NVMe Vendor Extension for performance improvements in a virtual environment
Rob Nelson
rlnelson at google.com
Wed Oct 1 15:25:26 PDT 2014
Second time's a charm. After some offline discussion with Keith, I'm sending
this patch out again. I've fixed up a few things here e.g. use
nvme_common_command
instead of new struct.
This patch provides a mechanism to reduce the number of MMIO doorbell
writes for the NVMe driver. When running in a virtualized environment
like QEMU, the cost of an MMIO is quite hefy here. The main idea for
the patch is provide the device two memory location locations: 1) to
store the doorbell values so they can be lookup without the doorbell
MMIO write 2) to store an event index.
I believe the doorbell value is obvious, the event index not so much.
Similar to the Virtio specificaiton, the device (virtual device) can
tell the driver not to write MMIO unless you are writing past this
value.
FYI: doorbell values are written by the nvme driver (guest OS) and the
event index is written by the virtual device (host OS).
The patch implements a new admin command that will communicate where
these two memory locations reside. If the command fails, the nvme
driver will work as before without any optimizations.
Just to give an idea on the performance boost with the vendor
extension: Running fio [1], a stock NVMe driver I get about 200K read
IOPs with my vendor patch I get about 1000K read IOPs. This was
running with a null device i.e. the backing device simply returned
success on every read IO request.
[1] Running on a 4 core machine:
fio --time_based --name=benchmark --runtime=30
--filename=/dev/nvme0n1 --nrfiles=1 --ioengine=libaio --iodepth=32
--direct=1 --invalidate=1 --verify=0 --verify_fatal=0 --numjobs=4
--rw=randread --blocksize=4k --randrepeat=0
Mihai, also has a patch for the qemu driver to use this vendor extension
as well. He was able to get 4x performance boost. I believe Keith you
got a little more performance gain with a more sane doorbell stride.
Git repo here:
https://github.com/rlnelson-git/linux-nvme
Patch is based off
http://git.infradead.org/users/willy/linux-nvme.git
Contributions in this patch from:
Eric Northup digitaleric at google.com
Frank Swiderski fes at google.com
Ted Tso tytso at google.com
Mihai Rusu dizzy at google.com
Keith Busch keith.busch at intel.com
diff --git a/drivers/block/nvme-core.c b/drivers/block/nvme-core.c
index 28aec2d..21acb93 100644
--- a/drivers/block/nvme-core.c
+++ b/drivers/block/nvme-core.c
@@ -62,6 +62,9 @@ static unsigned char retry_time = 30;
module_param(retry_time, byte, 0644);
MODULE_PARM_DESC(retry_time, "time in seconds to retry failed I/O");
+/* Google Vendor ID is not in include/linux/pci_ids.h */
+#define PCI_VENDOR_ID_GOOGLE 0x1AE0
+
static int nvme_major;
module_param(nvme_major, int, 0);
@@ -114,12 +117,16 @@ struct nvme_queue {
