RDMA: Stability and race condition fixes
Subsystems (2)
v3 cover RDMA: Stability and race condition fixes
This series addresses several stability issues in RDMA core and the mlx5 driver.
Patches 1-2 fix xarray race conditions in the mlx5 SRQ and DCT destroy paths where a concurrent create can reuse the same firmware object number right after firmware releases it, causing the destroy path to incorrectly erase the newly created entry.
The remaining patches are independent fixes.
---
Changes in v3:
- Dropped restrack destroy-ordering fixes (patches 1-6 in v2) to
rework them in a dedicated series based on code-review feedback
- Rebased onto latest for-next
- Link to v2: https://lore.kernel.org/r/[email protected]
Changes in v2:
- Added patch "RDMA/mlx5: Remove raw RSS QP restrack tracking" to
also suppress broken tracking for raw RSS QPs, which suffer from
the same silent failures as DCTs
- Link to v1: https://lore.kernel.org/r/[email protected]
---
Edward Srouji (2):
RDMA/mlx5: Fix UAF in SRQ destroy due to race with create RDMA/mlx5: Fix UAF in DCT destroy due to race with create
Maher Sanalla (1):
IB/core: Fix IPv6 netlink message size in ib_nl_ip_send_msg()
Michael Guralnik (2):
RDMA/core: Fix rereg_mr use-after-free race RDMA/mlx5: Fix null-ptr-deref in Raw Packet QP creation
drivers/infiniband/core/addr.c | 2 +- drivers/infiniband/core/uverbs_cmd.c | 9 +++++++-- drivers/infiniband/hw/mlx5/qp.c | 5 +++++ drivers/infiniband/hw/mlx5/qpc.c | 9 ++++++++- drivers/infiniband/hw/mlx5/srq_cmd.c | 9 ++++++++- 5 files changed, 29 insertions(+), 5 deletions(-)
---
base-commit: 9091e3b59f2bef11c0a841096327565ae0ca220b
change-id: 20260325-security-bug-fixes-6fdef22d9412
Best regards,
Signature · 1 line
Edward Srouji <[email protected]>
PATCH v3 1/5 RDMA/mlx5: Fix UAF in SRQ destroy due to race with create
A race condition exists between mlx5_cmd_destroy_srq() and mlx5_cmd_create_srq() that can lead to a use-after-free (UAF) [1].
After destroy_srq_split() releases the SRQ to firmware, the SRQN can be immediately reallocated for a new SRQ being created concurrently. If the create path stores the new SRQ in the xarray before the destroy path erases it, the destroy will incorrectly delete the new SRQ's entry. Later accesses then hit freed memory.
Fix by replacing the unconditional xa_erase_irq() with xa_cmpxchg_irq() that only erases the entry if it hasn't already been replaced (still contains XA_ZERO_ENTRY), preserving any newly created SRQ.
