On Tue, Oct 29, 2024 at 11:58:40AM -0700, Gwendal Grignou wrote: > On Tue, Oct 29, 2024 at 12:41 AM Christoph Hellwig <hch@xxxxxx> wrote: > > > > On Mon, Oct 28, 2024 at 07:42:36PM -0700, Gwendal Grignou wrote: > > > PCI_DEVICE(0x1217, 0x8760) (O2 Micro, Inc. FORESEE E2M2 NVMe SSD) > > > is a NMVe to eMMC bridge, that can be used with different eMMC > > > memory devices. > > > > Holy f**k, what an awful idea.. > > > > > The NVMe device name contains the eMMC device name, for instance: > > > `BAYHUB SanDisk-DA4128-91904055-128GB` > > > > > > The bridge is known to work with many eMMC devices, we need to limit > > > the queue depth once we know which eMMC device is behind the bridge. > > > > Please work with Tobert to quirk based on the identify data for "his" > > device to keep it quirked instead of regressing it. > > The issue is we would need to base the quirk on the model name > (subsys->model) that is not available in `nvme_id_table`. Beside, > `q_depth` is set in `nvme_pci_enable`, called at probe time before > calling `nvme_init_ctrl_finish` that will indirectly populate > `subsys`. > > Bob, to address the data corruption problem from user space, adding a > udev rule to set `queue/nr_requests` to 1 when `device/model` matches > the device used in the Steam Deck would most likely be too late in the > boot process, wouldn't it? I think that is too late. There's the module parameter, 'nvme.io_queue_depth=2', that accomplishes the same thing as this quirk, if adding kernel parameters isn't too inconvenient to use here. Alternatively, you could put the quirk in the core_quirks, which do take a model name. The pci driver would have to move this check until after init_ctrl_finish, as you noticed, but that looks okay to do. We just need to be careful to update both dev->q_depth and ctrl->sqsize after the "finish".
