hw/block/nvme: dulbe and dsm support
v4 cover hw/block/nvme: dulbe and dsm support
From: Klaus Jensen <[email protected]>
This adds support for the Deallocated or Unwritten Logical Block error recovery feature as well as the Dataset Management command.
I wanted to add support for the NPDG and NPDA fields such that the host could get a hint on how many blocks to request deallocation of for the deallocation to actually happen, but I cannot find a realiable way to get the actual block size of the underlying device. If it is an image on a file system we could typically use the host page size, but if it is a raw device, we might have 512 byte sectors that we can issue discards on. And QEMU doesn't seem to provide this without root privileges at least.
See the two patches for some gotchas.
I also integrated this into my zoned proposal. I'll spare you the v4, nobody cares anyway. But I put it in my repo[1] for posterity.
[1]: https://irrelevant.dk/g/pci-nvme.git/tag/?h=zoned-v4.
v4:
- Removed mixed declaration and code (Keith)
- Set NPDG and NPDA and account for the blockdev cluster size.
Klaus Jensen (2):
hw/block/nvme: add dulbe support hw/block/nvme: add the dataset management command
hw/block/nvme-ns.h | 4 + hw/block/nvme.h | 2 + include/block/nvme.h | 12 ++- hw/block/nvme-ns.c | 40 +++++++-- hw/block/nvme.c | 183 +++++++++++++++++++++++++++++++++++++++++- hw/block/trace-events | 4 + 6 files changed, 236 insertions(+), 9 deletions(-)
Signature · 1 line
2.28.0
PATCH v4 1/2 hw/block/nvme: add dulbe support
From: Klaus Jensen <[email protected]>
Add support for reporting the Deallocated or Unwritten Logical Block Error (DULBE).
Rely on the block status flags reported by the block layer and consider any block with the BDRV_BLOCK_ZERO flag to be deallocated.
Multiple factors affect when a Write Zeroes command result in deallocation of blocks.
* the underlying file system block size
* the blockdev format
* the 'discard' and 'logical_block_size' parameters
format | discard | wz (512b) wz (4kb) wz (64kb)
---------------------------------------------------
qcow2 ignore n n y
qcow2 unmap n n y
raw ignore n y y
raw unmap n y y
So, this works best with an image in raw format and 4kb LBAs, since holes can then be punched on a per-block basis (this assumes a file system with a 4kb block size, YMMV). A qcow2 image, uses a cluster size of 64kb by default and blocks will only be marked deallocated if a full cluster is zeroed or discarded. However, this *is* consistent with the spec since Write Zeroes "should" deallocate the block if the Deallocate attribute is set and "may" deallocate if the Deallocate attribute is not set. Thus, we always try to deallocate (the BDRV_REQ_MAY_UNMAP flag is always set).
--- hw/block/nvme-ns.h | 4 +++ include/block/nvme.h | 5 +++ hw/block/nvme-ns.c | 8 +++-- hw/block/nvme.c | 83 +++++++++++++++++++++++++++++++++++++++++-- hw/block/trace-events | 4 +++ 5 files changed, 99 insertions(+), 5 deletions(-)
hw/block/nvme-ns.h +4 -0
@@ -31,6 +31,10 @@ typedef struct NvmeNamespace {NvmeIdNs id_ns;NvmeNamespaceParams params; + +struct { +uint32_t err_rec; + } features; } NvmeNamespace;static inline uint32_t nvme_nsid (NvmeNamespace *ns)
include/block/nvme.h +5 -0
@@ -687,6 +687,7 @@ enum NvmeStatusCodes {NVME_E2E_REF_ERROR =0x0284 ,NVME_CMP_FAILURE =0x0285 ,NVME_ACCESS_DENIED =0x0286 , +NVME_DULB =0x0287 ,NVME_MORE =0x2000 ,NVME_DNR =0x4000 , NVME_NO_COMPLETE = 0xffff,