[1] RIP: 0010:mlx5_cmd_destroy_srq+0xd8/0x110 [mlx5_ib]
Code: 89 e1 ba 06 04 00 00 4c 89 f6 48 89 ef e8 80 19 70 e1 c6 83 a0 0f 00 00 00 fb 5b 44 89 e8 5d 41 5c 41 5d 41 5e c3 cc cc cc cc <0f> 0b 48 89 c2 83 e2 03 48 83 fa 02 75 08 48 3d 05 c0 ff ff 77 08 RSP: 0018:ff110001037b7d08 EFLAGS: 00010286 RAX: 0000000000000000 RBX: ff1100010bb9c000 RCX: 0000000000000000 RDX: 0000000000000000 RSI: 0000000000000000 RDI: ff110001037b7c90 RBP: ff1100010bb9cfa0 R08: 0000000000000000 R09: 0000000000000000 R10: ff110001037b7da0 R11: ff11000104f29580 R12: ff1100010e2ac090 R13: 000000000000000d R14: 0000000000000001 R15: ff11000105336300 FS: 00007fa24787c740(0000) GS:ff1100046eb8d000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fa247984e90 CR3: 0000000109d59005 CR4: 0000000000373eb0 Call Trace: <TASK> mlx5_ib_destroy_srq+0x25/0xa0 [mlx5_ib] ib_destroy_srq_user+0x21/0x90 [ib_core] uverbs_free_srq+0x1b/0x50 [ib_uverbs] destroy_hw_idr_uobject+0x1e/0x50 [ib_uverbs] uverbs_destroy_uobject+0x35/0x180 [ib_uverbs] __uverbs_cleanup_ufile+0xdd/0x140 [ib_uverbs] uverbs_destroy_ufile_hw+0x38/0xf0 [ib_uverbs] ib_uverbs_close+0x17/0xa0 [ib_uverbs] __fput+0xe0/0x2a0 __x64_sys_close+0x3a/0x80 do_syscall_64+0x55/0xac0 entry_SYSCALL_64_after_hwframe+0x76/0x7e RIP: 0033:0x7fa247984ea4 Code: 00 f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 80 3d a5 51 0e 00 00 74 13 b8 03 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 3c c3 0f 1f 00 55 48 89 e5 48 83 ec 10 89 7d RSP: 002b:00007ffecfa79498 EFLAGS: 00000202 ORIG_RAX: 0000000000000003 RAX: ffffffffffffffda RBX: 0000200000000080 RCX: 00007fa247984ea4 RDX: 0000000000000040 RSI: 0000200000000200 RDI: 0000000000000003 RBP: 00007ffecfa794e0 R08: 00007ffecfa794e0 R09: 00007ffecfa794e0 R10: 0000000000000000 R11: 0000000000000202 R12: 0000000000000001 R13: 0000000000000000 R14: 0000200000000000 R15: 0000200000000009 </TASK> ---[ end trace 0000000000000000 ]---
--- drivers/infiniband/hw/mlx5/srq_cmd.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
drivers/infiniband/hw/mlx5/srq_cmd.c +8 -1
@@ -683,7 +683,14 @@ int mlx5_cmd_destroy_srq(struct mlx5_ib_dev *dev, struct mlx5_core_srq *srq)xa_cmpxchg_irq (&table->array, srq->srqn,XA_ZERO_ENTRY , srq,0 );return err; } -xa_erase_irq (&table->array, srq->srqn); + +/* +* A race can occur where a concurrent create gets the same srqn +* (after hardware released it) and overwrites XA_ZERO_ENTRY with +* its new SRQ before we reach here. In that case, we must not erase +* the entry as it now belongs to the new SRQ. +*/ +xa_cmpxchg_irq (&table->array, srq->srqn,XA_ZERO_ENTRY ,NULL ,0 );mlx5_core_res_put (&srq->common);wait_for_completion (&srq->common.free);
Signature · 1 line
2.49.0
PATCH v3 2/5 RDMA/mlx5: Fix UAF in DCT destroy due to race with create
In reply to Edward SroujiA potential race condition exists between mlx5_core_destroy_dct() and mlx5_core_create_dct() that can lead to a use-after-free.
After _mlx5_core_destroy_dct() releases the DCT to firmware, the DCTN can be immediately reallocated for a new DCT being created concurrently. If the create path stores the new DCT in the xarray before the destroy path erases it, the destroy will incorrectly delete the new DCT's entry. Later accesses then hit freed memory.
Fix by replacing the unconditional xa_erase_irq() with xa_cmpxchg_irq() that only erases the entry if it hasn't already been replaced (still contains XA_ZERO_ENTRY), preserving any newly created DCT.
