[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <MWHPR21MB15934EAD302E27983E891CF2D7739@MWHPR21MB1593.namprd21.prod.outlook.com>
Date: Fri, 9 Apr 2021 15:38:14 +0000
From: Michael Kelley <mikelley@...rosoft.com>
To: "Andrea Parri (Microsoft)" <parri.andrea@...il.com>,
"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
CC: KY Srinivasan <kys@...rosoft.com>,
Haiyang Zhang <haiyangz@...rosoft.com>,
Stephen Hemminger <sthemmin@...rosoft.com>,
"wei.liu@...nel.org" <wei.liu@...nel.org>,
"davem@...emloft.net" <davem@...emloft.net>,
"kuba@...nel.org" <kuba@...nel.org>,
"jejb@...ux.ibm.com" <jejb@...ux.ibm.com>,
"martin.petersen@...cle.com" <martin.petersen@...cle.com>,
"linux-hyperv@...r.kernel.org" <linux-hyperv@...r.kernel.org>,
"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
"linux-scsi@...r.kernel.org" <linux-scsi@...r.kernel.org>
Subject: RE: [RFC PATCH hyperv-next] scsi: storvsc: Use blk_mq_unique_tag() to
generate requestIDs
From: Andrea Parri (Microsoft) <parri.andrea@...il.com> Sent: Thursday, April 8, 2021 9:13 AM
>
> Use blk_mq_unique_tag() to generate requestIDs for StorVSC, avoiding
> all issues with allocating enough entries in the VMbus requestor.
This looks good to me! I'm glad to see that the idea worked without
too much complexity.
See a few comments inline below.
>
> Suggested-by: Michael Kelley <mikelley@...rosoft.com>
> Signed-off-by: Andrea Parri (Microsoft) <parri.andrea@...il.com>
> ---
> drivers/hv/channel.c | 14 +++---
> drivers/hv/ring_buffer.c | 12 ++---
> drivers/net/hyperv/netvsc.c | 8 ++--
> drivers/net/hyperv/rndis_filter.c | 2 +
> drivers/scsi/storvsc_drv.c | 73 ++++++++++++++++++++++++++-----
> include/linux/hyperv.h | 13 +++++-
> 6 files changed, 92 insertions(+), 30 deletions(-)
>
> diff --git a/drivers/hv/channel.c b/drivers/hv/channel.c
> index db30be8f9ccea..f78e02ace51e8 100644
> --- a/drivers/hv/channel.c
> +++ b/drivers/hv/channel.c
> @@ -1121,15 +1121,14 @@ EXPORT_SYMBOL_GPL(vmbus_recvpacket_raw);
> * vmbus_next_request_id - Returns a new request id. It is also
> * the index at which the guest memory address is stored.
> * Uses a spin lock to avoid race conditions.
> - * @rqstor: Pointer to the requestor struct
> + * @channel: Pointer to the VMbus channel struct
> * @rqst_add: Guest memory address to be stored in the array
> */
> -u64 vmbus_next_request_id(struct vmbus_requestor *rqstor, u64 rqst_addr)
> +u64 vmbus_next_request_id(struct vmbus_channel *channel, u64 rqst_addr)
> {
> + struct vmbus_requestor *rqstor = &channel->requestor;
> unsigned long flags;
> u64 current_id;
> - const struct vmbus_channel *channel =
> - container_of(rqstor, const struct vmbus_channel, requestor);
>
> /* Check rqstor has been initialized */
> if (!channel->rqstor_size)
> @@ -1163,16 +1162,15 @@ EXPORT_SYMBOL_GPL(vmbus_next_request_id);
> /*
> * vmbus_request_addr - Returns the memory address stored at @trans_id
> * in @rqstor. Uses a spin lock to avoid race conditions.
> - * @rqstor: Pointer to the requestor struct
> + * @channel: Pointer to the VMbus channel struct
> * @trans_id: Request id sent back from Hyper-V. Becomes the requestor's
> * next request id.
