summaryrefslogtreecommitdiffstats
path: root/net/rds
AgeCommit message (Collapse)AuthorFilesLines
2020-02-16net/rds: Track user mapped pages through special APILeon Romanovsky1-12/+12
Convert net/rds to use the newly introduces pin_user_pages() API, which properly sets FOLL_PIN. Setting FOLL_PIN is now required for code that requires tracking of pinned pages. Note that this effectively changes the code's behavior: it now ultimately calls set_page_dirty_lock(), instead of set_page_dirty(). This is probably more accurate. As Christoph Hellwig put it, "set_page_dirty() is only safe if we are dealing with a file backed page where we have reference on the inode it hangs off." [1] [1] https://lore.kernel.org/r/20190723153640.GB720@lst.de Cc: Hans Westgaard Ry <hans.westgaard.ry@oracle.com> Cc: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: Leon Romanovsky <leonro@mellanox.com> Signed-off-by: John Hubbard <jhubbard@nvidia.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2020-01-18net/rds: Use prefetch for On-Demand-Paging MRHans Westgaard Ry1-0/+9
Try prefetching pages when using On-Demand-Paging MR using ib_advise_mr. Signed-off-by: Hans Westgaard Ry <hans.westgaard.ry@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
2020-01-18net/rds: Handle ODP mr registration/unregistrationHans Westgaard Ry7-56/+244
On-Demand-Paging MRs are registered using ib_reg_user_mr and unregistered with ib_dereg_mr. Signed-off-by: Hans Westgaard Ry <hans.westgaard.ry@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
2020-01-16net/rds: Detect need of On-Demand-Paging memory registrationHans Westgaard Ry1-2/+4
Add code to check if memory intended for RDMA is FS-DAX-memory. RDS will fail with error code EOPNOTSUPP if FS-DAX-memory is detected. Signed-off-by: Hans Westgaard Ry <hans.westgaard.ry@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
2019-11-16Merge git://git.kernel.org/pub/scm/linux/kernel/git/netdev/netDavid S. Miller1-8/+15
Lots of overlapping changes and parallel additions, stuff like that. Signed-off-by: David S. Miller <davem@davemloft.net>
2019-11-16rds: ib: update WR sizes when bringing up connectionDag Moxnes1-8/+15
Currently WR sizes are updated from rds_ib_sysctl_max_send_wr and rds_ib_sysctl_max_recv_wr when a connection is shut down. As a result, a connection being down while rds_ib_sysctl_max_send_wr or rds_ib_sysctl_max_recv_wr are updated, will not update the sizes when it comes back up. Move resizing of WRs to rds_ib_setup_qp so that connections will be setup with the most current WR sizes. Signed-off-by: Dag Moxnes <dag.moxnes@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-10-17net/rds: Remove unnecessary null checkYueHaibing1-2/+1
Null check before dma_pool_destroy is redundant, so remove it. This is detected by coccinelle. Signed-off-by: YueHaibing <yuehaibing@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-10-06net/rds: Add missing include fileYueHaibing2-0/+2
Fix build error: net/rds/ib_cm.c: In function rds_dma_hdrs_alloc: net/rds/ib_cm.c:475:13: error: implicit declaration of function dma_pool_zalloc; did you mean mempool_alloc? [-Werror=implicit-function-declaration] hdrs[i] = dma_pool_zalloc(pool, GFP_KERNEL, &hdr_daddrs[i]); ^~~~~~~~~~~~~~~ mempool_alloc net/rds/ib.c: In function rds_ib_dev_free: net/rds/ib.c:111:3: error: implicit declaration of function dma_pool_destroy; did you mean mempool_destroy? [-Werror=implicit-function-declaration] dma_pool_destroy(rds_ibdev->rid_hdrs_pool); ^~~~~~~~~~~~~~~~ mempool_destroy Reported-by: Hulk Robot <hulkci@huawei.com> Fixes: 9b17f5884be4 ("net/rds: Use DMA memory pool allocation for rds_header") Signed-off-by: YueHaibing <yuehaibing@huawei.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-10-05Merge git://git.kernel.org/pub/scm/linux/kernel/git/netdev/netDavid S. Miller1-3/+3
2019-10-03net/rds: Use DMA memory pool allocation for rds_headerKa-Cheong Poon5-61/+153
