Function dma_alloc_coherent() returns a NULL pointer if there is no enough memory. Function ocrdma_mbx_create_ah_tbl() defined in file drivers/infiniband/hw/ocrdma/ocrdma_hw.c will return 0 on success or negative error codes on failures. It calls function dma_alloc_coherent() twice and checks the return values against NULL (at lines 1681 and 1686). The control flow jumps to label "mem_err_ah" and returns the value of variable status. The value of status is 0 (see the check of variable status at line 1645). As a result, the caller of ocrdma_mbx_create_ah_tbl() will be misled to believe all goes well even the memory allocation fails. Maybe it is better to assign "-ENOMEM" to variable status before the jump instructions at lines 1682 and 1687, or simply initialize status with "-ENOMEM" rather than "0" at line 1645. Codes related to this bug are summarised as follows. ocrdma_mbx_create_ah_tbl @@ drivers/infiniband/hw/ocrdma/ocrdma_hw.c 1642 static int ocrdma_mbx_create_ah_tbl(struct ocrdma_dev *dev) 1643 { 1644 int i; 1645 int status = 0; // use "int status = -ENOMEM;" ? ... 1653 cmd = ocrdma_init_emb_mqe(OCRDMA_CMD_CREATE_AH_TBL, sizeof(*cmd)); 1654 if (!cmd) 1655 return status; ... 1678 dev->av_tbl.pbl.va = dma_alloc_coherent(&pdev->dev, PAGE_SIZE, 1679 &dev->av_tbl.pbl.pa, 1680 GFP_KERNEL); 1681 if (dev->av_tbl.pbl.va == NULL) 1682 goto mem_err; 1683 1684 dev->av_tbl.va = dma_alloc_coherent(&pdev->dev, dev->av_tbl.size, 1685 &pa, GFP_KERNEL); 1686 if (dev->av_tbl.va == NULL) 1687 goto mem_err_ah; ... 1706 return 0; 1707 1708 mbx_err: 1709 dma_free_coherent(&pdev->dev, dev->av_tbl.size, dev->av_tbl.va, 1710 dev->av_tbl.pa); 1711 dev->av_tbl.va = NULL; 1712 mem_err_ah: 1713 dma_free_coherent(&pdev->dev, PAGE_SIZE, dev->av_tbl.pbl.va, 1714 dev->av_tbl.pbl.pa); 1715 dev->av_tbl.pbl.va = NULL; 1716 dev->av_tbl.size = 0; 1717 mem_err: 1718 kfree(cmd); 1719 return status; 1720 } Thanks very much!
Created attachment 256427 [details] The patch fixes the bug The patch has been merged into the latest version of the Linux kernel. So I will close the bug.