kernel/linux.git/drivers/infiniband/core, branch v6.19.11

RDMA/rw: Fall back to direct SGE on MR pool exhaustion

2026-04-02T11:25:32+00:00

[ Upstream commit 00da250c21b074ea9494c375d0117b69e5b1d0a4 ] When IOMMU passthrough mode is active, ib_dma_map_sgtable_attrs() produces no coalescing: each scatterlist page maps 1:1 to a DMA entry, so sgt.nents equals the raw page count. A 1 MB transfer yields 256 DMA entries. If that count exceeds the device's max_sgl_rd threshold (an optimization hint from mlx5 firmware), rdma_rw_io_needs_mr() steers the operation into the MR registration path. Each such operation consumes one or more MRs from a pool sized at max_rdma_ctxs -- roughly one MR per concurrent context. Under write-intensive workloads that issue many concurrent RDMA READs, the pool is rapidly exhausted, ib_mr_pool_get() returns NULL, and rdma_rw_init_one_mr() returns -EAGAIN. Upper layer protocols treat this as a fatal DMA mapping failure and tear down the connection. The max_sgl_rd check is a performance optimization, not a correctness requirement: the device can handle large SGE counts via direct posting, just less efficiently than with MR registration. When the MR pool cannot satisfy a request, falling back to the direct SGE (map_wrs) path avoids the connection reset while preserving the MR optimization for the common case where pool resources are available. Add a fallback in rdma_rw_ctx_init() so that -EAGAIN from rdma_rw_init_mr_wrs() triggers direct SGE posting instead of propagating the error. iWARP devices, which mandate MR registration for RDMA READs, and force_mr debug mode continue to treat -EAGAIN as terminal. Fixes: 00bd1439f464 ("RDMA/rw: Support threshold for registration vs scattering to local pages") Signed-off-by: Chuck Lever Reviewed-by: Christoph Hellwig Link: https://patch.msgid.link/20260313194201.5818-2-cel@kernel.org Signed-off-by: Leon Romanovsky Signed-off-by: Sasha Levin

RDMA/umem: Fix double dma_buf_unpin in failure path

2026-03-04T12:20:57+00:00

[ Upstream commit 104016eb671e19709721c1b0048dd912dc2e96be ] In ib_umem_dmabuf_get_pinned_with_dma_device(), the call to ib_umem_dmabuf_map_pages() can fail. If this occurs, the dmabuf is immediately unpinned but the umem_dmabuf->pinned flag is still set. Then, when ib_umem_release() is called, it calls ib_umem_dmabuf_revoke() which will call dma_buf_unpin() again. Fix this by removing the immediate unpin upon failure and just let the ib_umem_release/revoke path handle it. This also ensures the proper unmap-unpin unwind ordering if the dmabuf_map_pages call happened to fail due to dma_resv_wait_timeout (and therefore has a non-NULL umem_dmabuf->sgt). Fixes: 1e4df4a21c5a ("RDMA/umem: Allow pinned dmabuf umem usage") Signed-off-by: Jacob Moroni Link: https://patch.msgid.link/20260224234153.1207849-1-jmoroni@google.com Signed-off-by: Leon Romanovsky Signed-off-by: Sasha Levin