@@ -903,6 +904,9 @@ enum NvmeIdCtrlLpa {#define NVME_AEC_NS_ATTR (aec) ((aec >> 8) & 0x1)#define NVME_AEC_FW_ACTIVATION (aec) ((aec >> 9) & 0x1) +#define NVME_ERR_REC_TLER (err_rec) (err_rec & 0xffff) +#define NVME_ERR_REC_DULBE (err_rec) (err_rec & 0x10000) +enum NvmeFeatureIds {NVME_ARBITRATION =0x1 ,NVME_POWER_MANAGEMENT =0x2 ,
@@ -1023,6 +1027,7 @@ enum NvmeNsIdentifierType {#define NVME_ID_NS_NSFEAT_THIN (nsfeat) ((nsfeat & 0x1)) +#define NVME_ID_NS_NSFEAT_DULBE (nsfeat) ((nsfeat >> 2) & 0x1)#define NVME_ID_NS_FLBAS_EXTENDED (flbas) ((flbas >> 4) & 0x1)#define NVME_ID_NS_FLBAS_INDEX (flbas) ((flbas & 0xf))#define NVME_ID_NS_MC_SEPARATE (mc) ((mc >> 1) & 0x1)
hw/block/nvme-ns.c +5 -3
@@ -33,9 +33,7 @@ static void nvme_ns_init(NvmeNamespace *ns)NvmeIdNs *id_ns = &ns->id_ns;int lba_index =NVME_ID_NS_FLBAS_INDEX (ns->id_ns.flbas); -if (blk_get_flags (ns->blkconf.blk) &BDRV_O_UNMAP ) { - ns->id_ns.dlfeat =0x9 ; - } + ns->id_ns.dlfeat =0x9 ; id_ns->lbaf[lba_index].ds =31 -clz32 (ns->blkconf.logical_block_size);
@@ -44,6 +42,9 @@ static void nvme_ns_init(NvmeNamespace *ns)/* no thin provisioning */ id_ns->ncap = id_ns->nsze; id_ns->nuse = id_ns->ncap; + +/* support DULBE */ + id_ns->nsfeat |=0x4 ; }static int nvme_ns_init_blk (NvmeCtrl *n,NvmeNamespace *ns,Error **errp)
@@ -92,6 +93,7 @@ int nvme_ns_setup(NvmeCtrl *n, NvmeNamespace *ns, Error **errp) }nvme_ns_init (ns); +if (nvme_register_namespace (n, ns, errp)) {return -1 ; }
hw/block/nvme.c +81 -2
135 lines not on this page · raw message
hw/block/trace-events +4 -0
@@ -43,6 +43,10 @@ pci_nvme_admin_cmd(uint16_t cid, uint16_t sqid, uint8_t opcode, const char *opna pci_nvme_rw(uint16_t cid, const char *verb, uint32_t nsid, uint32_t nlb, uint64_t count, uint64_t lba) "cid %"PRIu16" opname '%s' nsid %"PRIu32" nlb %"PRIu32" count %"PRIu64" lba 0x%"PRIx64"" pci_nvme_rw_cb(uint16_t cid, const char *blkname) "cid %"PRIu16" blk '%s'" pci_nvme_write_zeroes(uint16_t cid, uint32_t nsid, uint64_t slba, uint32_t nlb) "cid %"PRIu16" nsid %"PRIu32" slba %"PRIu64" nlb %"PRIu32"" +pci_nvme_block_status(int64_t offset, int64_t bytes, int64_t pnum, int ret, bool zeroed) "offset %"PRId64" bytes %"PRId64" pnum %"PRId64" ret 0x%x zeroed %d" +pci_nvme_dsm(uint16_t cid, uint32_t nsid, uint32_t nr, uint32_t attr) "cid %"PRIu16" nsid %"PRIu32" nr %"PRIu32" attr 0x%"PRIx32"" +pci_nvme_dsm_deallocate(uint16_t cid, uint32_t nsid, uint64_t slba, uint32_t nlb) "cid %"PRIu16" nsid %"PRIu32" slba %"PRIu64" nlb %"PRIu32"" +pci_nvme_aio_discard_cb(uint16_t cid) "cid %"PRIu16"" pci_nvme_create_sq(uint64_t addr, uint16_t sqid, uint16_t cqid, uint16_t qsize, uint16_t qflags) "create submission queue, addr=0x%"PRIx64", sqid=%"PRIu16", cqid=%"PRIu16", qsize=%"PRIu16", qflags=%"PRIu16"" pci_nvme_create_cq(uint64_t addr, uint16_t cqid, uint16_t vector, uint16_t size, uint16_t qflags, int ien) "create completion queue, addr=0x%"PRIx64", cqid=%"PRIu16", vector=%"PRIu16", qsize=%"PRIu16", qflags=%"PRIu16", ien=%d" pci_nvme_del_sq(uint16_t qid) "deleting submission queue sqid=%"PRIu16""
Signature · 1 line
2.28.0
PATCH v4 2/2 hw/block/nvme: add the dataset management command
In reply to Klaus JensenFrom: Klaus Jensen <[email protected]>
Add support for the Dataset Management command and the Deallocate attribute. Deallocation results in discards being sent to the underlying block device. Whether of not the blocks are actually deallocated is affected by the same factors as Write Zeroes (see previous commit).
format | discard | dsm (512b) dsm (4kb) dsm (64kb)
------------------------------------------------------
qcow2 ignore n n n
qcow2 unmap n n y
raw ignore n n n
raw unmap n y y
Again, a raw format and 4kb LBAs are preferable.