Currently, RDS calls ib_dma_alloc_coherent() to allocate a large piece of contiguous DMA coherent memory to store struct rds_header for sending/receiving packets. The memory allocated is then partitioned into struct rds_header. This is not necessary and can be costly at times when memory is fragmented. Instead, RDS should use the DMA memory pool interface to handle this. The DMA addresses of the pre- allocated headers are stored in an array. At send/receive ring initialization and refill time, this arrary is de-referenced to get the DMA addresses. This array is not accessed at send/receive packet processing. Suggested-by: Håkon Bugge <haakon.bugge@oracle.com> Signed-off-by: Ka-Cheong Poon <ka-cheong.poon@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-10-02net/rds: Log vendor error if send/recv Work requests failSudhakar Dindukurti2-4/+5
Log vendor error if work requests fail. Vendor error provides more information that is used for debugging the issue. Signed-off-by: Sudhakar Dindukurti <sudhakar.dindukurti@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-10-02net/rds: Fix error handling in rds_ib_add_one()Dotan Barak1-3/+3
rds_ibdev:ipaddr_list and rds_ibdev:conn_list are initialized after allocation some resources such as protection domain. If allocation of such resources fail, then these uninitialized variables are accessed in rds_ib_dev_free() in failure path. This can potentially crash the system. The code has been updated to initialize these variables very early in the function. Signed-off-by: Dotan Barak <dotanb@dev.mellanox.co.il> Signed-off-by: Sudhakar Dindukurti <sudhakar.dindukurti@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-27net/rds: Check laddr_check before calling itKa-Cheong Poon1-1/+4
In rds_bind(), laddr_check is called without checking if it is NULL or not. And rs_transport should be reset if rds_add_bound() fails. Fixes: c5c1a030a7db ("net/rds: An rds_sock is added too early to the hash table") Reported-by: syzbot+fae39afd2101a17ec624@syzkaller.appspotmail.com Signed-off-by: Ka-Cheong Poon <ka-cheong.poon@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-26net: Fix Kconfig indentationKrzysztof Kozlowski1-2/+2
Adjust indentation from spaces to tab (+optional two spaces) as in coding style with command like: $ sed -e 's/^ /\t/' -i */Kconfig Signed-off-by: Krzysztof Kozlowski <krzk@kernel.org> Acked-by: Sven Eckelmann <sven@narfation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-17Merge ra.kernel.org:/pub/scm/linux/kernel/git/netdev/netDavid S. Miller1-1/+1
Pull in bug fixes from 'net' tree for the merge window. Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15net/rds: Fix 'ib_evt_handler_call' element in 'rds_ib_stat_names'Gerd Rausch1-1/+1
All entries in 'rds_ib_stat_names' are stringified versions of the corresponding "struct rds_ib_statistics" element without the "s_"-prefix. Fix entry 'ib_evt_handler_call' to do the same. Fixes: f4f943c958a2 ("RDS: IB: ack more receive completions to improve performance") Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-15Merge git://git.kernel.org/pub/scm/linux/kernel/git/netdev/netDavid S. Miller1-22/+18
Minor overlapping changes in the btusb and ixgbe drivers. Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-11net/rds: An rds_sock is added too early to the hash tableKa-Cheong Poon1-22/+18
In rds_bind(), an rds_sock is added to the RDS bind hash table before rs_transport is set. This means that the socket can be found by the receive code path when rs_transport is NULL. And the receive code path de-references rs_transport for congestion update check. This can cause a panic. An rds_sock should not be added to the bind hash table before all the needed fields are set. Reported-by: syzbot+4b4f8163c2e246df3c4c@syzkaller.appspotmail.com Signed-off-by: Ka-Cheong Poon <ka-cheong.poon@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-05Convert usage of IN_MULTICAST to ipv4_is_multicastDave Taht3-6/+6
IN_MULTICAST's primary intent is as a uapi macro. Elsewhere in the kernel we use ipv4_is_multicast consistently. This patch unifies linux's multicast checks to use that function rather than this macro. Signed-off-by: Dave Taht <dave.taht@gmail.com> Reviewed-by: Toke Høiland-Jørgensen <toke@toke.dk> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-09-02Merge git://git.kernel.org/pub/scm/linux/kernel/git/netdev/netDavid S. Miller1-1/+4
