This is basically a copy of smb_extract_iter_to_rdma() moved
to common code.
Before we had the inconsistency we called ib_dma_unmap_single(),
while we mapped using ib_dma_map_page() in smb_set_sge().
Now ib_dma_unmap_page() is used for consistency.
It doesn't really matter as ib_dma_unmap_single() and
ib_dma_unmap_page() both operate
on dma_addr_t and dma_unmap_single_attrs() is just an
alias for dma_unmap_page_attrs().
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
The main logical differences are the following:
We now don't use smbdirect_connection_get_recv_io() on cleanup,
instead it uses list_for_each_entry_safe()...
For the smbdirect_recv_io payload we expose the whole payload including
the smbdirect_data_transfer header as documentation says data_offset = 0
and data_length != 0 would be valid, while the existing client code
requires data_offset >= 24.
The smbdirect_send_io cache includes header space for
sizeof(struct smbdirect_negotiate_resp) = 32 bytes
instead of sizeof(struct smbdirect_data_transfer) = 24 bytes.
If this ever becomes a problem, we can allocate separate
space for the smbdirect_negotiate_resp in the server.
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
This simplifies the code and allows us to share more code in common
with the server.
The only difference is that we use ib_dma_unmap_page() for all sges,
this simplifies the logic and doesn't matter as
ib_dma_unmap_single() and ib_dma_unmap_page() both operate
on dma_addr_t and dma_unmap_single_attrs() is just an
alias for dma_unmap_page_attrs().
We already had such an inconsistency before
as we called ib_dma_unmap_single(), while we mapped
using ib_dma_map_page() in smb_set_sge().
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
smbdirect_public.h contains functions which will be still be
eported when we move to an smbdirect.ko.
For now this uses the SMBDIRECT_USE_INLINE_C_FILES code path
and marks all function as '__maybe_unused static',
but this will make further changes easier.
Note this generates the following things from checkpatch.pl,
so I passed --ignore=FILE_PATH_CHANGES,EXPORT_SYMBOL,COMPLEX_MACRO
ERROR: Macros with complex values should be enclosed in parentheses
#514: FILE: fs/smb/common/smbdirect/smbdirect_public.h:18:
+#define __SMBDIRECT_PUBLIC__ __maybe_unused static
WARNING: EXPORT_SYMBOL(foo); should immediately follow its function/variable
#515: FILE: fs/smb/common/smbdirect/smbdirect_public.h:19:
+#define __SMBDIRECT_EXPORT_SYMBOL__(__sym)
WARNING: EXPORT_SYMBOL(foo); should immediately follow its function/variable
#518: FILE: fs/smb/common/smbdirect/smbdirect_public.h:22:
+#define __SMBDIRECT_EXPORT_SYMBOL__(__sym) EXPORT_SYMBOL_FOR_MODULES(__sym, "cifs,ksmbd")
This is exactly what we want here, so we should ignore the
checkpatch.pl problems.
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
This will be used by the server to handle new connections.
All rdma processing from rdma_accept() to RDMA_CM_EVENT_ESTABLISHED
as well as the waiting for the smbdirect negotiation request
and sending the negotiation response is done async
until we reach SMBDIRECT_SOCKET_CONNECTED.
Sync behaviour will be done by the server calling
smbdirect_conection_wait_for_connected() in order
to each SMBDIRECT_SOCKET_CONNECTED or an error.
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
This implements a fully async connect logic over
all rdma related operations: rdma_resolve_addr,
rdma_resolve_route and rdma_connect_locked
until we reach RDMA_CM_EVENT_ESTABLISHED,
followed by the smbdirect negotiation request/response
handling until we reach SMBDIRECT_SOCKET_CONNECTED.
smbdirect_connect_sync() is just a useful wrapper
around the async logic calling smbdirect_connect()
followed by smbdirect_connection_wait_for_connected(),
which only waits for SMBDIRECT_SOCKET_CONNECTED or
an error.
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
These will be used in order to initialize struct smbdirect_socket
with rdma.cm_id being valid from the start in order to hold
a reference to the correct net namespace, this will allow
us to implement async connecting and accepting logic in
the next steps.
