nvme-rdma/nvmet-rdma: Allocate sufficient RW ctxs to match hosts pgs len
Subsystems (3)
PATCH 1/1 nvme-rdma/nvmet-rdma: Allocate sufficient RW ctxs to match hosts pgs len
Current nvmet-rdma code allocates MR pool budget based on host's SQ size, assuming both host and target use the same "max_pages_per_mr" count. But if host's max_pages_per_mr is greater than target's, then target can run out of MRs while processing larger IO WRITEs.
That is, say host's SQ size is 100, then the MR pool budget allocated currently at target will also be 100 MRs. But 100 IO WRITE Requests with 256 sg_count(IO size above 1MB) require 200 MRs when target's "max_pages_per_mr" is 128.
The proposed patch enables host to advertise the max_fr_pages(via nvme_rdma_cm_req) such that target can allocate that many number of RW ctxs(if host's max_fr_pages is higher than target's).
--- drivers/nvme/host/rdma.c | 2 ++ drivers/nvme/target/rdma.c | 23 ++++++++++++++++++++--- include/linux/nvme-rdma.h | 4 +++- 3 files changed, 25 insertions(+), 4 deletions(-)
drivers/nvme/host/rdma.c +2 -0
@@ -1614,6 +1614,8 @@ static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue) priv.hsqsize =cpu_to_le16 (queue->ctrl->ctrl.sqsize); } + priv.hmax_fr_pages =cpu_to_le32 (ctrl->max_fr_pages); + ret =rdma_connect (queue->cm_id, ¶m);if (ret) { dev_err(ctrl->ctrl.device,
drivers/nvme/target/rdma.c +20 -3
@@ -98,6 +98,7 @@ struct nvmet_rdma_queue {int host_qid;int recv_queue_size;int send_queue_size; +int rdma_rw_ctxs_factor;struct list_head queue_list; };
@@ -1008,7 +1009,8 @@ static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue) qp_attr.qp_type =IB_QPT_RC ;/* +1 for drain */ qp_attr.cap.max_send_wr = queue->send_queue_size +1 ; - qp_attr.cap.max_rdma_ctxs = queue->send_queue_size; + qp_attr.cap.max_rdma_ctxs = queue->send_queue_size * + queue->rdma_rw_ctxs_factor; qp_attr.cap.max_send_sge =max (ndev->device->attrs.max_sge_rd, ndev->device->attrs.max_send_sge);
@@ -1094,6 +1096,7 @@ nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn, struct nvmet_rdma_queue *queue) {struct nvme_rdma_cm_req *req; +u32 host_fr_pages_len, tgt_fr_pages_len; req = (struct nvme_rdma_cm_req *)conn->private_data;if (!req || conn->private_data_len ==0 )
@@ -1111,6 +1114,19 @@ nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn, queue->recv_queue_size =le16_to_cpu (req->hsqsize) +1 ; queue->send_queue_size =le16_to_cpu (req->hrqsize); + host_fr_pages_len =le32_to_cpu (req->hmax_fr_pages), + tgt_fr_pages_len = queue->dev->device->attrs.max_fast_reg_page_list_len; + +if (host_fr_pages_len > tgt_fr_pages_len) +/* +* Allocate more RW contexts as more MRs are required when +* host_fr_pages_len is higher than target's. +*/ + queue->rdma_rw_ctxs_factor = +DIV_ROUND_UP (host_fr_pages_len, tgt_fr_pages_len); +else + queue->rdma_rw_ctxs_factor =1 ; +if (!queue->host_qid && queue->recv_queue_size >NVME_AQ_DEPTH )return NVME_RDMA_CM_INVALID_HSQSIZE ;
@@ -1147,6 +1163,9 @@ nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev, goto out_reject; } + queue->dev = ndev; + queue->cm_id = cm_id; + ret =nvmet_sq_init (&queue->nvme_sq);if (ret) { ret =NVME_RDMA_CM_NO_RSC ;
@@ -1162,8 +1181,6 @@ nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,* inside a CM callback would trigger a deadlock. (great API design..) */ INIT_WORK (&queue->release_work, nvmet_rdma_release_queue_work); - queue->dev = ndev; - queue->cm_id = cm_id;spin_lock_init (&queue->state_lock); queue->state =NVMET_RDMA_Q_CONNECTING ;