RDMA/core: Fix stale RoCE GIDs during netdev events at registration

2026-03-04T12:20:55+00:00

[ Upstream commit 9af0feae8016ba58ad7ff784a903404986b395b1 ] RoCE GID entries become stale when netdev properties change during the IB device registration window. This is reproducible with a udev rule that sets a MAC address when a VF netdev appears: ACTION=="add", SUBSYSTEM=="net", KERNEL=="eth4", \ RUN+="/sbin/ip link set eth4 address 88:22:33:44:55:66" After VF creation, show_gids displays GIDs derived from the original random MAC rather than the configured one. The root cause is a race between netdev event processing and device registration: CPU 0 (driver) CPU 1 (udev/workqueue) ────────────── ────────────────────── ib_register_device() ib_cache_setup_one() gid_table_setup_one() _gid_table_setup_one() ← GID table allocated rdma_roce_rescan_device() ← GIDs populated with OLD MAC ip link set eth4 addr NEW_MAC NETDEV_CHANGEADDR queued netdevice_event_work_handler() ib_enum_all_roce_netdevs() ← Iterates DEVICE_REGISTERED ← Device NOT marked yet, SKIP! enable_device_and_get() xa_set_mark(DEVICE_REGISTERED) ← Too late, event was lost The netdev event handler uses ib_enum_all_roce_netdevs() which only iterates devices marked DEVICE_REGISTERED. However, this mark is set late in the registration process, after the GID cache is already populated. Events arriving in this window are silently dropped. Fix this by introducing a new xarray mark DEVICE_GID_UPDATES that is set immediately after the GID table is allocated and initialized. Use the new mark in ib_enum_all_roce_netdevs() function to iterate devices instead of DEVICE_REGISTERED. This is safe because: - After _gid_table_setup_one(), all required structures exist (port_data, immutable, cache.gid) - The GID table mutex serializes concurrent access between the initial rescan and event handlers - Event handlers correctly update stale GIDs even when racing with rescan - The mark is cleared in ib_cache_cleanup_one() before teardown This also fixes similar races for IP address events (inetaddr_event, inet6addr_event) which use the same enumeration path. Fixes: 0df91bb67334 ("RDMA/devices: Use xarray to store the client_data") Signed-off-by: Jiri Pirko Link: https://patch.msgid.link/20260127093839.126291-1-jiri@resnulli.us Reported-by: syzbot+881d65229ca4f9ae8c84@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=881d65229ca4f9ae8c84 Signed-off-by: Leon Romanovsky Signed-off-by: Sasha Levin

RDMA/uverbs: Add __GFP_NOWARN to ib_uverbs_unmarshall_recv() kmalloc

2026-02-26T23:01:18+00:00

[ Upstream commit 58b604dfc7bb753f91bc0ccd3fa705e14e6edfb4 ] Since wqe_size in ib_uverbs_unmarshall_recv() is user-provided and already validated, but can still be large, add __GFP_NOWARN to suppress memory allocation warnings for large sizes, consistent with the similar fix in ib_uverbs_post_send(). Fixes: 67cdb40ca444 ("[IB] uverbs: Implement more commands") Signed-off-by: Yi Liu Link: https://patch.msgid.link/20260129094900.3517706-1-liuy22@mails.tsinghua.edu.cn Signed-off-by: Leon Romanovsky Signed-off-by: Sasha Levin

RDMA/core: add rdma_rw_max_sge() helper for SQ sizing

2026-02-26T23:01:18+00:00

[ Upstream commit afcae7d7b8a278a6c29e064f99e5bafd4ac1fb37 ] svc_rdma_accept() computes sc_sq_depth as the sum of rq_depth and the number of rdma_rw contexts (ctxts). This value is used to allocate the Send CQ and to initialize the sc_sq_avail credit pool. However, when the device uses memory registration for RDMA operations, rdma_rw_init_qp() inflates the QP's max_send_wr by a factor of three per context to account for REG and INV work requests. The Send CQ and credit pool remain sized for only one work request per context, causing Send Queue exhaustion under heavy NFS WRITE workloads. Introduce rdma_rw_max_sge() to compute the actual number of Send Queue entries required for a given number of rdma_rw contexts. Upper layer protocols call this helper before creating a Queue Pair so that their Send CQs and credit accounting match the QP's true capacity. Update svc_rdma_accept() to use rdma_rw_max_sge() when computing sc_sq_depth, ensuring the credit pool reflects the work requests that rdma_rw_init_qp() will reserve. Reviewed-by: Christoph Hellwig Fixes: 00bd1439f464 ("RDMA/rw: Support threshold for registration vs scattering to local pages") Signed-off-by: Chuck Lever Link: https://patch.msgid.link/20260128005400.25147-5-cel@kernel.org Signed-off-by: Leon Romanovsky Signed-off-by: Sasha Levin

