers/infiniband/core/cache.c > index c9e9fc81447e89..5c554ebd000e89 100644 > --- a/drivers/infiniband/core/cache.c > +++ b/drivers/infiniband/core/cache.c > @@ -1428,8 +1428,8 @@ int rdma_read_gid_l2_fields(const struct ib_gid_attr *attr, > } > EXPORT_SYMBOL(rdma_read_gid_l2_fields); > > -static int config_non_roce_gid_cache(struct ib_device *device, > - u32 port, int gid_tbl_len) > +static int config_non_roce_gid_cache(struct ib_device *device, u32 port, > + struct ib_port_attr *tprops) > { > struct ib_gid_attr gid_attr = {}; > struct ib_gid_table *table; > @@ -1441,7 +1441,7 @@ static int config_non_roce_gid_cache(struct ib_device *device, > table = rdma_gid_table(device, port); > > mutex_lock(&table->lock); > - for (i = 0; i < gid_tbl_len; ++i) { > + for (i = 0; i < tprops->gid_tbl_len; ++i) { > if (!device->ops.query_gid) > continue; > ret = device->ops.query_gid(device, port, i, &gid_attr.gid); > @@ -1452,6 +1452,8 @@ static int config_non_roce_gid_cache(struct ib_device *device, > goto err; > } > gid_attr.index = i; > + tprops->subnet_prefix = > + be64_to_cpu(gid_attr.global.subnet_prefix); > add_modify_gid(table, &gid_attr); > } > err: > @@ -1484,7 +1486,7 @@ ib_cache_update(struct ib_device *device, u32 port, bool update_gids, > > if (!rdma_protocol_roce(device, port) && update_gids) { > ret = config_non_roce_gid_cache(device, port, > - tprops->gid_tbl_len); > + tprops); > if (ret) > goto err; > } > Hi Jason, Thanks for the response! If the above change is to be made, there could arise a scenario in which: In case of a cache_update event, another application/module could try to call ib_query_port() and read subnet_prefix while the cache is still getting updated and the application/module could end up reading a stale value of subnet_prefix. I have a few questions: - How likely is it that an up and running Infiniband fabric would change the subnet_prefix? - Is it possible that different GIDs in the gid_table will have different values of subnet_prefix? >>> And I would much prefer things be re-organized so the cache can be >>> valid sooner to adding this variable. What is the earlier call that is >>> motivating this? >> >> During device load and when cache is yet to be updated, ib_query_port() >> should have a mechanism to identify if the cache entry is valid or invalid >> (uninitialized), we have added this variable just to ensure the validity of >> cache. > > Unless there is an actual user of ib_query_port() before > config_non_roce_gid_cache() that I can't see, don't bother, returning > 0 is fine. > > Jason > Hm! that makes sense, with the above change we wouldn't need to call device->ops.query_gid() from __ib_query_port() and can always read subnet_prefix using ib_get_cached_subnet_prefix(), if reading stale value during cache update event is not an issue. Thanks, Anand
On Fri, Jun 25, 2021 at 11:33:58AM +0530, Anand Khoje wrote: > > If the above change is to be made, there could arise a scenario in which: > In case of a cache_update event, another application/module could try to > call ib_query_port() and read subnet_prefix while the cache is still getting > updated and the application/module could end up reading a stale value of > subnet_prefix. Applications relying on this data must hook the event and update their state when the event fires. So long as ib_query_port returns the correct value in the event handler it is all OK. This whole thing is racy - the HW can change the subnet_prefix at anytime, this is just shuffling the race around. > - Is it possible that different GIDs in the gid_table will have different > values of subnet_prefix? Valid GIDs should have the same prefix Jason