r8152 conflicts are the NAPI fixes in 'net' overlapping with some tasklet stuff in net-next Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-27net/rds: Fix info leak in rds6_inc_info_copy()Ka-Cheong Poon1-1/+4
The rds6_inc_info_copy() function has a couple struct members which are leaking stack information. The ->tos field should hold actual information and the ->flags field needs to be zeroed out. Fixes: 3eb450367d08 ("rds: add type of service(tos) infrastructure") Fixes: b7ff8b1036f0 ("rds: Extend RDS API for IPv6 support") Reported-by: 黄ID蝴蝶 <butterflyhuangxx@gmail.com> Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Ka-Cheong Poon <ka-cheong.poon@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-27Merge git://git.kernel.org/pub/scm/linux/kernel/git/netdev/netDavid S. Miller4-8/+22
Minor conflict in r8169, bug fix had two versions in net and net-next, take the net-next hunks. Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-24net: rds: add service level support in rds-infoZhu Yanjun4-8/+22
>From IB specific 7.6.5 SERVICE LEVEL, Service Level (SL) is used to identify different flows within an IBA subnet. It is carried in the local route header of the packet. Before this commit, run "rds-info -I". The outputs are as below: " RDS IB Connections: LocalAddr RemoteAddr Tos SL LocalDev RemoteDev 192.2.95.3 192.2.95.1 2 0 fe80::21:28:1a:39 fe80::21:28:10:b9 192.2.95.3 192.2.95.1 1 0 fe80::21:28:1a:39 fe80::21:28:10:b9 192.2.95.3 192.2.95.1 0 0 fe80::21:28:1a:39 fe80::21:28:10:b9 " After this commit, the output is as below: " RDS IB Connections: LocalAddr RemoteAddr Tos SL LocalDev RemoteDev 192.2.95.3 192.2.95.1 2 2 fe80::21:28:1a:39 fe80::21:28:10:b9 192.2.95.3 192.2.95.1 1 1 fe80::21:28:1a:39 fe80::21:28:10:b9 192.2.95.3 192.2.95.1 0 0 fe80::21:28:1a:39 fe80::21:28:10:b9 " The commit fe3475af3bdf ("net: rds: add per rds connection cache statistics") adds cache_allocs in struct rds_info_rdma_connection as below: struct rds_info_rdma_connection { ... __u32 rdma_mr_max; __u32 rdma_mr_size; __u8 tos; __u32 cache_allocs; }; The peer struct in rds-tools of struct rds_info_rdma_connection is as below: struct rds_info_rdma_connection { ... uint32_t rdma_mr_max; uint32_t rdma_mr_size; uint8_t tos; uint8_t sl; uint32_t cache_allocs; }; The difference between userspace and kernel is the member variable sl. In the kernel struct, the member variable sl is missing. This will introduce risks. So it is necessary to use this commit to avoid this risk. Fixes: fe3475af3bdf ("net: rds: add per rds connection cache statistics") CC: Joe Jin <joe.jin@oracle.com> CC: JUNXIAO_BI <junxiao.bi@oracle.com> Suggested-by: Gerd Rausch <gerd.rausch@oracle.com> Signed-off-by: Zhu Yanjun <yanjun.zhu@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-23net/rds: Whitelist rdma_cookie and rx_tstamp for usercopyDag Moxnes3-15/+27
Add the RDMA cookie and RX timestamp to the usercopy whitelist. After the introduction of hardened usercopy whitelisting (https://lwn.net/Articles/727322/), a warning is displayed when the RDMA cookie or RX timestamp is copied to userspace: kernel: WARNING: CPU: 3 PID: 5750 at mm/usercopy.c:81 usercopy_warn+0x8e/0xa6 [...] kernel: Call Trace: kernel: __check_heap_object+0xb8/0x11b kernel: __check_object_size+0xe3/0x1bc kernel: put_cmsg+0x95/0x115 kernel: rds_recvmsg+0x43d/0x620 [rds] kernel: sock_recvmsg+0x43/0x4a kernel: ___sys_recvmsg+0xda/0x1e6 kernel: ? __handle_mm_fault+0xcae/0xf79 kernel: __sys_recvmsg+0x51/0x8a kernel: SyS_recvmsg+0x12/0x1c kernel: do_syscall_64+0x79/0x1ae When the whitelisting feature was introduced, the memory for the RDMA cookie and RX timestamp in RDS was not added to the whitelist, causing the warning above. Signed-off-by: Dag Moxnes <dag.moxnes@oracle.com> Tested-by: Jenny <jenny.x.xu@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-15rds: check for excessive looping in rds_send_xmitAndy Grover3-1/+14
Original commit from 2011 updated to include a change by Yuval Shaia <yuval.shaia@oracle.com> that adds a new statistic counter "send_stuck_rm" to capture the messages looping exessively in the send path. Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-15net/rds: Add a few missing rds_stat_names entriesGerd Rausch1-0/+2
