Skip to contentSkip to conversation

nvme: allow cmb and pmr emulation on same device

PATCH In review Andrzej Jakowski 3 participants 8 messages Jul 29, 2020 – Sep 28, 2020

lore
Share & tools
lore
#0 lore raw

PATCH v6 1/1 nvme: allow cmb and pmr emulation on same device

Resending series recently posted on mailing list related to nvme device extension with couple of fixes after review.

This patch series does following:
- Fixes problem where CMBS bit was not set in controller capabilities register, so support for CMB was not correctly advertised to guest. This is resend of patch that has been submitted and reviewed by Klaus [1]
- Introduces BAR4 sharing between MSI-X vectors and CMB. This allows to have CMB and PMR emulated on the same device. This extension was indicated by Keith [2]

v6:
- instead of using memory_region_to_absolute_addr() function local helper has been defined (nvme_cmb_to_absolute_addr()) to calculate absolute address of CMB in simplified way. Also a number of code style changes has been done (function rename, use literal instead of macro definition, etc.)

v5:
- fixed problem introduced in v4 where CMB buffer was represented as subregion of BAR4 memory region. In that case CMB address was used incorrectly as it was relative to BAR4 and not absolute. Appropriate changes were added to v5 to calculate CMB address properly ([6])

v4:
- modified BAR4 initialization, so now it consists of CMB, MSIX and PBA memory regions overlapping on top of it. This reduces patch complexity significantly (Klaus [5])

v3:
- code style fixes including: removal of spurious line break, moving define into define section and code alignment (Klaus [4])
- removed unintended code reintroduction (Klaus [4])

v2:
- rebase on Kevin's latest block branch (Klaus [3])
- improved comments section (Klaus [3])
- simplified calculation of BAR4 size (Klaus [3])

v1:
- initial push of the patch

[1]: https://lore.kernel.org/qemu-devel/[email protected]/
[2]: https://lore.kernel.org/qemu-devel/[email protected]/
[3]: https://lore.kernel.org/qemu-devel/[email protected]/
[4]: https://lore.kernel.org/qemu-devel/[email protected]/
[5]: https://lore.kernel.org/qemu-devel/[email protected]/
[6]: https://lore.kernel.org/qemu-devel/[email protected]/

#1 lore raw

PATCH v6 1/2 nvme: indicate CMB support through controller capabilities register

This patch sets CMBS bit in controller capabilities register when user configures NVMe driver with CMB support, so capabilites are correctly reported to guest OS.

---
 hw/block/nvme.c      |  1 +
 include/block/nvme.h | 10 +++++++---
 2 files changed, 8 insertions(+), 3 deletions(-)
hw/block/nvme.c +1 -0
index 841c18920c..43866b744f 100644
@@ -2198,6 +2198,7 @@ static void nvme_init_ctrl(NvmeCtrl *n, PCIDevice *pci_dev)
     NVME_CAP_SET_TO(n->bar.cap, 0xf);
     NVME_CAP_SET_CSS(n->bar.cap, 1);
     NVME_CAP_SET_MPSMAX(n->bar.cap, 4);
+    NVME_CAP_SET_CMBS(n->bar.cap, n->params.cmb_size_mb ? 1 : 0);
 
     n->bar.vs = NVME_SPEC_VER;
     n->bar.intmc = n->bar.intms = 0;
include/block/nvme.h +7 -3
index 370df7fc05..d641ca6649 100644
@@ -36,6 +36,7 @@ enum NvmeCapShift {
     CAP_MPSMIN_SHIFT   = 48,
     CAP_MPSMAX_SHIFT   = 52,
     CAP_PMR_SHIFT      = 56,
+    CAP_CMB_SHIFT      = 57,
 };
 
 enum NvmeCapMask {
@@ -49,6 +50,7 @@ enum NvmeCapMask {
     CAP_MPSMIN_MASK    = 0xf,
     CAP_MPSMAX_MASK    = 0xf,
     CAP_PMR_MASK       = 0x1,
+    CAP_CMB_MASK       = 0x1,
 };
 