> */
> -u64 vmbus_request_addr(struct vmbus_requestor *rqstor, u64 trans_id)
> +u64 vmbus_request_addr(struct vmbus_channel *channel, u64 trans_id)
> {
> + struct vmbus_requestor *rqstor = &channel->requestor;
> unsigned long flags;
> u64 req_addr;
> - const struct vmbus_channel *channel =
> - container_of(rqstor, const struct vmbus_channel, requestor);
>
> /* Check rqstor has been initialized */
> if (!channel->rqstor_size)
> diff --git a/drivers/hv/ring_buffer.c b/drivers/hv/ring_buffer.c
> index ecd82ebfd5bc4..46d8e038e4ee1 100644
> --- a/drivers/hv/ring_buffer.c
> +++ b/drivers/hv/ring_buffer.c
> @@ -310,10 +310,12 @@ int hv_ringbuffer_write(struct vmbus_channel *channel,
> */
>
> if (desc->flags == VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED) {
> - rqst_id = vmbus_next_request_id(&channel->requestor, requestid);
> - if (rqst_id == VMBUS_RQST_ERROR) {
> - spin_unlock_irqrestore(&outring_info->ring_lock, flags);
> - return -EAGAIN;
> + if (channel->next_request_id_callback != NULL) {
> + rqst_id = channel->next_request_id_callback(channel, requestid);
> + if (rqst_id == VMBUS_RQST_ERROR) {
> + spin_unlock_irqrestore(&outring_info->ring_lock, flags);
> + return -EAGAIN;
> + }
> }
> }
> desc = hv_get_ring_buffer(outring_info) + old_write;
> @@ -341,7 +343,7 @@ int hv_ringbuffer_write(struct vmbus_channel *channel,
> if (channel->rescind) {
> if (rqst_id != VMBUS_NO_RQSTOR) {
> /* Reclaim request ID to avoid leak of IDs */
> - vmbus_request_addr(&channel->requestor, rqst_id);
> + channel->request_addr_callback(channel, rqst_id);
> }
> return -ENODEV;
> }
> diff --git a/drivers/net/hyperv/netvsc.c b/drivers/net/hyperv/netvsc.c
> index c64cc7639c39c..1a221ce2d6fdc 100644
> --- a/drivers/net/hyperv/netvsc.c
> +++ b/drivers/net/hyperv/netvsc.c
> @@ -730,7 +730,7 @@ static void netvsc_send_tx_complete(struct net_device *ndev,
> int queue_sends;
> u64 cmd_rqst;
>
> - cmd_rqst = vmbus_request_addr(&channel->requestor, (u64)desc->trans_id);
> + cmd_rqst = channel->request_addr_callback(channel, (u64)desc->trans_id);
> if (cmd_rqst == VMBUS_RQST_ERROR) {
> netdev_err(ndev, "Incorrect transaction id\n");
> return;
> @@ -790,8 +790,8 @@ static void netvsc_send_completion(struct net_device *ndev,
>
> /* First check if this is a VMBUS completion without data payload */
> if (!msglen) {
> - cmd_rqst = vmbus_request_addr(&incoming_channel->requestor,
> - (u64)desc->trans_id);
> + cmd_rqst = incoming_channel->request_addr_callback(incoming_channel,
> + (u64)desc->trans_id);
> if (cmd_rqst == VMBUS_RQST_ERROR) {
> netdev_err(ndev, "Invalid transaction id\n");
> return;
> @@ -1602,6 +1602,8 @@ struct netvsc_device *netvsc_device_add(struct hv_device
> *device,
> netvsc_poll, NAPI_POLL_WEIGHT);
>
> /* Open the channel */
> + device->channel->next_request_id_callback = vmbus_next_request_id;
> + device->channel->request_addr_callback = vmbus_request_addr;
> device->channel->rqstor_size = netvsc_rqstor_size(netvsc_ring_bytes);
> ret = vmbus_open(device->channel, netvsc_ring_bytes,
> netvsc_ring_bytes, NULL, 0,
> diff --git a/drivers/net/hyperv/rndis_filter.c b/drivers/net/hyperv/rndis_filter.c
> index 123cc9d25f5ed..ebf34bf3f9075 100644
> --- a/drivers/net/hyperv/rndis_filter.c
> +++ b/drivers/net/hyperv/rndis_filter.c
> @@ -1259,6 +1259,8 @@ static void netvsc_sc_open(struct vmbus_channel *new_sc)
> /* Set the channel before opening.*/
> nvchan->channel = new_sc;
>
> + new_sc->next_request_id_callback = vmbus_next_request_id;
> + new_sc->request_addr_callback = vmbus_request_addr;
> new_sc->rqstor_size = netvsc_rqstor_size(netvsc_ring_bytes);
> ret = vmbus_open(new_sc, netvsc_ring_bytes,
> netvsc_ring_bytes, NULL, 0,
> diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c
> index 6bc5453cea8a7..1c05fabc06b04 100644
> --- a/drivers/scsi/storvsc_drv.c
> +++ b/drivers/scsi/storvsc_drv.c
> @@ -684,6 +684,62 @@ static void storvsc_change_target_cpu(struct vmbus_channel
> *channel, u32 old,
> spin_unlock_irqrestore(&stor_device->lock, flags);
> }
>
> +u64 storvsc_next_request_id(struct vmbus_channel *channel, u64 rqst_addr)
> +{
> + struct storvsc_cmd_request *request =
> + (struct storvsc_cmd_request *)(unsigned long)rqst_addr;
> + struct storvsc_device *stor_device;
> + struct hv_device *device;
> +
> + device = (channel->primary_channel != NULL) ?