diff --git a/drivers/infiniband/core/cache.c b/drivers/infiniband/core/cache.c index 2325171..88517b5 100644 --- a/drivers/infiniband/core/cache.c +++ b/drivers/infiniband/core/cache.c @@ -1466,6 +1466,7 @@ static int config_non_roce_gid_cache(struct ib_device *device, struct ib_port_attr *tprops = NULL; struct ib_pkey_cache *pkey_cache = NULL; struct ib_pkey_cache *old_pkey_cache = NULL; + union ib_gid gid; int i; int ret; @@ -1523,13 +1524,21 @@ static int config_non_roce_gid_cache(struct ib_device *device, device->port_data[port].cache.lmc = tprops->lmc; device->port_data[port].cache.port_state = tprops->state; - device->port_data[port].cache.subnet_prefix = tprops->subnet_prefix; + ret = rdma_query_gid(device, port, 0, &gid); + if (ret) { + write_unlock_irq(&device->cache_lock); + goto err; + } + + device->port_data[port].cache.subnet_prefix = + be64_to_cpu(gid.global.subnet_prefix); + write_unlock_irq(&device->cache_lock); if (enforce_security) ib_security_cache_change(device, port, - tprops->subnet_prefix); + be64_to_cpu(gid.global.subnet_prefix)); kfree(old_pkey_cache); kfree(tprops); @@ -1629,6 +1638,7 @@ int ib_cache_setup_one(struct ib_device *device) err = ib_cache_update(device, p, true, true, true); if (err) return err; + device->port_data[p].cache_is_initialized = 1; } return 0; diff --git a/drivers/infiniband/core/device.c b/drivers/infiniband/core/device.c index 7a617e4..76fbca2 100644 --- a/drivers/infiniband/core/device.c +++ b/drivers/infiniband/core/device.c @@ -2057,6 +2057,15 @@ static int __ib_query_port(struct ib_device *device, IB_LINK_LAYER_INFINIBAND) return 0; + if (!device->port_data[port_num].cache_is_initialized) + goto query_gid_from_device; + + ib_get_cached_subnet_prefix(device, port_num, + &port_attr->subnet_prefix); + + 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_verbs.h b/include/rdma/ib_verbs.h index c96d601..405f7da 100644 --- a/include/rdma/ib_verbs.h +++ b/include/rdma/ib_verbs.h @@ -2177,6 +2177,7 @@ struct ib_port_data { spinlock_t netdev_lock; + u8 cache_is_initialized:1; struct list_head pkey_list; struct ib_port_cache cache;
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 devices. In the situation of an event causing cache update, the subnet_prefix will get retrieved from newly updated GID cache in ib_cache_update(), so that we do not end up reading a stale value from cache via ib_query_port(). Fixes: fad61ad ("IB/core: Add subnet prefix to port info") Suggested-by: Leon Romanovsky <leonro@nvidia.com> Suggested-by: Aru Kolappan <aru.kolappan@oracle.com> Signed-off-by: Anand Khoje <anand.a.khoje@oracle.com> Signed-off-by: Haakon Bugge <haakon.bugge@oracle.com> --- v1 -> v2: - Split the v1 patch in 3 patches as per Leon's suggestion. v2 -> v3: - Added changes as per Mark Zhang's suggestion of clearing flags in git_table_cleanup_one(). v3 -> v4: - Removed the enum ib_port_data_flags and 8 byte flags from struct ib_port_data, and the set_bit()/clear_bit() API used to update this flag as that was not necessary. Done to keep the code simple. - Added code to read subnet_prefix from updated GID cache in the event of cache update. Prior to this change, ib_cache_update was reading the value for subnet_prefix via ib_query_port(), due to this patch, we ended up reading a stale cached value of subnet_prefix. v4 -> v5: - Removed the code to reset cache_is_initialised bit from cleanup as per Leon's suggestion. - Removed ib_cache_is_initialised() function. --- drivers/infiniband/core/cache.c | 14 ++++++++++++-- drivers/infiniband/core/device.c | 9 +++++++++ include/rdma/ib_verbs.h | 1 + 3 files changed, 22 insertions(+), 2 deletions(-)