 #define NVME_CAP_MQES(cap)  (((cap) >> CAP_MQES_SHIFT)   & CAP_MQES_MASK)
@@ -78,9 +80,11 @@ enum NvmeCapMask {
 #define NVME_CAP_SET_MPSMIN(cap, val) (cap |= (uint64_t)(val & CAP_MPSMIN_MASK)\
                                                            << CAP_MPSMIN_SHIFT)
 #define NVME_CAP_SET_MPSMAX(cap, val) (cap |= (uint64_t)(val & CAP_MPSMAX_MASK)\
-                                                            << CAP_MPSMAX_SHIFT)
-#define NVME_CAP_SET_PMRS(cap, val) (cap |= (uint64_t)(val & CAP_PMR_MASK)\
-                                                            << CAP_PMR_SHIFT)
+                                                           << CAP_MPSMAX_SHIFT)
+#define NVME_CAP_SET_PMRS(cap, val)   (cap |= (uint64_t)(val & CAP_PMR_MASK)   \
+                                                           << CAP_PMR_SHIFT)
+#define NVME_CAP_SET_CMBS(cap, val)   (cap |= (uint64_t)(val & CAP_CMB_MASK)   \
+                                                           << CAP_CMB_SHIFT)
 
 enum NvmeCcShift {
     CC_EN_SHIFT     = 0,
Signature · 1 line
2.25.4
#2 lore raw

PATCH v6 2/2 nvme: allow cmb and pmr to be enabled on same device

In reply to Andrzej Jakowski

So far it was not possible to have CMB and PMR emulated on the same device, because BAR2 was used exclusively either of PMR or CMB. This patch places CMB at BAR4 offset so it not conflicts with MSI-X vectors.

---
 hw/block/nvme.c      | 124 +++++++++++++++++++++++++++++--------------
 hw/block/nvme.h      |   1 +
 include/block/nvme.h |   4 +-
 3 files changed, 89 insertions(+), 40 deletions(-)
hw/block/nvme.c +85 -39
index 43866b744f..292bca445f 100644
@@ -22,12 +22,13 @@
  *              [pmrdev=<mem_backend_file_id>,] \
  *              max_ioqpairs=<N[optional]>
  *
- * Note cmb_size_mb denotes size of CMB in MB. CMB is assumed to be at
- * offset 0 in BAR2 and supports only WDS, RDS and SQS for now.
+ * Note cmb_size_mb denotes size of CMB in MB. CMB when configured is assumed
+ * to be resident in BAR4 at offset that is 2MiB aligned. When CMB is emulated
+ * on Linux guest it is recommended to make cmb_size_mb multiple of 2. Both
+ * size and alignment restrictions are imposed by Linux guest.
  *
- * cmb_size_mb= and pmrdev= options are mutually exclusive due to limitation
- * in available BAR's. cmb_size_mb= will take precedence over pmrdev= when
- * both provided.
+ * pmrdev is assumed to be resident in BAR2/BAR3. When configured it consumes
+ * whole BAR2/BAR3 exclusively.
  * Enabling pmr emulation can be achieved by pointing to memory-backend-file.
  * For example:
  * -object memory-backend-file,id=<mem_id>,share=on,mem-path=<file_path>, \
@@ -57,7 +58,6 @@
 #define NVME_MAX_IOQPAIRS 0xffff
 #define NVME_DB_SIZE  4
 #define NVME_SPEC_VER 0x00010300
-#define NVME_CMB_BIR 2
 #define NVME_PMR_BIR 2
 #define NVME_TEMPERATURE 0x143
 #define NVME_TEMPERATURE_WARNING 0x157
@@ -109,18 +109,25 @@ static uint16_t nvme_sqid(NvmeRequest *req)
     return le16_to_cpu(req->sq->sqid);
 }
 
+static inline hwaddr nvme_cmb_to_absolute_addr(NvmeCtrl *n)
+{
+    return n->bar4.addr + n->ctrl_mem.addr;
+}
+
 static bool nvme_addr_is_cmb(NvmeCtrl *n, hwaddr addr)
 {
-    hwaddr low = n->ctrl_mem.addr;
-    hwaddr hi  = n->ctrl_mem.addr + int128_get64(n->ctrl_mem.size);
+    hwaddr low = nvme_cmb_to_absolute_addr(n);
+    hwaddr hi  = low + int128_get64(n->ctrl_mem.size);
 
     return addr >= low && addr < hi;
 }
 
 static void nvme_addr_read(NvmeCtrl *n, hwaddr addr, void *buf, int size)
 {
+    hwaddr cmb_addr = nvme_cmb_to_absolute_addr(n);
+
     if (n->bar.cmbsz && nvme_addr_is_cmb(n, addr)) {
-        memcpy(buf, (void *)&n->cmbuf[addr - n->ctrl_mem.addr], size);
+        memcpy(buf, (void *)&n->cmbuf[addr - cmb_addr], size);
         return;
     }
 