--- drivers/infiniband/hw/mlx5/qpc.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
drivers/infiniband/hw/mlx5/qpc.c +8 -1
@@ -314,7 +314,14 @@ int mlx5_core_destroy_dct(struct mlx5_ib_dev *dev,xa_cmpxchg_irq (&table->dct_xa, dct->mqp.qpn,XA_ZERO_ENTRY , dct,0 );return err; } -xa_erase_irq (&table->dct_xa, dct->mqp.qpn); + +/* +* A race can occur where a concurrent create gets the same dctn +* (after hardware released it) and overwrites XA_ZERO_ENTRY with +* its new DCT before we reach here. In that case, we must not erase +* the entry as it now belongs to the new DCT. +*/ +xa_cmpxchg_irq (&table->dct_xa, dct->mqp.qpn,XA_ZERO_ENTRY ,NULL ,0 );return 0 ; }
Signature · 1 line
2.49.0
Jason Gunthorpe
maintainerRe: [PATCH rdma-next v3 2/5] RDMA/mlx5: Fix UAF in DCT destroy due to race with create
Quoted 28 lines · Edward Srouji
A potential race condition exists between
mlx5_core_destroy_dct()andmlx5_core_create_dct()that can lead to a use-after-free.After
_mlx5_core_destroy_dct()releases the DCT to firmware, the DCTN can be immediately reallocated for a new DCT being created concurrently. If the create path stores the new DCT in the xarray before the destroy path erases it, the destroy will incorrectly delete the new DCT's entry. Later accesses then hit freed memory.Fix by replacing the unconditional
xa_erase_irq()withxa_cmpxchg_irq()that only erases the entry if it hasn't already been replaced (still containsXA_ZERO_ENTRY), preserving any newly created DCT.Fixes afff24899846 ("RDMA/mlx5: Handle DCT QP logic separately from low level QP interface")Signed-off-by Edward Srouji <[email protected]>Reviewed-by Michael Guralnik <[email protected]>--- drivers/infiniband/hw/mlx5/qpc.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)drivers/infiniband/hw/mlx5/qpc.c +0
-0@@ -314,7 +314,14 @@ int mlx5_core_destroy_dct(struct mlx5_ib_dev *dev,xa_cmpxchg_irq (&table->dct_xa, dct->mqp.qpn,XA_ZERO_ENTRY , dct,0 );return err; }
drivers/infiniband/hw/mlx5/qpc.c int mlx5_core_destroy_dct(struct mlx5_ib_dev *dev, · line 325
Quoted 10 lines
drivers/infiniband/hw/mlx5/qpc.c +8
-1-xa_erase_irq (&table->dct_xa, dct->mqp.qpn); + +/* +* A race can occur where a concurrent create gets the same dctn +* (after hardware released it) and overwrites XA_ZERO_ENTRY with +* its new DCT before we reach here. In that case, we must not erase +* the entry as it now belongs to the new DCT. +*/ +xa_cmpxchg_irq (&table->dct_xa, dct->mqp.qpn,XA_ZERO_ENTRY ,NULL ,0 );return 0 ;
Sashiko is right about the
xa_init_flags(&table->dct_xa, XA_FLAGS_LOCK_IRQ);
Please make another patch for that.
The think about the xa_cmpxchg_irq().. I think that is not sane.. If the FW has returned the ID back to another thread then the FW is no longer allowed to fail this destruction, it is already destroyed.
In this case the xa will go wonky, but we are already pretty lost at that point.
The logic here is probably not great though, if the xa is mangled then we should still try to do the firmware call
Jason
PATCH v3 3/5 IB/core: Fix IPv6 netlink message size in ib_nl_ip_send_msg()
In reply to Edward SroujiFrom: Maher Sanalla <[email protected]>
When resolving an RDMA-CM IPv6 address, ib_nl_ip_send_msg() sends a netlink request to the userspace daemon to perform IP-to-GID resolution in certain cases. The function allocates the netlink message buffer using nla_total_size(sizeof(size)), which passes 8 bytes (the size of size_t) instead of 16 bytes (the size of an IPv6 address). This results in an 8-byte under-allocation.