In order to set the Namespace Preferred Deallocate Granularity and Alignment fields (NPDG and NPDA), choose a sane minimum discard granularity of 4kb. If we are using a passthru device supporting discard at a 512b granularity, user should set the discard_granularity property explicitly. NPDG and NPDA will also account for the cluster_size of the block driver if required (i.e. for QCOW2).
See NVM Express 1.3d, Section 6.7 ("Dataset Management command").
--- hw/block/nvme.h | 2 + include/block/nvme.h | 7 ++- hw/block/nvme-ns.c | 36 ++++++++++++++-- hw/block/nvme.c | 100 ++++++++++++++++++++++++++++++++++++++++++- 4 files changed, 139 insertions(+), 6 deletions(-)
hw/block/nvme.h +2 -0
@@ -28,6 +28,7 @@ typedef struct NvmeRequest {struct NvmeNamespace *ns;BlockAIOCB *aiocb;uint16_t status; +void *opaque;NvmeCqe cqe;NvmeCmd cmd;BlockAcctCookie acct;
@@ -60,6 +61,7 @@ static inline const char *nvme_io_opc_str(uint8_t opc)case NVME_CMD_WRITE :return "NVME_NVM_CMD_WRITE" ;case NVME_CMD_READ :return "NVME_NVM_CMD_READ" ;case NVME_CMD_WRITE_ZEROES :return "NVME_NVM_CMD_WRITE_ZEROES" ; +case NVME_CMD_DSM :return "NVME_NVM_CMD_DSM" ;default :return "NVME_NVM_CMD_UNKNOWN" ; } }
include/block/nvme.h +6 -1
@@ -990,7 +990,12 @@ typedef struct QEMU_PACKED NvmeIdNs {uint16_t nabspf;uint16_t noiob;uint8_t nvmcap[16 ]; -uint8_t rsvd64[40 ]; +uint16_t npwg; +uint16_t npwa; +uint16_t npdg; +uint16_t npda; +uint16_t nows; +uint8_t rsvd74[30 ];uint8_t nguid[16 ];uint64_t eui64;NvmeLBAF lbaf[16 ];
hw/block/nvme-ns.c +32 -4
@@ -28,10 +28,14 @@#include "nvme.h" #include "nvme-ns.h" -static void nvme_ns_init (NvmeNamespace *ns) +#define MIN_DISCARD_GRANULARITY (4 * KiB) + +static int nvme_ns_init (NvmeNamespace *ns,Error **errp) { +BlockDriverInfo bdi;NvmeIdNs *id_ns = &ns->id_ns;int lba_index =NVME_ID_NS_FLBAS_INDEX (ns->id_ns.flbas); +int npdg, ret; ns->id_ns.dlfeat =0x9 ;
@@ -43,8 +47,25 @@ static void nvme_ns_init(NvmeNamespace *ns) id_ns->ncap = id_ns->nsze; id_ns->nuse = id_ns->ncap; -/* support DULBE */ - id_ns->nsfeat |=0x4 ; +/* support DULBE and I/O optimization fields */ + id_ns->nsfeat |= (0x4 |0x10 ); + + npdg = ns->blkconf.discard_granularity / ns->blkconf.logical_block_size; + + ret =bdrv_get_info (blk_bs (ns->blkconf.blk), &bdi); +if (ret <0 ) { +error_setg_errno (errp, -ret,"could not get block driver info" ); +return ret; + } + +if (bdi.cluster_size && + bdi.cluster_size > ns->blkconf.discard_granularity) { + npdg = bdi.cluster_size / ns->blkconf.logical_block_size; + } + + id_ns->npda = id_ns->npdg = npdg -1 ; + +return 0 ; }static int nvme_ns_init_blk (NvmeCtrl *n,NvmeNamespace *ns,Error **errp)
@@ -59,6 +80,11 @@ static int nvme_ns_init_blk(NvmeCtrl *n, NvmeNamespace *ns, Error **errp)return -1 ; } +if (ns->blkconf.discard_granularity ==-1 ) { + ns->blkconf.discard_granularity = +MAX (ns->blkconf.logical_block_size,MIN_DISCARD_GRANULARITY ); + } + ns->size =blk_getlength (ns->blkconf.blk);if (ns->size <0 ) {error_setg_errno (errp, -ns->size,"could not get blockdev size" );
@@ -92,7 +118,9 @@ int nvme_ns_setup(NvmeCtrl *n, NvmeNamespace *ns, Error **errp)return -1 ; } -nvme_ns_init (ns); +if (nvme_ns_init (ns, errp)) { +return -1 ; + }if (nvme_register_namespace (n, ns, errp)) {return -1 ;
hw/block/nvme.c +99 -1
118 lines not on this page · raw message
Signature · 1 line
2.28.0
Re: [PATCH v4 2/2] hw/block/nvme: add the dataset management command
hw/block/nvme.c static void nvme_rw_cb(void *opaque, int ret) · line 1048
Quoted 5 lines · Klaus Jensen
hw/block/nvme.c +5
-0+if (--(*discards)) { + status =NVME_NO_COMPLETE ; + }else { +g_free (discards); + req->opaque =NULL ;
This case needs a
status = req->status;
So that we get the error set in the callback.