This comes with some related helper functions in
order to initialize the socket without the need
to access internals of struct smbdirect_socket:
smbdirect_socket_set_initial_parameters
smbdirect_socket_get_current_parameters
smbdirect_socket_set_kernel_settings
smbdirect_socket_set_custom_workqueue
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
This is basically a copy of the following functions in the server:
smb_direct_post_send => smbdirect_connection_post_send_wr
smb_direct_send_ctx_init => smbdirect_connection_send_batch_init
smb_direct_flush_send_list => smbdirect_connection_send_batch_flush
wait_for_send_lcredit => smbdirect_connection_wait_for_send_lcredit
wait_for_send_credits => smbdirect_connection_wait_for_send_credits
post_sendmsg => smbdirect_connection_post_send_io
smb_direct_post_send_data => smbdirect_connection_send_single_iter
smb_direct_send_iter => smbdirect_connection_send_iter
They will replace the server functions soon and will also be used in
the client as smbdirect_connection_send_single_iter() is very similar to
smbd_post_send_iter().
There's also a smbdirect_connection_send_wait_zero_pending()
helper that can be used together with
smbdirect_connection_send_single_iter() in the client
until it can use smbdirect_connection_send_iter(),
which can happen with David's refactoring.
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: David Howells <dhowells@redhat.com>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
This is basically a copy of recv_done() in client and server,
with the following additions:
- Only handling the SMBDIRECT_EXPECT_DATA_TRANSFER code path,
as we'll have separate functions for the negotiate messages.
- Using more helper variables
- Improved logging
- Add credits_requested == 0 error check
- Add data_offset not 8 bytes aligned error check
- Use disable_work(&sc->recv_io.posted.refill_work)
before smbdirect_connection_put_recv_io, when it
is followed by smbdirect_socket_schedule_cleanup()
This will be used on common between client and server in future
and replace the existing recv_done() functions.
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
This is basically contains the following functions copied from
the server: wait_for_rw_credits, calc_rw_credits, get_sg_list,
smb_direct_free_rdma_rw_msg, read_write_done, read_done,
write_done, smb_direct_rdma_xmit.
They got new names, some indentation/formatting changes,
some variable names are changed too.
They also only use struct smbdirect_socket instead of
struct smb_direct_transport.
But the logic is still the same. They will be used
by the server soon.
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
This is basically contains the following functions copied from
the client: destroy_mr_list, allocate_mr_list, register_mr_done,
smbd_mr_recovery_work, get_mr, smbd_iter_to_mr, smbd_register_mr
and smbd_deregister_mr.
They got new names, some indentation/formatting changes,
some variable names are changed too.
They also only use struct smbdirect_socket instead of
struct smbd_connection.
But the logic is still the same. They will be used
by the client soon.
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
This is basically a copy of smbd_post_send_credits() in the client
and smb_direct_post_recv_credits() in the server.
There are several improvements compared to the existing functions:
1. We calculate the number of missing posted buffers by getting the
difference between recv_io.credits.target and recv_io.posted.count.
Instead of the difference between recv_io.credits.target
and recv_io.credits.count, because recv_io.credits.count is
only updated once a message is send to the peer.
It was not really a problem before, because we have
a fixed number smbdirect_recv_io buffers, so the
loop terminated when smbdirect_connection_get_recv_io()
returns NULL.
But using recv_io.posted.count makes it easier to
understand.
2. In order to tell the peer about the newly posted buffer
and grant the credits, we only trigger the send immediate
when we're not granting only the last possible credit
(only one credit is missing to reach the desired target).
This is mostly a difference relative to the servers
smb_direct_post_recv_credits() implementation,
which should avoid useless ping pong messages.
Cc: Steve French <smfrench@gmail.com>
Cc: Tom Talpey <tom@talpey.com>
Cc: Long Li <longli@microsoft.com>
Cc: Namjae Jeon <linkinjeon@kernel.org>
Cc: linux-cifs@vger.kernel.org
Cc: samba-technical@lists.samba.org
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Acked-by: Namjae Jeon <linkinjeon@kernel.org>
Signed-off-by: Steve French <stfrench@microsoft.com>