RDMA/uverbs: Validate wqe_size before using it in ib_uverbs_post_send

2026-02-26T23:01:17+00:00

[ Upstream commit 1956f0a74ccf5dc9c3ef717f2985c3ed3400aab0 ] ib_uverbs_post_send() uses cmd.wqe_size from userspace without any validation before passing it to kmalloc() and using the allocated buffer as struct ib_uverbs_send_wr. If a user provides a small wqe_size value (e.g., 1), kmalloc() will succeed, but subsequent accesses to user_wr->opcode, user_wr->num_sge, and other fields will read beyond the allocated buffer, resulting in an out-of-bounds read from kernel heap memory. This could potentially leak sensitive kernel information to userspace. Additionally, providing an excessively large wqe_size can trigger a WARNING in the memory allocation path, as reported by syzkaller. This is inconsistent with ib_uverbs_unmarshall_recv() which properly validates that wqe_size >= sizeof(struct ib_uverbs_recv_wr) before proceeding. Add the same validation for ib_uverbs_post_send() to ensure wqe_size is at least sizeof(struct ib_uverbs_send_wr). Fixes: c3bea3d2dc53 ("RDMA/uverbs: Use the iterator for ib_uverbs_unmarshall_recv()") Signed-off-by: Yi Liu Link: https://patch.msgid.link/20260122142900.2356276-2-liuy22@mails.tsinghua.edu.cn Signed-off-by: Leon Romanovsky Signed-off-by: Sasha Levin

RDMA/iwcm: Fix workqueue list corruption by removing work_list

2026-02-26T23:01:16+00:00

[ Upstream commit 7874eeacfa42177565c01d5198726671acf7adf2 ] The commit e1168f0 ("RDMA/iwcm: Simplify cm_event_handler()") changed the work submission logic to unconditionally call queue_work() with the expectation that queue_work() would have no effect if work was already pending. The problem is that a free list of struct iwcm_work is used (for which struct work_struct is embedded), so each call to queue_work() is basically unique and therefore does indeed queue the work. This causes a problem in the work handler which walks the work_list until it's empty to process entries. This means that a single run of the work handler could process item N+1 and release it back to the free list while the actual workqueue entry is still queued. It could then get reused (INIT_WORK...) and lead to list corruption in the workqueue logic. Fix this by just removing the work_list. The workqueue already does this for us. This fixes the following error that was observed when stress testing with ucmatose on an Intel E830 in iWARP mode: [ 151.465780] list_del corruption. next->prev should be ffff9f0915c69c08, but was ffff9f0a1116be08. (next=ffff9f0a15b11c08) [ 151.466639] ------------[ cut here ]------------ [ 151.466986] kernel BUG at lib/list_debug.c:67! [ 151.467349] Oops: invalid opcode: 0000 [#1] SMP NOPTI [ 151.467753] CPU: 14 UID: 0 PID: 2306 Comm: kworker/u64:18 Not tainted 6.19.0-rc4+ #1 PREEMPT(voluntary) [ 151.468466] Hardware name: QEMU Ubuntu 24.04 PC (i440FX + PIIX, 1996), BIOS 1.16.3-debian-1.16.3-2 04/01/2014 [ 151.469192] Workqueue: 0x0 (iw_cm_wq) [ 151.469478] RIP: 0010:__list_del_entry_valid_or_report+0xf0/0x100 [ 151.469942] Code: c7 58 5f 4c b2 e8 10 50 aa ff 0f 0b 48 89 ef e8 36 57 cb ff 48 8b 55 08 48 89 e9 48 89 de 48 c7 c7 a8 5f 4c b2 e8 f0 4f aa ff <0f> 0b 66 2e 0f 1f 84 00 00 00 00 00 0f 1f 40 00 90 90 90 90 90 90 [ 151.471323] RSP: 0000:ffffb15644e7bd68 EFLAGS: 00010046 [ 151.471712] RAX: 000000000000006d RBX: ffff9f0915c69c08 RCX: 0000000000000027 [ 151.472243] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff9f0a37d9c600 [ 151.472768] RBP: ffff9f0a15b11c08 R08: 0000000000000000 R09: c0000000ffff7fff [ 151.473294] R10: 0000000000000001 R11: ffffb15644e7bba8 R12: ffff9f092339ee68 [ 151.473817] R13: ffff9f0900059c28 R14: ffff9f092339ee78 R15: 0000000000000000 [ 151.474344] FS: 0000000000000000(0000) GS:ffff9f0a847b5000(0000) knlGS:0000000000000000 [ 151.474934] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 151.475362] CR2: 0000559e233a9088 CR3: 000000020296b004 CR4: 0000000000770ef0 [ 151.475895] PKRU: 55555554 [ 151.476118] Call Trace: [ 151.476331] [ 151.476497] move_linked_works+0x49/0xa0 [ 151.476792] __pwq_activate_work.isra.46+0x2f/0xa0 [ 151.477151] pwq_dec_nr_in_flight+0x1e0/0x2f0 [ 151.477479] process_scheduled_works+0x1c8/0x410 [ 151.477823] worker_thread+0x125/0x260 [ 151.478108] ? __pfx_worker_thread+0x10/0x10 [ 151.478430] kthread+0xfe/0x240 [ 151.478671] ? __pfx_kthread+0x10/0x10 [ 151.478955] ? __pfx_kthread+0x10/0x10 [ 151.479240] ret_from_fork+0x208/0x270 [ 151.479523] ? __pfx_kthread+0x10/0x10 [ 151.479806] ret_from_fork_asm+0x1a/0x30 [ 151.480103] Fixes: e1168f09b331 ("RDMA/iwcm: Simplify cm_event_handler()") Signed-off-by: Jacob Moroni Link: https://patch.msgid.link/20260112020006.1352438-1-jmoroni@google.com Reviewed-by: Bart Van Assche Signed-off-by: Leon Romanovsky Signed-off-by: Sasha Levin