Otherwise, this looks fine. I am assuming everything still runs single threaded since this isn't using atomics.
Re: [PATCH v4 2/2] hw/block/nvme: add the dataset management command
Quoted 7 lines · Keith Busch
hw/block/nvme.c static void nvme_rw_cb(void *opaque, int ret) · line 1048
Quoted 5 lines · Klaus Jensen
hw/block/nvme.c +5
-0+if (--(*discards)) { + status =NVME_NO_COMPLETE ; + }else { +g_free (discards); + req->opaque =NULL ;
Quoted 6 lines
This case needs a
status = req->status;So that we get the error set in the callback.
There are no cases that result in a non-zero status code here. If an LBA range is invalid we simply continue with the next. In case the DMA transfer fails, we return the error directly and the normal path takes care of it. The else block is for when there are no pending aios for some reason (all invalid ranges or they completed immediately) - in that case we can just return NVME_SUCCESS directly.
Quoted 2 lines
Otherwise, this looks fine. I am assuming everything still runs single threaded since this isn't using atomics.
Yeah, all device code (including callbacks) run on the main thread.
Re: [PATCH v4 2/2] hw/block/nvme: add the dataset management command
Quoted 8 lines · Klaus Jensen
Quoted 7 lines · Keith Busch
Quoted 5 lines · Klaus Jensen
hw/block/nvme.c +5
-0+if (--(*discards)) { + status =NVME_NO_COMPLETE ; + }else { +g_free (discards); + req->opaque =NULL ;
Quoted 8 lines
Quoted 6 lines
This case needs a
status = req->status;So that we get the error set in the callback.
There are no cases that result in a non-zero status code here.
Your callback has a case that sets NVME_INTERNAL_DEV_ERROR status. That would get ignored if the final discard reference is dropped from the submission side.
+static void nvme_aio_discard_cb(void *opaque, int ret) +{
+ NvmeRequest *req = opaque;
+ int *discards = req->opaque;
+
+ trace_pci_nvme_aio_discard_cb(nvme_cid(req));
+
+ if (ret) {
+ req->status = NVME_INTERNAL_DEV_ERROR;
+ trace_pci_nvme_err_aio(nvme_cid(req), strerror(ret),
+ req->status);
+ }
Re: [PATCH v4 2/2] hw/block/nvme: add the dataset management command
Quoted 13 lines · Keith Busch
Quoted 12 lines · Klaus Jensen
Quoted 11 lines · Keith Busch
hw/block/nvme.c static void nvme_rw_cb(void *opaque, int ret) · line 1048
Quoted 5 lines · Klaus Jensen
hw/block/nvme.c +5
-0+if (--(*discards)) { + status =NVME_NO_COMPLETE ; + }else { +g_free (discards); + req->opaque =NULL ;This case needs a
status = req->status;
Quoted 9 lines
Quoted 4 lines
Quoted 2 lines
So that we get the error set in the callback.
There are no cases that result in a non-zero status code here.
Your callback has a case that sets
NVME_INTERNAL_DEV_ERRORstatus. That would get ignored if the final discard reference is dropped from the submission side.
Oh. Crap. You are right. Nice catch!
Quoted 12 lines
hw/block/nvme.c +12
-0+static void nvme_aio_discard_cb (void *opaque,int ret) +{ +NvmeRequest *req = opaque; +int *discards = req->opaque; + +trace_pci_nvme_aio_discard_cb (nvme_cid (req)); + +if (ret) { + req->status =NVME_INTERNAL_DEV_ERROR ; +trace_pci_nvme_err_aio (nvme_cid (req),strerror (ret), + req->status); + }