@@ -207,17 +214,18 @@ static uint16_t nvme_map_prp(QEMUSGList *qsg, QEMUIOVector *iov, uint64_t prp1,
                              uint64_t prp2, uint32_t len, NvmeCtrl *n)
 {
     hwaddr trans_len = n->page_size - (prp1 % n->page_size);
+    hwaddr cmb_addr = nvme_cmb_to_absolute_addr(n);
     trans_len = MIN(len, trans_len);
     int num_prps = (len >> n->page_bits) + 1;
 
     if (unlikely(!prp1)) {
         trace_pci_nvme_err_invalid_prp();
         return NVME_INVALID_FIELD | NVME_DNR;
-    } else if (n->bar.cmbsz && prp1 >= n->ctrl_mem.addr &&
-               prp1 < n->ctrl_mem.addr + int128_get64(n->ctrl_mem.size)) {
+    } else if (n->bar.cmbsz && prp1 >= cmb_addr &&
+               prp1 < cmb_addr + int128_get64(n->ctrl_mem.size)) {
         qsg->nsg = 0;
         qemu_iovec_init(iov, num_prps);
-        qemu_iovec_add(iov, (void *)&n->cmbuf[prp1 - n->ctrl_mem.addr], trans_len);
+        qemu_iovec_add(iov, (void *)&n->cmbuf[prp1 - cmb_addr], trans_len);
     } else {
         pci_dma_sglist_init(qsg, &n->parent_obj, num_prps);
         qemu_sglist_add(qsg, prp1, trans_len);
@@ -262,7 +270,7 @@ static uint16_t nvme_map_prp(QEMUSGList *qsg, QEMUIOVector *iov, uint64_t prp1,
                 if (qsg->nsg){
                     qemu_sglist_add(qsg, prp_ent, trans_len);
                 } else {
-                    qemu_iovec_add(iov, (void *)&n->cmbuf[prp_ent - n->ctrl_mem.addr], trans_len);
+                    qemu_iovec_add(iov, (void *)&n->cmbuf[prp_ent - cmb_addr], trans_len);
                 }
                 len -= trans_len;
                 i++;
@@ -275,7 +283,7 @@ static uint16_t nvme_map_prp(QEMUSGList *qsg, QEMUIOVector *iov, uint64_t prp1,
             if (qsg->nsg) {
                 qemu_sglist_add(qsg, prp2, len);
             } else {
-                qemu_iovec_add(iov, (void *)&n->cmbuf[prp2 - n->ctrl_mem.addr], trans_len);
+                qemu_iovec_add(iov, (void *)&n->cmbuf[prp2 - cmb_addr], trans_len);
             }
         }
     }
@@ -1980,7 +1988,7 @@ static void nvme_check_constraints(NvmeCtrl *n, Error **errp)
         return;
     }
 
-    if (!n->params.cmb_size_mb && n->pmrdev) {
+    if (n->pmrdev) {
         if (host_memory_backend_is_mapped(n->pmrdev)) {
             char *path = object_get_canonical_path_component(OBJECT(n->pmrdev));
             error_setg(errp, "can't use already busy memdev: %s", path);
@@ -2042,33 +2050,73 @@ static void nvme_init_namespace(NvmeCtrl *n, NvmeNamespace *ns, Error **errp)
     id_ns->nuse = id_ns->ncap;
 }
 
