vhost-scsi: Change def inline_sg_cnt and max_io_vqs for exadata

Upstream we set inline_sg_cnt and max_io_vqs based on their
original values and concerns about regressions:

max_io_vqs - If a user tries to create more than max_io_vqs then
the device setup will fail and QEMU will fail to start. Upstream
this is 128 because since mq was added this was the default. If
we change it too low, then VMs will fail to start.

For exadata, this is their first time using vhost-scsi so we can
set it low. Their primary concern is mem use so we set it to 1.

inline_sg_cnt - If a user has a really high perf setup (multiple
LUNs per vhost-scsi device with fast backend devices) then there
was a concern they could see a regression if we lower this too
low. Upstream it was left as the original value.

For exadata, their primary concern is mem use so we set this to
not preallocate any memory for command's scatterlists. We will
use whatever memory is available when performing IO.

Orabug: 38099086

Signed-off-by: Mike Christie <[email protected]>
Reviewed-by: Boris Ostrovsky <[email protected]>
Signed-off-by: Brian Maly <[email protected]>

Author: mikechristie

drivers/vhost/scsi.c

@@ -51,7 +51,7 @@
  */
 #define VHOST_SCSI_MAX_RESP_IOVS sizeof(struct virtio_scsi_cmd_resp)
 
-static unsigned int vhost_scsi_inline_sg_cnt = VHOST_SCSI_PREALLOC_SGLS;
+static unsigned int vhost_scsi_inline_sg_cnt = UINT_MAX;
 
 #ifdef CONFIG_ARCH_NO_SG_CHAIN
 static int vhost_scsi_set_inline_sg_cnt(const char *buf,
@@ -199,10 +199,11 @@ enum {
 };
 
 #define VHOST_SCSI_MAX_TARGET	256
+#define VHOST_SCSI_DEF_IO_VQ	128
 #define VHOST_SCSI_MAX_IO_VQ	1024
 #define VHOST_SCSI_MAX_EVENT	128
 
-static unsigned vhost_scsi_max_io_vqs = 128;
+static unsigned vhost_scsi_max_io_vqs = UINT_MAX;
 module_param_named(max_io_vqs, vhost_scsi_max_io_vqs, uint, 0644);
 MODULE_PARM_DESC(max_io_vqs, "Set the max number of IO virtqueues a vhost scsi device can support. The default is 128. The max is 1024.");
 
@@ -2811,6 +2812,35 @@ static int __init vhost_scsi_init(void)
 {
 	int ret = -ENOMEM;
 
+
+	if (static_branch_unlikely(&on_exadata)) {
+		/*
+		 * We can't change the upstream default because existing users
+		 * with more than 1 queue will hit failures during device
+		 * setup. Exadata has never used vhost-scsi before so it's ok
+		 * to start them at 1 so they can use as little mem as possible
+		 * initially.
+		 */
+		if (vhost_scsi_max_io_vqs == UINT_MAX)
+			vhost_scsi_max_io_vqs = 1;
+		/*
+		 * We have not changed the upstream default because there's
+		 * a chance of a perf regression if there are users doing
+		 * muiltiple LUNs per vhost-scsi device and using high perf
+		 * devices for the backends. For exadata we use 0, because
+		 * their primary concern is memory use.
+		 */
+		if (vhost_scsi_inline_sg_cnt == UINT_MAX)
+			vhost_scsi_inline_sg_cnt = 0;
+	} else {
+		if (vhost_scsi_max_io_vqs == UINT_MAX)
+			vhost_scsi_max_io_vqs = VHOST_SCSI_DEF_IO_VQ;
+
+		if (vhost_scsi_inline_sg_cnt == UINT_MAX)
+			vhost_scsi_inline_sg_cnt = VHOST_SCSI_PREALLOC_SGLS;
+	}
+
+
 	pr_debug("TCM_VHOST fabric module %s on %s/%s"
 		" on "UTS_RELEASE"\n", VHOST_SCSI_VERSION, utsname()->sysname,
 		utsname()->machine);