include/linux/nvme-rdma.h +3 -1
@@ -52,13 +52,15 @@ static inline const char *nvme_rdma_cm_msg(enum nvme_rdma_cm_status status)* @qid: queue Identifier for the Admin or I/O Queue * @hrqsize: host receive queue size to be created * @hsqsize: host send queue size to be created +* @hmax_fr_pages: host maximum pages per fast reg */ struct nvme_rdma_cm_req {__le16 recfmt;__le16 qid;__le16 hrqsize;__le16 hsqsize; -u8 rsvd[24 ]; +__le32 hmax_fr_pages; +u8 rsvd[20 ]; };/**
Signature · 1 line
2.23.0.rc0
Quoted 8 lines · Krishnamraju Eraparaju
include/linux/nvme-rdma.h +2
-0+++ b/include/linux/nvme-rdma.h@@ -52,13 +52,15 @@ static inline const char *nvme_rdma_cm_msg(enum nvme_rdma_cm_status status)* @qid: queue Identifier for the Admin or I/O Queue * @hrqsize: host receive queue size to be created * @hsqsize: host send queue size to be created +* @hmax_fr_pages: host maximum pages per fast reg
include/linux/nvme-rdma.h static inline const char *nvme_rdma_cm_msg(enum nvme_rdma_cm_status status) · line 64
Quoted 10 lines
include/linux/nvme-rdma.h +2
-1*/ struct nvme_rdma_cm_req {__le16 recfmt;__le16 qid;__le16 hrqsize;__le16 hsqsize; -u8 rsvd[24 ]; +__le32 hmax_fr_pages; +u8 rsvd[20 ]; };
This is an on the wire change - do you need to get approval from some standards body?
Jason
Sagi Grimberg
maintainerQuoted 14 lines
Quoted 14 lines
include/linux/nvme-rdma.h +2
-0+++ b/include/linux/nvme-rdma.h@@ -52,13 +52,15 @@ static inline const char *nvme_rdma_cm_msg(enum nvme_rdma_cm_status status)* @qid: queue Identifier for the Admin or I/O Queue * @hrqsize: host receive queue size to be created * @hsqsize: host send queue size to be created +* @hmax_fr_pages: host maximum pages per fast reg */ struct nvme_rdma_cm_req {__le16 recfmt;__le16 qid;__le16 hrqsize;
Quoted 8 lines
include/linux/nvme-rdma.h static inline const char *nvme_rdma_cm_msg(enum nvme_rdma_cm_status status) · line 64
Quoted 5 lines
include/linux/nvme-rdma.h +2
-1__le16 hsqsize; -u8 rsvd[24 ]; +__le32 hmax_fr_pages; +u8 rsvd[20 ]; };This is an on the wire change - do you need to get approval from some standards body?
Yes, this needs to go though the NVMe TWG for sure.
Sagi Grimberg
maintainerQuoted 10 lines
Current nvmet-rdma code allocates MR pool budget based on host's SQ size, assuming both host and target use the same "
max_pages_per_mr" count. But if host'smax_pages_per_mris greater than target's, then target can run out of MRs while processing larger IO WRITEs.That is, say host's SQ size is 100, then the MR pool budget allocated currently at target will also be 100 MRs. But 100 IO WRITE Requests with 256
sg_count(IO size above 1MB)require 200 MRs when target's "max_pages_per_mr" is 128.
The patch doesn't say if this is an actual bug you are seeing or theoretical.
Quoted 3 lines
The proposed patch enables host to advertise the
max_fr_pages(via nvme_rdma_cm_req)such that target can allocate that many number of RWctxs(if host's max_fr_pages is higher than target's).
As mentioned by Jason, this s a non-compatible change, if you want to introduce this you need to go through the standard and update the cm private_data layout (would mean that the fmt needs to increment as well to be backward compatible).