This is currently masked by nlmsg_new() over-allocation of the skb in its internal logic. However, the code remains incorrect.
Fix the issue by supplying the proper IPv6 address length to nla_total_size().
--- drivers/infiniband/core/addr.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/infiniband/core/addr.c +1 -1
@@ -150,7 +150,7 @@ static int ib_nl_ip_send_msg(struct rdma_dev_addr *dev_addr, attrtype =RDMA_NLA_F_MANDATORY |LS_NLA_TYPE_IPV6 ; } - len =nla_total_size (sizeof (size)); + len =nla_total_size (size); len +=NLMSG_ALIGN (sizeof (*header)); skb =nlmsg_new (len,GFP_KERNEL );
Signature · 1 line
2.49.0
PATCH v3 4/5 RDMA/core: Fix rereg_mr use-after-free race
In reply to Edward SroujiFrom: Michael Guralnik <[email protected]>
When a driver creates a new MR during rereg_user_mr, a race window exists between rdma_alloc_commit_uobject() for the new MR and the point where the code reads that MR to populate the response keys.
A concurrent rereg_mr or destroy_mr could destroy the MR in this window and cause UAF in the first thread.
Racing flow between two rereg_mr calls:
CPU0 CPU1
---- ----
rereg_user_mr(mr_handle)
uobj_get_write(mr_handle) -> mr0
mr1 = driver→rereg()
rdma_alloc_commit_uobject(mr1)
// mr1 replaced mr0 and is unlocked
uobj_put_destroy(mr0)
rereg_user_mr(mr_handle)
uobj_get_write(mr_handle) -> mr1
mr2 = driver→rereg()
rdma_alloc_commit_uobject(mr2)
// mr2 replaced mr1 and is unlocked
uobj_put_destroy(mr1)
// Destroys mr1! resp.lkey = mr1->lkey; // UAF - mr1 was freed!
resp.rkey = mr1->rkey; // UAF - mr1 was freed!
Fix by storing lkey/rkey in local variables before the new MR is unlocked and using the local variables to set the user response.
--- drivers/infiniband/core/uverbs_cmd.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-)
drivers/infiniband/core/uverbs_cmd.c +7 -2
@@ -778,6 +778,7 @@ static int ib_uverbs_rereg_mr(struct uverbs_attr_bundle *attrs)struct ib_pd *orig_pd;struct ib_pd *new_pd;struct ib_mr *new_mr; +u32 lkey, rkey; ret =uverbs_request (attrs, &cmd,sizeof (cmd));if (ret)
@@ -846,6 +847,8 @@ static int ib_uverbs_rereg_mr(struct uverbs_attr_bundle *attrs) new_mr->uobject = uobj;atomic_inc (&new_pd->usecnt); new_uobj->object = new_mr; + lkey = new_mr->lkey; + rkey = new_mr->rkey;rdma_restrack_new (&new_mr->res,RDMA_RESTRACK_MR );rdma_restrack_set_name (&new_mr->res,NULL );
@@ -871,11 +874,13 @@ static int ib_uverbs_rereg_mr(struct uverbs_attr_bundle *attrs) mr->iova = cmd.hca_va; mr->length = cmd.length; } + lkey = mr->lkey; + rkey = mr->rkey; }memset (&resp,0 ,sizeof (resp)); - resp.lkey = mr->lkey; - resp.rkey = mr->rkey; + resp.lkey = lkey; + resp.rkey = rkey; ret =uverbs_response (attrs, &resp,sizeof (resp));
Signature · 1 line
2.49.0
PATCH v3 5/5 RDMA/mlx5: Fix null-ptr-deref in Raw Packet QP creation
In reply to Edward SroujiFrom: Michael Guralnik <[email protected]>
Raw Packet QPs are unique in that they support separate send and receive queues, using 2 different user-provided buffers.
They can also be created with one of the queues having size 0, allowing a send-only or receive-only QP.
