| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Date: Mon, 15 Aug 2022 20:07:29 +0200
From: Greg Kroah-Hartman <gregkh@...uxfoundation.org>
To: linux-kernel@...r.kernel.org
Cc: Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
stable@...r.kernel.org, Hyunchul Lee <hyc.lee@...il.com>,
Namjae Jeon <linkinjeon@...nel.org>,
Steve French <stfrench@...rosoft.com>,
Sasha Levin <sashal@...nel.org>
Subject: [PATCH 5.18 1050/1095] ksmbd: smbd: introduce read/write credits for RDMA read/write
From: Hyunchul Lee <hyc.lee@...il.com>
[ Upstream commit ddbdc861e37c168cf2fb8a7b7477f5d18b4daf76 ]
SMB2_READ/SMB2_WRITE request has to be granted the number
of rw credits, the pages the request wants to transfer
/ the maximum pages which can be registered with one
MR to read and write a file.
And allocate enough RDMA resources for the maximum
number of rw credits allowed by ksmbd.
Signed-off-by: Hyunchul Lee <hyc.lee@...il.com>
Acked-by: Namjae Jeon <linkinjeon@...nel.org>
Signed-off-by: Steve French <stfrench@...rosoft.com>
Signed-off-by: Sasha Levin <sashal@...nel.org>
---
fs/ksmbd/transport_rdma.c | 120 ++++++++++++++++++++++----------------
1 file changed, 71 insertions(+), 49 deletions(-)
diff --git a/fs/ksmbd/transport_rdma.c b/fs/ksmbd/transport_rdma.c
index 479d279ee146..b44a5e584bac 100644
--- a/fs/ksmbd/transport_rdma.c
+++ b/fs/ksmbd/transport_rdma.c
@@ -80,9 +80,7 @@ static int smb_direct_max_fragmented_recv_size = 1024 * 1024;
/* The maximum single-message size which can be received */
static int smb_direct_max_receive_size = 8192;
-static int smb_direct_max_read_write_size = 524224;
-
-static int smb_direct_max_outstanding_rw_ops = 8;
+static int smb_direct_max_read_write_size = 8 * 1024 * 1024;
static LIST_HEAD(smb_direct_device_list);
static DEFINE_RWLOCK(smb_direct_device_lock);
@@ -147,10 +145,12 @@ struct smb_direct_transport {
atomic_t send_credits;
spinlock_t lock_new_recv_credits;
int new_recv_credits;
- atomic_t rw_avail_ops;
+ int max_rw_credits;
+ int pages_per_rw_credit;
+ atomic_t rw_credits;
wait_queue_head_t wait_send_credits;
- wait_queue_head_t wait_rw_avail_ops;
+ wait_queue_head_t wait_rw_credits;
mempool_t *sendmsg_mempool;
struct kmem_cache *sendmsg_cache;
@@ -377,7 +377,7 @@ static struct smb_direct_transport *alloc_transport(struct rdma_cm_id *cm_id)
t->reassembly_queue_length = 0;
init_waitqueue_head(&t->wait_reassembly_queue);
init_waitqueue_head(&t->wait_send_credits);
- init_waitqueue_head(&t->wait_rw_avail_ops);
+ init_waitqueue_head(&t->wait_rw_credits);
spin_lock_init(&t->receive_credit_lock);
spin_lock_init(&t->recvmsg_queue_lock);
@@ -984,18 +984,19 @@ static int smb_direct_flush_send_list(struct smb_direct_transport *t,
}
static int wait_for_credits(struct smb_direct_transport *t,
- wait_queue_head_t *waitq, atomic_t *credits)
+ wait_queue_head_t *waitq, atomic_t *total_credits,
+ int needed)
{
int ret;
do {
- if (atomic_dec_return(credits) >= 0)
+ if (atomic_sub_return(needed, total_credits) >= 0)
return 0;
- atomic_inc(credits);
+ atomic_add(needed, total_credits);
ret = wait_event_interruptible(*waitq,
- atomic_read(credits) > 0 ||
- t->status != SMB_DIRECT_CS_CONNECTED);
+ atomic_read(total_credits) >= needed ||
+ t->status != SMB_DIRECT_CS_CONNECTED);
if (t->status != SMB_DIRECT_CS_CONNECTED)
return -ENOTCONN;
@@ -1016,7 +1017,19 @@ static int wait_for_send_credits(struct smb_direct_transport *t,
return ret;
}
- return wait_for_credits(t, &t->wait_send_credits, &t->send_credits);