As a stop-gap, nvmet needs to limit the controller mdts to how much it can allocate based on the HCA capabilities (max_fast_reg_page_list_len).
Hi Sagi & Jason,
Thanks for the comments, please see inline.
Quoted 6 lines · Sagi Grimberg
Quoted 6 lines
Current nvmet-rdma code allocates MR pool budget based on host's SQ size, assuming both host and target use the same "
max_pages_per_mr" count. But if host'smax_pages_per_mris greater than target's, then target can run out of MRs while processing larger IO WRITEs.
Quoted 7 lines
Quoted 4 lines
That is, say host's SQ size is 100, then the MR pool budget allocated currently at target will also be 100 MRs. But 100 IO WRITE Requests with 256
sg_count(IO size above 1MB)require 200 MRs when target's "max_pages_per_mr" is 128.The patch doesn't say if this is an actual bug you are seeing or theoretical.
I've noticed this issue while running the below fio command:
fio --rw=randwrite --name=random --norandommap --ioengine=libaio --size=16m --group_reporting --exitall --fsync_on_close=1 --invalidate=1 --direct=1 --filename=/dev/nvme2n1 --iodepth=32 --numjobs=16 --unit_base=1 --bs=4m --kb_base=1000
Note: here NVMe Host is on SIW & Target is on iw_cxgb4 and the max_pages_per_mr supported by SIW and iw_cxgb4 are 255 and 128 respectively.
Traces on Target:
#cat /sys/kernel/debug/tracing/trace_pipe|grep -v "status=0x0" kworker/8:1H-2461 [008] .... 25476.995437: nvmet_req_complete: nvmet1:
disk=/dev/ram0, qid=1, cmdid=3, res=0xffff8b7f2ae534d0, status=0x6 kworker/8:1H-2461 [008] .... 25476.995467: nvmet_req_complete: nvmet1:
disk=/dev/ram0, qid=1, cmdid=4, res=0xffff8b7f2ae53700, status=0x6 kworker/8:1H-2461 [008] .... 25476.995511: nvmet_req_complete: nvmet1:
disk=/dev/ram0, qid=1, cmdid=1, res=0xffff8b7f2ae53980, status=0x6
Quoted 9 lines
Quoted 4 lines
The proposed patch enables host to advertise the
max_fr_pages(via nvme_rdma_cm_req)such that target can allocate that many number of RWctxs(if host's max_fr_pages is higher than target's).As mentioned by Jason, this s a non-compatible change, if you want to introduce this you need to go through the standard and update the cm
private_datalayout (would mean that the fmt needs to increment as well to be backward compatible).
Sure, will initiate a discussion at NVMe TWG about CM private_data format. Will update the response soon.
Quoted 5 lines
As a stop-gap, nvmet needs to limit the controller mdts to how much it can allocate based on the HCA capabilities (
max_fast_reg_page_list_len).
Sagi Grimberg
maintainerQuoted 10 lines
Quoted 3 lines
The patch doesn't say if this is an actual bug you are seeing or theoretical.
I've noticed this issue while running the below fio command:
fio --rw=randwrite --name=random --norandommap --ioengine=libaio --size=16m --group_reporting --exitall --fsync_on_close=1 --invalidate=1 --direct=1 --filename=/dev/nvme2n1 --iodepth=32 --numjobs=16 --unit_base=1 --bs=4m --kb_base=1000
Quoted 3 lines
Note: here NVMe Host is on SIW & Target is on
iw_cxgb4and themax_pages_per_mrsupported by SIW andiw_cxgb4are 255 and 128 respectively.
This needs to be documented in the change log.
Quoted 4 lines
Quoted 4 lines
Quoted 3 lines
The proposed patch enables host to advertise the
max_fr_pages(via nvme_rdma_cm_req)such that target can allocate that many number of RWctxs(if host's max_fr_pages is higher than target's).