In a previous commit, fields were added to "struct rds_statistics" but array "rds_stat_names" was not updated accordingly. Please note the inconsistent naming of the string representations that is done in the name of compatibility with the Oracle internal code-base. s_recv_bytes_added_to_socket -> "recv_bytes_added_to_sock" s_recv_bytes_removed_from_socket -> "recv_bytes_freed_fromsock" Fixes: 192a798f5299 ("RDS: add stat for socket recv memory usage") Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-15RDS: don't use GFP_ATOMIC for sk_alloc in rds_createChris Mason1-1/+1
Signed-off-by: Chris Mason <chris.mason@oracle.com> Signed-off-by: Bang Nguyen <bang.nguyen@oracle.com> Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Signed-off-by: Somasundaram Krishnasamy <somasundaram.krishnasamy@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-15RDS: limit the number of times we loop in rds_send_xmitChris Mason1-1/+11
This will kick the RDS worker thread if we have been looping too long. Original commit from 2012 updated to include a change by Venkat Venkatsubra <venkat.x.venkatsubra@oracle.com> that triggers "must_wake" if "rds_ib_recv_refill_one" fails. Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-08-15net/rds: Add RDS6_INFO_SOCKETS and RDS6_INFO_RECV_MESSAGES optionsKa-Cheong Poon1-3/+90
Add support of the socket options RDS6_INFO_SOCKETS and RDS6_INFO_RECV_MESSAGES which update the RDS_INFO_SOCKETS and RDS_INFO_RECV_MESSAGES options respectively. The old options work for IPv4 sockets only. Signed-off-by: Ka-Cheong Poon <ka-cheong.poon@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-07-27net: rds: Fix possible null-pointer dereferences in ↵Jia-Ju Bai1-1/+4
rds_rdma_cm_event_handler_cmn() In rds_rdma_cm_event_handler_cmn(), there are some if statements to check whether conn is NULL, such as on lines 65, 96 and 112. But conn is not checked before being used on line 108: trans->cm_connect_complete(conn, event); and on lines 140-143: rdsdebug("DISCONNECT event - dropping connection " "%pI6c->%pI6c\n", &conn->c_laddr, &conn->c_faddr); rds_conn_drop(conn); Thus, possible null-pointer dereferences may occur. To fix these bugs, conn is checked before being used. These bugs are found by a static analysis tool STCheck written by us. Signed-off-by: Jia-Ju Bai <baijiaju1990@gmail.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-07-19Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/netLinus Torvalds5-49/+109
Pull networking fixes from David Miller: 1) Fix AF_XDP cq entry leak, from Ilya Maximets. 2) Fix handling of PHY power-down on RTL8411B, from Heiner Kallweit. 3) Add some new PCI IDs to iwlwifi, from Ihab Zhaika. 4) Fix handling of neigh timers wrt. entries added by userspace, from Lorenzo Bianconi. 5) Various cases of missing of_node_put(), from Nishka Dasgupta. 6) The new NET_ACT_CT needs to depend upon NF_NAT, from Yue Haibing. 7) Various RDS layer fixes, from Gerd Rausch. 8) Fix some more fallout from TCQ_F_CAN_BYPASS generalization, from Cong Wang. 9) Fix FIB source validation checks over loopback, also from Cong Wang. 10) Use promisc for unsupported number of filters, from Justin Chen. 11) Missing sibling route unlink on failure in ipv6, from Ido Schimmel. * git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (90 commits) tcp: fix tcp_set_congestion_control() use from bpf hook ag71xx: fix return value check in ag71xx_probe() ag71xx: fix error return code in ag71xx_probe() usb: qmi_wwan: add D-Link DWM-222 A2 device ID bnxt_en: Fix VNIC accounting when enabling aRFS on 57500 chips. net: dsa: sja1105: Fix missing unlock on error in sk_buff() gve: replace kfree with kvfree selftests/bpf: fix test_xdp_noinline on s390 selftests/bpf: fix "valid read map access into a read-only array 1" on s390 net/mlx5: Replace kfree with kvfree MAINTAINERS: update netsec driver ipv6: Unlink sibling route in case of failure liquidio: Replace vmalloc + memset with vzalloc udp: Fix typo in net/ipv4/udp.c net: bcmgenet: use promisc for unsupported filters ipv6: rt6_check should return NULL if 'from' is NULL tipc: initialize 'validated' field of received packets selftests: add a test case for rp_filter fib: relax source validation check for loopback packets mlxsw: spectrum: Do not process learned records with a dummy FID ...