-static void nvme_init_cmb(NvmeCtrl *n, PCIDevice *pci_dev)
+static void nvme_init_bar4(PCIDevice *pci_dev, Error **errp)
 {
-    NVME_CMBLOC_SET_BIR(n->bar.cmbloc, NVME_CMB_BIR);
-    NVME_CMBLOC_SET_OFST(n->bar.cmbloc, 0);
+    NvmeCtrl *n = NVME(pci_dev);
+    int status;
+    uint64_t bar_size, cmb_offset = 0;
+    uint32_t msix_vectors;
+    uint32_t nvme_pba_offset;
+    uint32_t cmb_size_units;
+
+    msix_vectors = n->params.msix_qsize;
+    nvme_pba_offset = PCI_MSIX_ENTRY_SIZE * msix_vectors;
+    bar_size = nvme_pba_offset + QEMU_ALIGN_UP(msix_vectors, 64) / 8;
+
+    if (n->params.cmb_size_mb) {
+        NVME_CMBSZ_SET_SQS(n->bar.cmbsz, 1);
+        NVME_CMBSZ_SET_CQS(n->bar.cmbsz, 0);
+        NVME_CMBSZ_SET_LISTS(n->bar.cmbsz, 0);
+        NVME_CMBSZ_SET_RDS(n->bar.cmbsz, 1);
+        NVME_CMBSZ_SET_WDS(n->bar.cmbsz, 1);
+        NVME_CMBSZ_SET_SZU(n->bar.cmbsz, 2); /* MBs */
+        NVME_CMBSZ_SET_SZ(n->bar.cmbsz, n->params.cmb_size_mb);
+
+        cmb_size_units = NVME_CMBSZ_GETSIZEUNITS(n->bar.cmbsz);
+        /* Linux guest requires it to be 2MiB aligned */
+        cmb_offset = QEMU_ALIGN_UP(bar_size, 2 * MiB);
+
+        NVME_CMBLOC_SET_BIR(n->bar.cmbloc, 4);
+        NVME_CMBLOC_SET_OFST(n->bar.cmbloc, cmb_offset / cmb_size_units);
+
+        n->cmbuf = g_malloc0(NVME_CMBSZ_GETSIZE(n->bar.cmbsz));
 
-    NVME_CMBSZ_SET_SQS(n->bar.cmbsz, 1);
-    NVME_CMBSZ_SET_CQS(n->bar.cmbsz, 0);
-    NVME_CMBSZ_SET_LISTS(n->bar.cmbsz, 0);
-    NVME_CMBSZ_SET_RDS(n->bar.cmbsz, 1);
-    NVME_CMBSZ_SET_WDS(n->bar.cmbsz, 1);
-    NVME_CMBSZ_SET_SZU(n->bar.cmbsz, 2); /* MBs */
-    NVME_CMBSZ_SET_SZ(n->bar.cmbsz, n->params.cmb_size_mb);
+        bar_size += cmb_offset;
+        bar_size += NVME_CMBSZ_GETSIZE(n->bar.cmbsz);
+    }
+
+    bar_size = pow2ceil(bar_size);
 
-    n->cmbuf = g_malloc0(NVME_CMBSZ_GETSIZE(n->bar.cmbsz));
-    memory_region_init_io(&n->ctrl_mem, OBJECT(n), &nvme_cmb_ops, n,
-                          "nvme-cmb", NVME_CMBSZ_GETSIZE(n->bar.cmbsz));
-    pci_register_bar(pci_dev, NVME_CMBLOC_BIR(n->bar.cmbloc),
+    /*
+     * Create memory region for BAR4, then overlap cmb, msix and pba
+     * tables on top of it.
+     */
+    memory_region_init(&n->bar4, OBJECT(n), "nvme-bar4", bar_size);
+
+    if (n->params.cmb_size_mb) {
+        memory_region_init_io(&n->ctrl_mem, OBJECT(n), &nvme_cmb_ops, n,
+                              "nvme-cmb", NVME_CMBSZ_GETSIZE(n->bar.cmbsz));
+
+        memory_region_add_subregion(&n->bar4, cmb_offset, &n->ctrl_mem);
+    }
+
+    status = msix_init(pci_dev, n->params.msix_qsize,
+                       &n->bar4, 4, 0,
+                       &n->bar4, 4, nvme_pba_offset,
+                       0, errp);
+
+    if (status) {
+        return;
+    }
+
+    pci_register_bar(pci_dev, 4,
                      PCI_BASE_ADDRESS_SPACE_MEMORY |
                      PCI_BASE_ADDRESS_MEM_TYPE_64 |
-                     PCI_BASE_ADDRESS_MEM_PREFETCH, &n->ctrl_mem);
+                     PCI_BASE_ADDRESS_MEM_PREFETCH, &n->bar4);
 }
 