IB/cache: update gid cache on client reregister event

2026-02-26T23:01:14+00:00

[ Upstream commit ddd6c8c873e912cb1ead79def54de5e24ff71c80 ] Some HCAs (e.g: ConnectX4) do not trigger a IB_EVENT_GID_CHANGE on subnet prefix update from SM (PortInfo). Since the commit d58c23c92548 ("IB/core: Only update PKEY and GID caches on respective events"), the GID cache is updated exclusively on IB_EVENT_GID_CHANGE. If this event is not emitted, the subnet prefix in the IPoIB interface’s hardware address remains set to its default value (0xfe80000000000000). Then rdma_bind_addr() failed because it relies on hardware address to find the port GID (subnet_prefix + port GUID). This patch fixes this issue by updating the GID cache on IB_EVENT_CLIENT_REREGISTER event (emitted on PortInfo::ClientReregister=1). Fixes: d58c23c92548 ("IB/core: Only update PKEY and GID caches on respective events") Signed-off-by: Etienne AUJAMES Link: https://patch.msgid.link/aVUfsO58QIDn5bGX@eaujamesFR0130 Reviewed-by: Parav Pandit Signed-off-by: Leon Romanovsky Signed-off-by: Sasha Levin

RDMA/umad: Reject negative data_len in ib_umad_write

2026-02-26T23:00:39+00:00

commit 5551b02fdbfd85a325bb857f3a8f9c9f33397ed2 upstream. ib_umad_write computes data_len from user-controlled count and the MAD header sizes. With a mismatched user MAD header size and RMPP header length, data_len can become negative and reach ib_create_send_mad(). This can make the padding calculation exceed the segment size and trigger an out-of-bounds memset in alloc_send_rmpp_list(). Add an explicit check to reject negative data_len before creating the send buffer. KASAN splat: [ 211.363464] BUG: KASAN: slab-out-of-bounds in ib_create_send_mad+0xa01/0x11b0 [ 211.364077] Write of size 220 at addr ffff88800c3fa1f8 by task spray_thread/102 [ 211.365867] ib_create_send_mad+0xa01/0x11b0 [ 211.365887] ib_umad_write+0x853/0x1c80 Fixes: 2be8e3ee8efd ("IB/umad: Add P_Key index support") Signed-off-by: YunJe Shin Link: https://patch.msgid.link/20260203100628.1215408-1-ioerts@kookmin.ac.kr Signed-off-by: Leon Romanovsky Signed-off-by: Greg Kroah-Hartman

RDMA/core: always drop device refcount in ib_del_sub_device_and_put()

2025-12-21T11:45:17+00:00

Since nldev_deldev() (introduced by commit 060c642b2ab8 ("RDMA/nldev: Add support to add/delete a sub IB device through netlink") grabs a reference using ib_device_get_by_index() before calling ib_del_sub_device_and_put(), we need to drop that reference before returning -EOPNOTSUPP error. Reported-by: syzbot+881d65229ca4f9ae8c84@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=881d65229ca4f9ae8c84 Fixes: bca51197620a ("RDMA/core: Support IB sub device with type "SMI"") Signed-off-by: Tetsuo Handa Link: https://patch.msgid.link/80749a85-cbe2-460c-8451-42516013f9fa@I-love.SAKURA.ne.jp Reviewed-by: Parav Pandit Signed-off-by: Leon Romanovsky