2019-07-17net/rds: Initialize ic->i_fastreg_wrs upon allocationGerd Rausch1-1/+1
Otherwise, if an IB connection is torn down before "rds_ib_setup_qp" is called, the value of "ic->i_fastreg_wrs" is still at zero (as it wasn't initialized by "rds_ib_setup_qp"). Consequently "rds_ib_conn_path_shutdown" will spin forever, waiting for it to go back to "RDS_IB_DEFAULT_FR_WR", which of course will never happen as there are no outstanding work requests. Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-07-17net/rds: Keep track of and wait for FRWR segments in use upon shutdownGerd Rausch3-6/+45
Since "rds_ib_free_frmr" and "rds_ib_free_frmr_list" simply put the FRMR memory segments on the "drop_list" or "free_list", and it is the job of "rds_ib_flush_mr_pool" to reap those entries by ultimately issuing a "IB_WR_LOCAL_INV" work-request, we need to trigger and then wait for all those memory segments attached to a particular connection to be fully released before we can move on to release the QP, CQ, etc. So we make "rds_ib_conn_path_shutdown" wait for one more atomic_t called "i_fastreg_inuse_count" that keeps track of how many FRWR memory segments are out there marked "FRMR_IS_INUSE" (and also wake_up rds_ib_ring_empty_wait, as they go away). Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-07-17net/rds: Set fr_state only to FRMR_IS_FREE if IB_WR_LOCAL_INV had been ↵Gerd Rausch1-1/+2
successful Fix a bug where fr_state first goes to FRMR_IS_STALE, because of a failure of operation IB_WR_LOCAL_INV, but then gets set back to "FRMR_IS_FREE" uncoditionally, even though the operation failed. Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-07-17net/rds: Fix NULL/ERR_PTR inconsistencyGerd Rausch1-2/+2
Make function "rds_ib_try_reuse_ibmr" return NULL in case memory region could not be allocated, since callers simply check if the return value is not NULL. Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-07-17net/rds: Wait for the FRMR_IS_FREE (or FRMR_IS_STALE) transition after ↵Gerd Rausch2-27/+40
posting IB_WR_LOCAL_INV In order to: 1) avoid a silly bouncing between "clean_list" and "drop_list" triggered by function "rds_ib_reg_frmr" as it is releases frmr regions whose state is not "FRMR_IS_FREE" right away. 2) prevent an invalid access error in a race from a pending "IB_WR_LOCAL_INV" operation with a teardown ("dma_unmap_sg", "put_page") and de-registration ("ib_dereg_mr") of the corresponding memory region. Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-07-17net/rds: Get rid of "wait_clean_list_grace" and add lockingGerd Rausch2-38/+19
Waiting for activity on the "clean_list" to quiesce is no substitute for proper locking. We can have multiple threads competing for "llist_del_first" via "rds_ib_reuse_mr", and a single thread competing for "llist_del_all" and "llist_del_first" via "rds_ib_flush_mr_pool". Since "llist_del_first" depends on "list->first->next" not to change in the midst of the operation, simply waiting for all current calls to "rds_ib_reuse_mr" to quiesce across all CPUs is woefully inadequate: By the time "wait_clean_list_grace" is done iterating over all CPUs to see that there is no concurrent caller to "rds_ib_reuse_mr", a new caller may have just shown up on the first CPU. Furthermore, <linux/llist.h> explicitly calls out the need for locking: * Cases where locking is needed: * If we have multiple consumers with llist_del_first used in one consumer, * and llist_del_first or llist_del_all used in other consumers, * then a lock is needed. Also, while at it, drop the unused "pool" parameter from "list_to_llist_nodes". Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-07-17net/rds: Give fr_state a chance to transition to FRMR_IS_FREEGerd Rausch2-1/+27
In the context of FRMR (ib_frmr.c): Memory regions make it onto the "clean_list" via "rds_ib_flush_mr_pool", after the memory region has been posted for invalidation via "rds_ib_post_inv". At that point in time, "fr_state" may still be in state "FRMR_IS_INUSE", since the only place where "fr_state" transitions to "FRMR_IS_FREE" is in "rds_ib_mr_cqe_handler", which is triggered by a tasklet. So in case we notice that "fr_state != FRMR_IS_FREE" (see below), we wait for "fr_inv_done" to trigger with a maximum of 10msec. Then we check again, and only put the memory region onto the drop_list (via "rds_ib_free_frmr") in case the situation remains unchanged. This avoids the problem of memory-regions bouncing between "clean_list" and "drop_list" before they even have a chance to be properly invalidated. Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-07-15Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdmaLinus Torvalds1-4/+4
Pull rdma updates from Jason Gunthorpe: "A smaller cycle this time. Notably we see another new driver, 'Soft iWarp', and the deletion of an ancient unused driver for nes. - Revise and simplify the signature offload RDMA MR APIs - More progress on hoisting object allocation boiler plate code out of the drivers - Driver bug fixes and revisions for hns, hfi1, efa, cxgb4, qib, i40iw - Tree wide cleanups: struct_size, put_user_page, xarray, rst doc conversion - Removal of obsolete ib_ucm chardev and nes driver - netlink based discovery of chardevs and autoloading of the modules providing them - Move more of the rdamvt/hfi1 uapi to include/uapi/rdma - New driver 'siw' for software based iWarp running on top of netdev, much like rxe's software RoCE. - mlx5 feature to report events in their raw devx format to userspace - Expose per-object counters through rdma tool - Adaptive interrupt moderation for RDMA (DIM), sharing the DIM core from netdev" * tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (194 commits) RMDA/siw: Require a 64 bit arch RDMA/siw: Mark expected switch fall-throughs RDMA/core: Fix -Wunused-const-variable warnings rdma/siw: Remove set but not used variable 's' rdma/siw: Add missing dependencies on LIBCRC32C and DMA_VIRT_OPS RDMA/siw: Add missing rtnl_lock around access to ifa rdma/siw: Use proper enumerated type in map_cqe_status RDMA/siw: Remove unnecessary kthread create/destroy printouts IB/rdmavt: Fix variable shadowing issue in rvt_create_cq RDMA/core: Fix race when resolving IP address RDMA/core: Make rdma_counter.h compile stand alone IB/core: Work on the caller socket net namespace in nldev_newlink() RDMA/rxe: Fill in wc byte_len with IB_WC_RECV_RDMA_WITH_IMM RDMA/mlx5: Set RDMA DIM to be enabled by default RDMA/nldev: Added configuration of RDMA dynamic interrupt moderation to netlink RDMA/core: Provide RDMA DIM support for ULPs linux/dim: Implement RDMA adaptive moderation (DIM) IB/mlx5: Report correctly tag matching rendezvous capability docs: infiniband: add it to the driver-api bookset IB/mlx5: Implement VHCA tunnel mechanism in DEVX ...
