On Tue, Aug 26, 2025 at 11:14:19AM +0100, Vadim Fedorenko wrote: > On 26/08/2025 02:31, Yibo Dong wrote: > > On Mon, Aug 25, 2025 at 05:37:27PM +0100, Vadim Fedorenko wrote: > > > On 22/08/2025 03:34, Dong Yibo wrote: > > > > > > [...] > > > > +/** > > > > + * mucse_mbx_fw_post_req - Posts a mbx req to firmware and wait reply > > > > + * @hw: pointer to the HW structure > > > > + * @req: pointer to the cmd req structure > > > > + * @cookie: pointer to the req cookie > > > > + * > > > > + * mucse_mbx_fw_post_req posts a mbx req to firmware and wait for the > > > > + * reply. cookie->wait will be set in irq handler. > > > > + * > > > > + * @return: 0 on success, negative on failure > > > > + **/ > > > > +static int mucse_mbx_fw_post_req(struct mucse_hw *hw, > > > > + struct mbx_fw_cmd_req *req, > > > > + struct mbx_req_cookie *cookie) > > > > +{ > > > > + int len = le16_to_cpu(req->datalen); > > > > + int err; > > > > + > > > > + cookie->errcode = 0; > > > > + cookie->done = 0; > > > > + init_waitqueue_head(&cookie->wait); > > > > + err = mutex_lock_interruptible(&hw->mbx.lock); > > > > + if (err) > > > > + return err; > > > > + err = mucse_write_mbx_pf(hw, (u32 *)req, len); > > > > + if (err) > > > > + goto out; > > > > + /* if write succeeds, we must wait for firmware response or > > > > + * timeout to avoid using the already freed cookie->wait > > > > + */ > > > > + err = wait_event_timeout(cookie->wait, > > > > + cookie->done == 1, > > > > + cookie->timeout_jiffies); > > > > > > it's unclear to me, what part of the code is managing values of cookie > > > structure? I didn't get the reason why are you putting the address of > > > cookie structure into request which is then directly passed to the FW. > > > Is the FW supposed to change values in cookie? > > > > > > > cookie will be used in an irq-handler. like this: > > static int rnpgbe_mbx_fw_reply_handler(struct mucse *mucse, > > struct mbx_fw_cmd_reply *reply) > > { > > struct mbx_req_cookie *cookie; > > > > cookie = reply->cookie; > > > > if (cookie->priv_len > 0) > > memcpy(cookie->priv, reply->data, cookie->priv_len); > > cookie->done = 1; > > if (le16_to_cpu(reply->flags) & FLAGS_ERR) > > cookie->errcode = -EIO; > > else > > cookie->errcode = 0; > > wake_up(&cookie->wait); > > return 0; > > } > > That is why we must wait for firmware response. > > But irq is not added in this patch series. Maybe I should move all > > cookie relative codes to the patch will add irq? > > well, yes, in general it's better to introduce the code as a solid > solution. this way it's much easier to review > Ok, I will remove it in this series and add later. > > > > > > + > > > > + if (!err) > > > > + err = -ETIMEDOUT; > > > > + else > > > > + err = 0; > > > > + if (!err && cookie->errcode) > > > > + err = cookie->errcode; > > > > +out: > > > > + mutex_unlock(&hw->mbx.lock); > > > > + return err; > > > > +} > > > > > > [...] > > > > > > > +struct mbx_fw_cmd_req { > > > > + __le16 flags; > > > > + __le16 opcode; > > > > + __le16 datalen; > > > > + __le16 ret_value; > > > > + union { > > > > + struct { > > > > + __le32 cookie_lo; > > > > + __le32 cookie_hi; > > > > + }; > > > > + > > > > + void *cookie; > > > > + }; > > > > + __le32 reply_lo; > > > > + __le32 reply_hi; > > > > > > what do these 2 fields mean? are you going to provide reply's buffer > > > address directly to FW? > > > > > > > No, this is defined by fw. Some fw can access physical address. > > But I don't use it in this driver. > > FW can access physical address without previously configuring IOMMU? > How can that be? > memory is allocated by dma_alloc_coherent, and get physical address. Then fw use it. > > > > > > + union { > > > > + u8 data[32]; > > > > + struct { > > > > + __le32 version; > > > > + __le32 status; > > > > + } ifinsmod; > > > > + struct { > > > > + __le32 port_mask; > > > > + __le32 pfvf_num; > > > > + } get_mac_addr; > > > > + }; > > > > +} __packed; > > > > + > > > > +struct mbx_fw_cmd_reply { > > > > + __le16 flags; > > > > + __le16 opcode; > > > > + __le16 error_code; > > > > + __le16 datalen; > > > > + union { > > > > + struct { > > > > + __le32 cookie_lo; > > > > + __le32 cookie_hi; > > > > + }; > > > > + void *cookie; > > > > + }; > > > > > > This part looks like the request, apart from datalen and error_code are > > > swapped in the header. And it actually means that the FW will put back > > > the address of provided cookie into reply, right? If yes, then it > > > doesn't look correct at all... > > > > > > > It is yes. cookie is used in irq handler as show above. > > Sorry, I didn't understand 'the not correct' point? > > The example above showed that the irq handler uses some value received > from the device as a pointer to kernel memory. That's not safe, you > cannot be sure that provided value is valid pointer, and that it points > to previously allocated cookie structure. It is a clear way to corrupt > memory. > Yes. It is not safe, so I 'must wait_event_timeout before free cookie'.... But is there a safe way to do it? Maybe: ->allocate cookie -> map it to an unique id ->set the id to req->cookie ->receive response and check id valid? Then access cookie? Please give me some advice... > > > > > > + union { > > > > + u8 data[40]; > > > > + struct mac_addr { > > > > + __le32 ports; > > > > + struct _addr { > > > > + /* for macaddr:01:02:03:04:05:06 > > > > + * mac-hi=0x01020304 mac-lo=0x05060000 > > > > + */ > > > > + u8 mac[8]; > > > > + } addrs[4]; > > > > + } mac_addr; > > > > + struct hw_abilities hw_abilities; > > > > + }; > > > > +} __packed; > > > > >