When rdma_resolve_addr() is invoked with a broadcast destination on an
IPoIB interface, is_dst_local() inspects the resolved route and
incorrectly concludes that the address is local. As a result, the
resolution fails with -ENODEV.
The issue stems from using '&' to compare rt_type with RTN_LOCAL. The
RTN_* values form a sequential enum, not a bitmask (RTN_LOCAL=2,
RTN_BROADCAST=3). Thus, "rt_type & RTN_LOCAL" yields a non-zero result
for a broadcast route as well.
Replace '&' with '==' when comparing rt_type against RTN_LOCAL.
Link: https://patch.msgid.link/r/20260609-fix-rdma-resolve-addr-v1-1-449b8b4e6c09@nvidia.com
Cc: stable@vger.kernel.org
Fixes: c31e4038c9 ("RDMA/core: Use route entry flag to decide on loopback traffic")
Signed-off-by: Maher Sanalla <msanalla@nvidia.com>
Reviewed-by: Vlad Dumitrescu <vdumitrescu@nvidia.com>
Signed-off-by: Edward Srouji <edwards@nvidia.com>
Reviewed-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
bnxt_re_debugfs_add_pdev() allocates per-file private data for the CC
configuration debugfs entries. The loop that initializes those entries
uses rdev->cc_config_params immediately, so allocation failure would lead
to NULL pointer dereferences while setting up debugfs.
Debugfs is best-effort. If the CC configuration private data cannot be
allocated just stop.
Link: https://patch.msgid.link/r/20260606040644.13-1-ruoyuw560@gmail.com
Signed-off-by: Ruoyu Wang <ruoyuw560@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Implement the query port speed callback to report the port effective
bandwidth directly in 100 Mb/s granularity.
Link: https://patch.msgid.link/r/20260608083927.4116-1-tomsela@amazon.com
Reviewed-by: Michael Margolin <mrgolin@amazon.com>
Reviewed-by: Yonatan Nachum <ynachum@amazon.com>
Signed-off-by: Tom Sela <tomsela@amazon.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Add support for reporting 800 Gbps as 8X NDR and 1600 Gbps as 8X XDR
link speeds.
Link: https://patch.msgid.link/r/20260608083736.48454-1-tomsela@amazon.com
Reviewed-by: Michael Margolin <mrgolin@amazon.com>
Reviewed-by: Yonatan Nachum <ynachum@amazon.com>
Signed-off-by: Tom Sela <tomsela@amazon.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Sashiko noticed mlx4 was using whatever random access flags were provided
when IB_MR_REREG_ACCESS is not used. Since IB_MR_REREG_TRANS needs
access_flags it used the random ones which means it doesn't work sensibly
if userspace provides only IB_MR_REREG_TRANS.
Keep track of the current access_flag of the MR and use it if the user
does not specify one.
Also fixup a little confusion around mmr.access, it is the HW access flags
so the convert_access() was missing. But nothing reads this by the time
rereg_mr can happen.
Fixes: 9376932d0c ("IB/mlx4_ib: Add support for user MR re-registration")
Link: https://patch.msgid.link/r/0-v1-29ca7a402625+ddd6-mlx4_rereg_flags_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
usr_len is read from a network-supplied message field (le16_to_cpu)
and used to compute data_len = off - usr_len without validating that
usr_len <= off. A malicious RDMA client can send usr_len > off causing
an integer underflow, resulting in data_len wrapping to a huge size_t
value which is then passed to the rdma_ev callback as a memory length,
leading to out-of-bounds memory access.
Fix by reading and validating usr_len <= off before rtrs_srv_get_ops_ids()
in both process_read() and process_write(), ensuring the early return
path acquires no reference and has no resource leak.
Link: https://patch.msgid.link/r/20260608134802.5019-1-aurelien@hackers.camp
Reported-by: Aurelien DESBRIERES <aurelien@hackers.camp>
Reviewed-by: Md Haris Iqbal <haris.iqbal@ionos.com>
Signed-off-by: Aurelien DESBRIERES <aurelien@hackers.camp>
Assisted-by: Claude <claude-sonnet-4-6>
Acked-by: Md Haris Iqbal <haris.iqbal@ionos.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Remove the mlx5_mr_pdn() in pagefault_real_mr() by pushing the pdn up, all
the callers use 0 since they don't pass MLX5_PF_FLAGS_ENABLE except the
ioctl reg_mr path which can use the ioctl pd.