2019-07-09rds: avoid version downgrade to legitimate newer peer connectionsSantosh Shilimkar1-2/+4
Connections with legitimate tos values can get into usual connection race. It can result in consumer reject. We don't want tos value or protocol version to be demoted for such connections otherwise piers would end up different tos values which can results in no connection. Example a peer initiated connection with say tos 8 while usual connection racing can get downgraded to tos 0 which is not desirable. Patch fixes above issue introduced by commit commit d021fabf525f ("rds: rdma: add consumer reject") Reported-by: Yanjun Zhu <yanjun.zhu@oracle.com> Tested-by: Yanjun Zhu <yanjun.zhu@oracle.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
2019-07-09rds: Return proper "tos" value to user-spaceGerd Rausch1-0/+1
The proper "tos" value needs to be returned to user-space (sockopt RDS_INFO_CONNECTIONS). Fixes: 3eb450367d08 ("rds: add type of service(tos) infrastructure") Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Reviewed-by: Zhu Yanjun <yanjun.zhu@oracle.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
2019-07-09rds: Accept peer connection reject messages due to incompatible versionGerd Rausch1-2/+3
Prior to commit d021fabf525ff ("rds: rdma: add consumer reject") function "rds_rdma_cm_event_handler_cmn" would always honor a rejected connection attempt by issuing a "rds_conn_drop". The commit mentioned above added a "break", eliminating the "fallthrough" case and made the "rds_conn_drop" rather conditional: Now it only happens if a "consumer defined" reject (i.e. "rdma_reject") carries an integer-value of "1" inside "private_data": if (!conn) break; err = (int *)rdma_consumer_reject_data(cm_id, event, &len); if (!err || (err && ((*err) == RDS_RDMA_REJ_INCOMPAT))) { pr_warn("RDS/RDMA: conn <%pI6c, %pI6c> rejected, dropping connection\n", &conn->c_laddr, &conn->c_faddr); conn->c_proposed_version = RDS_PROTOCOL_COMPAT_VERSION; rds_conn_drop(conn); } rdsdebug("Connection rejected: %s\n", rdma_reject_msg(cm_id, event->status)); break; /* FALLTHROUGH */ A number of issues are worth mentioning here: #1) Previous versions of the RDS code simply rejected a connection by calling "rdma_reject(cm_id, NULL, 0);" So the value of the payload in "private_data" will not be "1", but "0". #2) Now the code has become dependent on host byte order and sizing. If one peer is big-endian, the other is little-endian, or there's a difference in sizeof(int) (e.g. ILP64 vs LP64), the *err check does not work as intended. #3) There is no check for "len" to see if the data behind *err is even valid. Luckily, it appears that the "rdma_reject(cm_id, NULL, 0)" will always carry 148 bytes of zeroized payload. But that should probably not be relied upon here. #4) With the added "break;", we might as well drop the misleading "/* FALLTHROUGH */" comment. This commit does _not_ address issue #2, as the sender would have to agree on a byte order as well. Here is the sequence of messages in this observed error-scenario: Host-A is pre-QoS changes (excluding the commit mentioned above) Host-B is post-QoS changes (including the commit mentioned above) #1 Host-B issues a connection request via function "rds_conn_path_transition" connection state transitions to "RDS_CONN_CONNECTING" #2 Host-A rejects the incompatible connection request (from #1) It does so by calling "rdma_reject(cm_id, NULL, 0);" #3 Host-B receives an "RDMA_CM_EVENT_REJECTED" event (from #2) But since the code is changed in the way described above, it won't drop the connection here, simply because "*err == 0". #4 Host-A issues a connection request #5 Host-B receives an "RDMA_CM_EVENT_CONNECT_REQUEST" event and ends up calling "rds_ib_cm_handle_connect". But since the state is already in "RDS_CONN_CONNECTING" (as of #1) it will end up issuing a "rdma_reject" without dropping the connection: if (rds_conn_state(conn) == RDS_CONN_CONNECTING) { /* Wait and see - our connect may still be succeeding */ rds_ib_stats_inc(s_ib_connect_raced); } goto out; #6 Host-A receives an "RDMA_CM_EVENT_REJECTED" event (from #5), drops the connection and tries again (goto #4) until it gives up. Tested-by: Zhu Yanjun <yanjun.zhu@oracle.com> Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
2019-07-09Revert "RDS: IB: split the mr registration and invalidation path"Gerd Rausch3-16/+8