> + channel->primary_channel->device_obj : channel->device_obj;
> + if (device == NULL)
> + return VMBUS_RQST_ERROR;
> +
> + stor_device = get_out_stor_device(device);
> + if (stor_device == NULL)
> + return VMBUS_RQST_ERROR;
> +
> + if (request == &stor_device->init_request)
> + return VMBUS_RQST_INIT;
> + if (request == &stor_device->reset_request)
> + return VMBUS_RQST_RESET;
Having to get the device and then the stor_device in order to detect the
init_request and reset_request special cases is unfortunate. So here's
an idea: The init_request and reset_request are used in a limited number
of specific places in the storvsc driver, and there are unique invocations
of vmbus_sendpacket() in those places. So rather than pass the address
of the request as the requestID parameter to vmbus_sendpacket(), pass
the sentinel value VMBUS_RQST_INIT or VMBUS_RQST_RESET. Then this
code can just detect those sentinel values as the rqst_addr input
parameter, and return them.
> +
> + return blk_mq_unique_tag(request->cmd->request);
> +}
> +
> +u64 storvsc_request_addr(struct vmbus_channel *channel, u64 rqst_id)
> +{
> + struct storvsc_cmd_request *request;
> + struct storvsc_device *stor_device;
> + struct hv_device *device;
> + struct Scsi_Host *shost;
> + struct scsi_cmnd *scmnd;
> +
> + device = (channel->primary_channel != NULL) ?
> + channel->primary_channel->device_obj : channel->device_obj;
> + if (device == NULL)
> + return VMBUS_RQST_ERROR;
> +
> + stor_device = get_out_stor_device(device);
> + if (stor_device == NULL)
> + return VMBUS_RQST_ERROR;
> +
> + if (rqst_id == VMBUS_RQST_INIT)
> + return (unsigned long)&stor_device->init_request;
> + if (rqst_id == VMBUS_RQST_RESET)
> + return (unsigned long)&stor_device->reset_request;
Unfortunately, the same simplification doesn't work here. And you need
stor_device anyway to get the scsi_host.
> +
> + shost = stor_device->host;
> +
> + scmnd = scsi_host_find_tag(shost, rqst_id);
> + if (scmnd == NULL)
> + return VMBUS_RQST_ERROR;
> +
> + request = (struct storvsc_cmd_request *)(unsigned long)scsi_cmd_priv(scmnd);
> + return (unsigned long)request;
The casts in the above two lines seem unnecessarily complex. 'request' is never
used as a pointer. So couldn't the last two lines just be:
return (unsigned long)scsi_cmd_priv(scmnd);
> +}
> +
> static void handle_sc_creation(struct vmbus_channel *new_sc)
> {
> struct hv_device *device = new_sc->primary_channel->device_obj;
> @@ -698,11 +754,8 @@ static void handle_sc_creation(struct vmbus_channel *new_sc)
>
> memset(&props, 0, sizeof(struct vmstorage_channel_properties));
>
> - /*
> - * The size of vmbus_requestor is an upper bound on the number of requests
> - * that can be in-progress at any one time across all channels.