Link: https://patch.msgid.link/r/9-v1-29ebd2c229b5+fd5-ib_mr_pd_jgg@nvidia.com
Assisted-by: Codex:gpt-5-5
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Keep pushing the pdn higher to remove more places touching mr->pd:
- XLT combinations that don't use PDN can just pass 0
- Use local pd values instead of mr->pd
- Implicit MR does not have inplace rereg, so the mr->pd is safe
Link: https://patch.msgid.link/r/8-v1-29ebd2c229b5+fd5-ib_mr_pd_jgg@nvidia.com
Assisted-by: Codex:gpt-5-5
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The rereg protocol expects the core code to change mr->pd and synchronize
that change with the atomics and syncs. The driver should not touch it.
mlx5 needed to update it in umr_rereg_pas() because
mlx5r_umr_update_mr_pas() required the updated mr->pd to build the
UMR.
Simply switch mlx5r_umr_update_mr_pas() to use the pdn directly from
the new pd and remove the mr->pd update.
Fixes: 56e11d628c ("IB/mlx5: Added support for re-registration of MRs")
Link: https://patch.msgid.link/r/7-v1-29ebd2c229b5+fd5-ib_mr_pd_jgg@nvidia.com
Assisted-by: Codex:gpt-5-5
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Instead of getting the pdn deep inside the umr code, pass it in from the
top. to_mpd(mr->ibmr.pd)->pdn is not safe due to the rereg races, so all
the call sites need some revision to obtain the pdn in a safe way.
Mark them with mlx5_mr_pdn(); following patches will go through and remove
these.
Cases where the XLT flags are known and do not require the PDN can pass 0,
such as for mlx5_ib_dmabuf_invalidate_cb().
Also extract the DMABUF data_direct special case from inside the UMR code
and into the only place that needs it, pagefault_dmabuf_mr(). The actual
mr was created directly without using the UMR flow. Ultimately this will
be moved into mlx5_ib_init_dmabuf_mr().
Link: https://patch.msgid.link/r/6-v1-29ebd2c229b5+fd5-ib_mr_pd_jgg@nvidia.com
Assisted-by: Codex:gpt-5-5
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The HW only processes mkc fields selected by mkey_mask.
pd, qpn and mkey_7_0 are never selected so they can be left as zero.
This removes a racy read of mr->pd.
Fixes: e73242aa14 ("RDMA/mlx5: Optimize DMABUF mkey page size")
Link: https://patch.msgid.link/r/5-v1-29ebd2c229b5+fd5-ib_mr_pd_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Sashiko points out that, due to rereg_mr, the PD is actually variable and
all the touches in nldev are racy.
Use mr->device instead of mr->pd->device.
Getting the PD restrack ID is more tricky. To avoid disturbing all the
happy paths, add an rdma_restrack_sync() operation which is sort of like
flush_workqueue() or synchronize_irq(): after it returns, all the old
nldev touches to the mr are gone and everything sees the new PD. This
makes it safe to reach into the PD pointer.
Fixes: da5c850782 ("RDMA/nldev: add driver-specific resource tracking")
Link: https://patch.msgid.link/r/4-v1-29ebd2c229b5+fd5-ib_mr_pd_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Due to all the child mkeys in the implicit ODP configuration we cannot
change anything in place for the parent mkey. Instead the whole thing
needs to be rebuilt if any change is requested. If the user does not
specify a translation then force the implicit values which will then fall
through the logic into mlx5_ib_reg_user_mr() to allocate a completely new
MR.
Since implicit children were also touching the mr->pd, this removes
another case where the access was racy.
Fixes: ef3642c4f5 ("RDMA/mlx5: Fix error unwinds for rereg_mr")
Link: https://sashiko.dev/#/patchset/20260427-security-bug-fixes-v3-0-4621fa52de0e%40nvidia.com?part=4
Link: https://patch.msgid.link/r/3-v1-29ebd2c229b5+fd5-ib_mr_pd_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
DMABUF generally relies on the ODP EQ mechanism to safely implement the
move semantics. ODP requires a device-global one time startup of the ODP
machinery when the first MR is created, and this was missed on the DMABUF
path.