u8 cqe_seen;
u8 q_suspended;
cpumask_var_t cpu_mask;
+ u32* sq_doorbell_addr;
+ u32* sq_eventidx_addr;
+ u32* cq_doorbell_addr;
+ u32* cq_eventidx_addr;
struct async_cmd_info cmdinfo;
unsigned long cmdid_data[];
};
/*
- * Check we didin't inadvertently grow the command struct
+ * Check we didn't inadvertently grow the command struct
*/
static inline void _nvme_check_size(void)
{
@@ -336,6 +343,30 @@ static void unlock_nvmeq(struct nvme_queue
*nvmeq) __releases(RCU)
rcu_read_unlock();
}
+static inline int nvme_need_event(u16 event_idx, u16 new_idx, u16 old) {
+ /* Borrowed from vring_need_event */
+ return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
+}
+
+static void write_doorbell(u16 value, u32 __iomem* q_db,
+ u32* db_addr, volatile u32* event_idx) {
+ u16 old_value;
+ if (!db_addr)
+ goto ring_doorbell;
+
+ old_value = *db_addr;
+ *db_addr = value;
+
+ rmb();
+ if (!nvme_need_event(*event_idx, value, old_value))
+ goto no_doorbell;
+
+ring_doorbell:
+ writel(value, q_db);
+no_doorbell:
+ return;
+}
+
/**
* nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
* @nvmeq: The queue to use
@@ -354,9 +385,12 @@ static int nvme_submit_cmd(struct nvme_queue
*nvmeq, struct nvme_command *cmd)
}
tail = nvmeq->sq_tail;
memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
+ if (nvmeq->sq_doorbell_addr)
+ wmb();
if (++tail == nvmeq->q_depth)
tail = 0;
- writel(tail, nvmeq->q_db);
+ write_doorbell(tail, nvmeq->q_db,
+ nvmeq->sq_doorbell_addr, nvmeq->sq_eventidx_addr);
nvmeq->sq_tail = tail;
spin_unlock_irqrestore(&nvmeq->q_lock, flags);
@@ -643,11 +677,13 @@ static int nvme_submit_discard(struct nvme_queue
*nvmeq, struct nvme_ns *ns,
cmnd->dsm.prp1 = cpu_to_le64(iod->first_dma);
cmnd->dsm.nr = 0;
cmnd->dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
+ if (nvmeq->sq_doorbell_addr)
+ wmb();
if (++nvmeq->sq_tail == nvmeq->q_depth)
nvmeq->sq_tail = 0;
- writel(nvmeq->sq_tail, nvmeq->q_db);
-
+ write_doorbell(nvmeq->sq_tail, nvmeq->q_db,
+ nvmeq->sq_doorbell_addr, nvmeq->sq_eventidx_addr);
return 0;
}
@@ -660,10 +696,13 @@ static int nvme_submit_flush(struct nvme_queue
*nvmeq, struct nvme_ns *ns,
cmnd->common.opcode = nvme_cmd_flush;
cmnd->common.command_id = cmdid;
cmnd->common.nsid = cpu_to_le32(ns->ns_id);
+ if (nvmeq->dev->pci_dev->vendor == PCI_VENDOR_ID_GOOGLE)
+ wmb();
if (++nvmeq->sq_tail == nvmeq->q_depth)
nvmeq->sq_tail = 0;
- writel(nvmeq->sq_tail, nvmeq->q_db);
+ write_doorbell(nvmeq->sq_tail, nvmeq->q_db,
+ nvmeq->sq_doorbell_addr, nvmeq->sq_eventidx_addr);
return 0;
}
@@ -709,10 +748,13 @@ static int nvme_submit_iod(struct nvme_queue
*nvmeq, struct nvme_iod *iod)
cpu_to_le16((bio->bi_iter.bi_size >> ns->lba_shift) - 1);
cmnd->rw.control = cpu_to_le16(control);
cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
+ if (nvmeq->dev->pci_dev->vendor == PCI_VENDOR_ID_GOOGLE)