+ return wait_for_credits(t, &t->wait_send_credits, &t->send_credits, 1);
+}
+
+static int wait_for_rw_credits(struct smb_direct_transport *t, int credits)
+{
+ return wait_for_credits(t, &t->wait_rw_credits, &t->rw_credits, credits);
+}
+
+static int calc_rw_credits(struct smb_direct_transport *t,
+ char *buf, unsigned int len)
+{
+ return DIV_ROUND_UP(get_buf_page_count(buf, len),
+ t->pages_per_rw_credit);
}
static int smb_direct_create_header(struct smb_direct_transport *t,
@@ -1332,8 +1345,8 @@ static void read_write_done(struct ib_cq *cq, struct ib_wc *wc,
smb_direct_disconnect_rdma_connection(t);
}
- if (atomic_inc_return(&t->rw_avail_ops) > 0)
- wake_up(&t->wait_rw_avail_ops);
+ if (atomic_inc_return(&t->rw_credits) > 0)
+ wake_up(&t->wait_rw_credits);
rdma_rw_ctx_destroy(&msg->rw_ctx, t->qp, t->qp->port,
msg->sg_list, msg->sgt.nents, dir);
@@ -1364,8 +1377,10 @@ static int smb_direct_rdma_xmit(struct smb_direct_transport *t,
struct ib_send_wr *first_wr = NULL;
u32 remote_key = le32_to_cpu(desc[0].token);
u64 remote_offset = le64_to_cpu(desc[0].offset);
+ int credits_needed;
- ret = wait_for_credits(t, &t->wait_rw_avail_ops, &t->rw_avail_ops);
+ credits_needed = calc_rw_credits(t, buf, buf_len);
+ ret = wait_for_rw_credits(t, credits_needed);
if (ret < 0)
return ret;
@@ -1373,7 +1388,7 @@ static int smb_direct_rdma_xmit(struct smb_direct_transport *t,
msg = kmalloc(offsetof(struct smb_direct_rdma_rw_msg, sg_list) +
sizeof(struct scatterlist) * SG_CHUNK_SIZE, GFP_KERNEL);
if (!msg) {
- atomic_inc(&t->rw_avail_ops);
+ atomic_add(credits_needed, &t->rw_credits);
return -ENOMEM;
}
@@ -1382,7 +1397,7 @@ static int smb_direct_rdma_xmit(struct smb_direct_transport *t,
get_buf_page_count(buf, buf_len),
msg->sg_list, SG_CHUNK_SIZE);
if (ret) {
- atomic_inc(&t->rw_avail_ops);
+ atomic_add(credits_needed, &t->rw_credits);
kfree(msg);
return -ENOMEM;
}
@@ -1418,7 +1433,7 @@ static int smb_direct_rdma_xmit(struct smb_direct_transport *t,
return 0;
err:
- atomic_inc(&t->rw_avail_ops);
+ atomic_add(credits_needed, &t->rw_credits);
if (first_wr)
rdma_rw_ctx_destroy(&msg->rw_ctx, t->qp, t->qp->port,
msg->sg_list, msg->sgt.nents,
@@ -1643,11 +1658,19 @@ static int smb_direct_prepare_negotiation(struct smb_direct_transport *t)
return ret;
}
+static unsigned int smb_direct_get_max_fr_pages(struct smb_direct_transport *t)
+{
+ return min_t(unsigned int,
+ t->cm_id->device->attrs.max_fast_reg_page_list_len,
+ 256);
+}
+
static int smb_direct_init_params(struct smb_direct_transport *t,
struct ib_qp_cap *cap)
{
struct ib_device *device = t->cm_id->device;
- int max_send_sges, max_pages, max_rw_wrs, max_send_wrs;
+ int max_send_sges, max_rw_wrs, max_send_wrs;
+ unsigned int max_sge_per_wr, wrs_per_credit;
/* need 2 more sge. because a SMB_DIRECT header will be mapped,
* and maybe a send buffer could be not page aligned.
@@ -1659,25 +1682,31 @@ static int smb_direct_init_params(struct smb_direct_transport *t,
return -EINVAL;
}
- /*
- * allow smb_direct_max_outstanding_rw_ops of in-flight RDMA
- * read/writes. HCA guarantees at least max_send_sge of sges for
- * a RDMA read/write work request, and if memory registration is used,
- * we need reg_mr, local_inv wrs for each read/write.
+ /* Calculate the number of work requests for RDMA R/W.
+ * The maximum number of pages which can be registered
+ * with one Memory region can be transferred with one
+ * R/W credit. And at least 4 work requests for each credit
+ * are needed for MR registration, RDMA R/W, local & remote
+ * MR invalidation.