Call mlx5r_odp_create_eq() when creating a ODP'able DMABUF.
The core code prevents using IB_ACCESS_ON_DEMAND unless the driver
advertises IB_ODP_SUPPORT, so until now, mlx5r_odp_create_eq() cannot be
called unless the device has ODP support.
However, DMABUF has no such protection and a second bug was allowing
DMABUFs to be created on non-ODP capable HW. Add a guard at the start of
mlx5r_odp_create_eq(). This is necessary here anyhow as the
dev->odp_eq_mutex is not initialized without IB_ODP_SUPPORT.
Link: https://patch.msgid.link/r/2-v1-29ebd2c229b5+fd5-ib_mr_pd_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Jumping to mlx5_ib_reg_user_mr() without IB_MR_REREG_TRANS set will use
garbage values for start, length, and iova. Recovering the original mr
parameters for ODP and DMABUF to properly recreate it is too hard in this
flow, so just fail it.
Fixes: ef3642c4f5 ("RDMA/mlx5: Fix error unwinds for rereg_mr")
Link: https://patch.msgid.link/r/1-v1-29ebd2c229b5+fd5-ib_mr_pd_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Disassociating a socket from an endpoint via siw_socket_disassoc() may
release the last reference on that endpoint and free it. Therefore, don't
clear the endpoints socket pointer after calling that function, but
within.
This fixes a:
BUG: KASAN: slab-use-after-free in siw_cm_work_handler (drivers/infiniband/sw/siw/siw_cm.c:1053 drivers/infiniband/sw/siw/siw_cm.c:1075)
which occurred after processing a malformed MPA request during connection
establishment, causing the new endpoint to be closed.
Fixes: 6c52fdc244 ("rdma/siw: connection management")
Link: https://patch.msgid.link/r/20260604160808.30948-1-bernard.metzler@linux.dev
Reported-by: Shuangpeng Bai <shuangpeng.kernel@gmail.com>
Signed-off-by: Bernard Metzler <bernard.metzler@linux.dev>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Initialize iwmr->access during initial user mem registration so
that it contains a valid value during a subsequent rereg_mr.
Otherwise, a rereg_mr that doesn't set IB_MR_REREG_ACCESS (for
example, one that only changes the PD) ends up clearing the
access flags in HW since iwmr->access is zero-initialized, which
is not intended.
Fixes: 5ac388db27 ("RDMA/irdma: Add support to re-register a memory region")
Link: https://patch.msgid.link/r/20260604154104.4035581-1-jmoroni@google.com
Signed-off-by: Jacob Moroni <jmoroni@google.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Sashiko pointed out an unrelated bug during a previous patch:
https://sashiko.dev/#/patchset/20260512183852.614045-1-jmoroni%40google.com
This change fixes the bug by eliminating the cqmr->split field which
was not being set properly and instead just checks the CQ resize
feature flag directly.
The cqmr->split field essentially tracks whether IRDMA_FEATURE_CQ_RESIZE
is set, but it was not being set until CQ creation time, which is _after_
CQ memory registration (the only other place where it is referenced).
As a result, it would always be false during MR registration and would
therefore cause irdma_handle_q_mem to populate cqmr->shadow even for GEN_2
HW and beyond:
cqmr->shadow = (dma_addr_t)arr[req->cq_pages];
The issue is that for GEN_2 and beyond, req->cq_pages may be exactly equal
to iwmr->page_cnt and therefore equal to the size of arr, which would cause
an OOB read by one.
Fixes: b48c24c2d7 ("RDMA/irdma: Implement device supported verb APIs")
Link: https://patch.msgid.link/r/20260602214423.1315105-2-jmoroni@google.com
Signed-off-by: Jacob Moroni <jmoroni@google.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The driver has the following invariants:
1. legacy_mode is only allowed on GEN_1 hardware (enforced
in irdma_alloc_ucontext).
2. GEN_1 hardware does not set IRDMA_FEATURE_CQ_RESIZE or
IRDMA_FEATURE_RTS_AE. These feature flags are only set
for GEN_2 and GEN_3 hardware.
