]> git.itanic.dy.fi Git - linux-stable/commitdiff
net/smc: Fix possible access to freed memory in link clear
authorYacan Liu <liuyacan@corp.netease.com>
Tue, 6 Sep 2022 13:01:39 +0000 (21:01 +0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 15 Sep 2022 08:47:17 +0000 (10:47 +0200)
[ Upstream commit e9b1a4f867ae9c1dbd1d71cd09cbdb3239fb4968 ]

After modifying the QP to the Error state, all RX WR would be completed
with WC in IB_WC_WR_FLUSH_ERR status. Current implementation does not
wait for it is done, but destroy the QP and free the link group directly.
So there is a risk that accessing the freed memory in tasklet context.

Here is a crash example:

 BUG: unable to handle page fault for address: ffffffff8f220860
 #PF: supervisor write access in kernel mode
 #PF: error_code(0x0002) - not-present page
 PGD f7300e067 P4D f7300e067 PUD f7300f063 PMD 8c4e45063 PTE 800ffff08c9df060
 Oops: 0002 [#1] SMP PTI
 CPU: 1 PID: 0 Comm: swapper/1 Kdump: loaded Tainted: G S         OE     5.10.0-0607+ #23
 Hardware name: Inspur NF5280M4/YZMB-00689-101, BIOS 4.1.20 07/09/2018
 RIP: 0010:native_queued_spin_lock_slowpath+0x176/0x1b0
 Code: f3 90 48 8b 32 48 85 f6 74 f6 eb d5 c1 ee 12 83 e0 03 83 ee 01 48 c1 e0 05 48 63 f6 48 05 00 c8 02 00 48 03 04 f5 00 09 98 8e <48> 89 10 8b 42 08 85 c0 75 09 f3 90 8b 42 08 85 c0 74 f7 48 8b 32
 RSP: 0018:ffffb3b6c001ebd8 EFLAGS: 00010086
 RAX: ffffffff8f220860 RBX: 0000000000000246 RCX: 0000000000080000
 RDX: ffff91db1f86c800 RSI: 000000000000173c RDI: ffff91db62bace00
 RBP: ffff91db62bacc00 R08: 0000000000000000 R09: c00000010000028b
 R10: 0000000000055198 R11: ffffb3b6c001ea58 R12: ffff91db80e05010
 R13: 000000000000000a R14: 0000000000000006 R15: 0000000000000040
 FS:  0000000000000000(0000) GS:ffff91db1f840000(0000) knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: ffffffff8f220860 CR3: 00000001f9580004 CR4: 00000000003706e0
 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
 DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
 Call Trace:
  <IRQ>
  _raw_spin_lock_irqsave+0x30/0x40
  mlx5_ib_poll_cq+0x4c/0xc50 [mlx5_ib]
  smc_wr_rx_tasklet_fn+0x56/0xa0 [smc]
  tasklet_action_common.isra.21+0x66/0x100
  __do_softirq+0xd5/0x29c
  asm_call_irq_on_stack+0x12/0x20
  </IRQ>
  do_softirq_own_stack+0x37/0x40
  irq_exit_rcu+0x9d/0xa0
  sysvec_call_function_single+0x34/0x80
  asm_sysvec_call_function_single+0x12/0x20

Fixes: bd4ad57718cc ("smc: initialize IB transport incl. PD, MR, QP, CQ, event, WR")
Signed-off-by: Yacan Liu <liuyacan@corp.netease.com>
Reviewed-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
net/smc/smc_core.c
net/smc/smc_core.h
net/smc/smc_wr.c
net/smc/smc_wr.h