 static void nvme_init_pmr(NvmeCtrl *n, PCIDevice *pci_dev)
 {
-    /* Controller Capabilities register */
-    NVME_CAP_SET_PMRS(n->bar.cap, 1);
-
     /* PMR Capabities register */
     n->bar.pmrcap = 0;
     NVME_PMRCAP_SET_RDS(n->bar.pmrcap, 0);
@@ -2126,13 +2174,10 @@ static void nvme_init_pci(NvmeCtrl *n, PCIDevice *pci_dev, Error **errp)
                           n->reg_size);
     pci_register_bar(pci_dev, 0, PCI_BASE_ADDRESS_SPACE_MEMORY |
                      PCI_BASE_ADDRESS_MEM_TYPE_64, &n->iomem);
-    if (msix_init_exclusive_bar(pci_dev, n->params.msix_qsize, 4, errp)) {
-        return;
-    }
 
-    if (n->params.cmb_size_mb) {
-        nvme_init_cmb(n, pci_dev);
-    } else if (n->pmrdev) {
+    nvme_init_bar4(pci_dev, errp);
+
+    if (n->pmrdev) {
         nvme_init_pmr(n, pci_dev);
     }
 }
@@ -2199,6 +2244,7 @@ static void nvme_init_ctrl(NvmeCtrl *n, PCIDevice *pci_dev)
     NVME_CAP_SET_CSS(n->bar.cap, 1);
     NVME_CAP_SET_MPSMAX(n->bar.cap, 4);
     NVME_CAP_SET_CMBS(n->bar.cap, n->params.cmb_size_mb ? 1 : 0);
+    NVME_CAP_SET_PMRS(n->bar.cap, n->pmrdev ? 1 : 0);
 
     n->bar.vs = NVME_SPEC_VER;
     n->bar.intmc = n->bar.intms = 0;
hw/block/nvme.h +1 -0
index 0b6a8ae665..f291395cd0 100644
@@ -91,6 +91,7 @@ typedef struct NvmeCtrl {
     PCIDevice    parent_obj;
     MemoryRegion iomem;
     MemoryRegion ctrl_mem;
+    MemoryRegion bar4;
     NvmeBar      bar;
     BlockConf    conf;
     NvmeParams   params;
include/block/nvme.h +3 -1
index d641ca6649..77b59d18dd 100644
@@ -217,9 +217,11 @@ enum NvmeCmbszMask {
     (cmbsz |= (uint64_t)(val & CMBSZ_SZU_MASK) << CMBSZ_SZU_SHIFT)
 #define NVME_CMBSZ_SET_SZ(cmbsz, val)    \
     (cmbsz |= (uint64_t)(val & CMBSZ_SZ_MASK) << CMBSZ_SZ_SHIFT)
+#define NVME_CMBSZ_GETSIZEUNITS(cmbsz) \
+    (1 << (12 + 4 * NVME_CMBSZ_SZU(cmbsz)))
 
 #define NVME_CMBSZ_GETSIZE(cmbsz) \
-    (NVME_CMBSZ_SZ(cmbsz) * (1 << (12 + 4 * NVME_CMBSZ_SZU(cmbsz))))
+    (NVME_CMBSZ_SZ(cmbsz) * NVME_CMBSZ_GETSIZEUNITS(cmbsz))
 
 enum NvmePmrcapShift {
     PMRCAP_RDS_SHIFT      = 3,
Signature · 1 line
2.25.4
Reviewed-by on v6 2/2
#3 lore raw

Re: [PATCH v6 2/2] nvme: allow cmb and pmr to be enabled on same device

Quoted 6 lines · Andrzej Jakowski

So far it was not possible to have CMB and PMR emulated on the same device, because BAR2 was used exclusively either of PMR or CMB. This patch places CMB at BAR4 offset so it not conflicts with MSI-X vectors.

---

Well, I'm certainly happy now. LGTM!

#4 lore raw

Re: [PATCH v6 2/2] nvme: allow cmb and pmr to be enabled on same device

Quoted 5 lines · Klaus Jensen
Quoted 4 lines · Andrzej Jakowski

So far it was not possible to have CMB and PMR emulated on the same device, because BAR2 was used exclusively either of PMR or CMB. This patch places CMB at BAR4 offset so it not conflicts with MSI-X vectors.