+ wmb();
if (++nvmeq->sq_tail == nvmeq->q_depth)
nvmeq->sq_tail = 0;
- writel(nvmeq->sq_tail, nvmeq->q_db);
+ write_doorbell(nvmeq->sq_tail, nvmeq->q_db,
+ nvmeq->sq_doorbell_addr, nvmeq->sq_eventidx_addr);
return 0;
}
@@ -806,6 +848,8 @@ static int nvme_process_cq(struct nvme_queue *nvmeq)
for (;;) {
void *ctx;
nvme_completion_fn fn;
+ if (nvmeq->dev->pci_dev->vendor == PCI_VENDOR_ID_GOOGLE)
+ rmb();
struct nvme_completion cqe = nvmeq->cqes[head];
if ((le16_to_cpu(cqe.status) & 1) != phase)
break;
@@ -828,7 +872,8 @@ static int nvme_process_cq(struct nvme_queue *nvmeq)
if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
return 0;
- writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
+ write_doorbell(head, nvmeq->q_db + nvmeq->dev->db_stride,
+ nvmeq->cq_doorbell_addr, nvmeq->cq_eventidx_addr);
nvmeq->cq_head = head;
nvmeq->cq_phase = phase;
@@ -1071,6 +1116,17 @@ int nvme_identify(struct nvme_dev *dev,
unsigned nsid, unsigned cns,
return nvme_submit_admin_cmd(dev, &c, NULL);
}
+int nvme_doorbell_memory(struct nvme_dev *dev) {
+ struct nvme_command c;
+
+ memset(&c, 0, sizeof(c));
+ c.common.opcode = nvme_admin_doorbell_memory;
+ c.common.prp1 = cpu_to_le64(dev->doorbell);
+ c.common.prp2 = cpu_to_le64(dev->eventidx);
+
+ return nvme_submit_admin_cmd(dev, &c, NULL);
+}
+
int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
dma_addr_t dma_addr, u32 *result)
{
@@ -1319,6 +1375,14 @@ static struct nvme_queue
*nvme_alloc_queue(struct nvme_dev *dev, int qid,
nvmeq->cq_vector = vector;
nvmeq->qid = qid;
nvmeq->q_suspended = 1;
+ if (dev->db_mem && dev->ei_mem && qid) {
+ nvmeq->sq_doorbell_addr = &dev->db_mem[qid * 2 *
dev->db_stride];
+ nvmeq->cq_doorbell_addr =
+ &dev->db_mem[(qid * 2 + 1) * dev->db_stride];
+ nvmeq->sq_eventidx_addr = &dev->ei_mem[qid * 2 *
dev->db_stride];
+ nvmeq->cq_eventidx_addr =
+ &dev->ei_mem[(qid * 2 + 1) * dev->db_stride];
+ }
dev->queue_count++;
rcu_assign_pointer(dev->queues[qid], nvmeq);
@@ -1355,6 +1419,14 @@ static void nvme_init_queue(struct nvme_queue
*nvmeq, u16 qid)
nvmeq->cq_head = 0;
nvmeq->cq_phase = 1;
nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
+ if (dev->pci_dev->vendor == PCI_VENDOR_ID_GOOGLE && qid != 0) {
+ nvmeq->sq_doorbell_addr = &dev->db_mem[qid * 2 *
dev->db_stride];
+ nvmeq->cq_doorbell_addr =
+ &dev->db_mem[(qid * 2 + 1) * dev->db_stride];
+ nvmeq->sq_eventidx_addr = &dev->ei_mem[qid * 2 *
dev->db_stride];
+ nvmeq->cq_eventidx_addr =
+ &dev->ei_mem[(qid * 2 + 1) * dev->db_stride];
+ }
memset(nvmeq->cmdid_data, 0, extra);
memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
nvme_cancel_ios(nvmeq, false);
@@ -2304,6 +2376,17 @@ static int nvme_dev_add(struct nvme_dev *dev)
goto out;
}
+ if (pdev->vendor == PCI_VENDOR_ID_GOOGLE) {
+ res = nvme_doorbell_memory(dev);
+ if (res) {
+ // Free memory and continue on.