This reverts commit 56012459310a1dbcc55c2dbf5500a9f7571402cb. RDS kept spinning inside function "rds_ib_post_reg_frmr", waiting for "i_fastreg_wrs" to become incremented: while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) { atomic_inc(&ibmr->ic->i_fastreg_wrs); cpu_relax(); } Looking at the original commit: commit 56012459310a ("RDS: IB: split the mr registration and invalidation path") In there, the "rds_ib_mr_cqe_handler" was changed in the following way: void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc) if (frmr->fr_inv) { frmr->fr_state = FRMR_IS_FREE; frmr->fr_inv = false; atomic_inc(&ic->i_fastreg_wrs); } else { atomic_inc(&ic->i_fastunreg_wrs); } It looks like it's got it exactly backwards: Function "rds_ib_post_reg_frmr" keeps track of the outstanding requests via "i_fastreg_wrs". Function "rds_ib_post_inv" keeps track of the outstanding requests via "i_fastunreg_wrs" (post original commit). It also sets: frmr->fr_inv = true; However the completion handler "rds_ib_mr_cqe_handler" adjusts "i_fastreg_wrs" when "fr_inv" had been true, and adjusts "i_fastunreg_wrs" otherwise. The original commit was done in the name of performance: to remove the performance bottleneck No performance benefit could be observed with a fixed-up version of the original commit measured between two Oracle X7 servers, both equipped with Mellanox Connect-X5 HCAs. The prudent course of action is to revert this commit. Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
2019-07-09rds: fix reordering with composite message notificationSantosh Shilimkar4-30/+14
RDS composite message(rdma + control) user notification needs to be triggered once the full message is delivered and such a fix was added as part of commit 941f8d55f6d61 ("RDS: RDMA: Fix the composite message user notification"). But rds_send_remove_from_sock is missing data part notify check and hence at times the user don't get notification which isn't desirable. One way is to fix the rds_send_remove_from_sock to check of that case but considering the ordering complexity with completion handler and rdma + control messages are always dispatched back to back in same send context, just delaying the signaled completion on rmda work request also gets the desired behaviour. i.e Notifying application only after RDMA + control message send completes. So patch updates the earlier fix with this approach. The delay signaling completions of rdma op till the control message send completes fix was done by Venkat Venkatsubra in downstream kernel. Reviewed-and-tested-by: Zhu Yanjun <yanjun.zhu@oracle.com> Reviewed-by: Gerd Rausch <gerd.rausch@oracle.com> Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
2019-06-28Merge tag 'v5.2-rc6' into rdma.git for-nextJason Gunthorpe4-5/+11
For dependencies in next patches. Resolve conflicts: - Use uverbs_get_cleared_udata() with new cq allocation flow - Continue to delete nes despite SPDX conflict - Resolve list appends in mlx5_command_str() - Use u16 for vport_rule stuff - Resolve list appends in struct ib_client Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-06-07Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/netDavid S. Miller3-5/+10
Some ISDN files that got removed in net-next had some changes done in mainline, take the removals. Signed-off-by: David S. Miller <davem@davemloft.net>
2019-06-06net: rds: fix memory leak in rds_ib_flush_mr_poolZhu Yanjun1-4/+6
When the following tests last for several hours, the problem will occur. Server: rds-stress -r 1.1.1.16 -D 1M Client: rds-stress -r 1.1.1.14 -s 1.1.1.16 -D 1M -T 30 The following will occur. " Starting up.... tsks tx/s rx/s tx+rx K/s mbi K/s mbo K/s tx us/c rtt us cpu % 1 0 0 0.00 0.00 0.00 0.00 0.00 -1.00 1 0 0 0.00 0.00 0.00 0.00 0.00 -1.00 1 0 0 0.00 0.00 0.00 0.00 0.00 -1.00 1 0 0 0.00 0.00 0.00 0.00 0.00 -1.00 " >From vmcore, we can find that clean_list is NULL. >From the source code, rds_mr_flushd calls rds_ib_mr_pool_flush_worker. Then rds_ib_mr_pool_flush_worker calls " rds_ib_flush_mr_pool(pool, 0, NULL); " Then in function " int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **ibmr_ret) " ibmr_ret is NULL. In the source code, " ... list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail); if (ibmr_ret) *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode); /* more than one entry in llist nodes */ if (clean_nodes->next) llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list); ... " When ibmr_ret is NULL, llist_entry is not executed. clean_nodes->next instead of clean_nodes is added in clean_list. So clean_nodes is discarded. It can not be used again. The workqueue is executed periodically. So more and more clean_nodes are discarded. Finally the clean_list is NULL. Then this problem will occur. Fixes: 1bc144b62524 ("net, rds, Replace xlist in net/rds/xlist.h with llist") Signed-off-by: Zhu Yanjun <yanjun.zhu@oracle.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-06-05net: rds: fix memory leak when unload rds_rdmaZhu Yanjun2-1/+4
