On Thu, Mar 31, 2022 at 04:33:37PM +0800, Xiaoke Wang wrote: > > On Thu, 31 Mar 2022 16:21:43 +0800, xkernel.wang@xxxxxxxxxxx wrote: > >In fact, this is considering that we do not know where is the failure > >from. In rtw_os_xmit_resource_alloc(), the failure can from > > > >> pxmitbuf->pallocated_buf = kzalloc(alloc_sz, GFP_KERNEL); > > > > , but also can from > > > >> pxmitbuf->pxmit_urb[i] = usb_alloc_urb(0, GFP_KERNEL); > > > > So if we do not handle the current failed item and just skip it, then some > > memory may be ignored. > > > > Maybe this should be the problem of rtw_os_xmit_resource_alloc() as it > does properly handle the error from it. > Anyway, I will attempt to fix them. > It would be a lot easier to fix this code, if it were cleaner. It goes to a lot of work to ensure that the pointers are aligned at the 4 byte mark, but memory from kmalloc() or vmalloc() is already aligned at 8 bytes so there is no need. pxmitpriv->pallocated_frame_buf = vzalloc(NR_XMITFRAME * sizeof(struct xmit_frame) + 4); The + 4 is for alignment. pxmitpriv->pxmit_frame_buf = (u8 *)N_BYTE_ALIGMENT((size_t)(pxmitpriv->pallocated_frame_buf), 4); So then it stores pxmitpriv->pxmit_frame_buf which is the aligned version of the pointer and the ->pallocated_frame_buf pointer which it uses to free the memory. Delete pxmitpriv->pallocated_frame_buf. Just allocate it directly. Use vcalloc(). pxmitpriv->pxmit_frame_buf = vzalloc(NR_XMITFRAME * sizeof(struct xmit_frame); That stuff can all be done in one patch because it is: Patch 1: Remove unnecessary alignment hacks. This code goes to great lengths to ensure that the buffers are 4 bytes aligned. However that is not necessary as the vmalloc() and kmalloc() functions return memory that is already aligned at 8 bytes. 1) No need for the temporary pxmitpriv->pallocated_frame_buf pointer to store unaligned memory. 2) No need to allocate "+ 4" extra bytes for alignment. 3) No need to call N_BYTE_ALIGMENT(). Patch 2: Use vcalloc() instead of vzalloc() Patch 3: Declare the pxmitpriv->pxmit_frame_buf pointer as an array of structs instead of an array of u8. This allows us to remove some casting. Once the code is cleaner then the error handling will also be easier. regards, dan carpenter