The Raw Packet RQ umem is created in the common user QP creation path, which allows zero-length queues. Add a later validation of the RQ umem in Raw Packet QP creation path when an RQ was requested.
This prevents possible null-ptr dereference crashes, as seen in the below trace:
Oops: general protection fault, probably for non-canonical address 0xdffffc0000000006: 0000 [#1] SMP KASAN KASAN: null-ptr-deref in range [0x0000000000000030-0x0000000000000037] CPU: 6 UID: 0 PID: 3539 Comm: raw_packet_umem Not tainted 6.19.0-rc1+ #166 NONE Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.3-0-ga6ed6b701f0a-prebuilt.qemu.org 04/01/2014 RIP: 0010:__mlx5_umem_find_best_quantized_pgoff+0x37/0x280 [mlx5_ib] Code: ff df 41 57 49 89 ff 41 56 41 55 41 89 d5 41 54 4d 89 cc 4c 8d 4f 30 55 4c 89 ca 48 89 f5 53 48 c1 ea 03 48 89 cb 48 83 ec 18 <80> 3c 02 00 44 89 04 24 0f 85 01 02 00 00 48 ba 00 00 00 00 00 fc RSP: 0018:ff1100013966f4e0 EFLAGS: 00010282 RAX: dffffc0000000000 RBX: 00000000ffffffc0 RCX: 00000000ffffffc0 RDX: 0000000000000006 RSI: 00000ffffffff000 RDI: 0000000000000000 RBP: 00000ffffffff000 R08: 0000000000000040 R09: 0000000000000030 R10: 0000000000000000 R11: 0000000000000000 R12: ff1100013966f648 R13: 0000000000000005 R14: ff1100013966f980 R15: 0000000000000000 FS: 00007fae6c82f740(0000) GS:ff11000898ba1000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 0000200000000000 CR3: 000000010f96c005 CR4: 0000000000373eb0 Call Trace: <TASK> create_qp+0x747d/0xc740 [mlx5_ib] ? is_module_address+0x18/0x110 ? _create_user_qp.constprop.0+0x18e0/0x18e0 [mlx5_ib] ? __module_address+0x49/0x210 ? is_module_address+0x68/0x110 ? static_obj+0x67/0x90 ? lockdep_init_map_type+0x58/0x200 mlx5_ib_create_qp+0xc85/0x2620 [mlx5_ib] ? find_held_lock+0x2b/0x80 ? create_qp+0xc740/0xc740 [mlx5_ib] ? lock_release+0xcb/0x260 ? lockdep_init_map_type+0x58/0x200 ? __init_swait_queue_head+0xcb/0x150 create_qp.part.0+0x558/0x7c0 [ib_core] ib_create_qp_user+0xa0/0x4f0 [ib_core] ? rdma_lookup_get_uobject+0x1e4/0x400 [ib_uverbs] create_qp+0xe4f/0x1d10 [ib_uverbs] ? ib_uverbs_rereg_mr+0xd40/0xd40 [ib_uverbs] ? ib_uverbs_cq_event_handler+0x120/0x120 [ib_uverbs] ? __might_fault+0x81/0x100 ? lock_release+0xcb/0x260 ? _copy_from_user+0x3e/0x90 ib_uverbs_create_qp+0x10a/0x150 [ib_uverbs] ? ib_uverbs_ex_create_qp+0xe0/0xe0 [ib_uverbs] ? __might_fault+0x81/0x100 ? lock_release+0xcb/0x260 ib_uverbs_write+0x7e5/0xc90 [ib_uverbs] ? uverbs_devnode+0xc0/0xc0 [ib_uverbs] ? lock_acquire+0xfa/0x2b0 ? find_held_lock+0x2b/0x80 ? finish_task_switch.isra.0+0x189/0x6c0 vfs_write+0x1c0/0xf70 ? lockdep_hardirqs_on_prepare+0xde/0x170 ? kernel_write+0x5a0/0x5a0 ? __switch_to+0x527/0xe60 ? __schedule+0x10a3/0x3950 ? io_schedule_timeout+0x110/0x110 ksys_write+0x170/0x1c0 ? __x64_sys_read+0xb0/0xb0 ? trace_hardirqs_off.part.0+0x4e/0xe0 