Quoted 7 lines
Quoted 2 lines

---

Well, I'm certainly happy now. LGTM!

Are anyone willing to chip in with another review on this?

#5 lore raw

Re: [PATCH v6 2/2] nvme: allow cmb and pmr to be enabled on same device

Quoted 9 lines · Klaus Jensen
Quoted 8 lines · Klaus Jensen
Quoted 6 lines · Andrzej Jakowski

So far it was not possible to have CMB and PMR emulated on the same device, because BAR2 was used exclusively either of PMR or CMB. This patch places CMB at BAR4 offset so it not conflicts with MSI-X vectors.

---

Quoted 7 lines
Quoted 4 lines

Well, I'm certainly happy now. LGTM!

Are anyone willing to chip in with another review on this?

I think this patch is ready (and have been for some time) for inclusion, but would really like an additional review on this; preferably from Keith, since he is the one that originally mentioned that we could do something like this.

I've mentioned it before, but I would prefer that the MSI-X stuff was in BAR0 instead of mixing it with the CMB, but that's bikeshedding and my R-b still holds of course.

Reviewed-by
#6 lore raw

Re: [PATCH v6 2/2] nvme: allow cmb and pmr to be enabled on same device

Quoted 15 lines · Klaus Jensen
Quoted 14 lines · Klaus Jensen
Quoted 12 lines · Klaus Jensen
Quoted 6 lines · Andrzej Jakowski

So far it was not possible to have CMB and PMR emulated on the same device, because BAR2 was used exclusively either of PMR or CMB. This patch places CMB at BAR4 offset so it not conflicts with MSI-X vectors.

---

Well, I'm certainly happy now. LGTM!

Quoted 7 lines
Quoted 2 lines

Are anyone willing to chip in with another review on this?

I think this patch is ready (and have been for some time) for inclusion, but would really like an additional review on this; preferably from Keith, since he is the one that originally mentioned that we could do something like this.

Yes, this does look good to me too.

#7 lore raw
In reply to Andrzej Jakowski
Quoted 42 lines · Andrzej Jakowski

Resending series recently posted on mailing list related to nvme device extension with couple of fixes after review.

This patch series does following:
- Fixes problem where CMBS bit was not set in controller capabilities register, so support for CMB was not correctly advertised to guest. This is resend of patch that has been submitted and reviewed by Klaus [1]
- Introduces BAR4 sharing between MSI-X vectors and CMB. This allows to have CMB and PMR emulated on the same device. This extension was indicated by Keith [2]

v6:
- instead of using memory_region_to_absolute_addr() function local helper has been defined (nvme_cmb_to_absolute_addr()) to calculate absolute address of CMB in simplified way. Also a number of code style changes has been done (function rename, use literal instead of macro definition, etc.)

v5:
- fixed problem introduced in v4 where CMB buffer was represented as subregion of BAR4 memory region. In that case CMB address was used incorrectly as it was relative to BAR4 and not absolute. Appropriate changes were added to v5 to calculate CMB address properly ([6])

v4:
- modified BAR4 initialization, so now it consists of CMB, MSIX and PBA memory regions overlapping on top of it. This reduces patch complexity significantly (Klaus [5])

v3:
- code style fixes including: removal of spurious line break, moving define into define section and code alignment (Klaus [4])
- removed unintended code reintroduction (Klaus [4])

v2:
- rebase on Kevin's latest block branch (Klaus [3])
- improved comments section (Klaus [3])
- simplified calculation of BAR4 size (Klaus [3])

v1:
- initial push of the patch

Quoted 7 lines

[1]: https://lore.kernel.org/qemu-devel/[email protected]/
[2]: https://lore.kernel.org/qemu-devel/[email protected]/
[3]: https://lore.kernel.org/qemu-devel/[email protected]/
[4]: https://lore.kernel.org/qemu-devel/[email protected]/
[5]: https://lore.kernel.org/qemu-devel/[email protected]/
[6]: https://lore.kernel.org/qemu-devel/[email protected]/

Hi Andrzej,

Any chance you can respin this on git.infradead.org/qemu-nvme.git nvme-next branch?

Thanks,
Klaus