nvme: retain split access workaround for capability reads
authorArd Biesheuvel <ard.biesheuvel@linaro.org>
Thu, 3 Oct 2019 11:57:29 +0000 (13:57 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 29 Jan 2020 09:24:33 +0000 (10:24 +0100)
[ Upstream commit 3a8ecc935efabdad106b5e06d07b150c394b4465 ]

Commit 7fd8930f26be4

  "nvme: add a common helper to read Identify Controller data"

has re-introduced an issue that we have attempted to work around in the
past, in commit a310acd7a7ea ("NVMe: use split lo_hi_{read,write}q").

The problem is that some PCIe NVMe controllers do not implement 64-bit
outbound accesses correctly, which is why the commit above switched
to using lo_hi_[read|write]q for all 64-bit BAR accesses occuring in
the code.

In the mean time, the NVMe subsystem has been refactored, and now calls
into the PCIe support layer for NVMe via a .reg_read64() method, which
fails to use lo_hi_readq(), and thus reintroduces the problem that the
workaround above aimed to address.

Given that, at the moment, .reg_read64() is only used to read the
capability register [which is known to tolerate split reads], let's
switch .reg_read64() to lo_hi_readq() as well.

This fixes a boot issue on some ARM boxes with NVMe behind a Synopsys
DesignWare PCIe host controller.

Fixes: 7fd8930f26be4 ("nvme: add a common helper to read Identify Controller data")
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Sasha Levin <sashal@kernel.org>
drivers/nvme/host/pci.c

index 1ac4cec..e2bce93 100644 (file)
@@ -1863,7 +1863,7 @@ static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
 
 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
 {
-       *val = readq(to_nvme_dev(ctrl)->bar + off);
+       *val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
        return 0;
 }