Quoted 12 lines
Quoted 4 lines
As mentioned by Jason, this s a non-compatible change, if you want to introduce this you need to go through the standard and update the cm
private_datalayout (would mean that the fmt needs to increment as well to be backward compatible).Sure, will initiate a discussion at NVMe TWG about CM
private_dataformat. Will update the response soon.Quoted 5 lines
As a stop-gap, nvmet needs to limit the controller mdts to how much it can allocate based on the HCA capabilities (
max_fast_reg_page_list_len).
Sounds good, please look at capping mdts in the mean time.
Quoted 29 lines · Sagi Grimberg
Quoted 10 lines
Quoted 3 lines
The patch doesn't say if this is an actual bug you are seeing or theoretical.
I've noticed this issue while running the below fio command:
fio --rw=randwrite --name=random --norandommap --ioengine=libaio --size=16m --group_reporting --exitall --fsync_on_close=1 --invalidate=1 --direct=1 --filename=/dev/nvme2n1 --iodepth=32 --numjobs=16 --unit_base=1 --bs=4m --kb_base=1000Quoted 3 lines
Note: here NVMe Host is on SIW & Target is on
iw_cxgb4and themax_pages_per_mrsupported by SIW andiw_cxgb4are 255 and 128 respectively.This needs to be documented in the change log.
Quoted 13 lines
Quoted 8 lines
Quoted 3 lines
The proposed patch enables host to advertise the
max_fr_pages(via nvme_rdma_cm_req)such that target can allocate that many number of RWctxs(if host's max_fr_pages is higher than target's).As mentioned by Jason, this s a non-compatible change, if you want to introduce this you need to go through the standard and update the cm
private_datalayout (would mean that the fmt needs to increment as well to be backward compatible).Sure, will initiate a discussion at NVMe TWG about CM
private_dataformat. Will update the response soon.Quoted 2 lines
Quoted 5 lines
Quoted 3 lines
Quoted 3 lines
As a stop-gap, nvmet needs to limit the controller mdts to how much it can allocate based on the HCA capabilities (
max_fast_reg_page_list_len).Sounds good, please look at capping mdts in the mean time.
guys, see my patches from adding MD support.
I'm setting mdts per ctrl there.
we can merge it meantime for this issue.
Hi Sagi & Max Gurtovoy,
Thanks for your pointers regarding mdts.
Looks like fixing this issue through mdts is more natural than fixing through RDMA private data.
Issue is not occuring after appling the below patch(inspired by Max's patch "nvmet-rdma: Implement set_mdts controller op").
So any consensus about merging the fix upstream, to fix this specific issue?
(quoted diff) +0 -0
b/drivers/nvme/target/admin-cmd.c
index 56c21b50..0d468ab 100644
drivers/nvme/target/admin-cmd.c +0 -0
@@ -346,8 +346,12 @@ static void nvmet_execute_identify_ctrl(struct
nvmet_req *req)
/* we support multiple ports, multiples hosts and ANA: */
id->cmic = (1 << 0) | (1 << 1) | (1 << 3);- /* no limit on data transfer sizes for now */ - id->mdts = 0; + /* Limit MDTS according to transport capability */ + if (ctrl->ops->set_mdts)
+ id->mdts = ctrl->ops->set_mdts(ctrl);
+ else
+ id->mdts = 0;
+
id->cntlid = cpu_to_le16(ctrl->cntlid);
id->ver = cpu_to_le32(ctrl->subsys->ver);drivers/nvme/target/nvmet.h +1 -0
@@ -279,6 +279,7 @@ struct nvmet_fabrics_ops { struct nvmet_port *port, char *traddr);u16 (*install_queue)(struct nvmet_sq *nvme_sq);void (*discovery_chg)(struct nvmet_port *port); +u8 (*set_mdts)(struct nvmet_ctrl *ctrl); };#define NVMET_MAX_INLINE_BIOVEC 8
drivers/nvme/target/rdma.c +0 -0
@@ -1602,6 +1602,17 @@ static void nvmet_rdma_disc_port_addr(struct
nvmet_req *req,
}
}
+static u8 nvmet_rdma_set_mdts(struct nvmet_ctrl *ctrl) +{
+ struct nvmet_port *port = ctrl->port; + struct rdma_cm_id *cm_id = port->priv; + u32 max_pages; + + max_pages = cm_id->device->attrs.max_fast_reg_page_list_len; + /* Assume mpsmin == device_page_size == 4KB */ + return ilog2(max_pages);
+}
+
static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
.owner = THIS_MODULE,
.type = NVMF_TRTYPE_RDMA,
@@ -1612,6 +1623,7 @@ static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
.queue_response = nvmet_rdma_queue_response,
.delete_ctrl = nvmet_rdma_delete_ctrl,
.disc_traddr = nvmet_rdma_disc_port_addr,
+ .set_mdts = nvmet_rdma_set_mdts,};
static void nvmet_rdma_remove_one(struct ib_device *ib_device, void*client_data)
Thanks,
Krishna.