index f40f6ed0fbdb4450a17a5f9fab4117ac8aa78b38..1f3bb1f6b1f7b164878fbf4212293abaf7b4d31b 100644 (file)
@@ -755,6 +755,7 @@ int smcr_link_init(struct smc_link_group *lgr, struct smc_link *lnk,
        lnk->lgr = lgr;
        smc_lgr_hold(lgr); /* lgr_put in smcr_link_clear() */
        lnk->link_idx = link_idx;
+       lnk->wr_rx_id_compl = 0;
        smc_ibdev_cnt_inc(lnk);
        smcr_copy_dev_info_to_link(lnk);
        atomic_set(&lnk->conn_cnt, 0);
index 4cb03e94236481508c7b072e70bcc34cf4894a21..7b43a78c7f73abd38fd6c84d1636effb3e58d755 100644 (file)
@@ -115,8 +115,10 @@ struct smc_link {
        dma_addr_t              wr_rx_dma_addr; /* DMA address of wr_rx_bufs */
        dma_addr_t              wr_rx_v2_dma_addr; /* DMA address of v2 rx buf*/
        u64                     wr_rx_id;       /* seq # of last recv WR */
+       u64                     wr_rx_id_compl; /* seq # of last completed WR */
        u32                     wr_rx_cnt;      /* number of WR recv buffers */
        unsigned long           wr_rx_tstamp;   /* jiffies when last buf rx */
+       wait_queue_head_t       wr_rx_empty_wait; /* wait for RQ empty */
 
        struct ib_reg_wr        wr_reg;         /* WR register memory region */
        wait_queue_head_t       wr_reg_wait;    /* wait for wr_reg result */
index 26f8f240d9e8473e9901cc1ce8c4e253651b2943..b0678a417e09d73d28c6382acc7b6e07e324c501 100644 (file)
@@ -454,6 +454,7 @@ static inline void smc_wr_rx_process_cqes(struct ib_wc wc[], int num)
 
        for (i = 0; i < num; i++) {
                link = wc[i].qp->qp_context;
+               link->wr_rx_id_compl = wc[i].wr_id;
                if (wc[i].status == IB_WC_SUCCESS) {
                        link->wr_rx_tstamp = jiffies;
                        smc_wr_rx_demultiplex(&wc[i]);
@@ -465,6 +466,8 @@ static inline void smc_wr_rx_process_cqes(struct ib_wc wc[], int num)
                        case IB_WC_RNR_RETRY_EXC_ERR:
                        case IB_WC_WR_FLUSH_ERR:
                                smcr_link_down_cond_sched(link);
+                               if (link->wr_rx_id_compl == link->wr_rx_id)
+                                       wake_up(&link->wr_rx_empty_wait);
                                break;
                        default:
                                smc_wr_rx_post(link); /* refill WR RX */
@@ -639,6 +642,7 @@ void smc_wr_free_link(struct smc_link *lnk)
                return;
        ibdev = lnk->smcibdev->ibdev;
 
+       smc_wr_drain_cq(lnk);
        smc_wr_wakeup_reg_wait(lnk);
        smc_wr_wakeup_tx_wait(lnk);
 
@@ -889,6 +893,7 @@ int smc_wr_create_link(struct smc_link *lnk)
        atomic_set(&lnk->wr_tx_refcnt, 0);
        init_waitqueue_head(&lnk->wr_reg_wait);
        atomic_set(&lnk->wr_reg_refcnt, 0);
+       init_waitqueue_head(&lnk->wr_rx_empty_wait);
        return rc;
 
 dma_unmap:
index a54e90a1110fdb03449e150afe9420b0aa0d93a2..45e9b894d3f8a427c88950f213128887118bee46 100644 (file)
@@ -73,6 +73,11 @@ static inline void smc_wr_tx_link_put(struct smc_link *link)
                wake_up_all(&link->wr_tx_wait);
 }
 
+static inline void smc_wr_drain_cq(struct smc_link *lnk)
+{
+       wait_event(lnk->wr_rx_empty_wait, lnk->wr_rx_id_compl == lnk->wr_rx_id);
+}
+
 static inline void smc_wr_wakeup_tx_wait(struct smc_link *lnk)
 {
        wake_up_all(&lnk->wr_tx_wait);