When KASAN is enabled, after several rds connections are created, then "rmmod rds_rdma" is run. The following will appear. " BUG rds_ib_incoming (Not tainted): Objects remaining in rds_ib_incoming on __kmem_cache_shutdown() Call Trace: dump_stack+0x71/0xab slab_err+0xad/0xd0 __kmem_cache_shutdown+0x17d/0x370 shutdown_cache+0x17/0x130 kmem_cache_destroy+0x1df/0x210 rds_ib_recv_exit+0x11/0x20 [rds_rdma] rds_ib_exit+0x7a/0x90 [rds_rdma] __x64_sys_delete_module+0x224/0x2c0 ? __ia32_sys_delete_module+0x2c0/0x2c0 do_syscall_64+0x73/0x190 entry_SYSCALL_64_after_hwframe+0x44/0xa9 " This is rds connection memory leak. The root cause is: When "rmmod rds_rdma" is run, rds_ib_remove_one will call rds_ib_dev_shutdown to drop the rds connections. rds_ib_dev_shutdown will call rds_conn_drop to drop rds connections as below. " rds_conn_path_drop(&conn->c_path[0], false); " In the above, destroy is set to false. void rds_conn_path_drop(struct rds_conn_path *cp, bool destroy) { atomic_set(&cp->cp_state, RDS_CONN_ERROR); rcu_read_lock(); if (!destroy && rds_destroy_pending(cp->cp_conn)) { rcu_read_unlock(); return; } queue_work(rds_wq, &cp->cp_down_w); rcu_read_unlock(); } In the above function, destroy is set to false. rds_destroy_pending is called. This does not move rds connections to ib_nodev_conns. So destroy is set to true to move rds connections to ib_nodev_conns. In rds_ib_unregister_client, flush_workqueue is called to make rds_wq finsh shutdown rds connections. The function rds_ib_destroy_nodev_conns is called to shutdown rds connections finally. Then rds_ib_recv_exit is called to destroy slab. void rds_ib_recv_exit(void) { kmem_cache_destroy(rds_ib_incoming_slab); kmem_cache_destroy(rds_ib_frag_slab); } The above slab memory leak will not occur again. >From tests, 256 rds connections [root@ca-dev14 ~]# time rmmod rds_rdma real 0m16.522s user 0m0.000s sys 0m8.152s 512 rds connections [root@ca-dev14 ~]# time rmmod rds_rdma real 0m32.054s user 0m0.000s sys 0m15.568s To rmmod rds_rdma with 256 rds connections, about 16 seconds are needed. And with 512 rds connections, about 32 seconds are needed. >From ftrace, when one rds connection is destroyed, " 19) | rds_conn_destroy [rds]() { 19) 7.782 us | rds_conn_path_drop [rds](); 15) | rds_shutdown_worker [rds]() { 15) | rds_conn_shutdown [rds]() { 15) 1.651 us | rds_send_path_reset [rds](); 15) 7.195 us | } 15) + 11.434 us | } 19) 2.285 us | rds_cong_remove_conn [rds](); 19) * 24062.76 us | } " So if many rds connections will be destroyed, this function rds_ib_destroy_nodev_conns uses most of time. Suggested-by: Håkon Bugge <haakon.bugge@oracle.com> Signed-off-by: Zhu Yanjun <yanjun.zhu@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-06-05net: rds: add per rds connection cache statisticsZhu Yanjun1-0/+2
The variable cache_allocs is to indicate how many frags (KiB) are in one rds connection frag cache. The command "rds-info -Iv" will output the rds connection cache statistics as below: " RDS IB Connections: LocalAddr RemoteAddr Tos SL LocalDev RemoteDev 1.1.1.14 1.1.1.14 58 255 fe80::2:c903:a:7a31 fe80::2:c903:a:7a31 send_wr=256, recv_wr=1024, send_sge=8, rdma_mr_max=4096, rdma_mr_size=257, cache_allocs=12 " This means that there are about 12KiB frag in this rds connection frag cache. Since rds.h in rds-tools is not related with the kernel rds.h, the change in kernel rds.h does not affect rds-tools. rds-info in rds-tools 2.0.5 and 2.0.6 is tested with this commit. It works well. Signed-off-by: Zhu Yanjun <yanjun.zhu@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-05-21rds: Don't check return value from destroy CQLeon Romanovsky1-4/+4
There is no value in checking ib_destroy_cq() result and skipping to clear struct ic fields. This connection needs to be reinitialized anyway. Signed-off-by: Leon Romanovsky <leonro@mellanox.com> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>