Updated nroff-generated man pages

Signed-off-by: OFIWG Bot <[email protected]>

man/man7/fi_cxi.7

@@ -1,7 +1,7 @@
 .\"t
 .\" Automatically generated by Pandoc 2.9.2.1
 .\"
-.TH "fi_cxi" "7" "2024\-11\-20" "Libfabric Programmer\[cq]s Manual" "#VERSION#"
+.TH "fi_cxi" "7" "2024\-11\-25" "Libfabric Programmer\[cq]s Manual" "#VERSION#"
 .hy
 .SH NAME
 .PP
@@ -442,53 +442,25 @@ hybrid RX match modes increase Request buffer space using the variables
 \f[I]FI_CXI_REQ_*\f[R].
 .SS Message Ordering
 .PP
-The CXI provider supports the following ordering rules:
-.IP \[bu] 2
-All message Send operations are always ordered.
-.IP \[bu] 2
-RMA Writes may be ordered by specifying \f[I]FI_ORDER_RMA_WAW\f[R].
-.IP \[bu] 2
-AMOs may be ordered by specifying
-\f[I]FI_ORDER_AMO_{WAW|WAR|RAW|RAR}\f[R].
-.IP \[bu] 2
-RMA Writes may be ordered with respect to AMOs by specifying
-\f[I]FI_ORDER_WAW\f[R].
-Fetching AMOs may be used to perform short reads that are ordered with
-respect to RMA Writes.
+Supported message ordering: FI_ORDER_SAS, FI_ORDER_WAW,
+FI_ORDER_RMA_WAW, FI_ORDER_RMA_RAR, FI_ORDER_ATOMIC_WAW, and
+FI_ORDER_ATOMIC_RAR.
+.PP
+Note: Any FI_ORDER_*_{WAR,RAW} are not supported.
+.PP
+Note: Relaxing the message ordering may result in improved performance.
+.SS Target Ordering
 .PP
 Ordered RMA size limits are set as follows:
 .IP \[bu] 2
 \f[I]max_order_waw_size\f[R] is -1.
-RMA Writes and non-fetching AMOs of any size are ordered with respect to
-each other.
-.IP \[bu] 2
-\f[I]max_order_raw_size\f[R] is -1.
-Fetching AMOs of any size are ordered with respect to RMA Writes and
-non-fetching AMOs.
-.IP \[bu] 2
-\f[I]max_order_war_size\f[R] is -1.
-RMA Writes and non-fetching AMOs of any size are ordered with respect to
-fetching AMOs.
-.SS PCIe Ordering
-.PP
-Generally, PCIe writes are strictly ordered.
-As an optimization, PCIe TLPs may have the Relaxed Order (RO) bit set to
-allow writes to be reordered.
-Cassini sets the RO bit in PCIe TLPs when possible.
-Cassini sets PCIe RO as follows:
-.IP \[bu] 2
-Ordering of messaging operations is established using completion events.
-Therefore, all PCIe TLPs related to two-sided message payloads will have
-RO set.
-.IP \[bu] 2
-Every PCIe TLP associated with an unordered RMA or AMO operation will
-have RO cleared.
-.IP \[bu] 2
-PCIe TLPs associated with the last packet of an ordered RMA or AMO
-operation will have RO cleared.
-.IP \[bu] 2
-PCIe TLPs associated with the body packets (all except the last packet
-of an operation) of an ordered RMA operation will have RO set.
+RMA Writes and AMO writes of any size are ordered with respect to each
+other.
+.PP
+Note: Due to FI_ORDER_*_{WAR,RAW} not being supported,
+max_order_{raw,war}_size are forced to zero.
+.PP
+Note: Relaxing the target ordering may result in improved performance.
 .SS Translation
 .PP
 The CXI provider supports two translation mechanisms: Address
@@ -1172,6 +1144,11 @@ offloading are met.
 .PP
 The CXI provider checks for the following environment variables:
 .TP
+\f[I]FI_CXI_MR_TARGET_ORDERING\f[R]
+MR target ordering (i.e.\ PCI ordering).
+Options: default, strict, or relaxed.
+Recommendation is to leave at default behavior.
+.TP
 \f[I]FI_CXI_ODP\f[R]
 Enables on-demand paging.
 If disabled, all DMA buffers are pinned.
@@ -1836,12 +1813,6 @@ if (ret)
 \f[R]
 .fi
 .PP
-When an endpoint does not support FI_FENCE (e.g.\ optimized MR), a
-provider specific transmit flag, FI_CXI_WEAK_FENCE, may be specified on
-an alias EP to issue a FENCE operation to create a data ordering point
-for the alias.
-This is supported for one-sided operations only.
-.PP
 Alias EP must be closed prior to closing the original EP.
 .SS PCIe Atomics
 .PP