Therefore, legacy_mode is always false if IRDMA_FEATURE_CQ_RESIZE
or IRDMA_FEATURE_RTS_AE is set, so remove the redundant checks.
Link: https://patch.msgid.link/r/20260602214423.1315105-1-jmoroni@google.com
Signed-off-by: Jacob Moroni <jmoroni@google.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In drivers/infiniband/sw/siw/siw_qp_rx.c, siw_proc_rresp() places each
inbound Read Response DDP segment at sge->laddr + wqe->processed and then
accumulates wqe->processed, but it never checks the running total against
the sink buffer length on continuation segments. siw_check_sge() resolves
and validates the sink memory only on the first fragment (the if (!*mem)
branch), and siw_rresp_check_ntoh() compares the cumulative length against
wqe->bytes only on the final segment (the !frx->more_ddp_segs guard).
A connected siw peer that answers an outstanding RREAD with Read Response
segments that keep the DDP Last flag clear, carrying more total payload
than the RREAD requested, drives wqe->processed past the validated sink
buffer; the next siw_rx_data() call writes out of bounds at
sge->laddr + wqe->processed. siw runs iWARP over ordinary routable TCP,
so the peer is the remote end of an established RDMA connection and needs
no local privilege.
Bound every segment before placement, exactly as siw_proc_send() and
siw_proc_write() already do for their tagged and untagged paths, and
terminate the connection with a base-or-bounds DDP error when the
Read Response would overrun the sink buffer.
This is the second receive-path length fix for this file. A separate
change rejects an MPA FPDU length that underflows the per-fragment
remainder in the header decode; that guard does not cover this case,
because here each individual segment length is self-consistent and only
the accumulated placement offset overruns the buffer.
Fixes: 8b6a361b8c ("rdma/siw: receive path")
Link: https://patch.msgid.link/r/20260602194700.2273758-1-michael.bommarito@gmail.com
Cc: stable@vger.kernel.org
Assisted-by: Claude:claude-opus-4-8
Signed-off-by: Michael Bommarito <michael.bommarito@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In mlx5_ib_enable_lb(), dev->lb.enabled was unconditionally set
to true even if mlx5_nic_vport_update_local_lb() failed.
Fix this by only setting dev->lb.enabled on success. On failure,
roll back the reference counters and return the error.
Link: https://patch.msgid.link/r/20260601095818.2227-1-lirongqing@baidu.com
Signed-off-by: Li RongQing <lirongqing@baidu.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
__mlx5_ib_add() currently returns -ENOMEM on any stage initialization
failure, losing the actual error code returned by the init function.
This makes it impossible for callers to distinguish between different
failure reasons (e.g. -EINVAL, -EIO, -EOPNOTSUPP) and leads to
misleading error handling.
Fix it by returning the actual error code stored in 'err'.
Link: https://patch.msgid.link/r/20260601095654.2178-1-lirongqing@baidu.com
Signed-off-by: Li RongQing <lirongqing@baidu.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Codex pointed out that cm_init_av_by_path() can call cm_set_av_port()
which takes a reference on the cm device, but then can immediately return
error if ib_init_ah_attr_from_path() fails.
Since callers like ib_send_cm_req() put the av on the stack this leaks
that cm device reference.
Re-order cm_init_av_by_path() so it doesn't touch the av until it has done
all its failable work, and then update the av in one shot so it is either
left alone or fully init'd.
Sashiko also pointed out that the cm_destroy_av() prior to
cm_init_av_by_path() is harmful as it leaves the AV broken in the error
case and thus the REJ won't send. Since cm_init_av_by_path() is now atomic
it is safe to delete the cm_destroy_av(). On succees the av from
cm_init_av_for_response() is cleaned up by cm_init_av_by_path(), on
failure the 'goto rejected' guarentees the av is destroyed during
ib_destroy_cm_id().