> - */
> - new_sc->rqstor_size = scsi_driver.can_queue;
> + new_sc->next_request_id_callback = storvsc_next_request_id;
> + new_sc->request_addr_callback = storvsc_request_addr;
>
> ret = vmbus_open(new_sc,
> storvsc_ringbuffer_size,
> @@ -1255,8 +1308,7 @@ static void storvsc_on_channel_callback(void *context)
> struct storvsc_cmd_request *request;
> u64 cmd_rqst;
>
> - cmd_rqst = vmbus_request_addr(&channel->requestor,
> - desc->trans_id);
> + cmd_rqst = channel->request_addr_callback(channel, desc->trans_id);
Here's another thought: You don't really need to set the channel request_addr_callback
function and then indirect through it here. You know the specific function that storvsc
is using, so could call it directly. The other reason to set request_addr_callback is so
that at the end of hv_ringbuffer_write() you can reclaim an allocated requestID if the
rescind flag is set. But there's nothing allocated that needs to be reclaimed in the storvsc
case, so leaving request_addr_callback as NULL is OK (but hv_ringbuffer_write would
have to check for the NULL).
Then if you do that, the logic in storvsc_request_addr() can effectively go inline in
here. And that logic can take advantage of the fact that stor_device is already determined
outside the foreach_vmbus_pkt() loop. The scsi_host could be calculated outside the loop
as well, leaving the detection of init_request and reset_request, and the call to
scsi_host_find_tag() as the only things to do.
This approach is a bit asymmetrical, but it would save some processing in this interrupt
handling code. So something to consider.
> if (cmd_rqst == VMBUS_RQST_ERROR) {
> dev_err(&device->device,
> "Incorrect transaction id\n");
> @@ -1290,11 +1342,8 @@ static int storvsc_connect_to_vsp(struct hv_device *device, u32
> ring_size,
>
> memset(&props, 0, sizeof(struct vmstorage_channel_properties));
>
> - /*
> - * The size of vmbus_requestor is an upper bound on the number of requests
> - * that can be in-progress at any one time across all channels.
> - */
> - device->channel->rqstor_size = scsi_driver.can_queue;
> + device->channel->next_request_id_callback = storvsc_next_request_id;
> + device->channel->request_addr_callback = storvsc_request_addr;
>
> ret = vmbus_open(device->channel,
> ring_size,
> diff --git a/include/linux/hyperv.h b/include/linux/hyperv.h
> index 2c18c8e768efe..5692ffa60e022 100644
> --- a/include/linux/hyperv.h
> +++ b/include/linux/hyperv.h
> @@ -779,7 +779,11 @@ struct vmbus_requestor {
>
> #define VMBUS_NO_RQSTOR U64_MAX
> #define VMBUS_RQST_ERROR (U64_MAX - 1)
> +/* NetVSC-specific */
It is netvsc specific at the moment. But if we harden other
drivers, they are likely to use the same generic requestID
allocator, and hence need the same sentinel value.
> #define VMBUS_RQST_ID_NO_RESPONSE (U64_MAX - 2)
> +/* StorVSC-specific */
> +#define VMBUS_RQST_INIT (U64_MAX - 2)
> +#define VMBUS_RQST_RESET (U64_MAX - 3)
>
> struct vmbus_device {
> u16 dev_type;
> @@ -1007,13 +1011,18 @@ struct vmbus_channel {
> u32 fuzz_testing_interrupt_delay;
> u32 fuzz_testing_message_delay;
>
> + /* callback to generate a request ID from a request address */
> + u64 (*next_request_id_callback)(struct vmbus_channel *channel, u64 rqst_addr);
> + /* callback to retrieve a request address from a request ID */
> + u64 (*request_addr_callback)(struct vmbus_channel *channel, u64 rqst_id);
> +
> /* request/transaction ids for VMBus */
> struct vmbus_requestor requestor;
> u32 rqstor_size;
> };
>
> -u64 vmbus_next_request_id(struct vmbus_requestor *rqstor, u64 rqst_addr);
> -u64 vmbus_request_addr(struct vmbus_requestor *rqstor, u64 trans_id);
> +u64 vmbus_next_request_id(struct vmbus_channel *channel, u64 rqst_addr);
> +u64 vmbus_request_addr(struct vmbus_channel *channel, u64 trans_id);
>
> static inline bool is_hvsock_channel(const struct vmbus_channel *c)
> {
> --
> 2.25.1
Powered by blists - more mailing lists