*/
t->max_rdma_rw_size = smb_direct_max_read_write_size;
- max_pages = DIV_ROUND_UP(t->max_rdma_rw_size, PAGE_SIZE) + 1;
- max_rw_wrs = DIV_ROUND_UP(max_pages, SMB_DIRECT_MAX_SEND_SGES);
- max_rw_wrs += rdma_rw_mr_factor(device, t->cm_id->port_num,
- max_pages) * 2;
- max_rw_wrs *= smb_direct_max_outstanding_rw_ops;
+ t->pages_per_rw_credit = smb_direct_get_max_fr_pages(t);
+ t->max_rw_credits = DIV_ROUND_UP(t->max_rdma_rw_size,
+ (t->pages_per_rw_credit - 1) *
+ PAGE_SIZE);
+
+ max_sge_per_wr = min_t(unsigned int, device->attrs.max_send_sge,
+ device->attrs.max_sge_rd);
+ wrs_per_credit = max_t(unsigned int, 4,
+ DIV_ROUND_UP(t->pages_per_rw_credit,
+ max_sge_per_wr) + 1);
+ max_rw_wrs = t->max_rw_credits * wrs_per_credit;
max_send_wrs = smb_direct_send_credit_target + max_rw_wrs;
if (max_send_wrs > device->attrs.max_cqe ||
max_send_wrs > device->attrs.max_qp_wr) {
- pr_err("consider lowering send_credit_target = %d, or max_outstanding_rw_ops = %d\n",
- smb_direct_send_credit_target,
- smb_direct_max_outstanding_rw_ops);
+ pr_err("consider lowering send_credit_target = %d\n",
+ smb_direct_send_credit_target);
pr_err("Possible CQE overrun, device reporting max_cqe %d max_qp_wr %d\n",
device->attrs.max_cqe, device->attrs.max_qp_wr);
return -EINVAL;
@@ -1712,7 +1741,7 @@ static int smb_direct_init_params(struct smb_direct_transport *t,
t->send_credit_target = smb_direct_send_credit_target;
atomic_set(&t->send_credits, 0);
- atomic_set(&t->rw_avail_ops, smb_direct_max_outstanding_rw_ops);
+ atomic_set(&t->rw_credits, t->max_rw_credits);
t->max_send_size = smb_direct_max_send_size;
t->max_recv_size = smb_direct_max_receive_size;
@@ -1720,12 +1749,10 @@ static int smb_direct_init_params(struct smb_direct_transport *t,
cap->max_send_wr = max_send_wrs;
cap->max_recv_wr = t->recv_credit_max;
- cap->max_send_sge = SMB_DIRECT_MAX_SEND_SGES;
+ cap->max_send_sge = max_sge_per_wr;
cap->max_recv_sge = SMB_DIRECT_MAX_RECV_SGES;
cap->max_inline_data = 0;
- cap->max_rdma_ctxs =
- rdma_rw_mr_factor(device, t->cm_id->port_num, max_pages) *
- smb_direct_max_outstanding_rw_ops;
+ cap->max_rdma_ctxs = t->max_rw_credits;
return 0;
}
@@ -1818,7 +1845,8 @@ static int smb_direct_create_qpair(struct smb_direct_transport *t,
}
t->send_cq = ib_alloc_cq(t->cm_id->device, t,
- t->send_credit_target, 0, IB_POLL_WORKQUEUE);
+ smb_direct_send_credit_target + cap->max_rdma_ctxs,
+ 0, IB_POLL_WORKQUEUE);
if (IS_ERR(t->send_cq)) {
pr_err("Can't create RDMA send CQ\n");
ret = PTR_ERR(t->send_cq);
@@ -1827,8 +1855,7 @@ static int smb_direct_create_qpair(struct smb_direct_transport *t,
}
t->recv_cq = ib_alloc_cq(t->cm_id->device, t,
- cap->max_send_wr + cap->max_rdma_ctxs,
- 0, IB_POLL_WORKQUEUE);
+ t->recv_credit_max, 0, IB_POLL_WORKQUEUE);
if (IS_ERR(t->recv_cq)) {
pr_err("Can't create RDMA recv CQ\n");
ret = PTR_ERR(t->recv_cq);
@@ -1857,17 +1884,12 @@ static int smb_direct_create_qpair(struct smb_direct_transport *t,
pages_per_rw = DIV_ROUND_UP(t->max_rdma_rw_size, PAGE_SIZE) + 1;
if (pages_per_rw > t->cm_id->device->attrs.max_sgl_rd) {
- int pages_per_mr, mr_count;
-
- pages_per_mr = min_t(int, pages_per_rw,
- t->cm_id->device->attrs.max_fast_reg_page_list_len);
- mr_count = DIV_ROUND_UP(pages_per_rw, pages_per_mr) *
- atomic_read(&t->rw_avail_ops);
- ret = ib_mr_pool_init(t->qp, &t->qp->rdma_mrs, mr_count,
- IB_MR_TYPE_MEM_REG, pages_per_mr, 0);
+ ret = ib_mr_pool_init(t->qp, &t->qp->rdma_mrs,
+ t->max_rw_credits, IB_MR_TYPE_MEM_REG,
+ t->pages_per_rw_credit, 0);
if (ret) {
pr_err("failed to init mr pool count %d pages %d\n",
- mr_count, pages_per_mr);
+ t->max_rw_credits, t->pages_per_rw_credit);
goto err;
}
}
--
2.35.1
Powered by blists - more mailing lists