Fixes: 76039ac909 ("IB/cm: Protect cm_dev, cm_ports and mad_agent with kref and lock")
Link: https://patch.msgid.link/r/0-v1-38292501f539+14f-ib_cm_av_leak_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
clang warns about a function missing a printf attribute:
include/rdma/rdma_vt.h:457:47: error: diagnostic behavior may be improved by adding the 'format(printf, 2, 3)' attribute to the declaration of 'rvt_set_ibdev_name' [-Werror,-Wmissing-format-attribute]
447 | static inline void rvt_set_ibdev_name(struct rvt_dev_info *rdi,
| __attribute__((format(printf, 2, 3)))
448 | const char *fmt, const char *name,
449 | const int unit)
The helper was originally added as an abstraction for the hfi1 and
qib drivers needing the same thing, but now qib is gone, and hfi1
is the only remaining user of rdma_vt.
Avoid the warning and allow the compiler to check the format string by
open-coding the helper and directly assigning the device name.
Fixes: 5084c8ff21 ("IB/{rdmavt, hfi1, qib}: Self determine driver name")
Link: https://patch.msgid.link/r/20260602140453.3542427-1-arnd@kernel.org
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Kees Cook <kees@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Sashiko points out the roundup_pow_of_two() only uses unsigned long but
dma_addr_t can be u64.
Change this algorithm to be simpler, compute the page size, if any page
size is found and it results in a single block then it is contiguous.
Link: https://patch.msgid.link/r/3-v1-88303e9e509f+f7-ib_umem_types_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Several corner cases, especially important on 32 bits:
- umem->iova is u64, the function argument should pass in u64 or
iova will be truncated
- Check that the length is not too large for the iova
- Check that lengths > 4G don't overflow the GENMASK
Link: https://patch.msgid.link/r/2-v1-88303e9e509f+f7-ib_umem_types_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This op hasn't followed the normal pattern of passing NULL for udata when
invoked by the kernel. Instead the kernel caller creates a dummy ib_udata
on the stack and passes that in. It does not seem to currently be a bug,
but this flow should be modernized to use the new API flow and in the
process accept NULL as well.
Only mlx4 uses an input request structure, have every other driver call
ib_is_udata_in_empty() to enforce the lack of request structs.
Use ib_respond_empty_udata() in every driver that does not use a response
struct.
Ensure a check for NULL udata before calling ib_respond_udata() in
bnxt_re, efa, and mlx5.
Make mlx4 safe to be called with NULL.
Link: https://patch.msgid.link/r/2-v1-922fa8e828ba+f7-ib_udata_stack_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Sashiko points out the udata for destruction has to be created using
uverbs_get_cleared_udata(). Move it to ib_core_uverbs.c so that the core
qp code can call it. Rework the call chain to pass the struct
uverbs_attr_bundle right up to the driver op callback.
Fixes a possible wild stack reference in drivers during error unwinding,
mlx5 can call rdma_udata_to_drv_context() from destroy_qp() when
destroying a QP.
Fixes: 00a79d6b99 ("RDMA/core: Configure selinux QP during creation")
Link: https://patch.msgid.link/r/1-v1-922fa8e828ba+f7-ib_udata_stack_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
When we generate completion for SQ the opcode while being properly read
from ring buffer is ignored when written back to completion. Seems
to be a simple typo.
Link: https://patch.msgid.link/r/ahjB87k54bYdFbft@grain
Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>
Reviewed-by: Jacob Moroni <jmoroni@google.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
setup_root_hem() reads the first entry of head->root and checks
the returned pointer against NULL:
root_hem = list_first_entry(&head->root,
struct hns_roce_hem_item, list);
if (!root_hem)
return -ENOMEM;
list_first_entry() never returns NULL. On an empty list it returns
container_of(head, ..., list), a non-NULL garbage pointer that
aliases the head. So the check is dead.
The only caller adds an entry to head.root right before invoking
setup_root_hem():
list_add(&root_hem->list, &head.root);
ret = setup_root_hem(..., &head, ...);
So head.root is guaranteed non-empty on entry. Drop the check.
Link: https://patch.msgid.link/r/20260526054653.2054800-1-maoyixie.tju@gmail.com
Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Maoyi Xie <maoyixie.tju@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
For non-SRQ QPs, the responder reads WQE fields directly from the
shared queue buffer mapped into userspace. This allows a malicious
user to modify fields like num_sge or sge entries while the kernel
is processing the WQE, leading to out-of-bounds reads in
rxe_resp_check_length() and copy_data().