do_syscall_64+0x70/0x1360 entry_SYSCALL_64_after_hwframe+0x4b/0x53 RIP: 0033:0x7fae6ca3118d Code: ff c3 66 2e 0f 1f 84 00 00 00 00 00 90 f3 0f 1e fa 48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f 05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 5b cc 0c 00 f7 d8 64 89 01 48 RSP: 002b:00007ffe678ca308 EFLAGS: 00000213 ORIG_RAX: 0000000000000001 RAX: ffffffffffffffda RBX: 00007ffe678ca448 RCX: 00007fae6ca3118d RDX: 0000000000000070 RSI: 0000200000000280 RDI: 0000000000000003 RBP: 00007ffe678ca320 R08: 00000000ffffffff R09: 00007fae6c8ec5b8 R10: 0000000000000064 R11: 0000000000000213 R12: 0000000000000001 R13: 0000000000000000 R14: 00007fae6cb71000 R15: 0000000000404df0 </TASK> Modules linked in: mlx5_ib mlx5_fwctl mlx5_core bonding ip6_gre ip6_tunnel tunnel6 ip_gre gre rdma_ucm ib_uverbs rdma_cm iw_cm ib_ipoib ib_cm ib_umad ib_core rpcsec_gss_krb5 auth_rpcgss oid_registry overlay nfnetlink zram zsmalloc fuse scsi_transport_iscsi [last unloaded: mlx5_core] ---[ end trace 0000000000000000 ]--- RIP: 0010:__mlx5_umem_find_best_quantized_pgoff+0x37/0x280 [mlx5_ib]
--- drivers/infiniband/hw/mlx5/qp.c | 5 +++++ 1 file changed, 5 insertions(+)
drivers/infiniband/hw/mlx5/qp.c +5 -0
@@ -1603,6 +1603,11 @@ static int create_raw_packet_qp(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp, }if (qp->rq.wqe_cnt) { +if (!rq->base.ubuffer.umem) { + err = -EINVAL ; + goto err_destroy_sq; + } + rq->base.container_mibqp = qp;if (qp->flags &IB_QP_CREATE_CVLAN_STRIPPING )
Signature · 1 line
2.49.0
Quoted 29 lines · Edward Srouji
This series addresses several stability issues in RDMA core and the mlx5 driver.
Patches 1-2 fix xarray race conditions in the mlx5 SRQ and DCT destroy paths where a concurrent create can reuse the same firmware object number right after firmware releases it, causing the destroy path to incorrectly erase the newly created entry.
The remaining patches are independent fixes.
Signed-off-by Edward Srouji <[email protected]>---
Changes in v3:
- Dropped restrack destroy-ordering fixes (patches 1-6 in v2) to
rework them in a dedicated series based on code-review feedback
- Rebased onto latest for-next
- Link to v2: https://lore.kernel.org/r/[email protected]Changes in v2:
- Added patch "RDMA/mlx5: Remove raw RSS QP restrack tracking" to
also suppress broken tracking for raw RSS QPs, which suffer from
the same silent failures as DCTs
- Link to v1: https://lore.kernel.org/r/[email protected]---
Edward Srouji (2):
RDMA/mlx5: Fix UAF in SRQ destroy due to race with create RDMA/mlx5: Fix UAF in DCT destroy due to race with create
Quoted 6 lines
Maher Sanalla (1):
IB/core: Fix IPv6 netlink message size inib_nl_ip_send_msg()Michael Guralnik (2):
RDMA/core: Fixrereg_mruse-after-free race RDMA/mlx5: Fix null-ptr-deref in Raw Packet QP creation
Applied to for-rc
Jason