Quoted 36 lines · Max Gurtovoy
Quoted 34 lines · Sagi Grimberg
Quoted 10 lines
Quoted 3 lines
The patch doesn't say if this is an actual bug you are seeing or theoretical.
I've noticed this issue while running the below fio command:
fio --rw=randwrite --name=random --norandommap --ioengine=libaio --size=16m --group_reporting --exitall --fsync_on_close=1 --invalidate=1 --direct=1 --filename=/dev/nvme2n1 --iodepth=32 --numjobs=16 --unit_base=1 --bs=4m --kb_base=1000Quoted 3 lines
Note: here NVMe Host is on SIW & Target is on
iw_cxgb4and themax_pages_per_mrsupported by SIW andiw_cxgb4are 255 and 128 respectively.This needs to be documented in the change log.
Quoted 17 lines
Quoted 8 lines
Quoted 3 lines
The proposed patch enables host to advertise the
max_fr_pages(via nvme_rdma_cm_req)such that target can allocate that many number of RWctxs(if host's max_fr_pages is higher than target's).As mentioned by Jason, this s a non-compatible change, if you want to introduce this you need to go through the standard and update the cm
private_datalayout (would mean that the fmt needs to increment as well to be backward compatible).Sure, will initiate a discussion at NVMe TWG about CM
private_dataformat.
Will update the response soon.Quoted 5 lines
As a stop-gap, nvmet needs to limit the controller mdts to how much it can allocate based on the HCA capabilities (
max_fast_reg_page_list_len).
Quoted 9 lines
Quoted 1 line
Sounds good, please look at capping mdts in the mean time.
guys, see my patches from adding MD support.
I'm setting mdts per ctrl there.
we can merge it meantime for this issue.
_______________________________________________
linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme
Sagi Grimberg
maintainerQuoted 5 lines
Hi Sagi & Max Gurtovoy,
Thanks for your pointers regarding mdts.
Quoted 8 lines
Looks like fixing this issue through mdts is more natural than fixing through RDMA private data.
Issue is not occuring after appling the below
patch(inspired by Max's patch "nvmet-rdma: Implement set_mdts controller op").So any consensus about merging the fix upstream, to fix this specific issue?
I think we can do this asap.
Max, please send a patch for this.
Thanks,
Quoted 11 lines · Sagi Grimberg
Quoted 11 lines
Hi Sagi & Max Gurtovoy,
Thanks for your pointers regarding mdts.
Looks like fixing this issue through mdts is more natural than fixing through RDMA private data.
Issue is not occuring after appling the below
patch(inspired by Max's patch "nvmet-rdma: Implement set_mdts controller op").
Quoted 6 lines
Quoted 2 lines
So any consensus about merging the fix upstream, to fix this specific issue?
I think we can do this asap.
Max, please send a patch for this.
sure, I'll do that.
But since nvmet/rdma uses RW API that may use multiple MR's per IO request we need to implement the set_mdts differently.
We need to update the value of max_rdma_ctxs that set the MR pool size. We can limit the RDMA transport to support upto 1 or 2 MB and get the factor we need for setting the max_rdma_ctxs.
I'll send a fix soon.
Quoted 2 lines
Thanks,