Introduce get_recv_wqe() that validates num_sge and copies the WQE
to a kernel-local buffer before processing, matching the approach
already used for SRQ WQEs in get_srq_wqe(). The srq_wqe buffer is
reused since SRQ and non-SRQ paths are mutually exclusive per QP.
Fixes: 8700e3e7c4 ("Soft RoCE driver")
Link: https://patch.msgid.link/r/20260518215040.1598586-3-tristan@talencesecurity.com
Signed-off-by: Tristan Madani <tristan@talencesecurity.com>
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
get_srq_wqe() reads wqe->dma.num_sge from the shared receive queue
buffer, which is mapped into userspace. It validates num_sge against
max_sge, but then re-reads the same field to calculate the memcpy
size. A concurrent userspace thread can modify num_sge between
validation and use, causing a heap buffer overflow when copying the
WQE into qp->resp.srq_wqe.
Read num_sge into a local variable and use it for both the bounds
check and the size calculation.
Fixes: 8700e3e7c4 ("Soft RoCE driver")
Link: https://patch.msgid.link/r/20260518215040.1598586-2-tristan@talencesecurity.com
Signed-off-by: Tristan Madani <tristan@talencesecurity.com>
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
When a device requires DMA bounce buffering inside a Confidential
Computing guest, __ib_umem_get_va() cannot work. The DMA mapping layer
redirects all mappings through swiotlb bounce buffers, so the device
receives DMA addresses pointing to bounce buffer memory rather than the
user's pages. Since RDMA devices access registered memory directly without
CPU involvement, there is no opportunity for swiotlb to synchronize
between the bounce buffer and the original pages.
The registration would already fail later on, since the umem mapping is
requested with DMA_ATTR_REQUIRE_COHERENT and gets rejected under
is_swiotlb_force_bounce() with -EIO. Fail early with -EOPNOTSUPP instead,
so the user gets a specific error code to react to.
Link: https://patch.msgid.link/r/20260517141311.2409230-3-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In CoCo guests, guest memory is encrypted and untrusted (T=0) devices
cannot DMA to it directly; such transfers must go through unencrypted
bounce buffers. RDMA registers user pages for direct device access,
bypassing the DMA layer and thus any bouncing, so registered memory does
not work in this configuration.
Until trusted (T=1) device detection is available, conservatively flag
every device attached to a CoCo guest. Expose the condition to userspace
as IB_UVERBS_DEVICE_CC_DMA_BOUNCE in device_cap_flags_ex so applications
can avoid memory registration and fall back to copying buffers through
send/recv.
Link: https://patch.msgid.link/r/20260517141311.2409230-2-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Add an optional mlx5 driver-namespace UMEM attribute on QP
create so userspace can supply the doorbell record umem
explicitly, symmetric to the CQ side. Resolve it inside
mlx5_ib_db_map_user() and use it as a private DBR page when
present; otherwise take the existing UHW share-or-pin path
that preserves per-page DBR sharing across CQ/QP/SRQ in the
same process.
Add mlx5's first UVERBS_OBJECT_QP UAPI definition chain to
attach the new attr.
Link: https://patch.msgid.link/r/20260529134312.2836341-17-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Add an optional mlx5 driver-namespace UMEM attribute on CQ
create so userspace can supply the doorbell record buffer
explicitly. mlx5_ib_db_map_user() resolves the attribute (or
falls back to the legacy UHW VA) into a struct
ib_uverbs_buffer_desc and runs a unified lookup-then-pin:
VA-typed descriptors share a per-page umem across CQ/QP/SRQ
in the same process, FD-typed descriptors are pinned per call.
Link: https://patch.msgid.link/r/20260529134312.2836341-16-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
ib_umem_is_contiguous() is defined under #ifdef
CONFIG_INFINIBAND_USER_MEM, but the #else branch lacks a stub.
Add the missing inline to fix potential broken build.
Fixes: c897c2c8b8 ("RDMA/core: Add umem "is_contiguous" and "start_dma_addr" helpers")
Link: https://patch.msgid.link/r/20260529134312.2836341-15-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Use the per-attribute UMEM helpers to pin QP buffer umems on
demand. The QP-type predicate selects between the BUF and RQ_BUF
attrs; raw-packet SQ uses its own dedicated SQ_BUF attr.