+ dma_free_coherent(&pdev->dev, 8192,
dev->db_mem, dev->doorbell);
+ dma_free_coherent(&pdev->dev, 8192,
dev->ei_mem, dev->doorbell);
+ dev->db_mem = 0;
+ dev->ei_mem = 0;
+ }
+ }
+
ctrl = mem;
nn = le32_to_cpup(&ctrl->nn);
dev->oncs = le16_to_cpup(&ctrl->oncs);
@@ -2332,7 +2415,6 @@ static int nvme_dev_add(struct nvme_dev *dev)
dma_addr + 4096, NULL);
if (res)
memset(mem + 4096, 0, 4096);
-
ns = nvme_alloc_ns(dev, i, mem, mem + 4096);
if (ns)
list_add_tail(&ns->list, &dev->namespaces);
@@ -2346,6 +2428,10 @@ static int nvme_dev_add(struct nvme_dev *dev)
return res;
}
+static int nvme_vendor_memory_size(struct nvme_dev *dev) {
+ return ((num_possible_cpus() + 1) * 8 * dev->db_stride);
+}
+
static int nvme_dev_map(struct nvme_dev *dev)
{
u64 cap;
@@ -2377,8 +2463,25 @@ static int nvme_dev_map(struct nvme_dev *dev)
dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
dev->dbs = ((void __iomem *)dev->bar) + 4096;
+ if (pdev->vendor == PCI_VENDOR_ID_GOOGLE) {
+ int mem_size = nvme_vendor_memory_size(dev);
+ dev->db_mem = dma_alloc_coherent(&pdev->dev, mem_size,
&dev->doorbell, GFP_KERNEL);
+ if (!dev->db_mem) {
+ result = -ENOMEM;
+ goto unmap;
+ }
+ dev->ei_mem = dma_alloc_coherent(&pdev->dev, mem_size,
&dev->eventidx, GFP_KERNEL);
+ if (!dev->ei_mem) {
+ result = -ENOMEM;
+ goto dma_free;
+ }
+ }
+
return 0;
+ dma_free:
+ dma_free_coherent(&pdev->dev, nvme_vendor_memory_size(dev),
dev->db_mem, dev->doorbell);
+ dev->db_mem = NULL;
unmap:
iounmap(dev->bar);
dev->bar = NULL;
@@ -2391,6 +2494,11 @@ static int nvme_dev_map(struct nvme_dev *dev)
static void nvme_dev_unmap(struct nvme_dev *dev)
{
+ int mem_size = nvme_vendor_memory_size(dev);
+ if (!dev->db_mem)
+ dma_free_coherent(&dev->pci_dev->dev, mem_size,
dev->db_mem, dev->doorbell);
+ if (!dev->ei_mem)
+ dma_free_coherent(&dev->pci_dev->dev, mem_size,
dev->ei_mem, dev->eventidx);
if (dev->pci_dev->msi_enabled)
pci_disable_msi(dev->pci_dev);
else if (dev->pci_dev->msix_enabled)
diff --git a/include/linux/nvme.h b/include/linux/nvme.h
index ed09074..5a67c79 100644
--- a/include/linux/nvme.h
+++ b/include/linux/nvme.h
@@ -104,6 +104,10 @@ struct nvme_dev {
u8 event_limit;
u8 vwc;
u8 initialized;
+ u32* db_mem;
+ dma_addr_t doorbell;
+ u32* ei_mem;
+ dma_addr_t eventidx;
};
/*
diff --git a/include/uapi/linux/nvme.h b/include/uapi/linux/nvme.h
index 134518b..e5d6e1f 100644
--- a/include/uapi/linux/nvme.h
+++ b/include/uapi/linux/nvme.h
@@ -285,6 +285,7 @@ enum nvme_admin_opcode {
nvme_admin_format_nvm = 0x80,
nvme_admin_security_send = 0x81,
nvme_admin_security_recv = 0x82,
+ nvme_admin_doorbell_memory = 0xC0,
};
enum {
@@ -469,6 +470,7 @@ enum {
NVME_SC_BAD_ATTRIBUTES = 0x180,
NVME_SC_INVALID_PI = 0x181,
NVME_SC_READ_ONLY = 0x182,
+ NVME_SC_DOORBELL_MEMORY_INVALID = 0x1C0,
NVME_SC_WRITE_FAULT = 0x280,
NVME_SC_READ_ERROR = 0x281,
NVME_SC_GUARD_CHECK = 0x282,
--
cheers, Rob
More information about the Linux-nvme
mailing list