Link: https://patch.msgid.link/r/20260529134312.2836341-14-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Apply the per-attribute UMEM model to the QP create method. Add
three optional UMEM attributes that drivers pick from based on
how their user ABI lays out the QP rings:
- CREATE_QP_BUF_UMEM is a single user buffer that backs both
the SQ and RQ of one QP. This is the common case where
userspace pins one contiguous WQE region for the QP.
- CREATE_QP_SQ_BUF_UMEM and CREATE_QP_RQ_BUF_UMEM are a pair
of user buffers backing the SQ and RQ independently, used
when the two rings live in physically distinct user
allocations and must be pinned and addressed separately.
Existing drivers would map their current umems as follows:
- mlx5: BUF for normal QPs (one ucmd->buf_addr covers SQ+RQ);
for IB_QPT_RAW_PACKET and IB_QP_CREATE_SOURCE_QPN, the RQ
side comes from ucmd->buf_addr (RQ-sized) via RQ_BUF and
the SQ from ucmd->sq_buf_addr via SQ_BUF.
- mlx4: BUF, single ucmd.buf_addr covering SQ+RQ.
- hns: BUF, single ucmd.buf_addr covering SQ + ext-SGE + RQ.
- erdma: BUF, single ureq.qbuf_va sliced by the kernel into
SQ at offset 0 and RQ at rq_offset.
- bnxt_re: SQ_BUF (ureq->qpsva) + RQ_BUF (ureq->qprva, the
RQ side is skipped when the QP uses an SRQ).
- vmw_pvrdma: SQ_BUF (sbuf_addr) + RQ_BUF (rbuf_addr, the RQ
side is skipped when the QP uses an SRQ).
- qedr: SQ_BUF (sq_addr) + RQ_BUF (rq_addr) for whichever
side the QP type actually has (no SQ for XRC_TGT/GSI; no
RQ for XRC_INI/XRC_TGT/SRQ).
- ionic: SQ_BUF (req.sq.addr) + RQ_BUF (req.rq.addr); both
are skipped when the rings are placed in CMB instead of
host memory.
- mana: raw-packet QP uses SQ_BUF (sq_buf_addr) only; the RC
path uses multiple per-queue user buffers (ucmd.queue_buf[])
that do not fit the SQ/RQ pair semantics of these attrs and
stays on the legacy UHW path.
- efa, irdma, hfi1, ocrdma, mthca, cxgb4 and usnic do not pin
a QP WQE buffer via umem; none of these attributes apply.
Link: https://patch.msgid.link/r/20260529134312.2836341-13-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Now that all drivers use helper to get umem and manage the lifetime,
legacy umem field in struct ib_cq is no longer needed. Remove it
along with ib_umem_get_cq_tmp() helper that populated it and both
error and destroy paths.
Link: https://patch.msgid.link/r/20260529134312.2836341-12-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Pin the user CQ buffer with ib_umem_get_cq_buf() and take
ownership of the umem in the driver; fall back to
ib_umem_get_va() for the legacy UHW VA path. Apply the same
ownership pattern to the resize path.
Link: https://patch.msgid.link/r/20260529134312.2836341-11-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Pin the user CQ buffer with ib_umem_get_cq_buf_or_va() and take
ownership of the umem in the driver. Apply the same ownership
pattern to the resize path.
Link: https://patch.msgid.link/r/20260529134312.2836341-10-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Pin the user CQ buffer with ib_umem_get_cq_buf_or_va() and take
ownership of the umem in the driver. Apply the same ownership
pattern to the resize path.
Link: https://patch.msgid.link/r/20260529134312.2836341-9-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Pin the user CQ buffer with ib_umem_get_cq_buf() and take
ownership of the umem in the driver. Fall back to the
existing kernel-DMA path on NULL.
Link: https://patch.msgid.link/r/20260529134312.2836341-8-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Add UVERBS_ATTR_CREATE_CQ_BUF_UMEM and two driver-facing
wrappers, ib_umem_get_cq_buf() and ib_umem_get_cq_buf_or_va(),
that pin a CQ buffer umem from it. The wrappers reuse the
existing legacy CQ buffer-attr filler.
Link: https://patch.msgid.link/r/20260529134312.2836341-7-jiri@resnulli.us
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
